Research Appraisals
Evidence-based critical appraisals of the latest medical research, systematically evaluated using Oxford CEBM methodology.
Showing 26 appraisals
Neural networks : the official journal of the International Neural Network Society
An approximation theory perspective on machine learning
A central problem in machine learning is often formulated as follows: Given a dataset [Formula: see text] , which is a sample drawn from an unknown probability distribution, the goal is to construct a functional model f such that f(x) ≈ y for any (x, y) drawn from the same distribution. Neural networks and kernel-based methods are commonly employed for this task due to their capacity for fast and parallel computation. The approximation capabilities, or expressive power, of these methods have been extensively studied over the past 35 years. In this paper, we will present examples of key ideas in this area found in the literature. We will discuss emerging trends in machine learning including the role of shallow/deep networks, approximation on manifolds, physics-informed neural surrogates, neural operators, and transformer architectures. Despite function approximation being a fundamental problem in machine learning, approximation theory does not play a central role in the theoretical foundations of the field. One unfortunate consequence of this disconnect is that it is often unclear how well trained models will generalize to unseen or unlabeled data. In this review, we examine some of the shortcomings of the current machine learning framework and explore the reasons for the gap between approximation theory and machine learning practice. We will then review some of recent work that achieves function approximation on unknown manifolds without the need to learn specific manifold features, such as the eigen-decomposition of the Laplace-Beltrami operator or atlas construction. In many machine learning problems, particularly classification tasks, the labels yj are drawn from a finite set of values. We summarize another recent paper that establishes a deep connection between signal separation problems and classification problems, proposing that classification tasks should be approached as instances of signal separation. We conclude by identifying several open research problems that warrant further investigation.
2 Aug 2026
Read appraisal →Cell reports methods
Data modeling the interplay between single-cell shape, single-cell protein expression, and tissue state.
While cell shape fundamentally governs tissue function, the underlying links between single-cell shape and protein expression have been difficult to resolve due to limitations in imaging multiplexing and population averaging. Here, we use multiplexed imaging and machine learning to investigate the coupling of cell shape and protein expression in heterogeneous human tissues. Our analysis establishes a universal, bi-directional link between a cell's shape and its protein expression across cell types, diseases, and disease states, amplified for cell state markers. Machine learning interpretability shows that shape features can potentially generate hypotheses of protein functions. Screening all protein-cell type pairs identified a subpopulation of large, p53-positive tumor cells across two cancers. Shape properties further enhanced graph neural network disease state prediction. Our results open the door to unraveling the intricate connections between protein expression, cell shape, tissue organization, and tissue state in a physiological context.
23 July 2026
Read appraisal →eLife
Disinformation elicits learning biases.
In open societies, disinformation is often considered a threat to the very fabric of democracy. However, we know little about how disinformation exerts its impact, especially its influence on individual learning processes. Guided by the notion that disinformation exerts its pernicious effects by capitalizing on learning biases, we ask which aspects of learning from potential disinformation align with ideal 'Bayesian' principles, and which exhibit biases deviating from these standards. To this end, we harnessed a reinforcement learning framework, offering computationally tractable models capable of estimating latent aspects of a learning process as well as identifying biases in learning. In two experiments, participants completed a two-armed bandit task, where they repeatedly chose between two lotteries and received outcome-feedback from sources of varying credibility, who occasionally disseminated disinformation by lying about true choice outcome (e.g., reporting non-reward when a reward was truly earned or vice versa). Computational modelling indicated that learning increased in tandem with source credibility, consistent with ideal-Bayesian principles. However, we also observed striking biases reflecting divergence from idealized Bayesian learning patterns. Notably, in one experiment individuals learned from sources that should have been ignored, as these were known to be fully unreliable. Additionally, the presence of disinformation elicited exaggerated learning from trustworthy information (akin to jumping to conclusions) and exacerbated a normalized measure of 'positivity bias' whereby individuals self-servingly boost their learning from positive, relative to negative, choice feedback. Thus, in the face of disinformation we identify specific cognitive mechanisms underlying learning biases, with potential implications for societal strategies aimed at mitigating its harmful impacts.
