Research Appraisals
Evidence-based critical appraisals of the latest medical research, systematically evaluated using Oxford CEBM methodology.
Showing 103 appraisals
Current opinion in chemical biology
Prediction of protein-protein interactions and co-complex models with deep learning
Protein-protein interactions (PPIs) are fundamental to cellular processes, and essential for understanding biological function and disease mechanisms. In this review, we emphasize recent deep learning-based methods for protein interaction study. Focusing on three closely related tasks of proteome-wide PPI prediction, PPI interface prediction, and PPI co-complex structure prediction, we discuss how emerging concepts and computation approaches have evolved to shape these fields We categorize recent approaches according to their methodological paradigms, summarize their strengths and limitations, and further explore diverse biological and biomedical applications, highlighting how computational methods in PPI prediction, PPI interface prediction, and PPI structure prediction jointly contribute to understanding of complex biological systems.
3 Aug 2026
Read appraisal →Journal of chemical information and modeling
ProphDR: An Interpretable Deep Learning Model for Predicting Cancer Drug Response via Multi-Omics and Cross-Attention Mechanisms
Predicting cancer drug responses (CDRs) accurately remains a significant challenge due to the complexity of tumor biology and the limitations of existing "black-box" machine learning models. To address this, we propose ProphDR, an interpretable deep learning framework that integrates multiomics data and drug structural information using a hierarchical attention mechanism. ProphDR incorporates a Criss-Cross Gene-level Multiomics Integration (CGMI) module to capture gene-level features and a cross-attention (CA) module to model drug-gene interactions. Evaluated on datasets from GDSC and CCLE, ProphDR achieves state-of-the-art performance in predicting ln(IC50) values (PCC = 0.938, RMSE = 0.978) and classifying drug sensitivity (AUC = 0.981). It also demonstrates strong generalizability in cold-start scenarios involving unseen drugs or cell lines. Crucially, ProphDR generates biologically interpretable attention maps that highlight key pharmacophores and resistance-related genes such as ERBB2 (HER2), consistent with established mechanisms in NSCLC and BRCA. These insights bridge genomic features with phenotypic outcomes, offering valuable guidance for target prioritization and drug repurposing. ProphDR represents a robust and explainable AI tool for advancing precision oncology.
28 July 2026
Read appraisal →PloS one
Multi-view graph-regularized deep metric subspace clustering network
Multi-view subspace clustering has progressed significantly by using deep neural networks to handle nonlinear data representations. A recent advancement, the Multi-view Self-Expressive Subspace Clustering (MSESC) network, achieves markedly higher computational efficiency by substituting the traditional self-expression layer with a deep metric learning approach. Nevertheless, MSESC still suffers from two notable limitations: it fails to adequately capture the high-order geometric structures inherent in multi-view data, and it lacks effective guidance from the underlying clustering distribution. To overcome these shortcomings, we propose a novel framework termed Multi-View Graph Regularized Deep Metric Subspace Clustering (MVGR-DMSC). The proposed method introduces two key components into MSESC to enhance the discriminability of representations. First, a dual-order graph regularization module is devised to maintain both first-order and second-order manifold structures, thereby allowing the model to capture more complex local geometric relationships. Second, an adaptive view-weighted deep clustering module is incorporated, which employs the Kullback-Leibler divergence to guide representation learning while dynamically adjusting the contributions of different views. Through evaluations on five benchmark datasets, we show that MVGR-DMSC consistently yields better results than several state-of-the-art approaches, including the direct baseline MSESC, in both accuracy and robustness.
26 July 2026
Read appraisal →New biotechnology
Harnessing AI to decode protein kinases: Structural, functional, and therapeutic design perspectives.
Recent advances in Artificial Intelligence (AI) are reshaping kinase research by uncovering complex regulatory mechanisms and accelerating drug discovery. These advances have enabled the capture of elusive allosteric transitions and transient cryptic pockets by exploring dynamic conformational landscapes, revealing features critical for understanding ligand interactions and enzymatic regulation. AI-driven structure prediction tools, such as AlphaFold (AF)-based models, offer high-resolution insights into kinase conformations and complex assembly, thereby clarifying receptor activation in signaling pathways. Moreover, cutting-edge AI frameworks provide fresh perspectives on substrate prediction, cofactor mapping, and mutation-induced structural changes. Integrative strategies that combine Machine Learning (ML), molecular docking, Molecular Dynamics (MD) simulation, and experimental validation have further streamlined the discovery of novel kinase inhibitors and related studies. Collectively, these AI-enhanced approaches deepen our understanding of kinase-mediated signaling-from developmental processes to oncogenic transformation-and highlight the powerful synergy between computational and experimental research in advancing therapeutic innovations.
25 July 2026
Read appraisal →ACS synthetic biology
LysePred: A Multiscale Convolutional Neural Network for Predicting Hemolytic Activity of Antimicrobial Peptides
Antimicrobial peptides (AMPs) represent promising alternatives to conventional antibiotics, yet hemolytic toxicity remains a critical barrier to clinical translation, with approximately 70% of known AMPs exhibiting high or moderate hemolytic activity. Existing computational prediction methods are often constrained by deficiencies, including high computational complexities, the inability to capture multiscale sequence patterns, and insufficient generalization across diverse datasets. We present LysePred, a multiscale convolutional neural network to address these deficiencies concurrently by employing parallel branches with exponentially spaced kernel sizes to simultaneously capture local amino acid motifs (bigrams, 4-g) and longer-range amphipathic patterns (8- to 32-g). LysePred achieves top-tier performance on six benchmark datasets, exceeding the performance of the second-best method by 9.13% in MCC and 3.65% in ACC on average, while maintaining exceptional stability (MCC CV < 6.92%, ACC CV < 2.73%). Furthermore, independent validation on the HemoPI2 dataset demonstrates that LysePred delivers highly competitive results with a computationally parsimonious design (∼0.55 M parameters). This represents a reduction in parameter density of 1 to 2 orders of magnitude compared to Transformer-based approaches, such as the 8M-parameter ESM-2 or the 110M-parameter BERT Base while maintaining linear complexity for high-throughput screening. Ablation studies validate that the multiscale architecture contributes meaningfully to performance, with single-scale variants showing up to 13.35% MCC degradation. Interpretability analyses via t-SNE visualization and SHAP feature importance reveal that LysePred learns biologically meaningful representations integrating both local sequence motifs and global structural patterns. LysePred offers a practical, efficient, and interpretable tool for rapid hemolytic toxicity prediction in antimicrobial peptide development. Code and data are available at https://github.com/lincubator/LysePred.
20 July 2026
Read appraisal →European journal of medicinal chemistry
Accurate and interpretable ADMET prediction: Integrating structural, geometric, and global molecular context representations
Predicting absorption, distribution, metabolism, excretion, and toxicity (ADMET) profiles early in drug discovery is essential to avoid late-stage failures. However, most current computational models depend on single-view molecular representations. This restricts them from capturing the complex interplay between 2D topology, 3D conformations, and specific functional groups that collectively drive structure-pharmacokinetic relationships. Here, we present LGSM, a comprehensive deep learning framework for integrative molecular property prediction. By fusing structural (2D topology), geometric (3D conformational descriptors), and sequence-derived semantic representations, LGSM provides accurate ADMET predictions. Specifically, a language model-based encoding of SMILES sequences is utilized to capture the global molecular context, seamlessly complementing local functional group characteristics. By fusing these different levels of information, our architecture accounts for both local chemical environments and broader spatial relationships. We evaluated this approach across four public ADMET datasets. To mimic real-world drug design scenarios where new chemotypes are constantly explored, we applied a strict scaffold split. Under these conditions, LGSM consistently outperformed standard computational baselines. Beyond predictive accuracy, we prioritized chemical interpretability. By combining the model's internal analysis with molecular docking, we show that LGSM successfully flags key substructures responsible for metabolic liabilities or transporter recognition. This transparency makes the framework a practical tool to guide rational lead optimization in medicinal chemistry.
6 July 2026
Read appraisal →The Journal of physiology
A deep learning-enabled toolkit for the 3D segmentation of ventricular cardiomyocytes
Segmentation of cardiomyocytes in microscopic 3D volumes is key to our understanding of cardiac (patho-)physiology; however, it poses substantial experimental and analytical challenges. Therefore, researchers often resort to inferring 3D information from 2D segmentations, which can lead to biased and incorrect conclusions. Deep learning-based methods are showing promise with respect to robustly segmenting objects in volumes acquired using various imaging modalities; yet, they have not been applied to high-resolution 3D cardiomyocyte segmentations, and suitable open-source tools and datasets are lacking. Here, we present a deep learning-enabled toolkit for segmentation of individual cardiomyocytes in 3D confocal microscopy volumes. We include a dataset of 73 volumes with expert annotations, covering seven species, including mouse, human, and elephant, and containing samples generated under different experimental conditions, such as post-myocardial infarction and ex vivo slice cultures. The toolkit additionally contains an image restoration workflow to address imaging-related artefacts, such as spatially varying blur. Our automatic cardiomyocyte segmentation workflow achieved an adapted Rand error of 0.063 ± 0.034 (∼94% voxel-pair agreement) on the test set. Our semi-automatic workflow reached a throughput of 3 cells min-1 on a challenging, previously unseen dataset. The toolkit and data are open-source and accessible through a dedicated graphical user interface. In summary, we provide an accessible toolkit enabling researchers to extract quantitative data on cardiomyocyte microstructure from 3D confocal image stacks of cardiac tissue. Given the size and diversity of our dataset, we expect our methods to perform well across species and experimental conditions, facilitating high-quality 3D reconstructions of large numbers of individual cardiomyocytes. KEY POINTS: 3D cardiomyocyte microstructure is a key determinant of cardiac function in health and disease. However, reliable extraction and quantification of 3D cardiomyocyte cytoarchitecture pose significant experimental and computational challenges. We present an effective experimental protocol and a deep learning-enabled toolkit for sample preparation and 3D analysis of cardiomyocyte morphology in ventricular myocardium. Our method is validated across seven species (mouse to human) and in samples prepared in diverse experimental conditions from a range of models, including myocardial infarction and ex vivo tissue culture, highlighting the robustness and versatility of our workflow. Our open-source dataset and toolkit enable large-scale analyses and extraction of realistic 3D geometries of ventricular microstructure. These can be used to explore a host of research questions and provide a new resource for modelling cardiac function at the cellular level.
4 July 2026
Read appraisal →Journal of chemical information and modeling
GraFSyn: An Interpretable Deep Learning Framework for Anticancer Drug Synergy via Graphlet Fingerprints
Predicting drug synergy is important for accelerating the discovery of effective anticancer combination therapies. Synergy intrinsically depends on the precise interactions of key chemical substructures within specific cellular environments. However, current molecular graph-based computational methods typically rely on implicit atom-level feature aggregation, which may obscure the topological representation of critical chemical substructures and limit structural traceability. Therefore, we present Graphlet Fingerprint-based Synergy prediction (GraFSyn), a deep learning framework for anticancer drug synergy prediction that uses graphlet fingerprints to encode drugs as explicit connected substructure units, preserving predefined chemical substructures and their topological identity. We further introduce a Dynamic Multi-Scale Convolution (DMSC) module to learn informative representations from high-dimensional and sparse graphlet features. The framework also includes an interaction module to capture context-dependent interactions between drug substructures and cell line gene expression. On the Merck and AstraZeneca benchmark data sets, GraFSyn achieved ROC-AUC/PR-AUC values of 0.972/0.912 and 0.823/0.906, respectively, outperforming representative baseline methods. In addition, attributed signals can be mapped back to specific pharmacophoric regions, supporting substructure-level analysis of synergistic interactions. In general, GraFSyn provides an accurate and structurally traceable approach for anticancer drug combination screening.
23 June 2026
Read appraisal →Science advances
Discovery of TYR inhibitors from de novo molecular generation to dual-track lead optimization: 'Competition' between AI and chemists.
This study introduces a unified framework combining artificial intelligence (AI)-directed de novo molecular generation with dual-track lead optimization-comprising expert-guided strategies and AI-driven pathways-to discover tyrosinase (TYR) inhibitors for hyperpigmentation disorders. Using a reinforcement learning (RL)-based generative model, the lead compound AI10 was identified. Subsequent optimization followed two parallel routes. The expert-guided approach yielded AI10-m15 as the most potent TYR inhibitor, with notable antipigmentation activity and excellent cellular safety profiles. In contrast, the AI-driven pathway explored broader chemical spaces, generating unconventional chemotypes, exemplified by the potent TYR inhibitor AI10-a2, highlighting AI's capacity to uncover nonintuitive activity cliffs despite greater output variability. Systematic comparison revealed that the AI model offers exploratory diversity, whereas expert-guided optimization provides predictable improvements in activity and developability. In summary, starting from an AI-generated lead and subsequently integrating both expert-guided and AI-driven structural optimization strategies, these findings further underscore that combining AI technologies with experts' medicinal chemistry insights can substantially accelerate the discovery of viable candidate compounds.
21 June 2026
Read appraisal →Science advances
PathTIGR: A pathway topology-informed graph representation learning framework for immunotherapy response prediction
Immunotherapy has revolutionized cancer treatment, yet substantial inter-patient response heterogeneity limits therapeutic benefit to specific patient subsets. Here, we present PathTIGR, a pathway topology-informed graph representation learning framework that systematically integrates biological pathway network topology knowledge with genome variation information for immunotherapy response prediction. PathTIGR uses a three-component design: (i) pathway graph encoder with multihead attention embeding pathway topology knowledge and cancer genomic variants to pathway representation, (ii) transformer module capturing pathway regulatory dependencies, and (iii) multilayer perceptron synthesizing pathway-level representations to predict immunotherapy response. This architecture enables PathTIGR to capture complex molecular interactions underlying immunotherapy response. Comprehensive validation across multiple independent immunotherapy cohorts demonstrates that PathTIGR achieves superior predictive performance compared to established biomarkers and state-of-the-art deep learning approaches while maintaining biological interpretability through identification of key signatures underlying response heterogeneity. PathTIGR represents an interpretable graph-based learning framework that enhances immunotherapy response prediction and elucidates molecular determinants of therapeutic efficacy, thereby facilitating the advancement of precision cancer immunotherapy.
14 June 2026
Read appraisal →Journal of ovarian research
Deep learning-based prediction of gene expression from histopathology identifies NR5A1 as a candidate biomarker and druggable target in high-grade serous ovarian carcinoma
BACKGROUND: High-grade serous ovarian cancer (HGSOC) is the most lethal ovarian cancer subtype, responsible for ~ 70% of ovarian cancer-related deaths and marked by late-stage diagnosis and frequent platinum resistance. Although transcriptomic profiling enables molecular stratification and prediction of therapeutic response; routine clinical use of this approach is limited by cost and logistical constraints. Computational pathology analysis offers a scalable alternative by inferring transcriptional states directly from routine hematoxylin and eosin (H&E) whole-slide images (WSIs). METHODS: Paired H&E WSIs and RNA-sequencing data from the TCGA-OV cohort, including 1,371 diagnostic H&E WSIs retrieved for preprocessing and quality control, were used to develop a self-supervised virtual-transcriptomics framework based on Momentum Contrast v2 (MoCo v2) and multi-output Random Forest regression. Model performance was assessed using patient-level five-fold cross-validation. Candidate genes were evaluated by reverse transcription quantitative polymerase chain reaction (RT-qPCR) in an independent cohort of 10 HGSOC tumors, including 4 platinum responders and 6 non-responders. RESULTS: The model predicted expression of approximately 6,400 protein-coding genes, achieving a genome-wide mean Pearson correlation of r = 0.36, with more than 300 genes showing stronger image-expression coupling (r > 0.44). RT-qPCR analysis of 18 candidate genes revealed substantial inter-patient heterogeneity. NR5A1 exhibited the highest expression variability (coefficient of variation [CV] = 1.486) and significantly higher expression in platinum-responsive tumors than in non-responders (mean 2⁻ΔCt = 0.263 vs. 0.013; p < 0.05). Exploratory in silico docking and molecular dynamics (MD) analyses suggested structurally stable binding interactions between Steroidogenic Factor-1 (SF-1/NR5A1) and the natural plant compound cubebin. CONCLUSION: This study demonstrates that histological architecture contains measurable transcriptomic information that can support scalable biomarker prioritization from routine diagnostic histology in HGSOC. NR5A1 represents a hypothesis-generating candidate biomarker and structurally tractable target for future experimental studies. Future validation in larger, multi-center cohorts will be essential to confirm model robustness, biological relevance, and potential clinical utility.
14 June 2026
Read appraisal →Journal of medicinal chemistry
Integrating AI in Medicinal Chemistry for Accelerated Drug Discovery: A Comprehensive SAR (CSAR) Optimization Strategy and Discovery of Potent ALDH3A1 Inhibitors
Developing potent, selective small-molecule inhibitors remains a major challenge in drug discovery. ALDH3A1, a detoxifying aldehyde dehydrogenase isoform implicated in cancer and neurodegeneration, is a promising yet underexplored therapeutic target. To accelerate inhibitor optimization, we developed an AI-guided, reaction-based hit-to-lead workflow combining sequential reaction enumeration, pharmacophore-informed docking, and predictive modeling to support scalable SAR expansion. Applied to ALDH3A1, two rounds of enumeration using Enamine building blocks generated about 250,000 virtual analogues. Combined deep learning and docking-based triage prioritized 150 compounds for synthesis, leading to a roughly 1,000-fold improvement in biochemical potency from 1.41 μM to 1 nM for NCATS-SM0707, together with 4 nM cellular activity for NCATS-SM0708. Crucially, this methodology can be expanded by applying various chemical reactions at different positions. This study highlights CSAR as a scalable, generalizable complementary strategy to accelerate hit optimization through reaction-based enumeration and AI-guided prioritization of larger synthetically accessible chemical space.
