Pub Date : 2026-10-20DOI: 10.1021/acs.jcim.6c02413
Wen Zhang,Yuxuan Tian,Hang Qiu
Drug combination therapy has emerged as a promising strategy for improving therapeutic efficacy. While hypergraphs have demonstrated substantial potential for high-order drug synergy modeling, existing approaches still suffer from two critical limitations: (1) reliance on locally determined aggregation limits the utilization of multihop topological context and broader structural information across the hypergraph; (2) homogeneous message passing overlooks the structural heterogeneity of drugs and cell lines within hyperedges. To address these issues, we propose GLH2Syn, a global–local dual-view heterogeneous hypergraph learning framework for drug synergy prediction. Specifically, GLH2Syn operates on a heterogeneous hypergraph in which each drug–drug–cell line triplet forms a hyperedge, incorporating two complementary views: a global diffusion view and a local topology view. The global diffusion view performs diffusion over the node–hyperedge topology to derive globally informed incidence weights, thereby incorporating structural context into local aggregation. The local topology view captures fine-grained structural heterogeneity within hyperedges through structure-aware node-hyperedge propagation, distinguishing the heterogeneous roles of drugs and cell lines in high-order interactions. Furthermore, we design a joint optimization objective integrating drug synergy prediction with similarity reconstruction to preserve both predictive accuracy and biological consistency. Extensive experiments on two public data sets demonstrate that GLH2Syn consistently outperforms 11 competitive baselines on both classification and regression tasks across three test scenarios. Case studies further validate its potential in discovering novel synergistic drug combinations.
{"title":"GLH2Syn: Global–Local Dual-View Heterogeneous Hypergraph Learning for Drug Synergy Prediction","authors":"Wen Zhang,Yuxuan Tian,Hang Qiu","doi":"10.1021/acs.jcim.6c02413","DOIUrl":"https://doi.org/10.1021/acs.jcim.6c02413","url":null,"abstract":"Drug combination therapy has emerged as a promising strategy for improving therapeutic efficacy. While hypergraphs have demonstrated substantial potential for high-order drug synergy modeling, existing approaches still suffer from two critical limitations: (1) reliance on locally determined aggregation limits the utilization of multihop topological context and broader structural information across the hypergraph; (2) homogeneous message passing overlooks the structural heterogeneity of drugs and cell lines within hyperedges. To address these issues, we propose GLH2Syn, a global–local dual-view heterogeneous hypergraph learning framework for drug synergy prediction. Specifically, GLH2Syn operates on a heterogeneous hypergraph in which each drug–drug–cell line triplet forms a hyperedge, incorporating two complementary views: a global diffusion view and a local topology view. The global diffusion view performs diffusion over the node–hyperedge topology to derive globally informed incidence weights, thereby incorporating structural context into local aggregation. The local topology view captures fine-grained structural heterogeneity within hyperedges through structure-aware node-hyperedge propagation, distinguishing the heterogeneous roles of drugs and cell lines in high-order interactions. Furthermore, we design a joint optimization objective integrating drug synergy prediction with similarity reconstruction to preserve both predictive accuracy and biological consistency. Extensive experiments on two public data sets demonstrate that GLH2Syn consistently outperforms 11 competitive baselines on both classification and regression tasks across three test scenarios. Case studies further validate its potential in discovering novel synergistic drug combinations.","PeriodicalId":44,"journal":{"name":"Journal of Chemical Information and Modeling ","volume":"11 1","pages":""},"PeriodicalIF":5.6,"publicationDate":"2026-10-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149221628","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"化学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-10-20DOI: 10.1021/acs.jcim.6c02169
Leon Obendorf,Niklas Piet Doering,Petra Knaus,Gerhard Wolber
AI-driven cofolding models have become central tools for predicting protein–ligand complexes, yet it remains unclear whether the resulting structures recover the correct protein conformation or merely place the ligand plausibly within an otherwise mispredicted fold. Here, we benchmark four state-of-the-art cofolding models, AlphaFold3, RoseTTAFold3, Boltz-2, and Chai-1, and show that correct ligand placement is systematically decoupled from accurate prediction of the surrounding protein conformation. As test systems, we use kinases and Class A G protein-coupled receptors (GPCRs), families whose pharmacology depends on well-defined conformational states. To enable a prospective assessment, we evaluate complexes released after the training cutoffs of all tested models, together with one earlier nucleotide-bound structure as a control. We find that state-defining structural features, including kinase activation-loop, DFG/DYG geometries, and GPCR intracellular conformations such as ICL3, are frequently mispredicted even when ligand root-mean-square deviation (RMSD) is low. Cofolding adaptively remodels the orthosteric pocket around the bound ligand, but this local adaptation rarely propagates to the correct global conformation. Incorporating state-annotated templates and filtered multiple sequence alignments (MSAs) improves conformational recovery in some cases but has little effect in others, and model confidence metrics correlate with small-molecule placement but cannot report whether the local pocket geometry was recovered. Furthermore, while orthosteric ligand placement is generally reliable, the single allosteric complex in our set is misplaced by all models, even under explicit guidance. These results identify conformational decoupling as a fundamental limitation of current cofolding approaches across the systems evaluated, with direct implications for state-selective and structure-based drug design.