28 June 2026
Read appraisal →Journal of chemical information and modeling
SSPSPredictor: A Sequence- and Structure-Based Deep Learning Model for Predicting Phase-Separating Proteins
Liquid-liquid phase separation (LLPS) underlies the formation of biomolecular liquid condensates (also termed membraneless organelles, MLOs), which are essential for spatially organizing various biochemical processes within cells. Proteins that play a key role in driving condensate formation are termed phase-separating proteins (PSPs). Given that experimental identification of PSPs remains labor-intensive and time-consuming, multiple computational tools have been developed based on empirical features or deep learning. In this study, we propose SSPSPredictor, a novel multimodal predictive model for PSPs with folded or intrinsically disordered structures, leveraging the fusion of sequence information from a protein language model ESM-2 and structural insights from a graph neural network GVP. Compared with existing tools, SSPSPredictor achieves a balanced performance in identifying PSPs, predicting relative LLPS propensities, and recognizing key regions that drive LLPS. More importantly, SSPSPredictor exhibits good interpretability in identifying driving regions along protein sequences, although no relevant supervision was provided during training. Further predictive analysis of the human proteome using SSPSPredictor reveals that the proportion of intrinsically disordered proteins (IDPs) undergoing LLPS is significantly higher than that of the folded proteins. In addition, pathogenic variants, especially those located in disordered regions, exhibit a higher LLPS propensity than other mutations, uncovering a link between LLPS and diseases at the amino acid level. An online Web server for SSPSPredictor (https://bio-comp.ucas.ac.cn/SSPSPredictor/) and its source code (https://github.com/ucas-biocomp/SSPSPredictor) are available.
23 June 2026
Read appraisal →Journal of medicinal chemistry
Developing Predictive Models by Sharing Predictions - An Investigation of a Federated Learning Approach for ADMET Predictions
Machine learning models for ADMET prediction benefit from large, diverse data sets, yet such data are typically siloed across organizations. Federated learning (FL) enables collaborative modeling while preserving data privacy. Here, we investigate a student-teacher model (STM) framework in which organizations train internal models on proprietary data and share predictions on a public data set to generate pseudolabels for a centralized student model. As a proof of concept, 11 pharmaceutical companies contributed predictions for rat steady-state volume of distribution, yielding a pseudolabeled data set of ∼133,000 compounds. The resulting student model achieved performance comparable to individual teacher models on an external test set (RMSE ≈ 0.51 vs 0.47-0.61). Compared with FL approaches such as MELLODY and Effiris, STM offers a simpler workflow that avoids direct data sharing or iterative collaboration, providing a practical and scalable framework for secure cross-company model development.
13 June 2026
Read appraisal →Nature communications
Adaptive Riemannian optimization for multi-scale diffeomorphic matching
Image matching is a fundamental task in quantitative biomedical and biological image analyses, enabling researchers to compare, integrate, and interpret imaging data across subjects, time points, modalities, and experimental conditions. Existing state-of-the-art registration methods are slow due to inefficient implementations and poor convergence rates because of the ill-conditioned nature of the optimization problem. Deep learning methods offer fast inference but require extensive training time, substantial inference memory, and fail to generalize across long-tailed distributions or diverse image modalities, necessitating costly retraining. We address these challenges by proposing FireANTs, a training-free, GPU-accelerated, multi-scale adaptive Riemannian optimization algorithm for fast and accurate dense diffeomorphic image matching. FireANTs more than doubles the speed of the community standard ANTs registration tool on a CPU, and is two orders of magnitude faster on a GPU. On the GPU, FireANTs performs competitively with deep learning methods on inference runtime while consuming up to 10 × less memory. FireANTs demonstrates robustness on a wide variety of matching problems across modalities, species, and organs, without any domain-specific training or tuning. Our framework allows hyperparameter grid search studies with less resources and time compared to traditional and deep learning registration algorithms alike.
12 June 2026
Read appraisal →Bioorganic chemistry
Multi-relational knowledge graph for drug-drug interaction prediction via dual aggregation and collaborative optimization
Identifying potential drug-drug interactions is crucial in clinical care and new drug development, as mutual interference between drugs can lead to adverse reactions. Recently, computational methods have been widely employed for predicting DDIs. However, these approaches often encounter high computational complexity during the process of node feature representation and frequently overlook the interactivity and polysemy of entities. Effectively handling heterogeneous information remains a significant challenge. To cope with these limitations, we propose a dual aggregation and collaborative optimization learning framework，named MRACO，for DDIs prediction based on a multi-relational knowledge graph. MRACO utilizes structural information from these graphs, efficiently learning interactivity information across various relationship types. Additionally, MRACO employs dual aggregation operations to explicitly encode and aggregate multi-type information, capturing diverse semantic relationships of drug nodes. Moreover, MRACO simplifies complex computational processes by utilizing the collaborative loss optimization function, eliminating redundant information and enhancing model robustness. By integrating deep-level interactive information, MRACO leverages rich contextual data from knowledge graphs, enhancing prediction stability. Experiments demonstrate the effective of MRACO feature extraction capabilities in multi-relational networks, highlighting its understanding of the underlying mechanisms of DDIs.