13 June 2026
Read appraisal →Journal of medicinal chemistry
MEG-mod: A Multiview Enhanced Graph Neural Network for Knockdown Efficiency Prediction of Chemically Modified siRNA
Chemical modification is essential for improving the stability and knockdown efficiency of siRNAs. However, the combinatorial complexity of modification types and positions makes rational design difficult. Here, we propose MEG-mod, a deep learning framework for predicting the knockdown efficiency of chemically modified siRNAs. By integrating a literature-reading agent, manual curation, and public database resources, we construct an expanded data set. MEG-mod jointly models sequence context, physicochemical properties, chemical modification features, and duplex structural relationships. A structure-aware Transformer-based graph neural network captures base-pairing and adjacency dependencies within double-stranded siRNAs, while a modification-base fusion module models context-dependent modification effects. MEG-mod outperforms existing methods, achieving a pearson correlation coefficient of 0.9171. The attention mechanism enables interpretable analysis of key modification positions and modification types, and provides recommendations consistent with reported experimental designs. MEG-mod is available as a public web server (https://lmmd.ecust.edu.cn/megmod/), providing a practical tool for prediction and candidate prioritization of chemically modified siRNAs.
13 June 2026
Read appraisal →Journal of theoretical biology
Subtypes estimation in single-cell multi-omics data via S-multi-SNE
Single-cell RNA sequencing (scRNA-seq) has rapidly become an essential method in modern biology, as it allows global characterisation of transcriptomics in individual cells from different tissues, or organisms. scRNA-seq data are collected from a single individual organism and they are composed as mRNA count matrices with rows representing the cells and columns the genes. New technologies further allow multi-omics data to be obtained on the same set of single cells. In this paper, multi-SNE and S-multi-SNE, two multi-view dimensionality reduction approaches, are adapted for visualization and classification of cellular subtypes, two important challenges in the analysis of scRNA-seq data. S-multi-SNE has been adapted for cell subtype estimation of scRNA-seq data. In a series of experiments we illustrate that it is possible to identify cell subtypes by leveraging information from individuals from the same species, from different species, and when data are generated by different sequencing technologies. In the conducted analyses, we show that S-multi-SNE consistently outperformed several other machine learning techniques and, in most comparisons, surpassed existing reference-based single-cell classification algorithms. We further illustrate how two single-cell multi-omics datasets, scATAC-seq and scRNA-seq, can be integrated together through S-multi-SNE for improved cell subtype identification.
7 June 2026
Read appraisal →PLoS computational biology
A prototype-augmented graph representation learning framework for identifying brain disorder-associated genes and facilitating drug repurposing
Many genetic loci were identified as associated with neuropsychiatric disorders and neurodegenerative disorders by Genome-wide association studies (GWAS). How these loci impact these diseases is unclear. Advances in deep-learning approaches and multi-omics data have the potential to link GWAS findings with disease mechanisms. Here, we proposed the Multi-omics Graph Transformer Network (MOGT), a semi-supervised graph neural network that leverages graph representation learning to model biological networks derived from multi-omics data to predict disease-associated genes. MOGT outperforms the current approaches in disease gene prediction for two psychiatric disorders and three neurodegenerative/neurological diseases. High-risk genes (HRGs) for Parkinson's disease (PD) predicted by MOGT were used to drug discovery by integrating with the CMAP database. Finally, 10 drugs were identified as potential candidates. Among them, the effect of drug UK-356618 was experimentally verified in a primary neuron model, showing that UK-356618 reversed the abnormal expression of PD-associated genes and improved the cell-level phenotypes of PD. Together, these results indicate that MOGT can be used to identify HRGs for brain disorders, and these predicted HRGs provide high-level insights into the mechanisms and treatments of brain disorders.
1 June 2026
Read appraisal →Journal of computer-aided molecular design
Target-aware molecule SMILES generation using a large language model with retrieval-augmented generation, multi-turn memory, and a predictive model
Given the vastness of chemical space and the cost and time requirements of high-throughput screening, resource-efficient computational strategies are needed to prioritize candidate drug molecules. Here, we evaluated whether an open-weight large language model (LLM) augmented with retrieval-augmented generation (RAG), multi-turn memory, and a pretrained drug-target interaction predictor can generate target-aware molecules without task-specific retraining. Protein-ligand-pKi examples retrieved from BindingDB, Davis, and KIBA were used as contextual guidance, while DeepPurpose provided the optimization signal during iterative SMILES refinement. Our approach produced a statistically significant increase in predicted pKi across successive multi-turn memory iterations, indicating that retrieval- and memory-guided refinement may improve target-conditioned molecular generation. The proposed framework also showed favorable molecular generation characteristics, with novelty reaching 100%, diversity up to 0.882, and uniqueness up to 1.0, suggesting that retrieval did not reduce the process to simple reproduction of known ligands but instead supported exploration of new regions of chemical space. In addition, the approach yielded supportive docking results consistent with the generation of chemically relevant candidate ligands. However, the increase in predicted affinity was accompanied by trade-offs in molecular quality, including reduced validity and drug-likeness in later iterations. Taken together, these findings suggest that this framework provides a flexible and comparatively resource-efficient strategy for target-aware de novo molecular design, while further multi-objective optimization and independent validation remain necessary.
31 May 2026
Read appraisal →mAbs
Prediction of antibody non-specificity using protein language models and biophysical parameters
The development of therapeutic antibodies requires optimizing target binding affinity and pharmacodynamics, while ensuring high developability potential, including minimizing non-specific binding. In this study, we address this problem by predicting antibody non-specificity by two complementary approaches: (1) antibody sequence embeddings by protein language models (PLMs) and (2) a comprehensive set of sequence-based biophysical descriptors. We benchmark the PLM embeddings against interpretable, sequence-derived biophysical descriptors and use fragment-specific models (variable heavy (VH) and variable light (VL) chains, concatenated regions, and individual complementary-determining regions (CDRs) to identify region-level determinants of non-specificity in functional antibodies that bind defined targets. These models were trained on previously published human and mouse antibody data and tested on three public datasets. We show that non-specificity is best predicted from the VH domain and heavy-chain CDRs. These region-level analyses show that heavy-chain features, especially H-CDR3, dominate the predictive signal. The top performing PLM, a VH domain-based Evolutionary Scale Modeling 1 v LogisticReg model, resulted in 10-fold cross-validation accuracy of up to 71%. While predictive accuracy is comparable to classical machine-learning baselines, PLM-based embeddings provide sequence-context representations that yield consistent region-level attribution and provide complementarity to the classical models by enhancing the reliability of predicted non‑specificity scores when used in combination. Our biophysical descriptor-based analysis identified the isoelectric point as a key driver of non-specificity, consistent with previous reports. Our findings highlight the importance of biophysical properties in predicting antibody non-specificity and highlight the potential of PLMs for the development of antibody-based therapeutics. These conclusions are robust to alternative class definitions and are not driven by VH-VL mutational bias. We illustrate the generalizability and practical use of the PLM approach by extending it to therapeutic antibodies and nanobodies, providing a tool for early-stage developability assessment, and we make the models publicly available.
29 May 2026
Read appraisal →Journal of medical Internet research
Biocomputing: Beyond the Hype
Biocomputing is a nascent but rapidly developing field at the intersection of biology and computer science. In this News and Perspectives article, JMIR Correspondent Simon Spichak reports on its current and potential applications for health care research and beyond.
29 May 2026
Read appraisal →Journal of chemical information and modeling
Machine Learning-Driven Simulations of the SARS-CoV-2 Fitness Landscape from Deep Mutational Scanning Experiments
Predicting protein variant effects is a key challenge in preparing for pathogenic viral strains, understanding mutation-linked diseases, and designing new proteins. Protein sequence-structure-function relationships are difficult to model due to complex allosteric and epistatic effects. To investigate efficient modeling strategies, we trained supervised machine learning (ML) models with deep mutational scanning (DMS) libraries of SARS-CoV-2 receptor binding domain (RBD) sequences labeled with angiotensin converting enzyme 2 (ACE2) binding affinity. These models demonstrate superior performance predicting combinatorial mutation effects compared to adding or averaging the effects of point mutations and exhibit strong extrapolative performance ranking omicron variants when training only near wild type (WT) variants. We characterize the RBD fitness landscape by combining ML with Markov Chain Monte Carlo simulations to predict evolutionary patterns from the WT sequence. These generate comparable sequence profiles to high-fitness sequences in DMS data and predict mutations in unseen omicron variants. These models provide insight into the relationship between RBD sequence elements and offer a new perspective on the use of DMS to predict emerging viral strains, which we anticipate will be applicable to other evolutionary prediction tasks. To facilitate application and future development of this strategy, we introduce Mavenets: https://github.com/SztainLab/mavenets.
27 May 2026
Read appraisal →Journal of chemical information and modeling
Integrating QSAR-Machine Learning, Biochemical Assays, and Molecular Dynamics for the Discovery of JAK2 Inhibitors in Cervical Cancer
Cervical cancer remains a major global health challenge, where dysregulated JAK2 signaling constitutes a key molecular driver. Nevertheless, selective small-molecule JAK2 inhibitors for HPV-positive cervical cancer are still limited. Here, we integrated biochemical assays, QSAR-machine learning, and molecular dynamics simulations to identify potent JAK2 inhibitors. A series of naphthalene-based derivatives, including hydroxynaphthalenamide and phosphorylated dihydronaphthylamide analogs, were evaluated for cytotoxicity in HeLa cells and JAK2 kinase inhibition. Several compounds exhibited selective cytotoxicity with minimal activity toward normal fibroblasts, among which 2q and 2s showed low-nanomolar JAK2 inhibition and strong apoptosis induction through suppression of the JAK2/STAT3/STAT5 pro-tumorigenic signaling pathway. To accelerate hit identification, a QSAR-machine learning (QSAR-ML) framework was employed to prioritize 13 newly designed derivatives. Among three ensemble boosting models, the Categorical Boosting (CB) model demonstrated the strongest predictive capability, achieving a high R2 of 0.955 for the training set and a low RMSE of 0.156 for the test set. This model successfully identified five active candidates with strong prediction-experiment agreement (MAPE = 2.6-14.4%), with D4 and D13 meeting drug-likeness criteria and displaying potent nanomolar JAK2 inhibition. Finally, 1-μs molecular dynamics simulations revealed that hydrophobic contacts and hydrogen bonding cooperatively stabilize these inhibitors within the JAK2 ATP-binding pocket. Collectively, these findings establish a QSAR-ML-guided strategy for accelerating JAK2 inhibitor discovery and highlight naphthalene-based scaffolds as promising leads for targeted cervical cancer therapy.
27 May 2026
Read appraisal →Journal of chemical theory and computation
MolXProt: A Cross-Attention Transformer-Based Graph Neural Network for Protein-Ligand Binding Affinity Prediction
Accurate and fast prediction of drug-target binding affinities (DTAs) is key for drug discovery; however, many methods, such as docking and empirical scoring, fail when generalizing to unseen cases. In this study, we introduced the MolXProt architecture, a novel transformer-based graph neural network that integrates graph ligand representations with protein language models using bidirectional multihead cross-attention. Our model is shown to be scalable to over 100000 protein-ligand pairs of mixed data sets, achieving 50% of predictions within ±1.0 kcal/mol and 80% within ±2.0 kcal/mol. We show that the architecture can explicitly learn residue-atom interactions while being computationally friendly via protein-token compression. By mapping token-residue interactions, we demonstrated that the model learns key binding pocket residues in benchmark complexes, such as CDK2-Staurosporine and DHFR-Methotrexate, but under-represents the hydrogen-bonding networks. Our calibration bias analysis revealed that the model overpredicted strong binders and underpredicted weak binders, which are linked to data imbalances and heteroscedastic noise. A simple posthoc isotonic correction partially mitigated the bias. Latent space analysis showed that the model learned continuous binding affinity manifolds without split leakage. Our work highlights a novel architecture that offers unique insights into binding mechanisms via transformer-based cross-attention and is computationally inexpensive.
27 May 2026
Read appraisal →Journal of chemical theory and computation
Predicting and Decoding Allosteric Binding Sites Using Protein Language Models and Structure-Based Machine Learning: An Energy Landscape-Guided Explainable AI Framework
Computational prediction of allosteric binding sites in protein structures remains a persistent challenge, as these regulatory pockets evade detection by both sequence-based and structure-based algorithms. Both computational and physical origins of this predictive asymmetry remain insufficiently understood. In this study, we systematically examine the determinants of binding site predictability using a dual framework that integrates a fine-tuned protein language model and the structure-based method P2Rank as complementary tools probing a diverse data set of 453 human kinases, together with a physics-based interpretability layer derived from energy landscape frustration analysis. Both predictors exhibit a sharp and reproducible dichotomy on protein kinases, in which orthosteric ATP-binding sites can be identified with high precision, whereas allosteric sites are detected with substantially lower confidence across kinase structures and distinct conformational states. To decode this divergence, we deploy energy landscape-based explainable AI approach that integrates local frustration analysis as an independent physical interpretability layer, mapping predictive behavior to the underlying energetic organization of protein structures. This analysis reveals that predictive success is governed by the local energetic embedding of binding sites within the protein energy landscape. Orthosteric pockets are located in minimally frustrated basins that generate strong evolutionary and structural signatures, whereas allosteric pockets occupy predominantly neutrally frustrated zones associated with conformational plasticity and reduced evolutionary constraint. By integrating the prediction results with energy landscape analysis, our framework converts predictive performance into physically interpretable descriptors of binding site organization in protein kinases. These results establish energy landscape frustration as a potentially important determinant of algorithmic visibility and an interpretability layer providing a feasible strategy for diagnosing the limits of current prediction methods.
27 May 2026
Read appraisal →Journal of computer-aided molecular design
ADMET-vault: an interactive framework for real-time ADMET prediction and molecular optimization
Drug discovery pipelines face persistent challenges, as candidates that show promising results in early laboratory studies are found ineffective in clinical trials-often because their absorption, distribution, metabolism, excretion, and toxicity (ADMET) profiles were not well understood early enough. In this study, an interactive machine-learning based platform, ADMET-Vault, is presented, which brings ADMET prediction directly into the molecular design process, allowing researchers to assess and optimize drug candidates during the molecule design phase itself. ADMET-Vault integrates multiple molecular representations including physicochemical descriptors, molecular fingerprints, and graph neural network embeddings. The platform evaluated 12 benchmark datasets from Therapeutics Data Commons using scaffold-based validation. ADMET-Vault demonstrated consistent predictive performance across multiple ADMET endpoints. Strong results were observed for clearance related properties, intestinal absorption, hepatotoxicity, and aqueous solubility. Moreover, regression showed mean absolute error value that ranged from 0.28 (Caco-2 permeability, expressed as log cm/s) to 7.68 (plasma protein binding rate, expressed as % bound), with Spearman's rank correlation coefficients between 0.44 and 0.66. Classification tasks showed area under the receiver operating characteristic curve values ranged between 0.87 and 0.99. ADMET-Vault's main contribution is not just in its predictive accuracy. It combines reliable machine learning models with a live molecular editor. This allows researchers to see how a structural change impacts drug-like properties instantly. This interactive design-predict-optimize loop fills a long-standing gap in early-stage drug discovery workflows.
27 May 2026
Read appraisal →Journal of computer-aided molecular design
In silico discovery of novel small-molecule PD-L1 inhibitors through a multi-stage computational workflow integrating machine learning and molecular dynamics
Cancer immunotherapy targeting the PD-1/PD-L1 pathway has transformed modern oncology; however, developing small-molecule inhibitors as viable alternatives to monoclonal antibodies remains a major challenge. In this study, an integrated computational framework is established in which machine learning, molecular docking, molecular dynamics simulations, and binding free-energy calculations are combined to enable the discovery and optimization of novel PD-L1 inhibitors. A validated ML-QSAR model was constructed using the XGBoost algorithm (R2_train = 0.925, R2_test = 0.743) on a dataset of 74 known inhibitors with consistent assay conditions. Through virtual screening of FDA-approved drugs, Pralatrexate was subsequently identified as a promising repurposing candidate, demonstrating a higher predicted binding affinity than the reference inhibitors. Structure-based modification of Pralatrexate yielded the derivative D1, which exhibited improved computational binding properties across all evaluation methods. Molecular dynamics simulations indicated that the D1-PD-L1 complex achieved greater stability than both the free protein and the reference complex, with reduced RMSD fluctuations and preserved key interactions with tyrosine residues Tyr56(A/B). MM-GBSA calculations further confirmed D1's superior binding affinity (-86.21 kcal/mol vs. - 73.65 kcal/mol for the reference), and predicted IC50 values suggested enhanced inhibitory potential. This multi-stage computational workflow effectively integrates machine learning predictions with atomic-level binding analyses, providing a robust platform for accelerated drug discovery. The optimized derivative D1 thus represents a promising candidate for experimental validation and further development as a potential cancer immunotherapeutic agent.
26 May 2026
Read appraisal →Proceedings of the National Academy of Sciences of the United States of America
Predictions from deep learning propose substantial protein-carbohydrate interplay
Noncovalent interaction between proteins and carbohydrates (sugars, glycans) is the basis for biological functions from metabolic regulation to intercellular recognition. It is a grand challenge to identify the protein-carbohydrate interactomes in organisms. Direct experiments would require extensive libraries of glycans to distinguish binding from nonbinding proteins. Computational screening of proteins for carbohydrate binding potential provides an attractive alternative. Current estimates propose that <5% of proteins bind carbohydrates, a number that is not well established. We therefore developed a neural network, "Protein interaction of Carbohydrates Predictor" (PiCAP), to predict whether a protein noncovalently binds to a carbohydrate. We trained PiCAP on a manually curated dataset of known carbohydrate binders and proteins that we identified as likely not to bind carbohydrates (transcription factors, cytoskeletal components, and small-molecule-binding proteins). PiCAP achieves 90% balanced accuracy on protein-level predictions of carbohydrate binding/nonbinding. Using the same datasets, we developed Carbohydrate Protein Site Identifier 2 (CAPSIF2) to predict protein residues that interact noncovalently with carbohydrates. CAPSIF2 achieves a Dice coefficient of 0.57 on residue-level predictions on our independent test dataset, outperforming previous models. To demonstrate the models' biological applicability, we investigated human cell surface proteins and further predicted the likelihood of carbohydrate binding in six proteomes (Escherichia coli, Mus musculus, Homo sapiens, Saccharomyces cerevisiae, Caenorhabditis elegans, Drosophila melanogaster). PiCAP predicts that ~35 to 40% of proteins in these proteomes bind carbohydrates, with 75% of extracellular and cell surface proteins predicted to bind. The PiCAP predicted binders are enriched for functions including growth factor receptor binding, inflammation, and cell-cell adhesion.