{"title":"Benchmarking Protein–Ligand Cofolding Models: Correct Ligand Placement Is Decoupled From Accurate Protein Conformation","authors":"Leon Obendorf,Niklas Piet Doering,Petra Knaus,Gerhard Wolber","doi":"10.1021/acs.jcim.6c02169","DOIUrl":"https://doi.org/10.1021/acs.jcim.6c02169","url":null,"abstract":"AI-driven cofolding models have become central tools for predicting protein–ligand complexes, yet it remains unclear whether the resulting structures recover the correct protein conformation or merely place the ligand plausibly within an otherwise mispredicted fold. Here, we benchmark four state-of-the-art cofolding models, AlphaFold3, RoseTTAFold3, Boltz-2, and Chai-1, and show that correct ligand placement is systematically decoupled from accurate prediction of the surrounding protein conformation. As test systems, we use kinases and Class A G protein-coupled receptors (GPCRs), families whose pharmacology depends on well-defined conformational states. To enable a prospective assessment, we evaluate complexes released after the training cutoffs of all tested models, together with one earlier nucleotide-bound structure as a control. We find that state-defining structural features, including kinase activation-loop, DFG/DYG geometries, and GPCR intracellular conformations such as ICL3, are frequently mispredicted even when ligand root-mean-square deviation (RMSD) is low. Cofolding adaptively remodels the orthosteric pocket around the bound ligand, but this local adaptation rarely propagates to the correct global conformation. Incorporating state-annotated templates and filtered multiple sequence alignments (MSAs) improves conformational recovery in some cases but has little effect in others, and model confidence metrics correlate with small-molecule placement but cannot report whether the local pocket geometry was recovered. Furthermore, while orthosteric ligand placement is generally reliable, the single allosteric complex in our set is misplaced by all models, even under explicit guidance. These results identify conformational decoupling as a fundamental limitation of current cofolding approaches across the systems evaluated, with direct implications for state-selective and structure-based drug design.","PeriodicalId":44,"journal":{"name":"Journal of Chemical Information and Modeling ","volume":"88 1","pages":""},"PeriodicalIF":5.6,"publicationDate":"2026-10-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149221703","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"化学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-10-20DOI: 10.1021/acs.jcim.6c02297
Tingkai Sun,Xiucai Ye,Tetsuya Sakurai
Federated molecular property prediction enables collaboration across private chemical libraries, but parameter averaging does not explicitly account for scaffold heterogeneity. FedSPL groups molecules using fixed descriptors and exchanges embedding prototypes alongside model parameters, enabling label-independent prototype matching while retaining supervised property prediction. Across five classification benchmarks and three heterogeneity levels, matched three-seed evaluation using masked, held-out server-model ROC-AUC showed higher mean final AUC than FedAvg in seven of 15 settings. Consistently positive differences on SIDER and mixed outcomes elsewhere indicate data set-dependent benefits, motivating validation-guided use of structural information in federated prototype learning.
