Journal of Chemical Information and Modeling 最新文献

{"title":"Pose Reproduction, Cross Docking, Database Enrichment, and Reverse Docking Benchmarks for DOCK6, Vina, and AutoDock 4 with Emphasis on Approved Drugs","authors":"Carissa P. Corbo,&nbsp;Scott R. Laverty,&nbsp;Robert C. Rizzo","doi":"10.1021/acs.jcim.6c01375","DOIUrl":"10.1021/acs.jcim.6c01375","url":null,"abstract":"<p>High-quality test sets to evaluate docking software helpthe communityunderstand strengths and weaknesses of different methods and programs.In this work, we introduce SB2025, a calculation-ready set of 1,244protein-ligand complexes derived from the PDB, curated using literature-basedligand protonation states, and freely available in DOCK6, AutoDockVina, and AutoDock 4 (AD4) formats. SB2025 includes 322 unique proteins,17 major protein classes, and ligands with a wide range of propertiesand flexibilities, making it a broad and challenging test set. FARMA2025,a specific subset of SB2025, contains 165 approved drugs bound totheir pharmacological targets and closely mirrors the distributionof FDA drug target classes for small molecules through 2023 (<em>r</em>² = 0.72). Comprehensive benchmarking was subsequentlyperformed to evaluate pose reproduction (PR), cross-docking (CD),database enrichment (DE), and reverse docking (RD) under standardizedconditions. Averaged PR success rate across triplicate runs for DOCK6was slightly improved compared to Vina (followed by AD4) using boththe large SB2025 test set and the more drug-focused FARMA2025 testset. Conversely, CD matrix success rates for AD4 were more favorablethan DOCK6 or Vina across 11 protein families. The programs all yieldedsubstantial early DE; however, the number of active ligands in commonbetween the programs within the top 1% of the database, with one exception,was surprisingly low. Challenging RD tests using FARMA2025, whichgauge the ability to correctly rank the cognate structure best, indicatethat generating the correct ligand pose is necessary but not sufficient.SB2025, FARMA2025, and example scripts and protocols for benchmarkingare available at ringo.ams.stonybrook.edu (downloads) and github.com/rizzolab.</p>","PeriodicalId":44,"journal":{"name":"Journal of Chemical Information and Modeling ","volume":"66 19","pages":"12816–12838"},"PeriodicalIF":6.4,"publicationDate":"2026-10-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149003490","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}

{"title":"Unraveling a Case Where the NAC Approach Fails─MD and QM Modeling of Regioselective cis-Stilbene Oxide Hydrolysis by Engineered Epoxide Hydrolases","authors":"Elvira Bombino,&nbsp;Hesam Arabnejad,&nbsp;Baoyan Liu,&nbsp;Peter A. Jekel,&nbsp;Xiang Sheng,&nbsp;Dick B. Janssen,&nbsp;Fahmi Himo,&nbsp;Hein J. Wijma","doi":"10.1021/acs.jcim.6c02029","DOIUrl":"10.1021/acs.jcim.6c02029","url":null,"abstract":"<p>Molecular dynamics(MD)-based methods to predict the enantioselectivityor product specificity of enzymes typically use a near attack conformation(NAC) approach. The computational efficiency of this ground-stateapproach, taking minutes to hours, allows to in silico screen thousandsof enzyme variants. During the computational design and screeningof new variants of limonene epoxide hydrolase for improved stereoselectivity,it was noticed that, while this approach gave satisfactory predictionaccuracy for most substrates, the enantioselectivity predictions werefar less accurate for variants designed for <em>cis</em>-stilbeneoxide. Here, to investigate whether inaccurate predictions with theNAC-based approach were typical for this substrate, first, a largerdata set was generated by experimentally testing and computationallypredicting the enantioselectivity of an additional 18 enzyme variantsfor <em>cis</em>-stilbene oxide. Subsequent MD simulationsconfirmed that a NAC analysis resulted in inaccurate enantioselectivitypredictions throughout the data set. Density functional theory (DFT)calculations using the B3LYP-D3(BJ) functional, with an active sitecluster model of &gt;300 atoms, provided a rationale for why the NACapproach gave poor predictions with <em>cis</em>-stilbeneoxide. The aromatic rings of the substrate influence the nucleophilicattack on the epoxide group, by providing a blend of stereoelectronicactivation and steric hindrance to the transition state, which isnot considered in the NAC approach. With this substrate, the DFT accuratelypredicted the enantioselectivity. The results offer valuable insightsinto the applicability of the NAC approach for predicting enzyme selectivity,suggesting that more expensive methodologies, which do explicitlyconsider the transition state structure, can also become requiredwhen bulky or stereoelectronically activating groups are bound tothe reacting atoms.</p>","PeriodicalId":44,"journal":{"name":"Journal of Chemical Information and Modeling ","volume":"66 19","pages":"12877–12890"},"PeriodicalIF":6.4,"publicationDate":"2026-10-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149021785","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}