26 May 2026
Read appraisal →PloS one
Computational models for the classification of antibody specificity using heavy chain features
BACKGROUND: Antibodies play a critical role in immune defense, with their antigen specificity primarily governed by the unique sequences of their heavy chains, rendering them invaluable tools in research and diagnostics. High-throughput sequencing technologies have facilitated comprehensive profiling of the immune repertoire, generating vast antibody sequence datasets that necessitate advanced analytical methods. METHODS: In this study, we utilized curated antibody sequences from NCBI databases to develop computational classification models for categorizing antibodies into predefined antigen classes. We extracted multifaceted features from the heavy chain sequences, encompassing physicochemical properties, structural composition, sequence order, and evolutionary information. These features were input into machine-learning classifiers to predict antigen specificity across five classes of antibodies: anti-dengue virus, anti-influenza virus, anti-tetanus bacillus, anti-SARS-CoV-2, and anti-Mycobacterium tuberculosis. RESULTS: Five tree-based machine-learning models were employed, with CatBoost achieving the highest accuracy of 0.7713. To further enhance predictive performance, we developed a stacking model leveraging multiple algorithms, resulting in an improved accuracy of 0.7803. Additionally, a Feature-Based Transformer deep-learning architecture was implemented, yielding an accuracy of 0.7399 and an F1-score of 0.6761. To elucidate the key determinants of antibody-antigen interactions, we applied the SHAP framework to assess feature importance. Among the top 30 contributing features, those representing sequence order and evolutionary information predominated, with amino acids such as cysteine (C), isoleucine (I), histidine (H), and phenylalanine (F) exhibiting notable SHAP values. Notably, cysteine (Cys) emerged as the most influential feature, underscoring its critical role in antibody structure and function. Specific antibodies contributed variably to these key features; for instance, the anti-tuberculosis antibody accounted for approximately 11% of a sequence order feature associated with alanine (A), while the anti-SARS-CoV-2 antibody contributed about 9.26% to a feature associated with isoleucine (I). CONCLUSIONS: Our study demonstrates the efficacy of machine-learning and deep-learning approaches in classifying antibodies into specific antigen categories, providing sequence-based insights into features associated with antibody specificity. These findings have significant implications for the mechanistic understanding, isolation, and development of potential therapeutic antibodies.
22 May 2026
Read appraisal →Cell systems
GFETM: Genome foundation-based embedded topic model for scATAC-seq modeling
Single-cell assay for transposase-accessible chromatin with sequencing (scATAC-seq) enables investigation of open chromatin landscapes at single-cell resolution. However, analyzing scATAC-seq data remains challenging due to inherent sparsity and noise. Genome foundation models (GFMs), pre-trained on extensive DNA sequence datasets, have demonstrated effectiveness in genome analysis. Because open chromatin regions (OCRs) harbor salient sequence features, we hypothesized that leveraging GFMs' sequence embeddings could enhance scATAC-seq modeling accuracy and generalizability. We introduce the genome foundation embedded topic model (GFETM), an interpretable deep learning framework combining GFMs with the embedded topic model (ETM) for scATAC-seq analysis. By integrating DNA sequence embeddings extracted by a GFM from OCRs, GFETM demonstrates superior accuracy and generalizability and captures cell-state-specific transcription factor (TF) activity with both zero-shot inference and attention-mechanism analysis. Finally, the topic mixtures inferred by GFETM reveal biologically meaningful epigenomic signatures of kidney diabetes. A record of this paper's transparent peer review process is included in the supplemental information.
22 May 2026
Read appraisal →Cell systems
Learning multi-cellular representations of single-cell transcriptomics data enables characterization of patient-level disease states
Single-cell RNA sequencing (scRNA-seq) is a prominent tool for studying human disease biology. The availability of massive scRNA-seq datasets and advanced machine learning has driven the development of single-cell foundation models that provide versatile cell representations based on expression. To understand disease states, we need to account for entire tissue ecosystems, simultaneously considering many different interacting cells. Here, we tackle this challenge by generating patient-level representations derived from multi-cellular expression contexts measured with scRNA-seq of tissues. We develop PaSCient, a machine learning model that employs a multi-level representation learning paradigm and provides importance scores at the cell and gene levels for fine-grained analysis of disease characteristics. We apply PaSCient to a large-scale scRNA-seq atlas of 12.5 million cells from over 2,700 patients. Comprehensive benchmarking demonstrates the superiority of PaSCient in disease classification and multiple downstream applications, including dimensionality reduction, feature prioritization, and patient subgroup discovery.
22 May 2026
Read appraisal →Cell reports methods
A tri-modal contrastive learning framework for protein representation learning
Protein foundation models, particularly protein language models, have shown strong success in learning meaningful protein representations using transformer architectures pretrained on large-scale datasets through self-supervised learning. These representations have proven effective for downstream tasks such as predicting protein functions and properties. However, most existing models focus solely on amino acid sequences, overlooking other informative modalities such as 3D structures and literature text. While some recent efforts incorporate multiple modalities, they often suffer from limitations in modality coverage or training strategy. To address this gap, we propose a multimodal pretraining framework that integrates three complementary modalities-protein sequences, structures, and literature text. Our method uses the sequence modality as an anchor and aligns the other two modalities to it via contrastive learning, enabling the model to capture richer and more holistic protein representations. Across a diverse set of downstream tasks, ProteinAligner outperforms state-of-the-art foundation models in predicting protein functions and properties.
20 May 2026
Read appraisal →Journal of virology
Is the molecular microenvironment of the latent HIV reservoir predictable using deep learning approaches?
Although current antiretroviral therapy effectively suppresses viral replication in infected individuals, infections remain incurable. One of the main reasons for this is the uncertainty of the molecular microenvironment of the latent HIV reservoir. If this microenvironment can be precisely targeted, the effectiveness of current antiretroviral therapy can be leveraged. To better depict the potent microenvironments of the latent HIV reservoir, it is crucial to first understand the mechanisms that make some viruses transcriptionally active and others silent in the first place. It is currently feasible to predict gene activity from DNA sequence, coupled with epigenetics, using deep learning approaches. Substantial progress has been made in creating multidimensional data sets related to the HIV latent reservoir. The implementation of deep learning to mine potent attributes, enabling the characterization and prediction of the microenvironment of the latent HIV reservoir, is thus imperative. This minireview discusses potent features that can be integrated into deep learning methodologies to predict the likelihood of the potential microenvironment from different tiered levels of the latent HIV reservoir organization, and proposes the principal workflow to tackle this question using deep learning approaches.
20 May 2026
Read appraisal →Cell reports methods
A lightweight TcrLM model predicts T cell receptor and epitope binding specificity.
Immune responses depend on specific interactions between T cell receptors (TCRs) and peptides presented by antigen-presenting cells (APCs). The vast diversity of the TCR repertoire makes accurate prediction of TCR-antigen binding specificity highly challenging. Here, we present a lightweight, masked language model, tcrLM, to address this problem. We pretrain tcrLM on a large-scale TCR CDR3 sequence dataset and use the pretrained encoder to extract informative features for pTCR binding prediction. tcrLM achieves competitive performance on hold-out and external test sets and shows comparatively robust zero-shot generalization relative to several baselines on a large, unseen-COVID-19 peptide set. The model effectively captures biochemical properties and positional preferences of amino acids within TCR sequences. In an exploratory melanoma cohort, the predicted TCR-neoantigen binding scores correlate with immunotherapy response and clinical outcomes. These results highlight the potential of tcrLM for advancing immunotherapy and personalized medicine.
20 May 2026
Read appraisal →Computers in biology and medicine
Generative deep learning-driven de novo design of targeted MAP4K6 inhibitors
The discovery of selective small-molecule inhibitors for pharmacologically underexplored kinases remains a critical barrier to precision drug development, particularly under data-sparse conditions, where ligand annotations are scarce. Here, we present MolEvoRNN, a generative deep learning (GDL) model that enables the de novo design of chemically valid, structurally diverse, and target-adapted compounds in low-data regimes. Through transfer learning and temperature-controlled sampling, MolEvoRNN efficiently navigates the kinase-relevant chemical space while preserving scaffold novelty, drug-likeness, and synthetic tractability. Applied to MAP4K6, a poorly characterized serine/threonine kinase implicated in hepatocellular carcinoma (HCC), MolEvoRNN generated >49,000 chemically valid molecules with >80% novel Bemis-Murcko scaffolds. MolEvoRNN outperformed the baseline GDL models across key generative metrics, including diversity (1.00), uniqueness (1.00), novelty (1.00), and internal diversity (0.869), demonstrating broad chemical exploration with minimal redundancy. A multilayered virtual screening cascade combining QSAR modeling, pharmacophore mapping, molecular docking, MM-GBSA binding energy estimation, and 1000 ns molecular dynamics simulations identified three lead candidates MINK-5730, MINK-13510, and MINK-16516 with high predicted affinity, favorable ADMET properties, and robust dynamic stability. Unlike end-to-end optimization pipelines, MolEvoRNN decouples generation from evaluation, promoting methodological transparency and rigorous post-hoc triaging. This study establishes a scalable, generalizable paradigm for rational small-molecule design in underexplored target spaces, bridging generative deep learning with structural pharmacology to accelerate early-phase therapeutic discovery.
17 May 2026
Read appraisal →ACS synthetic biology
Prompt-Only Gene-Circuit Modeling with a Large Language Model Simulates the synNotch Spheroids
Morphogenesis emerges from gene regulatory circuits that translate biochemical cues into spatially organized cell fates; yet, constructing quantitative simulations of these processes remains technically demanding. Here, we present a prompt-only pipeline in which a large language model (LLM) automatically generates CompuCell3D (CC3D) simulations directly from natural-language biological specifications. Using the canonical synNotch spheroid assay as a benchmark, we show that the LLM-generated code captures key features of contact-dependent signaling, cadherin-mediated adhesion, and multicellular sorting dynamics without manual programming. Baseline simulations exhibited the expected qualitative ordering of inducer-responder activation but revealed fragmentation into multiple spheroids, implicating premature reciprocal signaling as a kinetic failure mode. Incorporating a cumulative contact-duration threshold for synNotch activation rescued robust single-spheroid formation, suggesting that sustained cell-cell interactions are a critical design principle for this pattern. Our results establish conversational LLM guidance as a powerful accelerator for multiscale modeling, collapsing the steep learning curve of CC3D into a dialog-driven workflow. This framework enables nonspecialists to test gene-circuit hypotheses in silico and provides a generalizable route toward rapid, mechanism-level exploration of developmental design principles.
16 May 2026
Read appraisal →Journal of molecular modeling
Carbonic anhydrase under the computational lens: a review of advances, challenges, and future directions
CONTEXT: Carbonic anhydrases (CAs) are zinc metalloenzymes responsible for the catalysis of the reversible hydration of carbon dioxide and play pivotal roles in many physiological and pathological processes. The complexity of CA isoforms with distinct expression patterns and catalytic characteristics has aroused fierce interest in the development of isoform-selective inhibitors and activators for therapeutic purposes and diagnostics. In this review, a systematic overview of computational and theoretical methods utilized to determine the structure, mechanism, inhibition, and activation of CAs was given. It reviews the evolution and utilization of computational methods in deciphering catalytic and proton transfer mechanisms as well as active-site interactions into atomic resolution and also features recent developments in drug design and screening strategies engineered for CA isoforms, particularly those involved in cancers and Alzheimer's disease. This study can be a reference for multidisciplinary researchers and inspire the future rational design of CA modulators. METHODS: This review summarizes the literature that employed QM and QM/MM by using like B3LYP, BE0-D3BJ, and MD simulations, GROMACS and Desmond packages, and various force fields: OPLS-AA, AMBER14, and CHARMM36. Molecular docking was performed using such as Glide, AutoDock Vina, and GOLD. Machine learning and AI techniques including SVM, LightGBM, SHAP, and neural networks were also reviewed. This review addresses the limitations in these models as well as prospects with newly emerging instruments such as polarizable force fields and real-time DFT as well as explainable AI frameworks, too. Incorporating data-driven insights with physics-based simulations, the next era of work is primed to provide predictive, scalable platforms for rational design of next-generation CA modulators.
15 May 2026
Read appraisal →PLoS computational biology
Teaching artificial intelligence through drug-drug interaction clustering analysis: Integrating project-based learning and large language models.
In recent years, artificial intelligence (AI) has increasingly influenced daily life and scientific research. Traditionally, AI-related courses have targeted computer science majors, while systematic instructional opportunities for early-stage undergraduates from non-computing backgrounds remain limited. To bridge this gap, we developed an AI course that integrates project-based learning with large language models (LLMs). Specifically, we designed four progressive assignments based on our research project (i.e., drug-drug interaction network clustering analysis). The course does not require prior knowledge of pharmacology or programming. Instead, LLMs are used as assistive tools to support programming, data analysis, and result interpretation. Students engage in a complete workflow, including data curation, algorithm implementation, and critical evaluation of results. Preliminary feedback shows that this approach supports the development of problem-solving skills and increases student engagement. This study provides a framework for integrating LLMs into project-based learning. We believe that this teaching proposal will be valuable and inspiring for educators seeking to design or enrich similar courses.
15 May 2026
Read appraisal →Journal of computer-aided molecular design
Dynamic stability-driven machine learning improves binding pose identification on human serum albumin.
Accurate identification of ligand binding modes on human serum albumin (HSA) is important for understanding drug-protein interactions that influence pharmacokinetics, including site-specific competition and protein-binding displacement. However, because HSA is a large, multisite binding protein that permits diverse binding configurations, conventional molecular docking often struggles to distinguish crystal-like poses from nonspecific surface-binding poses using static scoring functions alone. Here, we developed a machine-learning framework for HSA binding-pose identification that incorporates dynamic stability information derived from short molecular dynamics (MD) simulations. Time-dependent interaction patterns observed during MD trajectories, including contact persistence, distance fluctuations, and interaction energy-related descriptors, were quantified as interpretable features and integrated with docking scores. Using Leave-One-PDB-Out cross-validation on 31 HSA-ligand crystal structures, models incorporating MD-derived features significantly outperformed the docking score-only model under both pose-level (root-mean-square deviation [RMSD] ≤ 2.5 Å) and site-level (RMSD ≤ 5.0 Å) criteria (Holm-corrected p < 0.001). External validation using a temporally independent nateglinide-HSA complex further supported the utility of the framework for discriminating plausible binding modes within the HSA setting. Notably, informative dynamic features were obtained from 5 ns MD simulations, indicating that early-stage dynamic behavior is sufficient to improve pose discrimination. These findings show that short MD simulations can provide practically useful and interpretable features for machine-learning-based binding-pose identification on HSA. While broader validation will be required to establish applicability beyond HSA, the proposed framework offers a practical strategy for improving binding-mode analysis in pharmacokinetically relevant HSA studies.
15 May 2026
Read appraisal →Molecular genetics and genomics : MGG
Comparison between a conventional tool and deep learning models for RNA velocity analysis of scRNA-Seq data
Single-cell RNA sequencing (scRNA-Seq) enables analysis of gene expression at single-cell resolution. RNA velocity analysis infers the temporal dynamics of transcriptional states from the relative abundances of spliced/unspliced mRNA quantified via scRNA-Seq. Classical RNA velocity approaches, such as scVelo, implement gene-specific kinetic modeling. Deep learning methods including DeepVelo, VeloVI, LatentVelo, SymVelo, and scTour are based on variational autoencoders (VAEs), which allow to enhance the robustness and accuracy by leveraging nonlinear latent representations. Here, we systematically evaluated the performance of deep learning RNA velocity tools by comparing with the scVelo dynamical model to access the possible advantages of VAE-base methods. For this purpose, public datasets (GSE149689 and GSE203233) were initially processed using a standard scRNA-Seq pipeline. Comparisons among results of selected velocity tools were conducted using cosine similarity of velocity vectors to assess directional concordance, and by mean squared error analysis of trajectory continuity for the deep learning models. Overall, VAE methods produced significant, richer, and more directionally coherent and consistent velocity fields than the classical model. Our findings indicate that deep learning models provide more consistent and biologically plausible cell-state trajectories, although at the expense of higher computational demands and reliance on accurate splicing quantification. Altogether, our results underscore the relevance of VAE-based frameworks to advance RNA velocity analysis while highlighting the need for careful preprocessing.
14 May 2026
Read appraisal →International journal of oral science
Machine learning-driven discovery of therapeutic nucleoside hydrogels for periodontitis
Supramolecular hydrogels hold significant potential in drug delivery and tissue engineering, with standing out for their unique properties. Despite their promise, predicting nucleoside bioactivity remains challenging. This study aims to predict the biological activity of nucleosides to guide the rational synthesis of hydrogels. Specifically, nine predictive models and databases for various biological activities were built with feature-selected machine learning methods including decision trees, logistic regression, random forest, and extreme gradient boosting. Then, the Molecular Bioactivity Specificity Index (MBSI) was introduced to gauge the primary bioactivity of nucleoside derivatives, and the Composite Molecular Attribute Score (CMAS) was devised to measure the overall performance of nucleoside derivatives. Subsequently, screening strategies for bioactive nucleoside hydrogels were established, and two candidate hydrogels (GMP and dGMP) with high hydrogel-forming ability, biocompatibility, and antibacterial activity were identified. Finally, two hydrogels were validated for antibacterial treatment of periodontitis. This study highlights the feasibility of ML-based strategies and MBSI/CMAS in rationally designing bioactive nucleoside hydrogels for biomedical applications. The discovery of GMP and dGMP hydrogels and their successful validation in periodontitis models highlight the potential of this strategy for developing targeted therapies for oral diseases.