{"title":"Structure Prototype Learning for Federated Molecular Property Prediction under Scaffold Heterogeneity","authors":"Tingkai Sun,Xiucai Ye,Tetsuya Sakurai","doi":"10.1021/acs.jcim.6c02297","DOIUrl":"https://doi.org/10.1021/acs.jcim.6c02297","url":null,"abstract":"Federated molecular property prediction enables collaboration across private chemical libraries, but parameter averaging does not explicitly account for scaffold heterogeneity. FedSPL groups molecules using fixed descriptors and exchanges embedding prototypes alongside model parameters, enabling label-independent prototype matching while retaining supervised property prediction. Across five classification benchmarks and three heterogeneity levels, matched three-seed evaluation using masked, held-out server-model ROC-AUC showed higher mean final AUC than FedAvg in seven of 15 settings. Consistently positive differences on SIDER and mixed outcomes elsewhere indicate data set-dependent benefits, motivating validation-guided use of structural information in federated prototype learning.","PeriodicalId":44,"journal":{"name":"Journal of Chemical Information and Modeling ","volume":"16 1","pages":""},"PeriodicalIF":5.6,"publicationDate":"2026-10-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149221629","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"化学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-10-20DOI: 10.1021/acs.jcim.6c00178
Nishat Nayla Labiba,Md Mushfiqul Islam,Lawrence O. Hall,David S. Simmons
Synthetic sequence-controlled polymers promise to transform polymer science by combining the chemical versatility of synthetic polymers with the precise sequence-mediated functionality of biological proteins. However, design of these materials has proven extraordinarily challenging, because they lack the massive data sets of closely related evolved molecules that accelerate design of proteins. Here, we report on an Artificial Intelligence strategy to reduce the amount of data necessary to accelerate the design of these materials. We focus on data connecting the repeat-unit-sequence of a model compatibilizer molecule to its ability to reduce the interfacial tension between distinct polymer domains. The optimal sequence of these molecules, which are essential for applications such as mixed-waste polymer recycling, depends strongly on variables such as concentration and chemical details of the polymer. We show that a deep neural network trained on low-fidelity data for sequence/interfacial tension relations at one set of conditions can be rapidly tuned to make higher-fidelity predictions at a distinct set of conditions, requiring far less data than would ordinarily be needed within the data sets we have tested. This priming using a pretrained model and then tuning it on unseen data from a different data set approach could allow a single low-fidelity parent data set to accelerate prediction and design in a larger set of related systems.
{"title":"Reducing Data Requirements for Sequence–Property Prediction in Copolymer Compatibilizers via Deep Neural Network Tuning","authors":"Nishat Nayla Labiba,Md Mushfiqul Islam,Lawrence O. Hall,David S. Simmons","doi":"10.1021/acs.jcim.6c00178","DOIUrl":"https://doi.org/10.1021/acs.jcim.6c00178","url":null,"abstract":"Synthetic sequence-controlled polymers promise to transform polymer science by combining the chemical versatility of synthetic polymers with the precise sequence-mediated functionality of biological proteins. However, design of these materials has proven extraordinarily challenging, because they lack the massive data sets of closely related evolved molecules that accelerate design of proteins. Here, we report on an Artificial Intelligence strategy to reduce the amount of data necessary to accelerate the design of these materials. We focus on data connecting the repeat-unit-sequence of a model compatibilizer molecule to its ability to reduce the interfacial tension between distinct polymer domains. The optimal sequence of these molecules, which are essential for applications such as mixed-waste polymer recycling, depends strongly on variables such as concentration and chemical details of the polymer. We show that a deep neural network trained on low-fidelity data for sequence/interfacial tension relations at one set of conditions can be rapidly tuned to make higher-fidelity predictions at a distinct set of conditions, requiring far less data than would ordinarily be needed within the data sets we have tested. This priming using a pretrained model and then tuning it on unseen data from a different data set approach could allow a single low-fidelity parent data set to accelerate prediction and design in a larger set of related systems.","PeriodicalId":44,"journal":{"name":"Journal of Chemical Information and Modeling ","volume":"56 1","pages":""},"PeriodicalIF":5.6,"publicationDate":"2026-10-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149218871","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"化学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-10-19DOI: 10.1021/acs.jcim.6c01938
Soyeon Jeon,Sung Wook Moon,Hyun Woo Kim
Drug efflux mediated by P-glycoprotein (P-gp) remains a major barrier to effective small-molecule therapeutics, limiting both bioavailability and central nervous system exposure. While machine learning models can predict P-gp substrates, translating these predictions into actionable molecular design remains largely unexplored, particularly in enabling the rational conversion of substrates into nonsubstrates to mitigate efflux liability. Here, we introduce the synergy-guided fragment genetic algorithm (SF-GA), a computational framework that transforms predictive insight into an interpretable molecular design strategy. By integrating fragment synergy analysis into the mutation stage of a genetic algorithm, SF-GA biases fragment replacement toward co-occurring substructures associated with nonsubstrate classification, enabling interpretability at the chemical structure level and property-directed molecular optimization. Compared with genetic-algorithm and reinforcement-learning baselines under a fixed function-evaluation budget, SF-GA achieves competitive optimization performance while producing molecules with the highest retrosynthetic accessibility among all methods tested. Unlike the baselines, every accepted mutation is reported together with the discarded and inserted fragments and their synergy scores, providing a fragment-level rationale for each structural change; ablation against an unguided and an importance-only control isolates the contribution of bit co-occurrence. The method enables scaffold-preserving and similarity-constrained editing that converts P-gp substrates into predicted nonsubstrates. Docking against two human P-gp structures reveals weaker predicted binding for molecules with higher predicted nonsubstrate probability, motivating future experimental evaluation of P-gp transport.