7 June 2026
Read appraisal →Physics in medicine and biology
Lattice peak optimization: a mixed-integer framework for geometry-adaptive lattice radiotherapy
Objective.Lattice radiotherapy (LATTICE) delivers spatially distributed high-dose peaks within the tumor volume while maintaining lower doses in surrounding valley regions. Determining feasible peak locations is typically performed using heuristic or manual approaches, which may limit the number and spatial distribution of deliverable peaks. This work introduces a lattice peak optimization (LPO) framework that jointly optimizes peak placement and dose distribution to identify the maximum number of geometrically feasible peaks within the target.Approach.Proton LATTICE planning is formulated as a mixed-integer optimization problem that selects an optimal subset of peaks from a large set of candidate locations within the target. Binary variables represent peak selection and continuous variables model spot weights. The formulation enforces geometric feasibility between peaks while optimizing dosimetric objectives to improve peak-to-valley dose ratio (PVDR) and reduce organs-at-risk (OAR) dose. The resulting nonconvex problem is solved using iterative convex relaxation within an alternating direction method of multipliers framework.Main results.LPO was evaluated on three clinical cases with 150-400 candidate peak locations, from which 4-13 peaks were selected. Compared with 50-90 randomly generated LATTICE configurations per case, LPO consistently achieved higher PVDR and improved OAR sparing. In an abdominal case, the composite objective value was 2.93 (worst random), 2.40 (median random), 1.90 (best random), and 1.95 (LPO), with similar trends observed across all cases.Significance.A geometry-adaptive, mixed-integer optimization framework for lattice peak placement is presented, demonstrating improved PVDR and OAR sparing relative to manual and random LATTICE approaches. The present study evaluates performance of the proposed framework in the context of proton LATTICE planning; however, the framework is mathematically modality-agnostic and could in principle be applied to other radiation modalities.
27 May 2026
Read appraisal →PloS one
Learning semantic similarity from sentence pairs using hybrid features centric approach and explainable siamese neural networks.
Semantic embeddings play an important role in modern natural language processing because they help models understand meaning beyond individual words. Accurate text similarity is essential for many applications such as search, automated scoring, summarization, and question answering. However, existing methods based on Term Frequency-Inverse Document Frequency (TF-IDF) or simple lexical overlap often fail when sentences differ in length, structure, or word choice. These methods are less in performance, especially when working with short or medium-length sentences where meaning is expressed in different ways. This study explores sentence-level similarity using a Siamese BiLSTM model that learns deep semantic patterns and relationships between two sentences. The model captures contextual meaning, word interactions, and paraphrastic variations more effectively than traditional approaches. Experimental results show that the proposed model achieves the highest performance among machine-learning regressors, with lower errors and improved stability. Compared to TF-IDF, cosine similarity, and feature-based regressors, the Siamese model provides more accurate judgments of semantic closeness with RMSE of 0.16 and MAE of 0.107. Feature-level analysis using TF-IDF, Jaccard similarity, and embedding distances further supports these findings. Explainable AI techniques (SHAP, LIME) confirm model transparency by highlighting meaningful semantic cues and distributing attention across important linguistic features.
23 May 2026
Read appraisal →Journal of pharmacokinetics and pharmacodynamics
Multi-objective optimization in population pharmacokinetic model selection and optimization: application of NSGA-II in pyDarwin
The selection of a "good" model usually involves a combination of objective and subjective criteria. Although many aspects of model quality can be expressed numerically, certain desirable characteristics remain difficult-or even impossible-to quantify precisely. Multi-objective optimization (MOO) provides a systematic way to handle this challenge by explicitly incorporating and balancing both objective (measurable) and subjective (judgment-based) considerations when choosing among candidate solutions. The generated Pareto front represents a set of non-dominated models where no single solution can be improved in one objective without sacrificing the performance in another objective. Using the non-dominated sorting genetic algorithm II (NSGA-II), an implementation of MOO, we simultaneously considered objective function value and number of estimated parameters as competing criteria. Concentration measurements of 17-DMAG, quetiapine, clozapine and ziprasidone were applied to build population pharmacokinetic models through traditional stepwise search, machine learning based single-objective hybrid genetic algorithm (SOHGA) and MOO. Local downhill search with MOO was also assessed in this study. While both objectives improved, models with lower objective function value generally contained more estimated parameters. The number of non-dominated solutions for DMAG, ziprasidone, clozapine, and quetiapine was 17, 9, 9, and 13, respectively. The optimal model selected by SOHGA appeared on the Pareto front for DMAG, ziprasidone and clozapine datasets. Overall, MOO provides objective transparency to the cost of tradeoffs between competing model objectives, allowing researchers to better contextualize subjective criteria (e.g., biological plausibility, improvements in diagnostic plots) when aligning model selection with clinical context.