14 May 2026
Read appraisal →Journal of molecular modeling
Exploring human GABA transporter 3 binders for epilepsy through quantum reactivity analysis, molecular dynamics, machine learning prediction, and network pharmacology
CONTEXT: Drug-resistant epilepsy (DRE) remains a major therapeutic challenge, affecting millions of patients globally who do not respond to conventional antiepileptic medications. Recent evidence highlights GABA transporter 3 (GAT3), a glial subtype responsible for regulating extracellular GABA levels, as a promising molecular target for seizure control. Despite its therapeutic potential, selective and potent GAT3 inhibitors remain largely unexplored. This study aims to identify novel GAT3 inhibitors through an integrative computational strategy combining quantum chemical, molecular dynamics, and machine learning approaches to accelerate antiepileptic drug discovery. The cryo-EM-resolved human GAT3-SNAP-5114 complex structure was used for virtual screening, leading to the identification of three highest-ranking compounds, viz. Comp_1, Comp_2, and Comp_3, from the diverse chemically optimized screening library. Analysis of their quantum chemistry profiles indicated HOMO-LUMO energy gap values of 4.14 eV, 4.21 eV, and 3.26 eV, respectively, suggesting them to be maximally reactive with favorable electronic properties for interacting with biological molecules. Molecular dynamics simulations were performed in a fully hydrated POPC lipid bilayer using the AMBER force field, providing a physiologically relevant environment for assessing ligand binding and transporter dynamic, revealed stable binding modes, with the highest structural integrity or least deviations in the structure with Comp_2, having RMSD values ranging from 2.5 to 3.0 Å, with strong hydrogen bonding, while the binding mode was more flexible with strong anchorage for Comp_3, with its mode involving diverse interactions in the complex. Principal component and free energy landscape analyses identified energetically favorable, ligand-induced conformational states. Machine learning-based pIC₅₀ predictions using the AdaBoost regressor yielded values of 7.95 (Comp_1), 7.98 (Comp_2), and 7.63 (Comp_3), all comparable or superior to the reference inhibitor (7.89). A novel approach for network pharmacology integrating detailed drug-gene interaction profiling was additionally used to improve systems-level understanding of the discovered compound. These findings validate the potential of Comp_1-3 as lead GAT3-targeting antiepileptic agents. METHODS: A docking-based virtual screening of a structurally diverse compound library was performed against the cryo-EM structure of human GAT3 in complex with the reference inhibitor SNAP-5114. The top-scoring ligands were subjected to density functional theory (DFT) analysis to assess electronic properties such as HOMO-LUMO energy gaps, reflecting potential binding reactivity. Lead candidates were refined, docked, and evaluated through 500 ns molecular dynamics (MD) simulations, analyzing structural stability via RMSD, RMSF, hydrogen bonding, and interaction network profiles. Principal component analysis and free energy landscape evaluations were used to explore ligand-induced conformational dynamics. Furthermore, machine learning (ML) models-particularly the AdaBoost algorithm-were employed to predict pIC₅₀ values, confirming bioactivity trends. Finally, network pharmacology analysis was conducted to elucidate gene targets and pathway interconnections, revealing the multi-pathway involvement of top candidates in seizure regulation. This integrative computational framework underscores the potential of GAT3-focused modelling and ML-guided screening for rational antiepileptic drug design.
14 May 2026
Read appraisal →Journal of computer-aided molecular design
Ensemble learning-guided discovery of anti-tuberculosis phytochemicals: computational prediction and mechanistic insights
Tuberculosis (TB) remains a major global health challenge driven by persistent Mycobacterium tuberculosis infection and increasing drug resistance. Phytochemicals represent a structurally diverse and underexplored chemical space for anti-TB drug discovery, yet systematic prioritization strategies integrating machine learning and structure-based validation are limited. A curated phenotypic anti-TB dataset of 425,180 compounds was used to train ensemble ExtraTrees models based on ECFP4 fingerprints and physicochemical descriptors. The models achieved strong predictive performance (ROC-AUC up to 0.983; MCC up to 0.871). SHAP analysis enabled mechanistic interpretation by identifying the key molecular descriptors and fingerprint features driving anti-TB activity predictions. The validated ensemble was applied to screen 4707 phytochemicals, yielding 3209 predicted actives, of which 778 satisfied applicability domain criteria. High-confidence candidates were subsequently evaluated by molecular docking against twelve structurally validated essential M. tuberculosis targets spanning cell wall biosynthesis, energy metabolism, nucleotide synthesis, and cofactor pathways. Docking analysis identified 486 phytochemicals with favorable predicted binding affinities, including 193 compounds exhibiting multi-target engagement. Several top-ranked candidates reproduced canonical interaction patterns of co-crystallized inhibitors, supporting mechanistic plausibility. This integrated chemoinformatics and structure-based framework enables robust prioritization of phytochemicals with biologically meaningful and multi-target antitubercular potential. The study provides a computationally grounded strategy for accelerating lead identification against drug-resistant TB.
13 May 2026
Read appraisal →Journal of chemical information and modeling
MechMemDyn: Coupling Frustration Analysis with Membrane Dynamics to Target the TREM2-DAP12 Complex Interface
The transmembrane complex formed by TREM2 (Triggering Receptor Expressed on Myeloid cells 2) and DAP12 (DNAX-activating protein of 12 kDa) constitutes a pivotal therapeutic target for Alzheimer's disease. However, the intrinsic plasticity of its interface poses a formidable challenge for structure-based discovery of protein-protein interaction stabilizers (PPI stabilizers). Conventional approaches, ranging from static molecular docking to geometric deep learning, fail to capture the subtle interfacial energetics required for stabilizing this assembly, often leading to erroneous activity predictions. Here, we present MechMemDyn, a novel predictive framework that uniquely integrates protein frustration analysis with membrane-embedded molecular dynamics (MD) simulations. To our knowledge, this is the first systematic application of frustration analysis to rationalize PPI-stabilizing activity and guide ligand design. By mapping the local frustration landscape, we identify critical, minimally frustrated contact networks at the protein interface, which are essential for stabilizing the PPI. We demonstrate that the ability of a ligand to dampen distance fluctuations within these key networks, a metric rooted in thermodynamic rigor, correlates strongly with experimental potency. This method outperforms conventional static docking, AI-driven dynamic docking approaches, and standard molecular simulations, providing a more accurate and reproducible basis for cross-ligand comparison. This work not only resolves the intractability of the TREM2-DAP12 complex but also establishes a physics-driven paradigm for targeting dynamic transmembrane interfaces via frustration-optimized PPI stabilizers.
13 May 2026
Read appraisal →Journal of chemical information and modeling
HLABrew for Human Leukocyte Antigen Class I-Presented Epitope Recognition and Mimotope Discovery
Accurate identification of human leukocyte antigen (HLA) class I-presented epitopes is essential for developing personalized vaccines and immunotherapies. Here, we present HLABrew, a unified deep learning framework for HLA class I epitope prediction and allele-specific mimotope design. By integrating the Transformer, variational autoencoder, and dual cross-attention, HLABrew achieves state-of-the-art performance with hierarchical and allele-aware representations. HLABrew accurately deconvolves multiallelic immunopeptidomics data to recover allele-specific binding motifs and expand epitope coverage. Leveraging its generative capacity, HLABrew further designs mimotopes guided by learned peptide-HLA binding preferences, with the potential to enhance antigen presentation and provide candidates for downstream immunogenicity evaluation.
13 May 2026
Read appraisal →Journal of chemical information and modeling
A Transformer for Reaction-Aware Compound Explorations with GFlowNet in QSAR-Guided Molecular Design
In drug discovery tasks, achieving a balance between high biological activity toward therapeutic targets and synthetic chemical feasibility is critically important. While the recently proposed deep learning-based molecular generation models have enabled explorations of vast chemical spaces, most existing approaches do not consider synthetic routes for generated compounds. To address this issue, TRACE-GFN is proposed for molecular optimization; this method incorporates chemical reaction pathways into a quantitative structure-activity relationship (QSAR)-guided molecular design procedure. The method integrates a transformer model to explicitly learn chemical reactions with a generative flow network (GFlowNet) that efficiently samples diverse candidates. In benchmark experiments involving dopamine receptor D2 (DRD2), AKT serine/threonine kinase 1 (AKT1), and C-X-C motif chemokine receptor 4 (CXCR4), TRACE-GFN demonstrated the ability to identify compounds with high QSAR values while maintaining strong diversity, outperforming the existing molecular generation models. These results demonstrate that the proposed model can efficiently explore promising compounds while accounting for real-world chemical reactions. The source code is publicly available under an MIT license at https://github.com/sekijima-lab/TRACE-GFN.
13 May 2026
Read appraisal →Journal of chemical information and modeling
mDADGAN: Predicting miRNA-Drug Response Associations Using lncRNAs and a Diffusion-Based Generative Adversarial Network
MicroRNAs (miRNAs) are key regulators of anticancer drug response, including sensitivity and resistance. Accurate prediction of miRNA-drug response associations is important for understanding response mechanisms and guiding targeted therapy. Existing prediction methods are difficult to fully capture complex high-order biological interactions and are sensitive to data sparsity and class imbalance. In this paper, we present an miRNA-drug response association prediction method called mDADGAN by constructing a diffusion-based generative adversarial network. First, we curated a large-scale and biologically informative data set from four public data resources by aggregating functional associations to facilitate model training. Second, we built a miRNA-lncRNA-drug interaction network by integrating lncRNA into the miRNA-drug network to address data sparsity and employed heterogeneous graph convolution to extract latent features from the interactions among miRNAs, lncRNAs, and drugs. Third, we applied diffusion-based perturbation to denoise and enhance the robustness of feature representations. Finally, we performed cost-sensitive learning to optimize the objective function in an adversarial training process to address class imbalance. The experimental results on four data sets demonstrated that, compared with the existing six methods, mDADGAN achieved higher values for AUC, AUPR, ACC, Recall, and F1-score. The case study demonstrates that integrating lncRNA information effectively enhances miRNA-drug response association prediction. mDADGAN can not only accurately identify known associations in the data sets but also predict previously unreported candidate associations supported by published evidence. Overall, our method mDADGAN has the ability to identify potential miRNA-drug response associations and can provide guidance for future biological experiments to enhance experimental efficiency and reduce experimental costs.
13 May 2026
Read appraisal →Journal of chemical information and modeling
scFNSA: A Factorized Node-Set Attentive Framework for Single-Cell Multi-Omics Integration
Single-cell multi-omics technologies enable simultaneous interrogation of transcriptional and epigenomic states at single-cell resolution, providing powerful means to dissect cellular heterogeneity and regulatory mechanisms. However, effective integration of single-cell multi-omics data remains challenging due to extreme sparsity, high dimensionality, and cross-modal heterogeneity, which often lead to distorted similarity structures, information loss, and limited biological interpretability in existing methods. Here, we propose scFNSA, a factorized node-set attentive framework for integrative analysis of single-cell multi-omics data that explicitly decouples modality-specific feature learning from graph-based relational modeling. scFNSA first learns aligned low-dimensional representations for individual modalities through multi-view variational inference and subsequently models complex cellular dependencies via attention-guided graph learning with structured node masking. This design enables robust cross-modal structural alignment while effectively mitigating noise propagation and oversmoothing commonly encountered in deep graph neural networks. Across multiple single-cell multi-omics datasets, scFNSA consistently improves integrative representation quality and cell type resolution compared with state-of-the-art approaches. By decoupling feature extraction from relational inference, scFNSA provides a robust and interpretable framework for single-cell multi-omics integration, facilitating more accurate characterization of cellular states and underlying regulatory landscapes.
13 May 2026
Read appraisal →Journal of chemical information and modeling
DRHIN: An Integrated and Interactive Web Server for Drug Repositioning
Drug repositioning (DR) identifies new therapeutic uses for approved drugs, reducing development burdens and offering safer treatment options for patients. While high-throughput technologies generate complex, large-scale multiomics data, existing DR tools struggle to comprehensively analyze the resulting biological networks. To address this challenge, we present DRHIN, an integrated, interactive web server for DR over heterogeneous information networks (HINs) using advanced deep learning techniques. DRHIN integrates transcriptomics, proteomics, and microbiome data, incorporating eight biological entities and 19 association types to build diverse HINs and elucidate the underlying molecular mechanisms. It includes 19 state-of-the-art graph representation algorithms, enabling flexible training, comparison, and evaluation of heterogeneous network data. The platform provides a code-free portal supporting three key predictive tasks: discovering drug-disease associations, repurposing existing drugs for new indications, and identifying potential therapies for specific diseases, making analyses accessible and reproducible. Leveraging high-performance computing, DRHIN efficiently processes million-scale networks, ensuring practical applicability in real-world scenarios. The web server is freely accessible at http://drhin.tianshanzw.cn.
13 May 2026
Read appraisal →Journal of chemical information and modeling
Artificial Intelligence for Predicting Small-Molecule Bioactive Conformations
Precise prediction of bioactive conformations represents a central challenge in drug discovery. Traditional computational methods often rely on the assumption that the global minimum energy state corresponds to the bioactive conformation─a hypothesis with inherent limitations. Data-driven artificial intelligence (AI) models have revolutionized research methodologies in this domain. By learning the implicit probability distributions of conformations, AI demonstrates the potential to achieve high-precision predictions of bioactive structures. This review systematically surveys technical advancements ranging from geometric regression to generative and conditional generative models, analyzing their effectiveness in benchmarks and de novo design experiments. We deeply explore critical challenges, including data sparsity, sampling efficiency, generation capability, and interpretability. Finally, we provide an outlook on bridging the gap from "structural prediction" to "functional design" through the construction of dynamic data sets, the development of efficient algorithms, and the implementation of human-in-the-loop collaborative decision-making systems.
13 May 2026
Read appraisal →Neuroscience
In silico neuronal morphology classification: A systematic review
Advances in connectomics and the characterization of neuronal diversity have been fundamental to understanding how the brain works. Defining a taxonomy is still challenging and requires complex computational methods. In this paper, we present a systematic review of the state-of-the-art methods for neuronal morphology classification, shedding light on trends and future directions. We address three core research questions: (1) What computational methods are used to classify neuronal morphologies? (2) What kind of representation do the classification models work on? (3) How are the classification models evaluated? We applied a structured protocol including peer-reviewed studies published since 2018. We collected 840 papers from five databases, published between 2018 and 2024. We assessed risk of bias using predefined quality criteria on study design, methodology, and reporting; 35 studies met these standards and were included to extract information and consolidate the results. The results show that Artificial Neural Networks are currently the main method for classifying neurons based on morphometric features (21 papers). Representations based on microscopy images for extracting morphometric features (30) are commonly used. Accuracy (29) is the predominant evaluation measure. Our research findings contribute to the body of knowledge by providing an up-to-date perspective on neuronal classification, which can guide future research in neuroscience and related fields.
11 May 2026
Read appraisal →Journal of computer-aided molecular design
NeuroPred-GMC: a dual-branch deep learning architecture for neuropeptide prediction based on gated dilated convolutional network and multi-scale convolutional network
Neuropeptides are multifunctional signaling molecules in the nervous system. By modulating synaptic transmission and integrating physiological systems, they influence a broad range of functions from pain perception to emotional regulation. Predicting neuropeptides can rapidly expand the library of potential therapeutic targets, thereby providing novel candidate molecules for drug development in areas such as analgesics, anti-anxiety medications, and weight-loss drugs. Traditional experimental methods are extremely time-consuming, labor-intensive, these promising alternative computational methods have emerged. In this study, a dual-branch deep learning architecture for neuropeptide prediction known as NeuroPred-GMC are built up based on gated dilated convolutional network with ESM-2 feature representation and multi-scale convolutional network with Prot-T5 feature representation. Dilated convolution exponentially enlarges the receptive field via increased dilation rates, gating mechanism enables dynamic, selective feature enhancement and noise suppression, and multi-scale convolution captures multi-level contextual information. On the independence test set, the accuracy of 93.24%, Sn of 93.69%, Sp of 92.79%, Pre of 92.86%, MCC of 0.8649 and the auROC of 0.9667 are obtained. The experimental results through cross-validation and independent test demonstrate that the proposed model has good robustness and generalizability, and can serve as a supplemental candidate predictor. The source datasets and codes can be freely available at https://github.com/yunyunliang88/NeuroPred-GMC .
11 May 2026
Read appraisal →Genome research
Incorporating valuable prior knowledge to improve deep learning prediction of genetic perturbation responses
Genetic perturbation response prediction plays a critical role in virtual cell research, yet the performance of current deep learning models still leaves room for improvement. In this study, we present a prior-guided response inference model (PRIM) that leverages a valuable priori knowledge of gene expression in control cells to model the effects of perturbations at the cellular level. This allows PRIM to predict the amount of change in gene expression after perturbation to better approximate the real situation. Compared with existing deep learning approaches, PRIM achieves superior performance across multiple data sets and notable advantages in predicting combinatorial perturbation responses. Moreover, it is more lightweight than current deep learning models and enables faster forward inference. Importantly, PRIM effectively captures nonadditive genetic interactions and shows the potential to uncover associations between combinatorial perturbations and new biologically meaningful phenotypes. These findings provide new insights into the application of deep learning for predicting cellular responses to genetic perturbations.