p -糖蛋白（P-gp）介导的药物外排仍然是有效小分子治疗的主要障碍，限制了生物利用度和中枢神经系统暴露。虽然机器学习模型可以预测P-gp底物，但将这些预测转化为可操作的分子设计在很大程度上仍未被探索，特别是在使底物合理转化为非底物以减轻外排责任方面。在这里，我们介绍了协同引导片段遗传算法（SF-GA），这是一个将预测性洞察力转化为可解释的分子设计策略的计算框架。通过将片段协同分析集成到遗传算法的突变阶段，SF-GA将片段替换偏向于与非底物分类相关的共同发生的亚结构，从而实现化学结构水平的可解释性和属性导向的分子优化。在固定的功能评估预算下，与遗传算法和强化学习基线相比，SF-GA实现了竞争性优化性能，同时产生的分子在所有测试方法中具有最高的反合成可达性。与基线不同，每个可接受的突变都与丢弃和插入的片段及其协同作用评分一起报告，为每个结构变化提供片段级的基本原理；针对非导向和仅重要控制的消融可以隔离钻头共发生的影响。该方法实现了支架保留和相似性约束编辑，将P-gp底物转化为预测的非底物。与两种人类P-gp结构的对接表明，预测的非底物概率较高的分子结合较弱，这为未来P-gp转运的实验评估提供了动力。
{"title":"Synergy-Guided Molecular Redesign of Predicted P-Glycoprotein Substrates toward Synthetically Feasible Non-Substrates","authors":"Soyeon Jeon,Sung Wook Moon,Hyun Woo Kim","doi":"10.1021/acs.jcim.6c01938","DOIUrl":"https://doi.org/10.1021/acs.jcim.6c01938","url":null,"abstract":"Drug efflux mediated by P-glycoprotein (P-gp) remains a major barrier to effective small-molecule therapeutics, limiting both bioavailability and central nervous system exposure. While machine learning models can predict P-gp substrates, translating these predictions into actionable molecular design remains largely unexplored, particularly in enabling the rational conversion of substrates into nonsubstrates to mitigate efflux liability. Here, we introduce the synergy-guided fragment genetic algorithm (SF-GA), a computational framework that transforms predictive insight into an interpretable molecular design strategy. By integrating fragment synergy analysis into the mutation stage of a genetic algorithm, SF-GA biases fragment replacement toward co-occurring substructures associated with nonsubstrate classification, enabling interpretability at the chemical structure level and property-directed molecular optimization. Compared with genetic-algorithm and reinforcement-learning baselines under a fixed function-evaluation budget, SF-GA achieves competitive optimization performance while producing molecules with the highest retrosynthetic accessibility among all methods tested. Unlike the baselines, every accepted mutation is reported together with the discarded and inserted fragments and their synergy scores, providing a fragment-level rationale for each structural change; ablation against an unguided and an importance-only control isolates the contribution of bit co-occurrence. The method enables scaffold-preserving and similarity-constrained editing that converts P-gp substrates into predicted nonsubstrates. Docking against two human P-gp structures reveals weaker predicted binding for molecules with higher predicted nonsubstrate probability, motivating future experimental evaluation of P-gp transport.","PeriodicalId":44,"journal":{"name":"Journal of Chemical Information and Modeling ","volume":"11 1","pages":""},"PeriodicalIF":5.6,"publicationDate":"2026-10-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149218796","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"化学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Peptide therapeutics are promising drug candidates because of their high activity and specificity. However, their potential toxicity poses a major hurdle, making accurate computational toxicity predictions urgently necessary for early safety screening. Existing methods largely rely on sequence data or simple graph representations, thereby limiting their ability to capture higher-order physicochemical interactions. To address this, we propose MBFGNK, an interpretable multimodal geometric learning framework, for peptide toxicity prediction. It integrates sequence representations from pretrained protein language models, 3D structural features via an equivariant graph neural network (EGNN), and physicochemical hypergraph representations capturing global, local, and biologically synergistic residue interactions. A modality attention gate adaptively fuses these heterogeneous data types, while a Chebyshev polynomial-based Kolmogorov–Arnold network (ChebyKAN) models complex nonlinear toxicity patterns. We evaluated MBFGNK using 5-fold cross-validation and an independent test set. The experimental results show that the ensemble achieved an AUC of 0.9687, an MCC of 0.8449, an Acc of 0.9221, and an F1 of 0.9205 on the independent test set, suggesting improved performance relative to previously reported baseline results on the same benchmark. Integrated gradients and modality-level analyses further identified prediction-informative motifs, residues, and spatial regions, providing interpretable evidence consistent with known characteristics of toxic peptides. These results suggest that MBFGNK is an effective and interpretable framework for peptide toxicity prediction and may assist in downstream safety-oriented peptide screening.
