Research Appraisals
Evidence-based critical appraisals of the latest medical research, systematically evaluated using Oxford CEBM methodology.
Showing 13 appraisals
PLoS computational biology
Twelve quick tips for AI-assisted coding in science.
While AI coding tools have demonstrated potential to accelerate software development, their use in scientific computing raises critical questions about code quality and scientific validity. In this paper, we provide twelve practical tips for AI-assisted coding that balance the capabilities of AI with the demands of scientific and methodological rigor. We address how AI can be leveraged strategically throughout the development cycle with four key themes: problem preparation and understanding, managing context and interaction, testing and validation, and code quality assurance and iterative improvement. These principles serve to emphasize maintaining human agency in coding decisions, establishing robust validation procedures, and preserving the domain expertise essential for methodologically sound research. These tips are intended to help researchers harness AI's transformative potential for faster software development while ensuring that their code meets the standards of reliability, reproducibility, and scientific validity that research integrity demands.
29 July 2026
Read appraisal →Journal of chemical information and modeling
Protein Language Model-Based Fitness Estimates Facilitate Resistance Mutation Identification
Drug resistance is a major challenge in cancer therapy. Cancer cells with pre-existing or acquired mutations that confer resistance to a given drug treatment outgrow the susceptible cell population and cause cancer recurrence after an initial successful treatment response. Knowledge about resistance mutations before they occur in the clinic could prevent unnecessary patient treatment with ineffective drugs, in clinical trials as well as clinical practice, or potentially speed up the development of follow-up compounds. Here, we focused on on-target amino acid mutations that confer resistance to an inhibitor compound with a known binding mode. We evaluated whether a combination of physics-based free energy perturbation (FEP) affinity estimates and protein language model-based protein fitness estimates could improve the in silico identification of resistance mutations. Validation was done with data from deep mutational scanning (DMS) experiments that tested for resistance to single amino acid mutations. A public data set testing ERK2 resistance against the inhibitor SCH772984 and an internal data set testing resistance of an EGFR_exon20 mutant against a Bayer small molecule inhibitor were used. Our results show that protein fitness estimates can facilitate the identification of resistance mutations by filtering mutations with a low estimated fitness. Even though FEP has flagged such mutations as affinity-decreasing and thus potentially resistant, they were not resistant according to the DMS experiment and therefore correctly filtered out. This indicates that protein language model-based protein fitness estimates could be a computationally efficient method to filter mutations without having to model the negative impact of mutations on native function or protein stability, which is error-prone and computationally expensive.
28 July 2026
Read appraisal →Genome biology
IDEAL-Age: an interpretable deep learning framework for single-cell resolution profiling of immunological aging
Immunosenescence increases susceptibility to infection and reduces vaccine responsiveness, yet bulk transcriptomic clocks obscure the cellular heterogeneity underlying this process. Here, we present IDEAL-Age, an interpretable deep learning framework that operates directly on single-cell PBMC transcriptomes. Benchmarking against 35 methods across independent cohorts demonstrates superior predictive performance. The framework's interpretability uncovers linear and non-linear gene contribution trajectories that reveal phase-specific physiological transitions, and identifies youth-associated or aging-associated cellular roles. Application to systemic lupus erythematosus reveals accelerated immunological aging driven by interferon-associated monocyte shifts. IDEAL-Age establishes a high-resolution computational framework for deciphering systemic immune aging.
27 July 2026
Read appraisal →Cell
Deep learning of functional perturbations from condensate morphology
Biomolecular condensates compartmentalize the interior of cells to organize complex functions, yet linking molecular interactions within condensates to their mesoscale organization remains a major challenge. To bridge this gap, we developed a neural-network-based framework-Deep-Phase (deep learning of phase-separated condensates)-that uses microscopy images to directly measure condensate morphology changes resulting from pharmacological alterations in associated biochemical processes. We use Deep-Phase to precisely quantify time- and concentration-dependent structural perturbations to the multiphase nucleolus and show that they are tightly coupled to potencies of drugs inhibiting ribosomal RNA (rRNA) transcription and processing. Applying Deep-Phase in a chemical screen, we identify a unique nucleolar morphology and discover a role for a DNA topoisomerase in rRNA processing. Mechanistic studies of this morphology provide insights into how the interfaces between nucleolar sub-compartments are maintained. We demonstrate Deep-Phase's adaptability to diverse cell lines, labeling techniques, and condensates, offering a powerful platform for connecting molecular pathways to cellular mesoscale organization.