11 May 2026
Read appraisal →Briefings in bioinformatics
Comprehensive evaluation of artificial intelligence-empowered approaches for protein-aptamer complex prediction
Drug discovery is a time-consuming, expensive, and high-risk process. Recent advances in artificial intelligence (AI) have enabled major breakthroughs in small-molecule and protein therapeutics. However, AI-driven design of aptamer drugs remains largely unexplored. Aptamers are short (15-100 nt) single-stranded DNAs or RNAs that exhibit high binding affinity, high specificity, and low immunogenicity, making them promising candidates for disease (such as cancer) therapeutics. Compared with protein-ligand or protein-protein systems, protein-aptamer complexes are under-represented in public structural databases, and aptamers themselves are highly flexible and relatively large molecules. These characteristics present distinct challenges for AI-based structural modeling. Here, we systematically evaluate recent AI frameworks, including AlphaFold3, Chai-1, Boltz-2, and RoseTTAFold2NA, along with a template-based approach, in predicting protein-aptamer complex structures and estimating binding free energies. We establish an independent benchmark to assess their performance in structural accuracy, stability, and energetic consistency. This study provides a foundation for the application of AI in aptamer drug design and offers a reference framework for future research in nucleic-acid therapeutics and biomolecular modeling.
10 May 2026
Read appraisal →Briefings in bioinformatics
MLN2SVG: domain-aware spatially variable gene detection using contrastive variational autoencoder and multi-level neighbor search
Spatial transcriptomics (ST) technologies have transformed our ability to examine gene expression within intact tissues, yet accurately identifying spatially variable genes (SVGs) remains challenging due to spatial heterogeneity, data sparsity, and incomplete modeling of domain-level dependencies. To address these limitations, we propose MLN2SVG, a domain-aware framework that integrates contrastive variational autoencoding with a multi-level neighbor (MLN) search algorithm to jointly learn tissue domains and SVGs. MLN2SVG constructs a weighted spatial graph to capture both local and long-range spatial relationships, employing a deep contrastive variational autoencoder to align augmented and original data representations while preserving biological diversity. The MLN algorithm dynamically expands neighborhood connectivity to mitigate sparsity and enhance domain coherence. Across multiple human and mouse ST datasets, including dorsolateral prefrontal cortex, breast cancer, and brain tissues, MLN2SVG consistently outperformed existing methods in clustering accuracy, robustness, and biological interpretability. Notably, in breast cancer tissues, MLN2SVG uncovers fine-grained spatial organization of tertiary lymphoid structures, delineating region-specific immune architectures spanning intratumoral, tumor-edge, and extratumoral compartments. Through the integration of spatial domain discovery and SVG detection, MLN2SVG delivers a robust and biologically interpretable framework for uncovering the molecular and structural complexity of tissue organization.
10 May 2026
Read appraisal →Protein science : a publication of the Protein Society
AlphaFold-driven discovery of oxysterol-binding protein-related protein-phosphoinositide 3-, 4-, and 5-phosphatase interactions using new generation confidence scores
Non-vesicular lipid transport contributes to the regulation of membrane composition and organelle function at membrane contact sites. OSBP-related proteins (ORPs) are central to this process, yet their interaction networks remain incompletely defined. Here, we systematically screened potential interactions between ORPs and phosphoinositide 3-, 4-, and 5-phosphatases using AlphaPulldown2, AlphaFold2-Multimer, and AlphaFold3. We established a protocol for model generation by combining AlphaFold2-Multimer predictions (including five-replicates) with an AlphaPulldown2 interaction screen across around 200 protein pairs, and with AlphaFold3 predictions including lipid-bound and multimeric assemblies. Interface confidence was assessed for consistency using the weighted ipTM + pTM metric, actifpTM, new generation ipSAE scoring, and FoldSeek-Multimer clustering. We further evaluated the protein pairs' biological plausibility based on subcellular localization data, in silico membrane insertion, evolutionary conservation via ConSurf, and protein binding interface analysis using the deep learning tool PeSTo. This integrative protocol uncovered functionally conserved binding modes in the SAC1 lipid phosphatase with the ORP family, particularly with ORP11, and predicted functionally relevant protein-lipid interfaces.
9 May 2026
Read appraisal →Computers in biology and medicine
A decoction-inspired genetic algorithm and PPO reinforcement learning for intelligent molecular discovery: Anti-colorectal cancer candidates from the tiao-pi AnChang formula
BACKGROUND: Colorectal cancer (CRC) remains among the leading causes of cancer-related mortality worldwide. The TiaoPi AnChang Decoction (TPACD), a traditional Chinese herbal formula, has shown potential anti-CRC activity in preclinical studies (in vitro and in vivo); however, its active chemical basis and key target interactions remain unclear. The complex, multicomponent nature of TPACD poses significant challenges for systematically identifying and optimizing compounds. PURPOSE: This study aimed to establish an intelligent molecular discovery framework that integrates a decoction-inspired genetic algorithm (EGA) with proximal policy optimization (PPO)-based reinforcement learning to simulate the dynamic transformation and recombination processes underlying traditional herbal decoctions and to generate and prioritize in silico candidate molecules for further anti-CRC evaluation. STUDY DESIGN: A parallel molecular generation strategy was developed, comprising three algorithmic pathways, namely, the basic molecular generation algorithm (BMGA), enhanced molecular generation algorithm (EMGA), and intelligent molecular generation algorithm (IMGA), which represent progressive levels of structural diversity, biological relevance, and optimization capacity. METHODS: The EGA simulated the selection-transformation-recombination principles of decoction through adaptive mutation, fragment recombination, and a diversity-preserving selection procedure. PPO-based reinforcement learning further refined the molecular properties via a reward-guided exploration of the chemical space. The BMGA enabled baseline molecular generation; the EMGA incorporated affinity-based selection; and the IMGA achieved multiobjective optimization, integrating drug likeness, novelty, and toxicity filtering. The candidate molecules were validated by affinity prediction, QED scoring, and molecular docking. RESULTS: All three methods generated chemically valid and pharmacologically enhanced molecules relative to the original TPACD components. The EMGA provided improved biological affinity and exploration efficiency, whereas the IMGA achieved the highest overall performance level, with superior QED and docking stability, and predicted the CRC inhibition potential of the molecules. CONCLUSION: The proposed EGA-PPO framework effectively connects traditional decoction principles with modern AI-based drug design. The tri-algorithmic system (BMGA-EMGA-IMGA) provides a scalable and interpretable strategy for de novo molecular discovery, offering promising leads for acquiring natural product-derived CRC therapeutics.
9 May 2026
Read appraisal →Annual review of biomedical engineering
Physics-Informed Machine Learning in Biomedical Science and Engineering
Physics-informed machine learning (PIML) is emerging as a potentially transformative paradigm for modeling complex biomedical systems by integrating parameterized physical laws with data-driven methods. Here, we review three main classes of PIML frameworks: physics-informed neural networks (PINNs), neural ordinary differential equations (NODEs), and neural operators (NOs), highlighting their growing role in biomedical science and engineering. We begin with PINNs, which embed governing equations into deep learning models and have been successfully applied to biosolid and biofluid mechanics, mechanobiology, and medical imaging, among other areas. We then review NODEs, which offer continuous-time modeling, especially suited to dynamic physiological systems, pharmacokinetics, and cell signaling. Finally, we discuss deep NOs as powerful tools for learning mappings between function spaces, enabling efficient simulations across multiscale and spatially heterogeneous biological domains. Throughout, we emphasize applications where physical interpretability, data scarcity, or system complexity make conventional black-box learning insufficient. We conclude by identifying open challenges and future directions for advancing PIML in biomedical science and engineering, including issues of uncertainty quantification, generalization, and integration of PIML and large language models.
5 May 2026
Read appraisal →Medical image analysis
PH2ST: Prompt-guided hypergraph learning for spatial transcriptomics prediction in whole slide images
Spatial Transcriptomics (ST) reveals the spatial distribution of gene expression in tissues, offering critical insights into biological processes and disease mechanisms. However, the high cost, limited coverage, and technical complexity of current ST technologies restrict their widespread use in clinical and research settings, making obtaining high-resolution transcriptomic profiles across large tissue areas challenging. Predicting ST from H&E-stained histology images has emerged as a promising alternative to address these limitations but remains challenging due to the heterogeneous relationship between histomorphology and gene expression, which is affected by substantial variability across patients and tissue sections. In response, we propose PH2ST, a prompt-guided hypergraph learning framework, which leverages limited ST signals to guide multi-scale histological representation learning for accurate and robust spatial gene expression prediction. Extensive evaluations on two public ST datasets and multiple prompt sampling strategies simulating real-world scenarios demonstrate that PH2ST not only outperforms existing state-of-the-art methods, but also shows strong potential for practical applications such as imputing missing spots, ST super-resolution, and local-to-global prediction, highlighting its value for scalable and cost-effective spatial gene expression mapping in biomedical contexts.
5 May 2026
Read appraisal →Journal of proteome research
NetMHCIIphosPan: A Machine Learning Tool for Predicting HLA Class II Antigen Presentation of Phosphorylated Peptides
Phosphorylated peptides presented by human leukocyte antigen (HLA) class II molecules play pivotal roles in immune regulation, yet their characterization and prediction remain challenging due to data noise and limited HLA coverage. Here, we introduce NetMHCIIphosPan, a prediction method for HLA-II antigen presentation of phosphorylated peptides, developed using mass spectrometry (MS)-based immunopeptidomics data sets. Employing a refined peptide identification workflow, we reanalyzed earlier HLA-II phospholigand data sets and trained predictive models, achieving superior performance compared to models trained on the original data. Binding motif analysis revealed that HLA-specific preferences for phospholigands closely aligned with those of unmodified ligands. Incorporating unmodified ligands into training further enhanced predictive accuracy, particularly for HLA-DP and HLA-DQ molecules. NetMHCIIphosPan outperformed existing tools, such as NetMHCIIpan-4.3 and MixMHC2pred-1.3, for prediction of HLA antigen presentation of phosphorylated peptides, demonstrating robustness and utility. This work establishes NetMHCIIphosPan as a state-of-the-art tool for understanding the HLA-II phospholigandome, with potential applications in immunotherapy and vaccine design.
4 May 2026
Read appraisal →Journal of molecular graphics & modelling
DeepHybridCPI: A hybrid deep learning framework for compound-protein interaction prediction
In bioinformatics, deep learning-based methods for Compound-Protein Interaction (CPI) prediction play a vital role in virtual screening, drug discovery, and drug repositioning. Recent improvements in computational methods have shown great possibility to save costs of experiment and speed up target identification. Nevertheless, the current CPI forecasting methods remain severely limited. Many rely on shallow Graph Neural Networks (GNNs) that struggle to capture the global structural context of compounds, while conventional Convolutional Neural Networks (CNNs) focus primarily on local sequence motifs and fail to model long-range dependencies in proteins. Even though a number of recent architectures strive to solve these problems by adding complexity to models, or by adding complex modules, these additions often cause significant computational overhead. To overcome these challenges, we propose DeepHybridCPI, a hybrid deep learning framework designed for accurate and efficient CPI prediction. Our hybrid model integrates a multiscale, densely connected GNN to extract compound features capturing both local substructures and global molecular topology, and employs CNNs with Long Short-Term Memory (LSTM) networks to model both local motifs and extended dependencies in protein sequences. The learned compound and protein representations are fused into a unified latent space to enable effective interaction modeling. Experimental evaluations on benchmark Human and C. elegans datasets demonstrate that DeepHybridCPI consistently outperforms existing state-of-the-art baseline methods in terms of AUC, Precision, and Recall. These findings highlight the importance of combining multiscale compound representations with hybrid sequence encoders within a single unified framework, providing a promising avenue for accelerating computational drug discovery. We release our source code and dataset at: https://github.com/jamshaidwarraich/DeepHybridCPI.
4 May 2026
Read appraisal →Computer methods in biomechanics and biomedical engineering
Multiple instance learning method based on convolutional neural network and self-attention for early cancer detection.
Early cancer detection using T-cell receptor sequencing (TCR-seq) and multiple instances learning methods has shown significant effectiveness. We introduce a multiple instance learning method based on convolutional neural networks and self-attention (MICA). First, MICA preprocesses TCR-seq using word vectors and then extracts features using convolutional neural networks. Second, MICA uses an enhanced self-attention mechanism to extract relational features of instances. Finally, MICA can extract the crucial TCR-seq. After cross-validation, MICA achieves an area under the curve (AUC) of 0.911 and 0.946 on the lung and thyroid cancer datasets, which are 7.1% and 2.1% higher than other methods, respectively.
4 May 2026
Read appraisal →Biotechnology advances
Computational methods for signal peptide prediction: From statistical models to deep learning
Signal peptides are short amino acid sequences located at the N-terminus of proteins. They guide newly synthesized proteins to their correct cellular destinations, playing a crucial role in protein localization and transport. Traditional experimental methods for identifying signal peptides are typically time-consuming, costly, and labor-intensive, driving rapid development of computational alternatives. Over the past two decades, researchers have proposed various computational approaches, with prediction accuracy continuously improving through evolution from early statistical and rule-based algorithms to deep learning. In this review, we systematically summarize these computational approaches, emphasizing methodological evolution and framework design. We compile representative computational methods, comparing their prediction outcomes and identifying existing limitations. Finally, we discuss current challenges and emerging opportunities, aiming to advance the development of computational frameworks characterized by unified evaluation, biologically grounded interpretation, and generative modeling.
4 May 2026
Read appraisal →Molecular informatics
ADME-DTI: Augmented Deep Meta Ensemble for Drug-Target Interaction Prediction
Drug-target interaction represents a critical focus area in computational drug discovery and pharmaceutical research. However, the process of identifying and analyzing these interactions is often resource-intensive, requiring extensive experimentation to evaluate the binding relationships between numerous drugs and their respective targets. This complexity is further compounded by the fact that a drug can inhibit multiple targets, and a target may also bind to various drugs. To address these issues, advanced deep learning models have been introduced as promising tools, offering the ability to accurately predict binding affinity and other bioactivity values to distinguish between potential drug-target interactions. The proposed model, named the Augmented Deep Meta Ensemble for Drug-Target Interaction (ADME-DTI), leverages multiple descriptors and fingerprint representations to extract meaningful insights from drug and protein data. These submodels from each representation are then combined with a deep learning architecture along with the metadata of the drug and target entries. The proposed approach has proved competitive performance against state-of-the-art models across diverse datasets and evaluation metrics, as evidenced by key metrics, namely r m 2 $r_m^2$ , concordance index, and mean squared error. Specifically, the model achieved mean squared error values of 0.186 (Davis), 0.118 (Kiba), 0.134 (DTC), 0.262 (Metz), 0.300 (ToxCast), and 0.791 (STITCH), all of which are publicly available benchmark datasets commonly used in drug-target interaction prediction tasks. The results highlight the model's effectiveness in reducing prediction errors and improving accuracy. Previous drug-target interaction research often relied on limited, non-diverse datasets, reducing generalizability. Few studies addressed drug-target interaction prediction as a regression task. Our novel deep learning metamodel integrates multiple models and representations, surpassing benchmarks and delivering greater prediction accuracy across varied datasets and evaluation metrics.
4 May 2026
Read appraisal →Nature reviews. Genetics
Interpretation, extrapolation and perturbation of single cells.
Single-cell analyses have transitioned from descriptive atlasing towards inferring causal effects and mechanistic relationships that capture cellular logic. Technological advances and the growing scale of observational and interventional datasets have fuelled the development of machine learning methods aimed at identifying such dependencies and extrapolating perturbation effects. Here, we review and connect these approaches according to their modelling concepts (including representation learning, causal inference, mechanistic discovery, disentanglement and population tracing), underlying assumptions and downstream tasks. We propose a unifying ontology to guide practitioners in selecting the most suitable methods for a given biological question, with detailed technical descriptions provided in an online resource . Finally, we identify promising computational directions and underexplored data properties that could pave the way for future developments.
4 May 2026
Read appraisal →Journal of computer-aided molecular design
Resolving the ambiguous binding site of quercetin at the calcineurin subunit junction using funnel metadynamics with deep learning collective variables
Calcineurin represents a prominent target for immunosuppressive drugs, where conventional macrocyclic inhibitors utilize an immunophilin-dependent mechanism for their inhibition but are consequently hindered by adverse effects and variable pharmacokinetics. The flavonoid quercetin has been shown to inhibit calcineurin in a non-competitive, immunophilin-independent manner, but its exact binding site remains ambiguous at the junction between calcineurin subunits A and B, with the open solvent-exposed nature of this region proving challenging to model. In this study, we employ funnel metadynamics with collective variables derived from a deep learning model to identify the binding site of quercetin. A selective strategy using multiple simulations with stricter funnel definitions and a smaller number of carefully chosen descriptors for model training proved more effective than a broad-based approach. These simulations were able to effectively distinguish the binding site of quercetin from three experimentally suggested sites, with a calculated free energy of binding of -8.36 ± 0.60 kcal/mol showing excellent agreement with experiment. Ligand-tryptophan distances similarly corroborated measurements from FRET assays with a r2 of 0.85. The corresponding binding pose showed that quercetin inserts itself into a channel between Arg122, Gly123, Tyr124 on one side and Phe160, Thr161, and Asn345 on the other, stabilizing its binding through a network of hydrogen bonds. The findings of this study provide insights into the modelling of challenging binding sites and ligands using deep learning driven metadynamics simulations and provide the foundations for rational development of immunophilin-independent inhibitors of calcineurin.