{"title":"A Multimodal Geometric Learning Framework for Interpretable Peptide Toxicity Prediction","authors":"Xin Li,Jinzhou Wu,Rui Wang,Donglin He,Yu-Juan Zhang","doi":"10.1021/acs.jcim.6c02423","DOIUrl":"https://doi.org/10.1021/acs.jcim.6c02423","url":null,"abstract":"Peptide therapeutics are promising drug candidates because of their high activity and specificity. However, their potential toxicity poses a major hurdle, making accurate computational toxicity predictions urgently necessary for early safety screening. Existing methods largely rely on sequence data or simple graph representations, thereby limiting their ability to capture higher-order physicochemical interactions. To address this, we propose MBFGNK, an interpretable multimodal geometric learning framework, for peptide toxicity prediction. It integrates sequence representations from pretrained protein language models, 3D structural features via an equivariant graph neural network (EGNN), and physicochemical hypergraph representations capturing global, local, and biologically synergistic residue interactions. A modality attention gate adaptively fuses these heterogeneous data types, while a Chebyshev polynomial-based Kolmogorov–Arnold network (ChebyKAN) models complex nonlinear toxicity patterns. We evaluated MBFGNK using 5-fold cross-validation and an independent test set. The experimental results show that the ensemble achieved an AUC of 0.9687, an MCC of 0.8449, an Acc of 0.9221, and an F1 of 0.9205 on the independent test set, suggesting improved performance relative to previously reported baseline results on the same benchmark. Integrated gradients and modality-level analyses further identified prediction-informative motifs, residues, and spatial regions, providing interpretable evidence consistent with known characteristics of toxic peptides. These results suggest that MBFGNK is an effective and interpretable framework for peptide toxicity prediction and may assist in downstream safety-oriented peptide screening.","PeriodicalId":44,"journal":{"name":"Journal of Chemical Information and Modeling ","volume":"45 1","pages":""},"PeriodicalIF":5.6,"publicationDate":"2026-10-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149218872","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"化学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Abstract Deep learning expands the chemical space accessible to antibiotic discovery, making interpretation of evaluation results increasingly important when only a few compounds can be prioritized from vast candidate libraries. AUPRC depends on test-set prevalence, while molecular splitting can alter training-to-test structural relatedness and test-set class composition, so when prevalence differs across splits, a cross-split AUPRC difference alone cannot quantify how much of the performance change is attributable to structural differences. Cross-split comparisons should therefore report test-set size, positive count, prevalence, splitting method, and the precision-recall summary metric, while direct model comparisons should use the same fixed independent test set whenever possible. Claims about practical screening utility should also consider top-k performance, early enrichment metrics, or prospective experimental hit rates under a prespecified budget, so that evidence remains aligned with the scientific claims it is intended to support.