18 May 2026
Read appraisal →Computer methods and programs in biomedicine
Geometry-aware PointNet for rapid prediction of cerebral aneurysm hemodynamics
BACKGROUND AND OBJECTIVE: Cerebral aneurysms affect 2-5% of the global population and pose a significant health risk upon rupture. While computational fluid dynamics provides detailed hemodynamic information for risk assessment, its high computational demands limit routine clinical use. This study aims to develop a deep learning model to rapidly predict three-dimensional velocity fields and wall shear stress at peak systole using aneurysm geometry. METHODS: We synthesized a dataset of 984 idealized middle cerebral artery bifurcation aneurysms. For each case, computational fluid dynamics simulations were conducted with pulsatile boundary conditions to generate ground-truth data, and peak-systolic snapshots were extracted. We developed a single-input point-cloud network augmented with a distance-to-wall feature to predict both velocity fields and wall shear stress. Model performance was evaluated using mean absolute error (MAE), normalized MAE (NMAE), and relative L2 error (rL2). RESULTS: On the test set, the model achieved velocity-field accuracy of NMAE 4.05% and rL2 19.2% over the full geometry domain, and WSS accuracy of NMAE 2.59% and rL2 23.9%. The mean inference time was approximately 1.6 seconds for velocity and 0.3 seconds for wall shear stress per case. Out-of-distribution evaluation on non-idealized geometries showed substantial zero-shot degradation (NMAE 19.1%, rL2 62.4%), while leave-one-out fine-tuning improved performance (NMAE 10.8%, rL2 37.0%). CONCLUSIONS: The proposed geometry-aware point-cloud surrogate provides fast peak-systolic hemodynamic prediction on idealized aneurysm geometries. However, out-of-distribution evaluation indicates that broader patient-specific training, physiological boundary conditions, and reliability assessment are required before routine clinical application.
17 May 2026
Read appraisal →Biological psychiatry
Multimodal Representation Learning for Parsing Biological Heterogeneity in Psychiatric Neuroimaging
For decades, psychiatric neuroimaging has searched for biomarkers of depression and other disorders, but they remain elusive in clinical practice. While the last 5 years have seen rapid progress, other large-scale correlative studies have found only small, unreliable links between brain measures and clinical symptoms. Growing evidence suggests that such limitations are not just about sample size but depend critically on how models represent data. This review traces a recent shift away from univariate methods to multivariate/multiview approaches that learn more effective representations of biological and symptom measures by flexibly learning multimodal latent representations. First, we review how linear multiview embedding methods have revealed reproducible biological depression subtypes but do not perform well in small samples or samples enriched for mild symptoms. Then, we consider newer work exploring more sophisticated representations for neuroimaging data, including deep-learning and graph-based representations, and multimodal extensions that uncover complex latent patterns that single-modality studies miss. Then, we review recent developments in foundation models, which, once trained on large corpora, can "transfer learn" readily to small clinical cohorts, potentially bringing the advantages of large-scale learning to small, privacy-limited data. Finally, we highlight emerging representation tools that treat the brain as a dynamic, stateful multivariate process. Taken together, these advances point to a future in which the value of neuroimaging will be determined not only by ever-larger sample sizes but also by data quality and by how well our algorithms capture the distributed, multimodal, and evolving nature of psychiatric disorders.
17 May 2026
Read appraisal →PloS one
TDGN: A text-guided dual-gated network for multimodal sentiment analysis
Multimodal Sentiment Analysis (MSA) aims to interpret emotions by integrating textual, acoustic, and visual information. However, the heterogeneous quality and weak correlation among nonverbal modalities often lead to unstable alignment and ineffective fusion. To address these challenges, we propose a Text-Guided Dual-Gated Network (TDGN) that introduces a hierarchical gating and text-anchored contrastive learning framework. During the alignment phase, a Text-Anchored Gated Attention (TGA) module employs text as a semantic anchor to guide fine-grained alignment between audio and visual modalities while suppressing noise and emphasizing salient emotional cues. For the fusion stage, a Dual-layer Gated Fusion (DGF) module performs intra-modal gating to refine modality-specific features and inter-modal gating to dynamically balance cross-modal contributions. Furthermore, we introduce a Text-Anchored Contrastive Learning (TACL), which constructs contrastive targets based on textual similarity anchors, ensuring the fused features maintain both modal consistency and modal diversity. Extensive experiments on the CMU-MOSI and CMU-MOSEI benchmarks demonstrate that TDGN achieves state-of-the-art performance, achieving higher accuracy and robustness. Ablation studies and visualization further validate the effectiveness of each component.