3 May 2026
Read appraisal →Scientific reports
scDecorr: feature decorrelation based representation learning enables self-supervised alignment of multiple single-cell experiments
Single-cell RNA sequencing (scRNA-seq) has revolutionized our understanding of cellular heterogeneity in complex biological systems. However, analyzing and integrating scRNA-seq data poses unique computational challenges due to sparsity, high variability, and technical batch effects. Here, we propose a novel framework called scDecorr for robust representation learning and data integration for scRNA-seq analysis. Our approach leverages the idea of feature decorrelation-based self-supervised learning (SSL) to obtain efficient low-dimensional representations of individual cells without relying on cell-type annotations. By maximizing similarity among distorted embeddings while decorrelating their components, scDecorr captures the biological signature while eliminating technical noise. Furthermore, scDecorr incorporates unsupervised domain adaptation to bridge the gap between batches with different distributions, enabling effective integration of scRNA-seq data from diverse sources. Our framework achieves domain-invariant representations by learning cell embeddings independently across domains and employing domain-specific batch normalization. We evaluate scDecorr on a variety of single-cell datasets and demonstrate its ability to integrate batches without losing the inherent biological variance, thereby facilitating optimal clustering. The representations generated by scDecorr also exhibit robustness in label transfer tasks, allowing for effective transfer of cell-type labels from reference to query datasets. Overall, scDecorr offers a powerful tool for efficient analysis and integration of large and complex scRNA-seq datasets, advancing our understanding of cellular processes and disease mechanisms. The code is available here https://github.com/hayatlab/scdecorr .
3 May 2026
Read appraisal →PLoS computational biology
MsgaBpred: A B-cell epitope predictor integrating AlphaFold3-predicted structures with multi-scale GCNs and pre-trained language model ESM-C
Accurate prediction of B-cell epitopes plays a key role in facilitating advancements in vaccines, therapeutics, and diagnostics. In contrast to labor-intensive experimental approaches, computational strategies provide a more economical and efficient means of identifying potential epitopes. Existing methods are often limited by their reliance on experimentally resolved protein structures or by the use of lower-accuracy predicted structures. Sequence-based approaches, while fast, largely fail to capture the 3D spatial context essential for conformational epitopes. With the breakthroughs achieved by AlphaFold3 in predicting protein structures, we present MsgaBpred, the model to apply AlphaFold3-derived structures to B-cell epitope identification. Given only a protein sequence, our model employs a multi-scale graph convolutional network and additive attention to capture complex structural dependencies without relying on experimentally determined structures. The multi-scale design allows for effective modeling of both local and global contexts by aggregating information across different neighborhood ranges. Additionally, we leverage ESM-C, a more expressive protein language model than ESM-2, to enhance feature representation for B-cell epitope prediction. Extensive evaluations across multiple benchmark datasets demonstrate that MsgaBpred achieves competitive and robust performance; notably, it yields a statistically significant improvement in AUC compared to existing state-of-the-art methods. Moreover, the modular and scalable architecture of MsgaBpred holds promise for broader applications, including the structural analysis of other biomolecular entities such as nucleic acids and carbohydrates.
3 May 2026
Read appraisal →Journal of chemical information and modeling
TriPDCL: A Tri-Pathway Prototype-Driven Contrastive Learning Framework for Cross-Modality Single-Cell Integration
The rapid development of single-cell sequencing technologies has laid a solid foundation for multiomics data integration. A central objective in multiomics integration is to attenuate cross-omic discrepancies stemming from technical batch effects and inherent biological variability, thereby facilitating a unified latent representation that supports robust downstream analyses. However, the intrinsic high sparsity of single-cell data and pronounced intercellular heterogeneity pose greater challenges for joint analysis. We foreground cellular heterogeneity and present TriPDCL, a prototype-based contrastive learning framework, which not only achieves effective transfer of heterogeneity information through an iterative prototype-learning update mechanism but also, by leveraging learnable prototype centers, enables precise construction of reliable positive-negative sample pairs. It tackles two principal challenges, the alignment of cross-modal heterogeneity and the robust learning, in multiomics analysis. Subsequently, comparative assessments conducted on five data sets in contrast to seven representative methods illustrate the superiority of the proposed approach.
2 May 2026
Read appraisal →Journal of chemical information and modeling
Explainability Methods from Machine Learning Detect Important Drugs' Atoms in Drug-Target Interactions
Predicting drug-target interactions (DTI) with graph neural networks (GNNs) is hindered by their lack of interpretability. To address this, we benchmark four explainable artificial intelligence (XAI) attribution methods on GNN models trained for kinase and G-protein-coupled receptors (GPCR) targets. We assess the methods' consistency through atom-level intersection over union (IoU) and validate their biological relevance by mapping attributed atoms to three-dimensional (3D) protein-ligand structures. While consistency across methods was modest, consensus attributions were highly enriched for atoms directly contacting the binding pocket─up to 76% within 2 Å in the kinase-inhibitor complexes. Notably, these attributed atoms were frequently found contacting experimentally important regulatory residues such as those in the DFG motif. This indicates that XAI methods, despite their disagreements, can identify chemically meaningful ligand features, providing a foundation for developing more interpretable GNNs in drug discovery.
2 May 2026
Read appraisal →Journal of chemical information and modeling
A Reinforcement Learning-Guided Genetic Algorithm Integrating Medicinal Chemistry-Inspired Molecular Transformations
Achieving optimal target activity while maintaining synthetic accessibility and drug-likeness represents a major challenge in computational drug discovery. Existing de novo generative models often yield chemically invalid or synthetically intractable structures and struggle to optimize multiple objectives simultaneously. Here, we introduce ALCHIMIA, an interpretable hybrid framework combining reinforcement learning (RL) and a genetic algorithm (GA), built based on a vocabulary of 33 medicinal chemistry-inspired molecular transformations. The RL component trains a policy network to prioritize transformation sequences that improve synthetic accessibility (SA) and the quantitative estimate of drug-likeness (QED) scores, embedding these constraints directly into molecular generation. The GA component applies the learned policy as a mutational operator within population-based optimization guided by molecular docking, enabling the exploration of diverse chemical lineages while converging toward high-affinity ligands. ALCHIMIA was applied to two different pharmacologically relevant targets: human Cannabinoid Receptor 2 (CB2R) and human Sigma nonopioid intracellular Receptor 1 (S1R). We considered three different scenarios: (i) unconstrained hit identification; (ii) scaffold-constrained lead optimization; and (iii) design of dual modulators. The framework generated chemically valid molecules with QED and SA scores comparable to or better than those obtained with random baselines and selected de novo design methods. By codifying typical medicinal chemistry actions as learnable transformations and coupling multiobjective optimization with GA-based diversity maintenance, ALCHIMIA, freely available as a GitHub repository (https://github.com/alberdom88/ALCHIMIA), provides a practical, interpretable, and scalable framework for molecular de novo design.
2 May 2026
Read appraisal →Scientific reports
LASSO-HHO two-stage hybrid gene selection framework for accurate Alzheimer's disease diagnosis
Selecting informative genes is essential for building accurate and efficient diagnostic models, especially for high-dimensional gene expression data such as those related to Alzheimer's disease. These datasets typically involve thousands of genes with limited samples, increasing the risk of overfitting and computational complexity. To address this, we propose a LASSO-HHO Gene Selection (LHGS) framework. LASSO is first applied to reduce dimensionality by filtering irrelevant genes. A conditional HHO-based optimization stage is then applied only when the LASSO-selected subset does not achieve sufficient accuracy or remains relatively large. Otherwise, the LASSO-selected features are directly used without further optimization. Experimental results show that the proposed method reduces the number of selected genes by up to 99.9% while maintaining high performance. The experimental results indicate that 100% accuracy can be achieved on specific datasets. In particular, GSE48350 and GSE36980 achieved 100% accuracy using LASSO alone, whereas GSE118553 and GSE132903 required the full LHGS framework to achieve the same performance. The framework also improves computational efficiency within a consistent experimental setup by reducing the optimization search space after LASSO filtering. Overall, LHGS provides a practical and efficient solution for gene selection in high-dimensional biomedical data.
28 Apr 2026
Read appraisal →Proceedings of the National Academy of Sciences of the United States of America
Machine learning multiscale collective cell dynamics: From single-cell characterization to multicellular monolayer modeling
Collective cell dynamics are fundamental to numerous physiological processes, including embryo development, tissue morphogenesis, immune response, and disease progression. Accurately modeling these dynamics across scales remains challenging, as traditional physics-based models usually rely on many unmeasurable parameters and unclear active physics, limiting their ability to capture both single-cell features and emergent multicellular behaviors. Here, we introduce a scale-adaptive hybrid machine learning (ML) framework, in which complementary physics-guided and physics-agnostic method each resolves an above key issue at single-cell and multicellular scales, respectively. The physics-guided ML method integrates physical models with experimental data to infer previously unmeasurable parameters, enabling more accurate characterization of single-cell shape and velocity features, but unknown active forces and inherent physical assumptions still constrain its ability to predict multicellular scale dynamics. In contrast, the physics-agnostic ML method bypasses explicit physical assumptions to directly model multicellular behaviors from historical state sequences. This method provides robust long-term predictions of coarse-grained density oscillations and waves observed experimentally, yet it suffers from high stochasticity when applied to single-cell scale modeling. By assigning each approach to the scale where it is most effective, our framework leverages their respective strengths while mitigating their limitations. This complementary strategy bridges the gap between theoretical modeling and experimental observations, offering a versatile computational paradigm for multiscale collective cell dynamics with broad potential in diverse physiological and pathological contexts.
27 Apr 2026
Read appraisal →Journal of medicinal chemistry
Machine Learning Approach to Anticancer Activity Prediction of Transition-Metal Complexes Based on a Large-Scale Experimental Database
In this work, we developed a straightforward data-driven approach to predict the cytotoxicity of metal complexes based entirely on their (metal + ligands) composition. To this end, we have manually curated MetalCytoToxDB─a comprehensive experimental database comprising 26,500 IC50 values for 7050 metal complexes against 754 cell lines from 1921 articles. Based on these, machine learning models were created to accurately assign the cytotoxicity class within the ruthenium and iridium subsets. Moreover, external validation of the best-performing model on the unseen data was carried out. The possibility of multimetal predictions was explored, enabling assessment of cytotoxicity among the complexes of metals, for which experimental data are relatively scarce. The interpretability and limitations of the developed models are discussed. Finally, a pipeline for the effective high-throughput computational screening of ruthenium complexes is proposed. The MetalCytoToxDB is available online for AI-assisted exploration at https://biometaldb.streamlit.app/.
25 Apr 2026
Read appraisal →Journal of medicinal chemistry
An Interpretable, Thermodynamics-Based Deep Learning Framework for Predicting and Optimizing Drug Membrane Permeability
Cellular membranes serve as selective barriers, and membrane permeability is crucial for drug pharmacokinetics. While in vitro and in vivo methods exist, predicting and designing membrane permeability remains challenging. We developed a thermodynamics-based deep learning framework to analyze the structure-permeability relationship, based on the concept that interactions between membrane lipids and small molecules influence permeability. We determined the membrane penetration thermodynamics of 8,239 compounds using coarse-grained molecular dynamics simulations and created interpretable graph neural network models to predict and design drug membrane permeability. As a proof-of-concept, we designed a novel nasal-administered melatonin analog, MT-A2, optimized for permeability. Compared with melatonin, MT-A2 showed superior nasal absorption, prolonged brain retention, and enhanced sleep efficacy. Our results provide a promising approach for predicting and designing membrane permeability, aiding in the development of drugs with better pharmacokinetics.
25 Apr 2026
Read appraisal →Science advances
Dual computational systems in the development and evolution of mammalian brains
Analyses of brain sizes across mammalian taxonomic groups reveal a consistent pattern of covariation between major brain components, including a robust inverse relationship between the limbic system and the neocortex. To find the functional basis of this relationship, we mapped the multidimensional representations of task-optimized artificial neural networks onto two-dimensional surfaces resembling the forebrain cortices. We found that networks optimized for visual, somatosensory, and auditory representations develop ordered spatiotopic maps where units draw information from localized regions of the sensory input. In contrast, networks optimized for olfactory and relational memory representations develop fractured maps with distributed patterns of information convergence. Evolutionary optimization of multimodal networks for varying task objectives results in inverse covariation between spatiotopic and disordered network components that compete for the representational space. These results suggest that the observed pattern of covariation between brain components reflects an essential computational duality in brain evolution.
24 Apr 2026
Read appraisal →Cell reports methods
A high-speed attention network for MHC-bound peptide identification and 3D modeling
We developed SwiftMHC, an ultra-fast and accurate structure-based framework for peptide-MHC (pMHC) modeling and binding affinity prediction. Using task-specific deep learning trained on physics-derived synthetic data, SwiftMHC predicts pMHC binding affinities in 0.009 s per case on a single A100 GPU when running in batch mode, offering improved speed compared with leading sequence-based tools such as netMHCpan and MHCflurry while maintaining competitive accuracy. In addition, SwiftMHC generates all-atom 3D pMHC structures with a median Cα-RMSD of 1.32 Å against crystallographic benchmarks, matching or exceeding state-of-the-art methods such as AlphaFold2-finetune but at a lower computational cost. Optimized for HLA-A∗02:01 9-mer peptides but readily extensible to other alleles, SwiftMHC unites structural insight with high-throughput scalability to accelerate safe and effective epitope discovery in cancer immunotherapy.
24 Apr 2026
Read appraisal →Analytical chemistry
scAClc: A Multi-Objective Adaptive Clustering Framework for Single-Cell Transcriptomics via Contrastive and Resolution-Aware Representation Learning
Single-cell RNA sequencing (scRNA-seq) enables whole-transcriptomic profiling at single-cell resolution, facilitating the construction of virtual cell representations that capture the full spectrum of cellular identities. Realizing this goal hinges on accurate clustering, which remains challenging due to data sparsity, high dimensionality, noise, and the need to specify cluster numbers a priori. We propose scAClc, a novel clustering framework featuring multiobjective optimization and adaptive resolution discovery, designed to address these limitations through three key innovations. First, a Hierarchical Gene Relevance Module integrates global gene variability with local neighborhood-specific signals to eliminate redundancy while retaining biologically informative features. Second, an Anchor-Centered Contrastive Learning Module adaptively selects representative anchors to guide embedding learning, promoting compact intracluster structure and clear intercluster separation. Third, based on the robust low-dimensional embedding, we propose a Self-Adaptive Resolution Discovery Module to automatically infer the number of clusters by jointly modeling intra- and intercluster distances. Extensive experiments on 15 real scRNA-seq data sets demonstrate that scAClc consistently outperforms six state-of-the-art methods across multiple evaluation metrics. Ablation studies further confirmed the complementary contributions of each module. In addition, interpretability analysis effectively mitigates the "black box" nature of clustering models and sheds light on the biological mechanisms underlying cell clustering. The source code is publicly available at https://github.com/scAClc/scAClc.
23 Apr 2026
Read appraisal →Biological cybernetics
Geometric Learning Dynamics
We present a unified geometric framework for modeling learning dynamics in physical, biological, and machine learning systems. The theory reveals three fundamental regimes, each emerging from the power-law relationship g ∝ κ α between the metric tensor g in the space of trainable variables and the noise covariance matrix κ . The quantum regime corresponds to α = 1 and describes Schrödinger-like dynamics that emerges from a discrete shift symmetry. The efficient learning regime corresponds to α = 1 2 and describes very fast machine learning algorithms. The equilibration regime corresponds to α = 0 and describes classical models of biological evolution. We argue that the emergence of the intermediate regime α = 1 2 is a key mechanism underlying the emergence of biological complexity.
20 Apr 2026
Read appraisal →Proceedings of the National Academy of Sciences of the United States of America
Making invisible excited-state structures of pro-interleukin-18 visible by combining NMR and machine learning
NMR spectroscopy is a powerful tool for studies of protein dynamics over timescales extending more than twelve orders of magnitude, with motion queried at many of the backbone and sidechain positions in the molecule of interest. NMR experiments can, in principle, provide atomic resolution descriptions of excited conformational states that are sparsely populated and transiently formed-i.e., invisible, and there are examples of such structures in the literature. However, in other cases, the NMR data are not sufficient for generating structural ensembles of rare and transient higher energy conformers of the protein of interest. One example is provided by the precursor form of the proinflammatory cytokine interleukin-18, pro-IL-18, studied here. Although NMR studies show that pro-IL-18 adopts two sparsely populated (<0.5%) and transiently formed (ms lifetimes) excited-state conformations in exchange with a highly populated ground state conformer and localize regions undergoing exchange to a pair of short β-strands that are preserved in at least one of the excited states, additional structural information is not forthcoming. Here, we develop a protocol whereby the NMR data are used to select alternative conformers of pro-IL-18 from ensembles predicted by the generative machine learning model AlphaFlow that are then evaluated through further NMR experiments. The method identifies distinct conformers that correspond to the pair of excited states reported by NMR, combining experiment with computation to characterize the pro-IL-18 energy landscape in ways that are not possible with either approach alone.
20 Apr 2026
Read appraisal →Cell systems
Context-informed subgraph foundation models enable interpretable protein-function prediction
Protein-function prediction is crucial for elucidating molecular mechanisms driving biological processes and therapeutics development. Despite numerous computational tools demonstrating promising performance, they fall short when predicting rare, uncharacterized functions or indirect activities. Here, we present COSMOS, a context-aware Gene Ontology (GO) subgraph mining system for protein-function prediction. By leveraging inductive subgraph foundation models and an enriched knowledge graph of protein-GO relationships, COSMOS performs zero-shot, few-shot, and low-homology protein-function prediction. Built on 7,923,952 functional semantic relationships, COSMOS demonstrates robust capabilities to (1) generate state-of-the-art predictions for GO classes with sparse or no experimental annotations, (2) provide interpretable functional subgraphs for transparent rationale analysis, and (3) deliver complementary benefits when integrated with existing embedding-based prediction methods. We anticipate that COSMOS will serve as a complementary approach to conventional protein annotation methods and an interpretable tool for predicting protein functions within underexplored GO classes, thereby advancing genomics and therapeutic research.
18 Apr 2026
Read appraisal →Proceedings of the National Academy of Sciences of the United States of America
Coalescence and translation: A language model for population genetics.