{"title":"Interpretive Limits of Cross-Split AUPRC Differences\u0000in Deep Learning-Based Antibiotic Screening","authors":"Zhijie Pan,Tianming Han,Wenchi Ge,Churong Wang,Qi Zhao","doi":"10.1021/acs.jcim.6c03208","DOIUrl":"https://doi.org/10.1021/acs.jcim.6c03208","url":null,"abstract":"Abstract Deep learning expands the chemical space accessible to antibiotic discovery, making interpretation of evaluation results increasingly important when only a few compounds can be prioritized from vast candidate libraries. AUPRC depends on test-set prevalence, while molecular splitting can alter training-to-test structural relatedness and test-set class composition, so when prevalence differs across splits, a cross-split AUPRC difference alone cannot quantify how much of the performance change is attributable to structural differences. Cross-split comparisons should therefore report test-set size, positive count, prevalence, splitting method, and the precision-recall summary metric, while direct model comparisons should use the same fixed independent test set whenever possible. Claims about practical screening utility should also consider top-k performance, early enrichment metrics, or prospective experimental hit rates under a prespecified budget, so that evidence remains aligned with the scientific claims it is intended to support.","PeriodicalId":44,"journal":{"name":"Journal of Chemical Information and Modeling ","volume":"27 1","pages":""},"PeriodicalIF":5.6,"publicationDate":"2026-10-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149210138","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"化学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-10-16DOI: 10.1021/acs.jcim.6c01975
Jia He,Le Chen,Wenqi Fang,Huiyun Wang,Yanwei Yang,Wen Yang,Yongqing Zhang,Tianhao Li
Abstract Drug–target interaction (DTI) prediction is essential for drug discovery, yet existing models often suffer from limited generalization when drugs or targets are unseen during training. This challenge is particularly severe in cold-start settings, where interaction records are unavailable for novel entities and effective semantic transfer between chemical and biological spaces is required. To address this problem, we propose CMA-DTI, a cross-modal semantic alignment framework for cold-start drug–target interaction prediction. CMA-DTI first extracts drug and protein representations from pretrained sequence models, and then it uses a dual-channel encoder to capture local functional patterns and global contextual dependencies. A bidirectional cross-modal alignment module is further introduced to model interactions between drug functional groups and protein binding regions, enabling fine-grained semantic transfer across heterogeneous modalities. Experiments on four benchmark data sets show that CMA-DTI achieves competitive performance in warm-start settings and improves generalization in challenging target cold-start and blind-start scenarios. Ablation studies and structural visualizations further support the contribution of the proposed modules and suggest residue-level interpretability for protein–ligand binding analysis.
{"title":"Cross-Modal Semantic\u0000Alignment for Cold-Start Drug–Target\u0000Interaction Prediction","authors":"Jia He,Le Chen,Wenqi Fang,Huiyun Wang,Yanwei Yang,Wen Yang,Yongqing Zhang,Tianhao Li","doi":"10.1021/acs.jcim.6c01975","DOIUrl":"https://doi.org/10.1021/acs.jcim.6c01975","url":null,"abstract":"Abstract Drug–target interaction (DTI) prediction is essential for drug discovery, yet existing models often suffer from limited generalization when drugs or targets are unseen during training. This challenge is particularly severe in cold-start settings, where interaction records are unavailable for novel entities and effective semantic transfer between chemical and biological spaces is required. To address this problem, we propose CMA-DTI, a cross-modal semantic alignment framework for cold-start drug–target interaction prediction. CMA-DTI first extracts drug and protein representations from pretrained sequence models, and then it uses a dual-channel encoder to capture local functional patterns and global contextual dependencies. A bidirectional cross-modal alignment module is further introduced to model interactions between drug functional groups and protein binding regions, enabling fine-grained semantic transfer across heterogeneous modalities. Experiments on four benchmark data sets show that CMA-DTI achieves competitive performance in warm-start settings and improves generalization in challenging target cold-start and blind-start scenarios. Ablation studies and structural visualizations further support the contribution of the proposed modules and suggest residue-level interpretability for protein–ligand binding analysis.","PeriodicalId":44,"journal":{"name":"Journal of Chemical Information and Modeling ","volume":"10 1","pages":""},"PeriodicalIF":5.6,"publicationDate":"2026-10-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149210104","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"化学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-10-16DOI: 10.1021/acs.jcim.6c01050
Victor J. Y. Lim,Yee-Song Law,Guangrong Peh,Clarence Tan,Charmaine Chia,Jhoann M. T. Miyajima,Nazreen Abdul Muthaliff,Pnelope Ng,Yoganathan Kanagasundaram,Wan Lin Yeo,Sebastian Maurer-Stroh,Yee Hwee Lim,Ee Lui Ang,Hao Fan