{"title":"ZeoAgent: Autonomous Design of Zeolite Frameworks through Pore-Topology-Guided Generation and Evaluation","authors":"Jing Ping,&nbsp;Zhendong Liu","doi":"10.1021/acs.jcim.6c02058","DOIUrl":"10.1021/acs.jcim.6c02058","url":null,"abstract":"<p>Autonomousagents are emerging as a promising approach to assistand accelerate materials design and discovery. For zeolites, an importantclass of crystalline porous materials, this paradigm could addressa long-standing challenge in framework discovery: the lack of rationalstrategies for coordinating framework generation with property objectives.Yet building such an autonomous agent system for zeolites remainschallenging, as performance is governed by periodic pore topologyand strict framework connectivity constraints. This requires the agentto jointly reason over pore architectures and generate valid frameworks.Here we present ZeoAgent, a zeolite-specific system that convertsnatural-language objectives into computational operations for zeolitestructure analysis and property prediction as well as framework design.ZeoAgent uses a point-cloud representation of pore architecture tolearn from existing zeolite frameworks and guide the generation offramework candidates, which are then evaluated against the designobjective within the same loop. We demonstrate ZeoAgent in representativedesign tasks, including the design of frameworks with prescribed structuralconstraints, enhanced diffusion performance, and pore dimensions suitedfor CO₂/CH₄ separation. This work establishesa route for goal-directed zeolite design, showing how AI agents canbe adapted to periodic porous materials governed by constrained poretopology.</p>","PeriodicalId":44,"journal":{"name":"Journal of Chemical Information and Modeling ","volume":"66 19","pages":"12694–12708"},"PeriodicalIF":6.4,"publicationDate":"2026-10-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149021787","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}

{"title":"Computer-Aided Drug Design for Type 2 Diabetes-Related Enzymes: SILCS-Based Fragment Mapping, Pharmacophore Modeling, and Virtual Screening","authors":"Esmat Mohammadi,&nbsp;Justin A. Lemkul","doi":"10.1021/acs.jcim.6c01733","DOIUrl":"10.1021/acs.jcim.6c01733","url":null,"abstract":"<p>Highly charged enzyme active sites present a challengefor conventionalSite Identification by Ligand Competitive Saturation (SILCS) becausestrongly interacting charged probes can suppress sampling of neutralfragments, while additive force fields do not explicitly account forelectronic polarization. To address these limitations, we developeda comprehensive SILCS-based computational framework and applied itto the N-terminal and C-terminal domains of maltase–glucoamylaseand to α-amylase. Both the conventional mixed-probe SILCS workflowand a modified neutral–charged protocol were evaluated usingthe CHARMM additive and Drude polarizable force fields. FragMap convergenceanalysis demonstrated reliable sampling across all systems, with theneutral–charged protocol alleviating competitive exclusioneffects observed in conventional simulations. In highly charged catalyticpockets, the conventional workflow was dominated by cationic probesampling, suppressing neutral and hydrogen-bonding features. Separationof charged and neutral probes restored chemically diverse interactionpatterns, improving representation of hydrogen-bond donor, acceptor,and hydrophobic hotspots. Incorporation of the Drude polarizable forcefield further enhanced resolution of electrostatic and hydrogen-bondinginteractions. Energy-based SILCS-Pharmacophore models were generatedfrom the neutral–charged FragMaps and used for virtual screeningof the Maybridge and ChemBridge libraries. Hits were refined usingSILCS-MC, ranked by ligand grid free energy (LGFE) and ligand efficiency(LE), and analyzed through scaffold- and fingerprint-based clustering.Integration of energetic and structural analyses enabled identificationof favorable scaffolds and functional group patterns associated withstrong binding efficiency. Together, this study develops, evaluates,and applies an improved SILCS workflow to three enzymes associatedwith type 2 diabetes (T2D), demonstrating its utility for characterizinghighly charged active sites and prioritizing candidate inhibitors.</p>","PeriodicalId":44,"journal":{"name":"Journal of Chemical Information and Modeling ","volume":"66 19","pages":"12969–12988"},"PeriodicalIF":6.4,"publicationDate":"2026-10-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149090866","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}