Probabilistic models such as the sequentially Markovian coalescent have long provided a powerful framework for population genetic inference, enabling reconstruction of demographic history and ancestral relationships from genomic data. However, these methods are inherently specialized, relying on predefined assumptions and/or limited scalability. Recent advances in simulation and deep learning provide an alternative approach: learning directly to generalize from synthetic genetic data to infer hidden evolutionary processes. Here we reframe the inference of coalescence times as a problem of translation between two biological languages: the sparse, observable patterns of mutation along the genome and the unobservable ancestral recombination graph that gave rise to them. Inspired by large language models, we develop cxt, a decoder-only transformer that autoregressively predicts coalescent events conditioned on local mutational context. We show that cxt performs competitively with state-of-the-art Markov Chain Monte Carlo-based likelihood models across a broad range of demographic scenarios, matching their accuracy in-distribution and approaching it in out-of-distribution, with the potential for improvement via fine-tuning. Trained on simulations spanning the stdpopsim catalog [J. R. Adrion et al., eLife 9, e54967 (2020); M. E. Lauterbur et al., eLife 12, e84874 (2023); G. Gower et al., Accessible, realistic genome simulation with selection using stdpopsim. bioRxiv [Preprint] (2025). https://doi.org/10.1101/2025.03.23.644823.], the model generalizes robustly and enables efficient inference at scale, producing over a million coalescence predictions in minutes. cxt also produces well-calibrated approximate posteriors, enabling principled uncertainty quantification. We apply cxt to population genomic data from both humans and mosquitoes, highlighting the model's ability to deal with the complexities of empirical data.
17 Apr 2026
Read appraisal →Physical chemistry chemical physics : PCCP
TransG4: an interpretable deep-learning approach for sequence-based G-quadruplex prediction
G-quadruplexes (G4) are non-canonical nucleic acid secondary structures formed in guanine-rich regions and have been shown to regulate diverse cellular processes such as gene expression, DNA replication, and telomere maintenance, with increasing evidence linking G4 to cancer and other human diseases. G4 predominantly emerge in guanine-rich regions and are implicated in a spectrum of molecular interactions and disease phenotypes, thus researchers are interested in the formation of G4. However, predicting the formation of G4 from nucleotide sequences is a persistent problem. Existing computational tools for G4 prediction are either rule-based on domain knowledge or rely on a single neural network model like a convolutional neural network (CNN), which lacks interpretability and struggles to capture long-range dependencies among bases. Here, we introduce TransG4, a novel neural network architecture that integrates a CNN, a transformer, and bidirectional gated recurrent units (BiGRUs) to identify potential G4 structures. TransG4 demonstrates strong predictive performance on both G4-seq and rG4-seq datasets, accurately predicting DNA mismatch scores and consistently outperforming existing methods in RNA RSR-ratio prediction. Attention-based interpretations further show that TransG4 captures biologically meaningful motifs consistent with canonical G4 structures, providing an interpretable and generalizable framework and representing a novel and impactful contribution to sequence-based G4 propensity prediction.
17 Apr 2026
Read appraisal →PloS one
Exploration of schizophrenia-associated gene modules using graph theory, co-expression networks, and dimensionality reduction.
The PsychENCODE consortium has generated a comprehensive multi-omic dataset from human brain samples, spanning healthy individuals and those with neuropsychiatric disorders. In this study, we focus exclusively on schizophrenia and analyze PsychENCODE transcriptomic data to construct gene regulatory and co-expression networks, aiming to uncover biologically and clinically relevant gene modules. We apply three independent analytical approaches to the same dataset. First, we preprocess the data using Surrogate Variable Analysis (SVA) to correct for latent variation and normalize gene expression. For graph-based analysis, we use Pearson correlation and the igraph package to construct gene co-expression networks and apply Prim's algorithm to generate minimum spanning trees (MST) and compute centrality measures. Next, we implement Weighted Gene Co-expression Network Analysis (WGCNA), assuming a scale-free topology, to identify modules associated with schizophrenia traits. Dimensionality reduction is performed using Principal Component Analysis (PCA), with visualization aided by t-distributed Stochastic Neighbor Embedding (t-SNE) and Multidimensional Scaling (MDS). Functional enrichment is carried out using Gene Ontology (GO) and KEGG pathway databases. Each method reveals distinct yet complementary biological signatures associated with schizophrenia. The igraph-based approach highlights differentially expressed genes (DEGs) with high centrality, uncovering hub genes involved in mRNA export, synaptic signaling, and ion transport. WGCNA identifies co-expression modules strongly correlated with schizophrenia diagnosis, enriched for immune response, histone modification, and mRNA surveillance pathways. PCA isolates key genes contributing to diagnostic variance, with enrichment in neurotransmitter release cycles and cytokine signaling. Collectively, these results underscore the involvement of immune, synaptic, and epigenetic processes in schizophrenia and demonstrate the power of using multiple, orthogonal analytical lenses.
17 Apr 2026
Read appraisal →Journal of molecular modeling
HopWD-DTA: a novel framework for drug-target affinity prediction fusing multi-hop neighborhoods and deep features
CONTEXT: Accurate prediction of drug-target affinity (DTA) is crucial for accelerating drug discovery, but it remains a significant challenge. While deep learning methods have shown promise, many existing models struggle with representing long-range protein relationships within protein-protein interaction (PPI) networks, often suffering from performance degradation as graph neural network layers deepen. To overcome this limitation, we developed HopWD-DTA, a novel framework designed to effectively capture both local and global protein features. The core innovation lies in integrating multi-hop neighborhood information from PPI networks with deep structural features of proteins and drugs. Comprehensive evaluations on the Davis, KIBA, and Human benchmark datasets demonstrate that HopWD-DTA achieves state-of-the-art performance, significantly improving DTA prediction accuracy over existing cutting-edge solutions. METHODS: The HopWD-DTA framework consists of separate branches for protein and drug feature extraction. For proteins, structural features from residue contact maps are first encoded using a Graph Convolutional Network (GCN). These features, along with InterPro annotations, are then embedded into a PPI network. We introduce a multi-hop neighborhood serialization technique, which generates a sequence of feature matrices representing different neighborhood scopes. This sequence is processed by a Variational Autoencoder (VAE) to learn a robust protein representation. For drugs, molecules are represented as graphs from SMILES strings and encoded via a GCN, followed by a novel Wide-and-Deep Path (WDPATH) module to capture both macroscopic and microscopic features. The final protein and drug features are concatenated and fed into a multilayer perceptron for affinity prediction. The model was implemented using PyTorch and RDKit.
15 Apr 2026
Read appraisal →Journal of chemical information and modeling
zERExtractor: An Automated Platform for Enzyme-Catalyzed Reaction Data Extraction from Scientific Literature
The rapid expansion of enzyme reaction literature has created a major bottleneck in database curation, leaving vast amounts of enzyme-substrate-condition relationships unstructured and inaccessible for DL-driven modeling. How to fully utilize the enzymatic reaction data has been an important task for future accurate enzyme activity prediction models. Current deep learning (DL)-based data extraction models heavily rely on large language models (LLMs) without a fidelity check and the ability to continuously evolve. To address these issues, we developed zERExtractor (Zelixir's Enzyme Reaction Data Extractor), an accuracy-oriented and extensible platform for extracting enzyme-catalyzed reaction data from scientific publications. This system offers a unified multimodal information extraction framework (covering molecular reaction diagrams, tables, and texts) to integrate enzymatic reaction descriptors into structured storage. We employ fine-tuned large LLMs together with DL in a human-in-the-loop pipeline that evolves through data fidelity validation by experts and active learning. Also, zERExtractor achieves 89.9% accuracy in table recognition and over 98% accuracy in molecular image recognition on synthetic data sets, outperforming the strongest baseline by more than 2% and consistently maintaining above 95% on realistic benchmarks. zERExtractor bridges the data gap in enzyme reaction data with a scalable framework for accurate multimodal extraction, advancing DL-driven enzyme modeling and enabling future applications in computational enzymology and biotechnology. The platform is publicly accessible online at https://zpaper.zelixir.com/.
14 Apr 2026
Read appraisal →Journal of chemical information and modeling
AI-MedCraft: A Strategy-Driven AI Platform for Multi-Objective Molecular Design
Small-molecule drug discovery requires balancing potency, selectivity, solubility, safety, and synthetic feasibility, among other interdependent properties. Yet most computational workflows address these properties sequentially or collapse them into fixed weighted scores. Although artificial intelligence has accelerated molecular design, many platforms struggle to handle explicit multiobjective trade-offs. Here, we present AI-MedCraft, a strategy-driven molecular design framework that applies adaptive, Pareto-guided reinforcement learning to optimize multiple objectives concurrently within a unified workflow. When structural information is available, physics-aware scoring is incorporated to support binding-competent designs. In a structure-based benchmark on the solubility rescue of the BTK inhibitor GDC-0834, AI-MedCraft achieves broader Pareto-front coverage than the scalar-reward reinforcement learning framework REINVENT 4 under matched objectives and computational cost. In a second case study, AI-MedCraft redesigns Efavirenz to retain HIV-1 reverse transcriptase engagement while reducing predicted 5-HT2A off-target binding, demonstrating its use in multiconstraint molecular optimization.
14 Apr 2026
Read appraisal →Journal of chemical information and modeling
ncProFormer: A CNN-enhanced Transformer for ncRNA Coding-Potential Prediction
Recent studies have revealed that some noncoding RNAs (ncRNAs) bear translational potential, and their encoded micropeptides have essential functions in multiple biological processes. However, accurate identification of coding-capable ncRNAs remains challenging due to weak translation signals, low conservation, and heterogeneous data distributions. Herein, we propose ncProFormer, a deep learning framework tailored for ncRNA coding-potential prediction. ncProFormer integrates the nucleic-acid language model GENA-LM to obtain contextual sequence embeddings, adopts an all-token representation strategy, and employs a convolutional neural network (CNN)-enhanced transformer encoder to jointly capture local nucleotide patterns and long-range dependencies. ncProFormer consistently outperformed the existing methods across the in-house human data set, the external validation data set, the public CPPred benchmark data set. More importantly, this study presents the first cross-species evaluation in ncRNA coding-potential prediction. Without retraining, ncProFormer maintained its strong predictive performance on mouse and rat data sets, showing that the learned biological representations are transferable and it is robust under the distributional shift and cross-species conditions. Collectively, these findings establish ncProFormer as an effective and generalizable framework for uncovering the coding potential of ncRNAs, thus offering a promising computational tool for characterizing ncRNA functions across diverse transcriptomic contexts.
14 Apr 2026
Read appraisal →Journal of chemical information and modeling
Compressing Chemistry Reveals Functional Groups
We introduce the first formal large-scale assessment of the utility of traditional chemical functional groups as used in chemical explanations. Our assessment employs a fundamental principle from computational learning theory: a good compression of data should reveal a good explanation. We introduce an unsupervised learning algorithm based on the Minimum Message Length (MML) principle that searches for substructures that compress around three million biologically relevant molecules. We demonstrate that the discovered substructures contain most human-curated functional groups as well as novel larger patterns with more specific functions. We also run our algorithm on 24 specific bioactivity prediction data sets to discover data set-specific functional groups. Fingerprints constructed from data set-specific functional groups are shown to significantly outperform other fingerprint representations, including the MACCS and Morgan fingerprint, when training ridge regression models on bioactivity regression tasks.
14 Apr 2026
Read appraisal →Journal of chemical information and modeling
ToxPLTC: Peptide Toxicity Prediction by Integrating Pretrained T5 Protein Language Model and Text Convolutional Neural Network
Peptide-based therapeutics show promising potential in treating a range of diseases, such as diabetes, cancer, and chronic pain. However, critical challenges, including peptide toxicity, immunogenicity, and stability deficiencies of peptides, have become major obstacles to their direct clinical application. Traditional toxicity testing methods based on a wet lab are not only costly but also time-consuming. In contrast, the classification model based on deep learning provides a new technical path for the efficient identification of peptide toxicity. In this study, we propose a deep learning framework called ToxPLTC, which employs the ProtT5 protein language model based on the Transformer architecture for pretraining peptide sequences, adopts the borderline SMOTE algorithm to handle an imbalanced training set data, and utilizes a text convolutional neural network combined with a fully connected layer for classification. Additionally, visualization analysis, motif analysis, and mutation-scan analysis are performed to understand the function of each module and enhance the interpretability of our model. The applicability domain is constructed based on the K-NN strategy to define the effective prediction range of our model to ensure the reliability of model predictions. The ToxPLTC model achieves a balanced accuracy of 93.02% on independent test set 1 and 88.04% on independent test set 2. Experimental results demonstrate that our model outperforms existing models on independent test sets and has good generalization ability. The ToxPLTC model possesses great potential as a valuable and robust tool for peptide-based drug development. The source data sets and codes can be available at the following GitHub repository: https://github.com/yunyunliang88/ToxPLTC.
14 Apr 2026
Read appraisal →Journal of chemical information and modeling
EviCYP: In Silico Prediction of Cytochrome P450 Substrates Based on Vector Quantization and Evidential Deep Learning
The accurate identification of cytochrome P450 (CYP) substrates is crucial in drug discovery and safety assessment, as these enzymes mediate the metabolism of most clinical drugs. However, existing computational models are often limited by data quality issues and lack the ability to quantify prediction uncertainty, hindering their reliable application. To address these challenges, we present EviCYP, a novel prediction framework that integrates evidential deep learning with vector quantization (VQ). We first constructed a high-quality data set by curating 4388 substrates and 2880 nonsubstrates from 1629 publications, and supplemented it with 3728 pseudonegative samples, resulting in 10,996 samples spanning nine major CYP isoforms. The EviCYP architecture processes multimodal molecular representations and enzyme sequences through dedicated encoders, compresses features via VQ to reduce redundancy, and employs an evidential layer to output both class probabilities and an uncertainty estimate. On an internal test set, EviCYP achieved an average AUROC of 0.9500. Notably, the model's uncertainty quantification is highly reliable, with high-uncertainty predictions strongly correlating with classification errors. This work provides a robust and trustworthy computational tool for CYP substrate prediction.
14 Apr 2026
Read appraisal →Journal of chemical information and modeling
Multiscale-Aware Graph Embedding Approach Uncovers LC-61, a Potent Anti-Leishmania infantum Compound
Visceral leishmaniasis caused by Leishmania infantum remains a lethal disease with few therapeutic options, necessitating innovative computational methods and approaches to accelerate drug discovery. Here, we present a graph neural network (GNN) framework incorporating well-established multiscale mechanisms to improve the identification of novel antileishmanial compounds. Across two classificatory antileishmanial data sets, our GNNs demonstrated significant improvements in predictive performance, with area under the receiver operating characteristic curve (AUC) increases of 2.2-29.2% on the imbalanced data set (activity cutoff: 1 μM) and 3.4-22.5% on the balanced data set (activity cutoff: 10 μM) compared to default GNNs. Subsequently, the framework was applied to screen a library of approximately 1.3 million compounds, pinpointing LC-61 as a potent antileishmanial agent with nanomolar activity against intracellular L. infantum (IC50 = 0.076 μM) and minimal cytotoxicity to macrophages (THP-1 CC50 = 157 μM). A comprehensive in vitro ADME profiling revealed that LC-61 combines high solubility at both acidic and physiological pH (>28 μg/mL), balanced lipophilicity (eLogD = 4.07), and favorable passive permeability (PAMPA = 4.86 × 10-6 cm/s), while exhibiting lower microsomal stability. Overall, our GNN framework effectively accelerated the discovery of LC-61, a novel and biologically validated hit suitable for hit-to-lead optimization.
14 Apr 2026
Read appraisal →International journal of pharmaceutics
From permeability to prediction: evolving strategies for evaluating oral drug absorption
Oral drug administration, preferred for patient compliance and convenience, encounters significant bioavailability challenges arising from complex physicochemical, physiological, and formulation-related barriers. Traditional models for evaluating absorption exhibit limitations in physiological relevance and predictive accuracy. This review critically examines evolving methodologies for oral drug absorption assessment, from classical permeability assays to emerging predictive platforms, emphasizing integration of mechanistic research with predictive modeling to advance intelligent platforms and clinical translation. While classical permeability models including parallel artificial membrane permeability assay (PAMPA), Caco-2, Ussing chambers, and intestinal perfusion remain indispensable for early screening, their limited physiological relevance and inability to elucidate mechanistic pathways constrain predictive accuracy. Mechanistic approaches such as lymphatic transport studies and advanced tracking techniques using fluorescence, Single Photon Emission Computed Tomography (SPECT), or Positron Emission Tomography (PET) imaging reveal critical biological pathways but suffer from scalability challenges. Emerging platforms provide transformative potential, as organoid-on-a-chip offer physiologically relevant and high-throughput alternatives, while integrated computational modeling enables multiscale prediction, from molecular interactions to systemic pharmacokinetics. Future oral absorption assessment requires converging in vitro biomimetics, computational modeling, and Artificial intelligence (AI) within unified platforms. Achieving this requires standardized in vitro models, robust data integration, and alignment with regulatory to accelerate translational drug development.