25 July 2026
Read appraisal →Cell reports methods
Spatial multi-omics imputation and embedding with SpaMIE
SpaMIE is a deep graph neural network framework designed to tackle the challenge of multi-section integration in spatial multi-omics (SMO) datasets with systematic missing modalities. Current SMO platforms face limitations such as high cost and limited throughput, leading to many large-scale spatial atlases relying on cost-effective mono-omics measurements while only a few sections are profiled with full multi-omics technologies. This results in heterogeneous modality coverage across tissue sections. SpaMIE offers a two-stage solution. In the first stage, it performs spatially informed cross-modal imputation, enabling accurate inference of missing modalities from mono-omics data. In the second stage, it integrates measured and imputed spatial multi-omics profiles across multiple tissue sections to learn a unified embedding. Benchmarking on simulated and experimental datasets shows that SpaMIE achieves accurate cross-modal imputation, robust multi-section integration, and improved spatial domain identification, providing a flexible and scalable solution for constructing and analyzing SMO atlases.
22 July 2026
Read appraisal →International journal of cancer
Drug repurposing in oncology: Bridging computational discovery to clinical application.
Drug repurposing, the identification of new therapeutic applications for existing drugs, has emerged as a pragmatic and cost-efficient strategy to accelerate oncology drug discovery. Faced with rising development costs, protracted timelines, and high attrition rates associated with traditional de novo drug development, repurposing leverages known pharmacokinetics, safety profiles, and manufacturing processes to expedite clinical translation. This review synthesizes current advances in computational and experimental methodologies and mechanistic insights that drive drug repurposing for cancer therapy. In silico strategies, including molecular docking, machine learning, transcriptomic-proteomic signature reversal, and network-based modeling, have enabled rapid identification of repurposable agents by mining multi-omics and historical pharmacological data. Experimental pipelines spanning high-throughput screening, phenotypic assays, biochemical validations, and functional animal models remain essential to establish efficacy and delineate mechanisms of action. Notably, repurposed drugs exhibit anticancer activity by modulating key pathways, including phosphatidylinositol 3-kinase/Ak strain transforming/mechanistic target of rapamycin, mitogen-activated protein kinase/extracellular signal-regulated kinase, wingless-related integration site/β-catenin signaling, as well as redox homeostasis and DNA response. Despite their promise, repurposed candidates face barriers including limited intellectual property protections, dose optimization challenges, and regulatory uncertainty. Moreover, clinical translation is often hindered by insufficient mechanistic understanding and a lack of predictive biomarkers. Integration of multi-omics datasets, explainable artificial intelligence, patient-derived organoids, and clustered regularly interspaced short palindromic repeats-based genetic screens now offers unprecedented precision in identifying context-specific drug effects and synthetic lethal interactions. With cancer causing one in six global deaths and marked by therapeutic resistance and molecular heterogeneity, drug repurposing provides a scalable solution. This approach bridged preclinical insight with clinical application, potentially transforming cancer therapeutics through rational, data-driven innovation.
15 July 2026
Read appraisal →Journal of chemical information and modeling
DeepKbhb: Context-Aware Prediction of Human Lysine β-Hydroxybutyrylation Sites
Lysine β-hydroxybutyrylation (Kbhb) is a metabolism-linked post-translational modification (PTM) that plays a critical role in regulating gene expression, stress responses, and disease progression. Despite its emerging biological significance, identifying Kbhb sites remains limited due to the cost and complexity of experimental methods. Prior work such as KbhbXG is constrained by its reliance on hand-crafted features and lacks the ability to model contextual dependencies within sequences. To address this challenge, we present DeepKbhb, a deep learning framework designed for human Kbhb site identification. By integrating sequence embeddings and six engineered descriptors through a bilinear attention network, DeepKbhb effectively captures position-dependent relationships essential for accurate Kbhb site prediction. On an independent test set, DeepKbhb achieved state-of-the-art performance with an accuracy of 0.856, an F1-score of 0.863, and a Matthews correlation coefficient of 0.716. Experimental results across multiple evaluation metrics confirm the superior performance of DeepKbhb, highlighting its potential as a valuable tool for advancing Kbhb-related functional and mechanistic studies. This capability can further support disease-oriented research, particularly in cancer, metabolic disorders, and immune regulation. Further sequence analyses revealed distinct local amino acid preferences, supporting the biological relevance of our model. The web interface is accessible at https://awi.cuhk.edu.cn/~DeepKbhb/.