{"title":"STELLAR: A Fragment-Based Docking Workflow for Highly Flexible Long-Chain Biopolymers","authors":"Alejandro Rodríguez-Martínez,&nbsp;Jochem Nelen,&nbsp;Miguel Carmena-Bargueño,&nbsp;Carlos Martínez-Cortés,&nbsp;Horacio Pérez-Sánchez","doi":"10.1021/acs.jcim.6c00987","DOIUrl":"10.1021/acs.jcim.6c00987","url":null,"abstract":"<p>Docking methods have improved significantly through optimizationusing generative approaches and systematic exploration of parameterconfigurations within docking software. However, docking large moleculessuch as peptides remains challenging, while the docking of large non-peptidicbiopolymers is still insufficiently explored. Existing tools oftenexhibit limitations in both accuracy and computational efficiencywhen applied to longer peptides or other polymers. Moreover, althoughgenerative deep learning and machine learning approaches have shownpromise, they still frequently lack robust ranking metrics, resultingin suboptimal performance. To address these limitations, we developedSTELLAR (Score-Tuning for Efficient Ranking of Large Ligands usingan Accurate and Refined Docking Configuration), a workflow designedto enable efficient and accurate docking of highly flexible long-chainbiopolymers exceeding 10 subunits. STELLAR implements a fragment-basedstrategy, decomposing the biopolymer into smaller units for individualdocking, followed by recomposition into full-length structures. Italso includes optimized algorithms for handling long-chain biopolymers.The pipeline integrates structural optimization steps using toolssuch as GNINA, RDKit, and GROMACS, ensuring physically realistic poses.STELLAR achieved RMSD values below 5 Å in validation experimentsusing benchmark complexes from Propedia and the Protein Data Bank(PDB). It offers a faster and more efficient alternative to severalstate-of-the-art tools across a range of peptide lengths. Additionally,it scales linearly in computational time, runs primarily on CPUs withminimal GPU usage, avoids reliance on generative approximations, providesan accurate metric for ligand ranking, and reduces computational costcompared to similar tools. Its design also supports high-throughputscreening of linear polymer–protein interactions by efficientlyreducing the complexity associated with high conformational flexibility.</p>","PeriodicalId":44,"journal":{"name":"Journal of Chemical Information and Modeling ","volume":"66 19","pages":"12645–12666"},"PeriodicalIF":6.4,"publicationDate":"2026-10-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148973940","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}

{"title":"Cholesterol Hot Spot Automated Mapping Protocol (CHAMP): Identifying Cholesterol-Binding Hot Spots in Membrane Proteins","authors":"David Sotillo-Núñez,&nbsp;Matteo Nardi Cesarini,&nbsp;Gian Marco Elisi,&nbsp;Mattia Bernetti,&nbsp;Giovanni Bottegoni","doi":"10.1021/acs.jcim.6c01419","DOIUrl":"10.1021/acs.jcim.6c01419","url":null,"abstract":"<p>Biological membranes are dynamic and heterogeneous environmentsthat actively influence protein structure and function. Cholesterol,in particular, has been widely reported to modulate membrane proteinsthrough binding to specific interaction sites. However, the identificationof these sites remains challenging due to the intrinsic complexityand dynamics of lipid bilayers, especially when multiple systems andmembrane environments are considered. Here, we present CHAMP (Cholesterolhot spot automated mapping protocol), an automated Python-based workflowthat combines coarse-grained molecular dynamics simulations with adual analysis based on contact persistence and spatial density toidentify cholesterol-binding hot spots at protein–membraneinterfaces. The protocol requires limited user intervention and isdesigned to enable consistent and reproducible analyses across differenttargets. We validated CHAMP on the serotonin transporter and the cannabinoidreceptor 1 (CB₁), where it recovers known cholesterol-bindingregions reported in experimental and computational studies. In theCB₁ system, the method also identifies a potential additionalinteraction site. To assess the generality of the approach, we appliedthe protocol to the CB₁ negative allosteric modulator Org27569,showing that it can capture relevant interaction hot spots beyondsterol molecules. Overall, CHAMP provides a practical framework formapping interaction hot spots at protein–membrane interfacesand may support large-scale and comparative studies of membrane proteins.Limitations of the current implementation and possible extensionsare discussed.</p>","PeriodicalId":44,"journal":{"name":"Journal of Chemical Information and Modeling ","volume":"66 19","pages":"12667–12681"},"PeriodicalIF":6.4,"publicationDate":"2026-10-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149021784","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}