14 May 2026
Read appraisal →PloS one
Narrative-affect discrepancy as a regulated degree of freedom in 351,734 relationship narratives
In naturalistic emotional narratives, the intensity of expressed affect does not scale proportionally with narrative structure. Using 351,734 English-language relationship narratives from online support communities (the ANEST Narrative-Affect Dataset, ANAD v1.1.0), we constructed a two-dimensional expressive space defined by narrative complexity (N) and linguistically inferred affective intensity (A), with their signed discrepancy ([Formula: see text]) treated as a derived coordinate. Rather than converging toward low discrepancy, human narratives occupied a broad but structured space consistent with trade-offs between relational exposure and cognitive effort. We identified four empirically separable regimes of expressive organization: coupled expression (non-extreme discrepancy; the complement of the extreme regimes), strategic understatement (high A, low N, D < 0), strategic overstatement (high N, low A, D > 0), and collapse (high A with limited narrative scaffolding). A data-anchored cost model (NCS) formalized these regimes as arising from the interaction of exposure risk and cognitive effort, not from discrepancy minimization per se. Coupled expression dominated the corpus (91.3%), while the remaining regimes formed smaller but non-negligible subpopulations (Understatement: n = 20,223; Collapse: n = 8,040; Overstatement: n = 2,223), indicating that extreme discrepancy configurations occur systematically rather than as isolated outliers. As a comparative probe, we projected an RLHF-aligned large language model into the same space using matched prompts and identical feature extraction. Because D is a deterministic function of N and A, expressive extent was quantified as convex hull area in the clipped [Formula: see text] plane. Under this procedure, the model occupied a markedly smaller region (approximately [Formula: see text] smaller hull area; bootstrap 95% CI [1.68, 1.70]; permutation p < 0.0001) and concentrated near low-discrepancy configurations, with sparse occupancy of extreme under- and overstatement regimes. Together, these findings suggest that narrative-affect discrepancy is a measurable and regulated dimension of emotional expression and provide a reproducible geometric basis for comparing expressive degrees of freedom across populations and systems.
14 May 2026
Read appraisal →Neural networks : the official journal of the International Neural Network Society
Semi-supervised classification and projection with adaptive flexible structure optimal graph
Graph-based semi-supervised learning (GSSL) has received more attention in recent years. Many existing methods obtain a fixed similarity graph based on the original data. Affected by the redundant information and noise of the original data, the obtained similarity graph is not optimal, which greatly affects the subsequent work. And the similarity graphs obtained by these methods are fully connected, so the local structure information of the data cannot be preserved. Moreover, many GSSL methods project the original data into a low-dimensional space through a linear mapping function. However, linear functions may not be appropriate for data embedded in a nonlinear manifold. In order to overcome these shortcomings mentioned above, we relax the linear mapping function and impose a ℓ0-norm constraint on the similarity graph to gain an adaptive flexible structure optimal graph. Furthermore, incorporating the principle of maximum separability, an efficient GSSL method is proposed, named semi-supervised classification and projection with adaptive flexible structure optimal graph (SAFSG). Combining the construction of similarity graphs and label propagation, SAFSG can simultaneously get an adaptive flexible structure optimal graph, a label prediction matrix and a projection matrix. In addition, an efficient iterative algorithm for optimizing SAFSG is proposed. Finally, we conduct experiments on more than ten benchmark datasets, and the experimental results show that SAFSG performs satisfactorily in both classification and projection.
9 May 2026
Read appraisal →Annual review of biomedical engineering
Digital Twins for Biofluids
Digital twins-virtual representations dynamically linked to physical systems-have the potential to transform biomedical engineering by enabling real-time prediction, optimization, and personalization in health and disease. In biofluids, digital twins offer a framework for integrating physics-based models with data from clinical imaging, sensors, and physiological measurements to support diagnostics, therapeutic planning, and device design. This article reviews modeling approaches used in the construction of digital twins for biofluid applications. We survey high-fidelity numerical methods alongside emerging machine learning techniques, highlighting their respective strengths and limitations. Key requirements for digital twins are discussed, emphasizing the bidirectional interaction between physical and virtual assets and the importance of selecting modeling strategies tailored to specific biomedical contexts. While notable progress has been made over the past decade, significant challenges remain, particularly in integrating multiphysics models with data-driven methods and in establishing standardized protocols for data acquisition, interoperability, and sharing.