12 Apr 2026
Read appraisal →Human mutation
Use of Single-Cell Data and scPagwas Analysis to Identify T Cell Subsets and Construct a Prognostic Model for Clear Cell Renal Cell Carcinoma
BACKGROUND: Clear cell renal cell carcinoma (KIRC), the most prevalent pathological renal cell carcinoma (RCC) subtype, makes up approximately 75%-84% of total cases. KIRC is characterized by high heterogeneity, high metastasis rates, and a poor prognosis. Its incidence rate has continued to rise in recent years. We sought to construct new prognostic models to optimize treatment decisions, improve clinical benefits, and explore potential therapeutic targets. METHODS: This study integrated various omics data including single-cell RNA seq (GSE171306), TCGA-KIRC, GWAS, and validation datasets (GSE29609 and E-MTAB-1980). The scPagwas algorithm combines GWAS with scRNA-seq to identify immune subgroups with high feature correlation. Key genes are identified through the combination of weighted correlation network analysis (WGCNA) with differentially expressed genes (DEGs). We built a clinical prognostic model by using machine learning algorithms and validated it through survival rate and receiver operating characteristic (ROC) analysis. We used cancer drug sensitivity genomics data to analyze drug sensitivity and performed molecular docking to identify potential therapeutic drugs. RESULTS: Using single-cell RNA seq data, we identified T cell subsets as characteristic cell subsets in KIRC through scPagwas analysis. In single-cell analysis, key genes in T cell subsets and genes with PCC values > 0.05 were combined with the core genes in DEGs and WGCNA modules, thus yielding 86 intersecting genes. These genes were significantly enriched in immune-related pathways. We established a clinical prognostic model containing seven risk genes. Low-risk patients exhibited substantial survival advantages. Time-dependent ROC analysis indicated the prognostic model's excellent clinical predictive value. Functional enrichment, immune infiltration, and somatic mutation analyses highlighted different biological mechanisms among risk populations. The SHAP values of the XGBoost and LightGBM machine learning algorithms indicated DOCK8 as a potential biomarker. Drug prediction and molecular docking predicted five potential drugs targeting DOCK8 (finasteride, nocodazole, palonosetron, pifithrin alpha, and topiramate). CONCLUSION: Our systematic analysis of the immune microenvironment, key genes, and prognosis of KIRC highlighted the critical roles of T cell subsets. We additionally established an effective clinical prognostic model. Our findings provide new insights and potential targets for the precise diagnosis and targeted KIRC therapy.
11 Apr 2026
Read appraisal →Cell genomics
Variant-resolved prediction of context-specific isoform variation with a graph-based attention model
In eukaryotes, most genes produce multiple transcript isoforms that diversify the transcriptome and proteome, serving as a key mechanism of functional regulation. Genetic variation can disrupt the RNA processing signals that shape isoform structure and abundance, yet modeling these effects at full-length isoform resolution remains challenging due to the complexity of transcript regulation. Here, we introduce Otari, an attention-based graph neural network framework trained on the human genomic sequence and long-read transcriptomes across 30 tissue types and brain regions. Otari predicts tissue-specific differential isoform abundance by integrating sequence-derived epigenetic and post-transcriptional signals, enabling isoform-resolved variant effect interpretation. Applied to large-scale variant datasets, including an autism cohort, Otari uncovers patterns of isoform dysregulation undetectable at the gene level, such as variant-driven perturbations in isoform abundance and microexon usage implicated in autism pathophysiology. We provide Otari as a resource for powering isoform-level analyses across tissues at scale.
11 Apr 2026
Read appraisal →Chemical communications (Cambridge, England)
Innovative integration of molecular docking and machine learning for drug discovery: from virtual screening to nanomolar inhibitors
This review highlights our group's systematic approach to integrating molecular docking, pharmacophore modeling, and machine learning methodologies for the rational discovery of bioactive leads. We describe innovative strategies including docking-based data augmentation, ligand-receptor contact fingerprints, genetic algorithm-guided feature selection, and SHAP-based model interpretation. These approaches have enabled the discovery of nanomolar inhibitors against multiple therapeutic targets including STAT3, TTK, LSD-1, and HER2. The presented workflow demonstrates how machine learning (ML) can be synergistically combined with traditional computer-aided drug design methods to achieve efficient scaffold hopping and identify novel chemotypes with potent biological activities.
10 Apr 2026
Read appraisal →Analytical chemistry
SIBioX: A Matrix Based Bioinformatics Analysis Tool Based on Swarm Intelligence Algorithm
Biological matrix data are essential for computational analysis, providing a structured framework to identify patterns and relationships in biological systems. Many other biological data types, including sequences, networks, and images, can be transformed into matrix representations through feature extraction and encoding. However, their high dimensionality complicates analysis, leading to increased computational complexity and the risk of overfitting, known as the curse of dimensionality. To address these challenges, we developed SIBioX, a matrix-based bioinformatics tool powered by swarm intelligence algorithms. It integrates 54 swarm intelligence methods, 5 conventional feature selection techniques, and 17 machine learning models, enabling comprehensive analysis of biological matrix data. With a user-friendly graphical interface, it supports operations such as feature normalization, selection, classification, clustering, statistical analysis, and data visualization. Additionally, it converts nonmatrix biological data, like gene and protein sequences, into matrix formats for further study. Experimental results demonstrate that SIBioX not only attains high accuracy in feature selection but also effectively reduces dimensionality, thereby streamlining bioinformatics workflows and promoting greater efficiency in biomedical research.
10 Apr 2026
Read appraisal →mAbs
Deep learning assessment of nativeness and pairing likelihood for antibody and nanobody design with AbNatiV2.
Most immune-system created antibodies balance good binding and stability with low toxicity and self-reactivity. Quantifying the nativeness of a candidate sequence - its likelihood of belonging to natural immune repertoires - has thus emerged as a valuable strategy for hit selection from synthetic libraries, optimization and humanization, and for guiding de novo design toward developable candidates. We previously introduced AbNatiV, a transformer-based VQ-VAE for nativeness assessment, which proved effective across multiple nanobody engineering tasks. However, AbNatiV1 operated on unpaired sequences, limiting applicability to conventional VH-VL antibodies. Moreover, its performance on nanobody nativeness was constrained by the limited number and diversity of nanobody repertoires available at the time. Here, we sequenced new camelid repertoires, curated additional recent datasets, and present AbNatiV2: an enhanced architecture comprising various models each trained on ≥20 million sequences. AbNatiV2 improves nanobody nativeness classification across held-out and diverse test sets, and more robustly detects nativeness changes upon CDR grafting. We also introduce p-AbNatiV2, a cross-attention model fine-tuned on 3.7 million paired human sequences. p-AbNatiV2 provides residue- and sequence-level humanness for VH/VL pairs and learns pairing-likelihood via noise-contrastive training. On held-out tests, it assigns the native pair a higher score in 74% of cases, substantially outperforming recent pairing models. Together, AbNatiV2 and p-AbNatiV2 extend nativeness assessment and engineering to both nanobodies and conventional antibodies, supporting design decisions at single-residue, Fv-sequence, and paired-domain levels. AbNatiV2 is available as downloadable software and webserver.
8 Apr 2026
Read appraisal →Proceedings of the National Academy of Sciences of the United States of America
Proteome-wide prediction of interactions between structured domains and peptide motifs reveals functionally coherent subnetworks
Many protein-protein interactions (PPIs) are mediated by the binding of short linear motifs (SLiMs) to peptide recognition domains (PRDs). Here, we describe PrePPI-SLiM, a proteome-scale computational pipeline that leverages data from the Eukaryotic Linear Motif (ELM) database to predict whether two proteins will form a peptide-mediated complex. The ELM database defines classes of protein-peptide interactions with SLiMs represented by sequence motifs and PRDs represented by Pfam domains. PrePPI-SLiM systematically evaluates all pairwise combinations of proteins within a proteome and identifies PRD-SLiM pairs that occur in the same ELM class. This evidence together with disorder prediction and sequence conservation of the motif are integrated in a naïve Bayes framework to assign a likelihood for complex formation. To obtain potential PDB templates for atomistic models of PrePPI-SLiM interactions, we associate individual PPI predictions with homologous PDB complexes involving the same PRD Pfam domain and SLIM, and obtain PDB templates for 92% of our high-confidence predictions. Moreover, studies with AF3Complex suggest that prior knowledge of the interacting PRD and SLiM, as provided here, is a critical starting point for creating a 3D model of the specific sequences of the PRD and SLiM query proteins. Finally, we demonstrate that clustering of the high-confidence PrePPI-SLiM interactome yields functionally coherent PPI networks that reveal mechanistic insights into cellular processes. The PrePPI webserver provides convenient access to high-confidence PrePPI-SLiM predictions, PDB templates for modeling, and functional networks.
8 Apr 2026
Read appraisal →Journal of computer-aided molecular design
Computer-aided discovery of CDK16 inhibitors: a docking-augmented machine learning regression modelling approach
Cyclin-dependent kinase 16 (CDK16), a serine/threonine protein kinase, is a critical regulator of cell cycle progression, vesicle trafficking, and apoptosis. Its dysregulation is implicated in the progression of aggressive cancers, including triple-negative breast, lung, and prostate cancer, where its overexpression correlates with poor prognosis. Despite its therapeutic promise, CDK16 remains an understudied kinase lacking selective inhibitors, underscoring the need for innovative discovery approaches. This study introduces a novel computational framework that leverages multiple docked poses to augment datasets for regression-based machine learning (ML) models targeting CDK16 inhibition. This data augmentation strategy incorporated docking scores, ligand-receptor contact fingerprints (LRCFs), and conformation-sensitive physicochemical descriptors as input features. To our knowledge, this is the first application of docked pose augmentation for a regression-based drug discovery model. A systematic evaluation of eight ML algorithms across multiple docking score consensus levels identified Gradient Boosted Trees as the optimal learner. The Genetic Function Algorithm was integrated to select a minimal set of descriptors, which improved model generalizability. The resulting validated ML-QSAR model guided the generation of a robust pharmacophore model, which was used for virtual screening of the NCI and OpnMe databases. Subsequent in vitro LanthaScreen kinase assays confirmed two novel and potent CDK16 inhibitors: compound H_28 (OpnMe code: BI-831266) with an IC50 of 3.5 µM, and H_33 (OpnMe code: BI-1282) with an IC50 of 5.8 µM. These findings highlight the efficacy of integrating docked pose augmentation with regression modeling and experimental validation, offering a robust framework for accelerating the discovery of targeted anticancer agents.
7 Apr 2026
Read appraisal →Neuroscience
Limitations and opportunities in multi-omics integration for neurodevelopmental, neurodegenerative and psychiatric disorders: A systematic review
Recent advances in high-throughput technologies have led to an increased generation of biological data across genomics, transcriptomics, proteomics, epigenomics, and metabolomics. However, a major challenge remains: effectively integrating these multi-omics datasets to allow a more holistic understanding of the complex, interconnected mechanisms underlying human diseases. Neurodevelopmental, neurodegenerative, and psychiatric disorders are particularly multifactorial and heterogeneous, making them candidates for multi-omics approaches. In this context, this systematic review assesses the current state of multi-omics integration in neurological research. Records retrieved from five major databases were processed, and 156 studies were included for further analysis. The most frequently studied conditions were Alzheimer's Disease, Depressive Disorder and Parkinson's Disease, with epigenomics-transcriptomics and metagenomics-metabolomics emerging as the most common omics pairings. The field remains dominated by studies integrating pairs of omics layers. Only a limited number of computational tools are currently being applied to the integration of more than two omics layers, highlighting a gap in comprehensive multi-omics modeling. Despite progress, key challenges persist, including data accessibility and the need for standardized frameworks to allow cross-study comparisons. Moreover, most computational findings lack experimental validation in wet-laboratory settings. Future research should address these challenges, develop scalable algorithms for integrating multi-omics data, and leverage large, open-access datasets. Integrating computational predictions with experimental validation could help researchers prioritize high-confidence biomarkers relevant to clinical applications. Collaborative efforts among bioinformaticians, clinicians, and experimentalists will be essential to translating these advances into clinically actionable solutions.
6 Apr 2026
Read appraisal →Journal of computer-aided molecular design
Identification of potential inhibitors of 3‑mercaptopyruvate sulfurtransferase with a deep-learning based screening of natural products
3-Mercaptopyruvate sulfurtransferase (3-MST), a key enzyme in sulfur metabolism, has recently gained attention as a potential anticancer target. However, reported 3-MST inhibitors remain limited, motivating the exploration of new scaffolds such as natural products. In this study, a library of 3744 natural products was virtually screened against human 3-MST using DiffDock (diffusion-model-based docking) followed by AutoDock Vina docking. Top-ranking candidates were further analyzed via molecular dynamics simulations and Molecular Mechanics Poisson-Boltzmann Surface Area binding free energy calculations. Methylophiopogonanone A (4), Daphnoretin (5), and L-asarinin (9) exhibited stable binding with favorable energetics, displaying binding free energies comparable to the reference ligand 7NC301. Binding mode analyses revealed that Methylophiopogonanone A primarily engaged in hydrophobic interactions, whereas Daphnoretin and L-asarinin formed extensive polar contacts, accompanied by higher desolvation penalties. In vitro cytotoxicity assays showed that Methylophiopogonanone A and L-asarinin reduced HCT116 cell viability by 40.3% and 26.3% at 25 µM, which is consistent with their inhibitory of 3-MST with IC50 = 5.83 ± 0.69 µM and IC50 = 19.11 ± 3.37 µM, respectively. These results suggested that the natural products identified in this study represent promising scaffolds for further optimization as potential 3-MST inhibitors. This work provides an AI-guided natural-product screening workflow for 3-MST and delivers prioritized inhibitor scaffolds for subsequent optimization and experimental validation.
6 Apr 2026
Read appraisal →Genetics
Mapping whole-organism genetic comorbidities across model Species using unified ontologies
Understanding how genetic variation contributes to organism-wide phenotypes is critical for identifying mechanisms of disease. Here, we present a computational approach to analyze whole-organism comorbidities associated with genes underlying non-obstructive azoospermia (NOA), the most severe form of male infertility. We curated 204 mouse genes with experimentally validated spermatogenic failure and mapped their orthologs and phenotype annotations across humans, M. musculus, D. rerio, D. melanogaster, and C. elegans using a unified cross-species phenotype structure. This framework integrates a newly developed reproductive phenotype ontology with standardized whole-body phenotype categories to enable direct cross-species comparisons, using thousands of genotype-phenotype associations stored in model organism databases. Our analysis shows that most NOA genes have conserved orthologs across models, and that perturbation of these genes is frequently associated with non-reproductive phenotypes. In particular, integumentary defects and neoplastic phenotypes recur across species, supporting a shared genetic basis for epidemiological links between male infertility and systemic disease. Gene-level comorbidity patterns are partially predictable from single-cell RNA-seq, whole-body gene expression, and Gene Ontology annotations, suggesting that fundamental biological constraints shape the systemic consequences of reproductive gene dysfunction. Clustering genes by comorbidity profiles further distinguished genes with isolated reproductive effects from those with broad organismal consequences. This work demonstrates the power of ontology-based cross-species analysis for identifying pleiotropic effects of Mendelian disease mutations, and provides a resource, CoMorbidity DataBase Mapper (CoMo DBM), for joint analysis of genotype-phenotype associations in model organism databases.
6 Apr 2026
Read appraisal →Vaccine
Applied immunoinformatics in modern vaccine design: a comprehensive review of available computational tools
The rapid evolution of computational biology has transformed vaccine design from a largely empirical process into a rational, data-driven discipline. At the center of this transformation lies immunoinformatics, which integrates immunology, molecular biology, and data science to accelerate the identification, evaluation, and optimization of vaccine candidates. By enabling the systematic transition from genomic data to viable immunotherapeutic candidates, these approaches offer an unprecedented capacity to shorten development timelines, reduce costs, and broaden population coverage compared to traditional trial-and-error vaccine pipelines. This review provides a state-of-the-art overview of the currently available immunoinformatic tool that underpins modern vaccine design. From identification of conserved and functionally essential pathogen targets, advances in structural prediction, epitope mapping, safety assessment, and systems-level analyses of host-pathogen interactions, immunoinformatics allows researchers to design vaccines with enhanced success rate. Special emphasis is placed on the shift from purely predictive models toward generative frameworks, where artificial intelligence and deep learning now support not only the evaluation, but also the rational engineering of optimized vaccine candidates. For this review, we systematically examined more than 250 computational tools and algorithms spanning every stage of the vaccine discovery pipeline. From this extensive survey, we distilled and discussed the approaches that currently demonstrate the highest performance and practical usability, thereby providing researchers with a curated overview of some the most reliable resources available.
5 Apr 2026
Read appraisal →Current opinion in genetics & development
Deep learning for psychiatric genomics: from tools to applications
The genetic architecture of psychiatric disorders is highly complex, with genome-wide association studies implicating thousands of risk loci. A central challenge is that most of these variants are located in noncoding regions, making it difficult to elucidate their regulatory consequences within the brain's intricate cellular landscape. The recent convergence of advanced artificial intelligence, particularly deep learning (DL), has catalyzed a paradigm shift by providing powerful tools to address this gap. This review traces the evolution of DL in genomics, beginning with task-specific models. We then examine the transformative impact of foundation models, pretrained neural networks that learn the 'language' of biology, including genomic language models, single-cell foundation models, and large language models originally trained on natural language. Finally, we survey applications to key problems in psychiatric genomics. We hope this review provides a comprehensive overview of recent advances in DL for genomics and serves as a bridge to help researchers in psychiatric genomics more effectively understand and apply these frontier methods to guide the development of novel therapeutic strategies for psychiatric disorders.
4 Apr 2026
Read appraisal →Cell
A generative deep learning approach to de novo antibiotic design
The antimicrobial resistance crisis necessitates structurally distinct antibiotics. While deep learning approaches can identify antibacterial compounds from existing libraries, structural novelty remains limited. Here, we developed a generative artificial intelligence framework for designing de novo antibiotics through two approaches: a fragment-based method to comprehensively screen >107 chemical fragments in silico against Neisseria gonorrhoeae or Staphylococcus aureus, subsequently expanding promising fragments, and an unconstrained de novo compound generation, each using genetic algorithms and variational autoencoders. Of 24 synthesized compounds, seven demonstrated selective antibacterial activity. Two lead compounds exhibited bactericidal efficacy against multidrug-resistant isolates with distinct mechanisms of action and reduced bacterial burden in vivo in mouse models of N. gonorrhoeae vaginal infection and methicillin-resistant S. aureus skin infection. We further validated structural analogs for both compound classes as antibacterial. Our approach enables the generative deep-learning-guided design of de novo antibiotics, providing a platform for mapping uncharted regions of chemical space.
1 Oct 2025
Read appraisal →