{"title":"Energy Landscape Sampling Reveals Ligand-Dependent Structural Dynamics of the Protease-Activated Receptor 1","authors":"Inken Kaja Schwerin,&nbsp;Shihai Jiang,&nbsp;Claudia Stäubert,&nbsp;Berend Isermann,&nbsp;Georg Künze","doi":"10.1021/acs.jcim.6c01315","DOIUrl":"10.1021/acs.jcim.6c01315","url":null,"abstract":"<p>Protease-activatedreceptor 1 (PAR1) is a class A G protein-coupledreceptor activated by proteolytic cleavage and exposure of a tetheredagonist. While existing structures have revealed inactive and activeconformations, the ligand-dependent conformational energy landscapegoverning PAR1 activation remains poorly understood. Here, we employedenhanced-sampling molecular dynamics simulations using the stringmethod with swarms of trajectories to quantify how extracellular ligandsand G-protein binding modulate PAR1 activation. Free-energy landscapeswere computed for five receptor states (apo, tethered agonist, thrombinreceptor-activating peptide (TRAP), vorapaxar (VPX), and VPX + Na⁺) using five transmembrane (TM) helix distance metrics ascollective variables. In the apo state, PAR1 samples inactive, intermediate,and active-like conformations, indicating substantial intrinsic receptorflexibility. VPX increases the free energy of active conformationsand restricts TM helix 5 and 6 rearrangements, while VPX + Na⁺ eliminates intermediate basins and stabilizes the inactivestate. In contrast, the tethered agonist stabilizes the active stateas the global minimum and promotes TM6 displacement of up to ∼1nm. The simulations reveal a continuous activation pathway linkingextracellular ligand engagement to intracellular motifs, includingthe connector, DRF, intracellular loop 3 (ICL3), and KRK motifs, withintracellular chloride ion coordination contributing to inactive-statestabilization. Gα_q binding reduces TM6 flexibilityand shifts the free-energy minimum toward more active-like conformations.Alanine mutagenesis and G_q-CASE BRET assays validate predictedroles of key residues in activation and G-protein coupling. Together,these results provide a detailed understanding of ligand- and transducer-dependentPAR1 signaling and inform the design of pathway-selective PAR1 modulators.</p>","PeriodicalId":44,"journal":{"name":"Journal of Chemical Information and Modeling ","volume":"66 19","pages":"12800–12815"},"PeriodicalIF":6.4,"publicationDate":"2026-10-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149021786","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}

{"title":"Graph-to-Graph Inverse Design of Surfactants under Multiobjective and Structural Constraints","authors":"Alexander W. Rogers,&nbsp;Ruediger Zillmer,&nbsp;Amanda Lane,&nbsp;Adam Kowalski,&nbsp;Dongda Zhang","doi":"10.1021/acs.jcim.6c02137","DOIUrl":"10.1021/acs.jcim.6c02137","url":null,"abstract":"<p>Rapid,sustainable redesign of large functional moleculesrequiresefficient exploration of vast chemical spaces subject to multipleproperties and constraints. Graph neural networks have shown strongperformance in molecular property prediction, yet their use for inversedesign remains relatively underexplored. Meanwhile, chemical languagetransformers capable of learning long-range relationships have emergedas a competing paradigm. To address this gap, we propose a novel motif-basedgraph-to-graph framework for the inverse design of surfactants, inwhich latent representations are optimized to meet target specifications.We demonstrate the framework on a challenging and industrially relevantproblem, targeting the critical micelle concentration (CMC), the air–watersurface tension at CMC, and the maximum surface excess concentration,while considering constraints on surfactant classes and compatibilitybetween amphiphiles and counterions. We observed that our frameworkcould accurately target multiple properties with greater accuracythan surfactant-property models from the literature. Additionally,compared with our previously proposed SMILES-based transformer, thegraph-to-graph framework significantly improves prediction accuracyand inverse design validity and expands the diversity of generatedsurfactant classes. Overall, these results establish the motif-basedgraph-to-graph framework as a more accurate, chemically expressive,and industrially applicable technology for the multiobjective inversedesign of large functional molecules.</p>","PeriodicalId":44,"journal":{"name":"Journal of Chemical Information and Modeling ","volume":"66 19","pages":"12492–12506"},"PeriodicalIF":6.4,"publicationDate":"2026-10-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149049609","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}