5 May 2026
Read appraisal →Journal of chemical information and modeling
Generative AI Uncovers Novel Chrebp/Txnip Axis Inhibitors with Potential Anti-inflammatory Activity
Type 2 diabetes is driven in part by metabolic inflammation, where activation of the Chrebp/Txnip axis promotes NLRP3 inflammasome assembly, leading to pancreatic β-cell dysfunction and pro-inflammatory cytokine release. Despite the therapeutic relevance of this pathway, the Chrebp-14-3-3 (regulatory-protein client) protein-protein interaction (PPI) remains largely underexplored, with only a limited number of small-molecule modulators reported to date. To address this gap, we developed an artificial intelligence-driven generative design framework for de novo discovery of selective PPI-targeting compounds. A conditional recurrent neural network (cRNN), implemented as a quantitative structure-property relationship-guided generative network (QSPR-GEN), was pretrained on a large, chemically diverse corpus to learn general SMILES syntax and structural priors, and subsequently fine-tuned on a curated, target-focused data set of approximately 5900 compounds, achieving high scaffold uniqueness (94.6%). Selectivity-oriented physicochemical descriptors were incorporated as conditional inputs to bias generation away from promiscuous chemotypes, while maintaining anchoring to a known active seed. Structure-based refinement was further applied by focusing on the noncanonical α-helical epitope unique to the Chrebp regulatory-protein interface, establishing a dual-layered strategy for selective PPI modulation. The integrated pipeline, combining virtual screening, molecular dynamics simulations, and MM/PBSA free-energy calculations, prioritized lead candidates with favorable binding energetics and pharmacokinetic profiles. In THP-1 macrophages under metabolic stress, the top candidate T7 markedly suppressed Txnip and NLRP3 expression, reduced IL-1β secretion, and attenuated pyroptotic cell death, outperforming a reference inhibitor. Collectively, this study presents a robust computational framework for the inverse design of challenging PPIs and demonstrates its utility through the identification and experimental validation of mechanistically precise lead compounds, exemplified by T2 and T7.
2 May 2026
Read appraisal →Journal of chemical information and modeling
DWPL-GCNMF: Structure-Aware Dynamic Weighted Pseudo-Label Learning for Adverse Drug Reaction Prediction
Adverse drug reaction (ADR) prediction is typically formulated as drug-ADR association learning on extremely sparse, positive-unlabeled matrices, which limits generalization and makes naive pseudolabel expansion prone to confirmation bias. Here, we propose DWPL-GCNMF, a structure-aware semisupervised framework that combines graph convolutional network (GCN)-based structural embeddings learned from a drug-protein knowledge graph with matrix factorization scoring and introduces a prediction-consistency-driven dynamic weighted pseudolabel learning strategy to expand training pairs while down-weighting unreliable pseudolabels; multiple base models are further fused to improve prediction stability. We evaluate DWPL-GCNMF on a DrugBank-derived benchmark (1,177 drugs × 4,247 ADRs; ∼2.6% observed associations) and an external SIDER benchmark (1,080 drugs × 5,579 ADRs; ∼2.3% observed associations) to assess robustness under different data sets and increased sparsity. DWPL-GCNMF consistently outperforms representative baselines, achieving AUPR/F1 of 0.6656/0.6194 on DrugBank and 0.6106/0.5816 on SIDER, while also improving top-K prioritization (e.g., Recall@15 up to 0.8078 and 0.8146, respectively), indicating that structure-aware representation learning with dynamically weighted pseudolabeling provides a robust approach for prioritizing potential ADRs under severe data sparsity.
2 May 2026
Read appraisal →Toxicology mechanisms and methods
Introducing AutoML framework for drug-drug interaction prediction: application of AutoGluon.
Drug-drug interactions (DDIs) remain a major concern in medication safety. Although advanced artificial intelligence methods such as deep learning have improved DDI prediction, their adoption is limited by the need for specialized expertise and complex model development. This study introduces the first application of the AutoGluon AutoML framework to DDI prediction using molecular features, aiming to automate and simplify model development. A curated subset of 100,000 drug pairs from DrugBank was used, employing three molecular representations: 2D molecular descriptors, 2048-bit Morgan fingerprints, and their combination. Models were trained using AutoGluon-Tabular with no manual hyperparameter tuning. The descriptor-only model achieved the best performance, with 84.4% test accuracy and an AUC of 0.916, outperforming fingerprint-based and hybrid models. Feature importance analysis identified key physicochemical and topological descriptors-such as drug-likeness, electrotopological indices, and hydrophobic surface area-as critical predictors of DDIs. These results demonstrate that AutoML can extract chemically meaningful patterns while reducing technical barriers. Overall, our results validate AutoGluon as a scalable approach to DDI prediction that provides chemically meaningful, feature-level interpretability, and lay the groundwork for future applications involving larger datasets and more complex chemical representations.