Abstract Recent advances in enzyme informatics have produced increasingly powerful general models for the prediction of protein function prediction. However, their performance remains limited for underexplored enzyme families where experimental data are scarce and heterologous expression remains challenging. Here, we present a computational-experimental workflow for prioritizing, characterizing, and modeling fungal unspecific peroxygenases (UPOs) as a case study in enzyme discovery under data-limited conditions. The workflow integrates large-scale sequence mining, expression-aware candidate prioritization, systematic experimental screening, and machine-learning-based prediction of enzyme–substrate activity. From more than 3000 mined candidate sequences, 59 putative UPOs were shortlisted, expressed in Escherichia coli across two batches, and tested against colorimetric substrates, yielding 27 UPOs with consistent expression and activity. A subset of 15 novel UPOs and 4 well-characterized reference enzyme sequences was then screened against a chemically diverse panel of 27 substrates to generate an enzyme–substrate activity data set. Using this data set, we developed two complementary machine learning protocols based on gradient boosting: (i) a structure-informed model incorporating homology modeling, docking, and interaction-derived features (modeling-docking-ML) and (ii) a latent-representation model utilizing joint protein/chemical language model embeddings (protein/chemical-language-model-ML). In retrospective validation, the models achieved performance with a Matthews correlation coefficient (MCC) of up to 55% and accuracy up to 78%. In selected prospective testing on previously unseen batch 2 enzymes, the models achieved an MCC of up to 48% and accuracy up to 70%. Taken together, these results demonstrate that targeted experimental data generation, combined with complementary structure-informed and embedding-based models, can support prioritization and activity prediction within poorly characterized enzyme families while highlighting the continued challenge of generalizing to broader chemical substrate space.
{"title":"An Integrated Computational-Experimental\u0000Workflow\u0000for Fungal Peroxygenase Discovery and Enzyme–Substrate Prediction","authors":"Victor J. Y. Lim,Yee-Song Law,Guangrong Peh,Clarence Tan,Charmaine Chia,Jhoann M. T. Miyajima,Nazreen Abdul Muthaliff,Pnelope Ng,Yoganathan Kanagasundaram,Wan Lin Yeo,Sebastian Maurer-Stroh,Yee Hwee Lim,Ee Lui Ang,Hao Fan","doi":"10.1021/acs.jcim.6c01050","DOIUrl":"https://doi.org/10.1021/acs.jcim.6c01050","url":null,"abstract":"Abstract Recent advances in enzyme informatics have produced increasingly powerful general models for the prediction of protein function prediction. However, their performance remains limited for underexplored enzyme families where experimental data are scarce and heterologous expression remains challenging. Here, we present a computational-experimental workflow for prioritizing, characterizing, and modeling fungal unspecific peroxygenases (UPOs) as a case study in enzyme discovery under data-limited conditions. The workflow integrates large-scale sequence mining, expression-aware candidate prioritization, systematic experimental screening, and machine-learning-based prediction of enzyme–substrate activity. From more than 3000 mined candidate sequences, 59 putative UPOs were shortlisted, expressed in Escherichia coli across two batches, and tested against colorimetric substrates, yielding 27 UPOs with consistent expression and activity. A subset of 15 novel UPOs and 4 well-characterized reference enzyme sequences was then screened against a chemically diverse panel of 27 substrates to generate an enzyme–substrate activity data set. Using this data set, we developed two complementary machine learning protocols based on gradient boosting: (i) a structure-informed model incorporating homology modeling, docking, and interaction-derived features (modeling-docking-ML) and (ii) a latent-representation model utilizing joint protein/chemical language model embeddings (protein/chemical-language-model-ML). In retrospective validation, the models achieved performance with a Matthews correlation coefficient (MCC) of up to 55% and accuracy up to 78%. In selected prospective testing on previously unseen batch 2 enzymes, the models achieved an MCC of up to 48% and accuracy up to 70%. Taken together, these results demonstrate that targeted experimental data generation, combined with complementary structure-informed and embedding-based models, can support prioritization and activity prediction within poorly characterized enzyme families while highlighting the continued challenge of generalizing to broader chemical substrate space.","PeriodicalId":44,"journal":{"name":"Journal of Chemical Information and Modeling ","volume":"160 1","pages":""},"PeriodicalIF":5.6,"publicationDate":"2026-10-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149210100","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"化学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-10-16DOI: 10.1021/acs.jcim.6c01542
Seungmin Yoon,Alex Dickson