{"title":"Computer-Aided Drug Design for Type 2 Diabetes-Related Enzymes: Polarizable and Nonpolarizable Simulations of Enzyme-Inhibitor Complexes","authors":"Esmat Mohammadi,&nbsp;Justin A. Lemkul","doi":"10.1021/acs.jcim.6c01734","DOIUrl":"10.1021/acs.jcim.6c01734","url":null,"abstract":"<p>Carbohydrate-digestingenzymes, including pancreatic α-amylaseand maltase-glucoamylase (MGAM), are important therapeutic targetsfor type 2 diabetes due to their role in regulating glucose production.In this study, molecular dynamics (MD) simulations were performedto investigate the stability and interaction patterns of inhibitorsidentified through SILCS-based virtual screening. Representative ligandsfor the N-terminal and C-terminal domains of MGAM (NtMGAM, CtMGAM),and α-amylase were parametrized using both the nonpolarizableCHARMM and polarizable Drude force fields based on quantum mechanical(QM) reference data. MD simulations of the apo enzymes show that theiroverall structures remain stable, with flexibility primarily localizedto regions surrounding the catalytic pockets. Simulations of the enzyme–ligandcomplexes demonstrate that the ligands remain associated with thebinding regions throughout the trajectories. Comparison between thetwo force fields reveals distinct interaction patterns: CHARMM interactionsare largely driven by a small number of dominant contacts, whereasthe Drude model captures a more extensive and cooperative interactionnetwork. Notably, the Drude model enables enhanced engagement withcatalytic residues in NtMGAM, whereas ligand binding in CtMGAM andα-amylase remains more flexible and less directly coupled tothe catalytic machinery. These results highlight the importance ofpolarization in describing enzyme–ligand interactions and suggestthat polarization can influence the balance and persistence of interactionsin a system-dependent manner. These findings provide further computationalsupport for the candidate inhibitors identified through computationalscreening.</p>","PeriodicalId":44,"journal":{"name":"Journal of Chemical Information and Modeling ","volume":"66 19","pages":"12955–12968"},"PeriodicalIF":6.4,"publicationDate":"2026-10-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149090868","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}

{"title":"Enzymatic Reaction Feasibility Classification Using Machine Learning Methods","authors":"Xin Wang,&nbsp;Hongyan Yin,&nbsp;Yekai Shen,&nbsp;Yushan Zhu,&nbsp;Igor V. Tetko,&nbsp;Aixia Yan","doi":"10.1021/acs.jcim.6c02294","DOIUrl":"10.1021/acs.jcim.6c02294","url":null,"abstract":"<p>With the advancement of computer-aided retrobiosynthesis,numerousbiosynthetic pathways have been predicted, exceeding the capacityof experimental validation. Effective classifiers are needed to identifyfeasible reactions. In this study, we collected 75,864 feasible reactionsfrom biocatalysis databases and generated an equal number of infeasiblereactions based on reaction rules. Following atom mapping focusedon reaction centers and systematic reaction preprocessing, three datasetsfor training were constructed: the “stereo” dataset,which retained reaction stereochemical information; the “non-stereo”dataset, which was a stereochemistry-agnostic version of the “stereo”dataset; and the “mixed” dataset, which comprised both.We established a total of 22 individual enzymatic reaction feasibilityclassification models, which include: eXtreme Gradient Boosting (XGBoost)and Deep Neural Network (DNN) models utilizing Reaction Fingerprints(RXNFP), Differential Reaction Fingerprint (DRFP), and our constructedCombined ECFP4 Reaction Fingerprints (c_ECFP4) for reaction representation,and Transformer models and fine-tuned ChemBERTa-77M-MLM (ChemMLM)models using reaction SMILES strings as the direct input. The resultsindicate that models utilizing the c_ECFP4 representation achievedthe highest predictive performance, which effectively captured underlyingenzymatic reaction mechanisms. Among them, Model 1A-M (based on XGBoostand “mixed” dataset) was identified as the optimal individualmodel, achieving Matthews Correlation Coefficient (MCC) values of0.865 and 0.853 and Area Under Curve (AUC) values of 0.981 and 0.980on the “stereo” and “non-stereo” testsets, respectively. Furthermore, a consensus model enzymatic reactionfeasibility classification (ERFC) integrating four reaction representationsfurther improved predictive performance, achieving MCC values of 0.894and 0.886 and an AUC of 0.986 on both test sets. Moreover, both models(Model 1A-M and ERFC) successfully validated a five-step biosyntheticpathway, demonstrating higher prediction accuracy than the previouslyreported DeepRFC and DORA-XGB models in identifying feasible reactions.All data, the individual Model 1A-M, and the consensus model ERFCare openly available, offering a reliable and flexible framework forpredicting enzymatic reaction feasibility in the presence or absenceof stereochemical information.</p>","PeriodicalId":44,"journal":{"name":"Journal of Chemical Information and Modeling ","volume":"66 19","pages":"12613–12623"},"PeriodicalIF":6.4,"publicationDate":"2026-10-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149090874","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}