14 July 2026
Read appraisal →Nucleic acids research
DeepKinomeWeb: a quantitative, panel-level platform for kinase inhibitor screening and selectivity profiling
Protein kinases are central targets in drug discovery, yet early-stage development of potent and selective inhibitors remains challenging due to high experimental costs and limited interpretability of large-scale screening data. Here, we present DeepKinomeWeb, an integrated web-based platform that transforms competition-based high-throughput screening data into actionable insights for kinase inhibitor prioritization. Built upon our previously validated deep learning regression model, DeepKinome, the platform enables quantitative prediction of kinase-inhibitor binding affinities and provides panel-level visualization of selectivity landscapes, selectivity metric calculations, and integrated structural and physicochemical analyses. Through its user-friendly interface, DeepKinomeWeb supports rational, data-driven decision-making for biologists and medicinal chemists, lowering the barrier to systematic selectivity assessment in kinase inhibitor discovery. DeepKinomeWeb is freely available to all users without any login requirement at https://str.kribb.re.kr/deepkinome.
13 July 2026
Read appraisal →Functional & integrative genomics
A multi-scale graph frequency network for structural and functional region analysis in spatial transcriptomics
Spatial transcriptomics enables the systematic exploration of how gene expression patterns are organized within intact tissues, yet effective analysis remains difficult due to the complexity of spatial dependencies and multi-scale tissue architectures. Here, we present the Spatial Graph Frequency Network (SGFN), a deep learning framework that integrates graph signal processing, graph attention, and contrastive learning to jointly model spatial topology and molecular features. Central to SGFN is a frequency-domain enhancement module that decomposes spatial graphs into multi-scale spectral components using the Laplacian eigenbasis, complemented by adaptive wavelet denoising when the retained graph-frequency sequence length permits valid decomposition. Evaluation across diverse biological systems-including the human dorsolateral prefrontal cortex, mouse brain, human breast cancer, osmFISH, MERFISH, STARmap, mouse embryonic development, head and neck angiosarcoma, and brain metastasis-shows that SGFN achieves improved or competitive performance relative to representative baseline methods in reference-based benchmarks, and identifies biologically coherent spatial or functional regions in unlabeled datasets supported by marker-gene, spatial-autocorrelation, cell-type-colocalization, and pathway-enrichment evidence. SGFN accurately reconstructed cortical layer architecture in the human brain, delineated immune and metabolic modules in tumors, and revealed spatiotemporal trajectories during embryogenesis. By combining interpretable frequency-domain representations with data-driven learning, SGFN provides a unified computational framework for decoding tissue organization and molecular heterogeneity, advancing the understanding of developmental, physiological, and pathological spatial systems.
7 July 2026
Read appraisal →Briefings in bioinformatics
Bridging local-global transmembrane protein contexts with contrastive pretraining for alignment-free pathogenicity prediction.
Predicting the pathogenic consequences of protein mutations is a cornerstone of precision medicine, yet it remains a formidable challenge for transmembrane proteins (TMPs), a clinically vital class of drug targets. Existing computational methods are often hampered by their reliance on evolutionary data and fail to model TMP-specific biophysical constraints. Here, we introduce Memo-Patho, a deep learning framework for robust, alignment-free pathogenicity prediction of TMP variants. The core innovation is a within-protein, label-informed supervised contrastive pretraining strategy that learns sequence-encoded biophysical signatures distinguishing pathogenic and benign variants by directly comparing them within the same protein context. By fusing sequence-level representations from protein language models with local structural proxies derived from sequence, Memo-Patho achieves accurate predictions without multiple sequence alignments or experimental structures. Across diverse TMP benchmarks and under protein-level group splits, Memo-Patho consistently outperforms leading predictors, achieving up to 0.93 accuracy, and it transfers to an independent KCNQ1 ion-channel cohort without re-training. Its resource-efficient, alignment-free design enables routine large-scale screening when evolutionary or structural data are sparse. Conceptually, Memo-Patho addresses a key gap by directly learning discriminative, sequence-anchored signatures pertinent to TMP-specific constraints, offering a principled and generalizable foundation for research-use clinical variant triage and proteome-wide mutation-effect modeling.