Abstract In recent years, deep generative models have been increasingly adopted for molecular optimization and design. In many such models, molecules are encoded into latent spaces that provide compact representations of their features. However, these latent spaces remain difficult to interpret in physical terms, and it is unclear how much structural and molecular information is retained under compression. This limitation becomes especially important when molecular behavior depends on specific three-dimensional arrangements embedded in an environmental context. Here, we introduce COMPRESS (Compression Of Molecular Physical fields into Reduced Spatial Sites), an optimization framework that represents a small molecule using a fixed number of physically parametrized spatial sites. Each site is defined by three spatial coordinates and three interaction parameters, yielding a fixed-size representation that remains directly interpretable in physical terms. To preserve molecular information, COMPRESS representations are optimized to reproduce the density, electrostatic, and van der Waals fields of the corresponding all-atom molecule on discretized three-dimensional grids. We systematically evaluate how molecular information is retained as the number of sites, K, is reduced. Our analysis covers benzene, 20 chemically diverse chemical scaffolds, benzene dimers, protein–ligand complexes, and 10,000 molecules chosen from the GEOM-Drugs data set. Across systems, field agreement deteriorates slowly at high resolution and then undergoes a rapid, pronounced transition at lower resolution. We further find that K = 25 conformers capture about 97% of the molecules from GEOM-Drugs under a loss threshold of 0.05, which we use as an indicator of fidelity. These results show that COMPRESS can provide a fixed-size, physically interpretable molecular representation, enabling compression to be evaluated directly using molecular interaction fields and resulting physical quantities. We also introduce a decoder that reconstructs all-atom structures directly from a COMPRESS representation, showing that the representation can support molecular generation in addition to field-preservation analysis. Finally, rationally modified COMPRESS representations through changes to site charge and size or the addition or removal of sites are shown to generate realistic chemical compounds upon decoding.
{"title":"Building a Physically\u0000Interpretable Compressed Chemical\u0000Space via Grid-Based Optimization","authors":"Seungmin Yoon,Alex Dickson","doi":"10.1021/acs.jcim.6c01542","DOIUrl":"https://doi.org/10.1021/acs.jcim.6c01542","url":null,"abstract":"Abstract In recent years, deep generative models have been increasingly adopted for molecular optimization and design. In many such models, molecules are encoded into latent spaces that provide compact representations of their features. However, these latent spaces remain difficult to interpret in physical terms, and it is unclear how much structural and molecular information is retained under compression. This limitation becomes especially important when molecular behavior depends on specific three-dimensional arrangements embedded in an environmental context. Here, we introduce COMPRESS (Compression Of Molecular Physical fields into Reduced Spatial Sites), an optimization framework that represents a small molecule using a fixed number of physically parametrized spatial sites. Each site is defined by three spatial coordinates and three interaction parameters, yielding a fixed-size representation that remains directly interpretable in physical terms. To preserve molecular information, COMPRESS representations are optimized to reproduce the density, electrostatic, and van der Waals fields of the corresponding all-atom molecule on discretized three-dimensional grids. We systematically evaluate how molecular information is retained as the number of sites, K, is reduced. Our analysis covers benzene, 20 chemically diverse chemical scaffolds, benzene dimers, protein–ligand complexes, and 10,000 molecules chosen from the GEOM-Drugs data set. Across systems, field agreement deteriorates slowly at high resolution and then undergoes a rapid, pronounced transition at lower resolution. We further find that K = 25 conformers capture about 97% of the molecules from GEOM-Drugs under a loss threshold of 0.05, which we use as an indicator of fidelity. These results show that COMPRESS can provide a fixed-size, physically interpretable molecular representation, enabling compression to be evaluated directly using molecular interaction fields and resulting physical quantities. We also introduce a decoder that reconstructs all-atom structures directly from a COMPRESS representation, showing that the representation can support molecular generation in addition to field-preservation analysis. Finally, rationally modified COMPRESS representations through changes to site charge and size or the addition or removal of sites are shown to generate realistic chemical compounds upon decoding.","PeriodicalId":44,"journal":{"name":"Journal of Chemical Information and Modeling ","volume":"17 1","pages":""},"PeriodicalIF":5.6,"publicationDate":"2026-10-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149210102","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"化学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：