15 Apr 2026
Read appraisal →International journal of pharmaceutics
Optimization of drug diffusion in drug-eluting stents for coronary artery based on deep reinforcement learning
Drug-eluting stents (DES) are extensively used to treat coronary artery disease, and improving their therapeutic efficacy remains a long-standing research objective. Existing optimization strategies for DES drug delivery lack adaptive, learning-based frameworks capable of simultaneously enhancing drug concentration and uniformity in the therapeutic domain while regulating levels in the non-treatment domain, and show limited capacity to address complex, nonlinear optimization problems. This study applies a deep reinforcement learning (DRL)-based framework to optimize drug diffusion in DES. By integrating neural networks with the proximal policy optimization (PPO) algorithm, the framework enables closed-loop, feedback-driven regulation of drug diffusion in real time and demonstrates robustness to stochastic disturbances and parametric variations. The method was systematically evaluated across three stent embedding configurations: half-embedded, fully embedded, and non-embedded. Results show the DRL agent successfully balances the three therapeutic objectives and adopts distinct control strategies for different configurations. All configurations exhibited improved overall performance. For the half-embedded configuration, the DRL policy increased therapeutic domain concentration by 6.99%, accompanied by a slight rise in the non-treatment domain and a minor decrease in uniformity within the therapeutic domain. Under the fully embedded configuration, concentrations in both domains remained essentially unchanged, whereas uniformity in the therapeutic domain improved by 8.14%. For the non-embedded configuration, both therapeutic domain concentration and uniformity increased (by 2.04% and 6.30%, respectively), with negligible change in the non-treatment domain. These results demonstrate the feasibility of DRL for adaptive DES optimization and highlight its potential for patient-specific applications and guidance in clinical selection of drug-embedding configurations.
12 Apr 2026
Read appraisal →PloS one
Adaptive Feature Fusion Gate and Gated Channel-Spatial Attention in CNN-Transformer Models for Music Genre Classification
With the rapid growth of music data, automatic music genre classification has become a critical task in music information retrieval. Traditional methods based on handcrafted features are increasingly inadequate when handling large-scale analysis. This paper proposes the Convolutional Neural Network-Gated Transformer Network (CT-GateNet), a hybrid architecture that integrates a gated channel-spatial attention mechanism with an adaptive feature fusion gating mechanism to achieve discriminative feature learning and efficient feature integration. To mitigate data scarcity, a data augmentation strategy based on a denoising diffusion probabilistic model is introduced. Experiments are conducted on three public music genre datasets: GTZAN, FMA-SMALL and FMA-Medium. The method achieves classification accuracies of 98.72%, 89.42%, and 69.07% on GTZAN, FMA-SMALL and FMA-Medium, respectively, demonstrating outstanding performance and robust generalization capabilities. These results validate CT-GateNet's effectiveness in music genre classification and provide valuable insights for audio classification research.
11 Apr 2026
Read appraisal →Proceedings of the National Academy of Sciences of the United States of America
Reconstruction of human metabolic models with large language models
Genome-scale metabolic models (GEMs) have become essential tools for understanding human metabolism. Here, we introduce Human2, a consensus human GEM with enhanced precision and biological relevance, which leverages large language models (LLMs) and GitHub Action checks to streamline automated, efficient, and collaborative curation. Human2 supports the reconstruction of tissue- and organ-specific models tailored to sex- and age-specific human groups. By integrating transcriptomic, proteomic, and kinetic data, we reveal distinct metabolic features across these groups, such as significant differences in arachidonic acid and leukotriene metabolism. The specific models were integrated into a dynamic whole-body framework, marking an enzyme-constrained dynamic model that simulates interorgan metabolite exchanges under varying nutritional states, from feeding to fasting. Our work highlights the transformative role of LLMs in GEM reconstruction and introduces a whole-body dynamic simulation that integrates kinetic data, offering a powerful resource for multiscale human metabolism modeling.
10 Apr 2026
Read appraisal →PloS one
Deep learning for adaptive chemotherapy: A DDPG-based approach to optimizing tumor-immune dynamics
In this article, we propose a deep reinforcement learning based chemotherapy regulation framework to realize personalized and dynamic optimization of cancer treatment. We use a nonlinear dynamic system to model the dynamic evolution of the tumor microenvironment including tumor cell, normal cell and immune cell interactions, with drug concentration serving as the control input variable. The Deep Deterministic Policy Gradient (DDPG) algorithm makes agents can study optimal dosing strategy in a continuous space of movement to inhibit tumor growth effectively and minimize damage to normal tissues. To make the strategy more stable, Gauss noise is added to the model to simulate physiological oscillations and uncertainties in the treatment reaction. Experimental results show that it can control the growth of tumor in various initial scenarios and the accumulation of the drug concentration with high flexibility and safety. Our technique provides a feasible technical way for precision, low toxicity adaptive chemotherapy.