6 July 2026
Read appraisal →Journal of advanced research
scMapNet: Marker-based cell type annotation of scRNA-seq data via vision transfer learning with tabular-to-image transformations
INTRODUCTION: Identifying cell types is a key step in single-cell RNA sequencing data analysis that aids in understanding cellular heterogeneity and facilitates downstream analyses such as those concerning cell-cell interactions and data integration. Cell-type annotation methods often rely on unsupervised clustering algorithms, followed by manual or automatic annotation via marker genes, which are prone to inefficiency and inconsistency. Supervised methods, while more automated and consistent, have gained remarkable attention because of the rapid growth of large-scale, high-quality single-cell datasets. However, these methods cannot effectively leverage cellular marker knowledge and much unlabelled data. OBJECTIVES: This study aims to introduce a novel deep learning method, scMapNet, which can sufficiently learn cellular marker knowledge and information from unlabelled data. METHODS: scMapNet is a self-supervised deep learning model based on masked autoencoders (MAE) and vision transformer (ViT), which adopts treemap transformations to leverage cell marker information and capture information by pretraining on large amounts of unlabelled data. RESULTS: scMapNet outperforms six competing methods across diverse datasets, excelling in accuracy and batch insensitivity. Moreover, scMapNet can effectively extract attention gene information, providing strong support for cell type identification. The models and codes are available at https://github.com/Yuz7/scMapNet. CONCLUSION: scMapNet shows significant performance in annotation accuracy and batch insensitivity. Besides, this method also has good interpretability and can provide biological insights for researchers.
3 July 2026
Read appraisal →Journal of advanced research
Machine learning for food flavor prediction and regulation: models, data integration, and future perspectives.
BACKGROUND: Flavor is a central attribute of food quality, shaping consumer preferences and market performance. Traditional evaluation methods, such as sensory panels and basic assays, are often constrained by subjectivity, low throughput, and limited scalability. With the rise of high-throughput technologies and multimodal datasets, machine learning (ML) has emerged as a promising tool for deciphering and regulating complex flavor systems. AIM OF REVIEW: This review examines current flavor detection techniques and the application of ML across diverse domains. It compares supervised learning models (SVM, DT), ensemble algorithms (XGBoost, LightGBM), and deep learning approaches (CNN, ANN). This review also discusses the contribution of three major data dimensions to flavor prediction, as well as future prospects in the field. ML enables precise flavor prediction, compound screening, and real-time process control. To support these tasks, researchers have developed integrated analytical systems that combine electronic nose (E-nose), electronic tongue (E-tongue), gas chromatography-mass spectrometry (GC-MS), and gas chromatography-ion mobility spectrometry (GC-IMS). Ensemble learning and deep learning models show strong performance when handling complex, nonlinear datasets. Explainable artificial intelligence (XAI) tools such as Shapley Additive Explanations (SHAP) improve model transparency by linking predictions to underlying features. ML models further enhance both prediction accuracy and generalizability. Innovations such as attention mechanisms, graph neural networks, and digital twins support dynamic flavor modulation. ML also aids in identifying key flavor compounds and genotype-phenotype relationships, accelerating breeding and formulation. KEY SCIENTIFIC CONCEPTS OF REVIEW: ML is opening up new technological avenues in flavor science, with significant potential to predict and control flavor formation mechanisms, verify product authenticity, and support the targeted design of flavor-active compounds that align with consumer expectations for sensory appeal.
2 July 2026
Read appraisal →Cell systems
Fusing imaging and metabolic modeling via multimodal deep learning in ovarian cancer
Integrating genotype (e.g., transcriptomics), phenotype (e.g., imaging), and tumor microenvironment (e.g., metabolomics) is crucial to elucidating the molecular basis of ovarian cancer. However, there is a lack of robust multimodal integration methods when only a limited number of common samples is available. Here, we generate patient-specific metabolic models starting from transcriptomics data and integrate them with imaging data. We show that this multimodal integration-never attempted before-improves survival estimation and enables a mechanistic interpretation of the predictions. We assess the robustness of our approach with different combinations of transcriptomics, fluxomics, and 3D computerized tomography (CT) imaging data, correctly stratifying patients based on risk. Fusing metabolic modeling with imaging and transcriptomics significantly improves model accuracy compared with widely used transcriptomics-imaging approaches and elucidates critical metabolic reactions. Our approach is general and can be applied to other cancer types where coupled imaging-transcriptomics data are available. A record of this paper's transparent peer review process is included in the supplemental information.
19 June 2026
Read appraisal →