9 Apr 2026
Read appraisal →Physiological measurement
StructEIT: realistic 3D EIT model generation from CT scans for deep learning applications
Objective.Artificial intelligence has significantly improved image reconstruction quality across various medical imaging modalities. However, its application in electrical impedance tomography (EIT) reconstruction remains limited, mainly due to the absence of comprehensivein vivodatasets that incorporate realistic anatomical geometries and conductivity distributions. This limitation constrains the development of supervised and data-driven reconstruction methods.Approach.To address this bottleneck, we developed StructEIT, an integrated EIT modeling framework for generating anatomically and biophysically realistic EIT simulation models. The framework incorporates three key components: (1) a structure extraction module, which automatically processes human CT scans to extract body contours and organ boundaries, thereby providing high-fidelity spatial geometry for three-dimensional finite element modeling; (2) a surface electrode attachment module, which enables flexible and accurate placement of electrodes on irregular body surfaces, supporting diverse configurations and ensuring precise definition of the electrode-tissue interface; and (3) a tissue property assignment module, which establishes frequency-dependent conductivity models for multiple organs, enabling physiologically realistic conductivity values across tissues.Main results and Significance.By bridging the gap between CT imaging and EIT, StructEIT facilitates flexible, realistic, and scalable generation of high-resolution EIT datasets. Using this framework, we constructed Chest-EIT, a thoracic EIT simulation dataset comprising over 1400 publicly available CT cases, with multiple electrode configurations provided for each case.
4 Apr 2026
Read appraisal →IEEE transactions on medical imaging
EviVLM: When Evidential Learning Meets Vision Language Model for Medical Image Segmentation
The disparity between image and text representations, often referred to as the modality gap, remains a significant obstacle for Vision Language Models (VLMs) in medical image segmentation. This gap complicates multi-modal fusion, thereby restricting segmentation performance. To address this challenge, we propose Evidence-driven Vision Language Model (EviVLM)-a novel paradigm that integrates Evidential Learning (EL) into VLMs to systematically measure and mitigate the modality gap for enhanced multi-modal fusion. To drive this paradigm, an Evidence Affinity Map Generator (EAMG) is proposed to collect complementary cross-modal evidences by learning a global cross-modal affinity map, thus refining modality-specific evidence embedding. An Evidence Differential Similarity Learning (EDSL) is further proposed to collect consistent cross-modal evidences by performing Bias-Variance Decomposition on differential matrix derived from bidirectional similarity matrices between image and text evidence embeddings. Finally, the subjective logic is used for mapping the collected evidences to opinions, and the Dempster-Shafer's theory based combination rule is introduced for opinion aggregation, thereby quantifying the modality gap and facilitating effective multi-modal integration. Experimental results on three public medical image segmentation datasets validate that the proposed EviVLM can achieve state-of-the-art performance. Code is available at: https://github.com/QingtaoPan/EviVLM.
4 Apr 2026
Read appraisal →Pharmacology research & perspectives
Improving Population Pharmacokinetic Modelling with Artificial Patients using Generative Artificial Intelligence
In population pharmacokinetics (PopPK), non-linear mixed effects (NLME) models are used to simultaneously describe a drug's pharmacokinetics (PK) and dynamics (PD) in a patient population using systems of ordinary differential equations. In this field, machine learning is mainly used for data preparation, hypothesis generation, predictive modeling, and model validation. Some approaches to integrate artificially generated information have already been explored, but real-world application is still limited. We therefore conducted a proof-of-concept study to analyze the ability of generative artificial intelligence (AI) to create artificial patient profiles to augment PopPK data sets and assess their influence on parameter estimates. We defined the pharmacokinetic parameters of a hypothetical drug and simulated the concentration curves of 20 patients. We then trained Wasserstein Generative Adversarial Networks (WGANs) with gradient penalty (GP) to generate artificial patients. The data distribution of original and artificial patients was statistically indistinguishable as shown by the Maximum Mean Discrepancy test. Therefore, the WGAN-GP is neither overfitted, that is producing only single instances of artificial patients, nor underfitted, that is producing unrealistic artificial patients. We then combined different shares of original and artificial patients in separate data sets to build and compare PopPK model estimates. Addition of artificial patients led to narrower confidence intervals, indicating more robust parameter estimates, and accentuated the allometric effect of weight on the volume of distribution. In conclusion, we provide a proof-of-concept that generative AI can be used to augment pharmacokinetic data sets, with preliminary evidence suggesting improved parameter estimation.
4 Apr 2026
Read appraisal →