Journal of Chemical Information and Modeling 最新文献

Universal machine learning interatomic potentials (uMLIPs)enableperforming condensed-phase molecular dynamics (MD) simulations withaccuracy approaching that of first-principles; however, their lackof explicit molecular topology limits bond-aware analysis and reconnectionto classical force fields. This study presents CoTAR, a hybrid graphneural network (GNN)–hidden Markov model (HMM) framework thatreconstructs bond connectivity and order, formal charges, and unpairedelectrons from atomic species and coordinates by combining learnedlocal environments, chemical constraints, and temporal smoothing.After joint fine-tuning with 20 labeled snapshots per benchmark system,CoTAR achieved category-macro bond and conditional bond-order F1 scoresof 0.992 and 0.998 and chemically valid snapshots of 82.1% across128 uMLIP systems. Zero-shot tests were successful for all testednonionic systems but revealed persistent formal-charge errors in ionicmixtures. For successfully reconnected systems, the reconstructed-topologysimulation results showed close agreement with the reference-topologyresults for the density, element-resolved diffusion, and intermolecularradial distribution functions. These results establish CoTAR as apractical route from topology-free uMLIP trajectories to chemicallycomplete molecular representations within the present scope of nonreactive,topology-preserving systems.

This letter comparesthe amino acid signatures of natural and artificialintelligence (AI)-predicted antimicrobial peptides. While naturalpeptides from different sources possess distinct signatures, the aminoacid compositions of peptides predicted from venoms, amphibians, insects,and bacteria shared a common feature with higher leucine, lysine,and arginine similar to those of man-made synthetic peptides. Thisstunning finding underscores the need of the right data set for improvingfuture AI prediction of genomic antimicrobial peptides.

Histone deacetylases (HDACs) are pivotal epigenetic regulatorsthat are aberrantly expressed in various cancers, making them prominenttargets for anticancer drug development. Herein, we present HDACiAP(https://hdac.kangsgo.cn), a meticulously curated, comprehensive database and analyticalplatform dedicated to HDAC inhibitors. HDACiAP currently encompasses32,721 compounds with 123,154 bioactivity records, integrating multidimensionaldata such as physicochemical properties, biological activities, toxicitypredictions, compound selectivity annotations, computationally generateddocking poses, and protein–ligand interaction profiles. Theplatform incorporates built-in molecular docking and integrated machinelearning and deep learning-based prediction modules, enabling usersto evaluate the potency of novel compounds. Additionally, HDACiAPoffers an intuitive visualization interface, multimodal search capabilities,and programmatic access via RESTful APIs. In summary, by providinga curated bioactivity resource together with computational dockingannotations and predictive models, HDACiAP offers a practical platformfor exploring HDAC inhibitor chemical space, supporting preliminaryvirtual screening, and facilitating AI-assisted discovery of HDAC-targetedcompounds.

RNAsubcellular localization determines the regulatory contextin which microRNAs (miRNAs), circular RNAs (circRNAs), and long noncodingRNAs (lncRNAs) exert their functions, yet these RNA classes differgreatly in sequence length and circularity, making fixed-length raw-sequencerepresentations inconvenient for cross-RNA comparison. Here, we presentSwinTransRNA, a unified framework that encodes each RNA as a length-normalized64 × 64 frequency chaos game representation (FCGR) of 6-mer composition,where every cell carries a fixed 6-mer identity and the matrix entriessum to one, enabling direct comparison across sequence lengths. Ahierarchical window-attention network captures localized interactionsamong neighboring composition regions and enlarges the receptive fieldthrough shifted windows and patch merging. On audited benchmark datasets for nuclear/cytoplasmic lncRNA, nuclear/cytoplasmic circRNA,and intracellular/extracellular miRNA classification, SwinTransRNAachieves the highest accuracy and F1 among the compared methods onall three tasks (0.748/0.723, 0.914/0.908, and 0.819/0.823 for lncRNA,miRNA, and circRNA, respectively), surpassing RNALight and RNALoc-LM.Ablations show consistent gains over one-hot Transformer baselinesand classical FCGR-based classifiers, while feature-space and curve-levelanalyses characterize the complementary contributions of the FCGRrepresentation and the window-attention backbone. The main noveltylies in combining an order-traceable FCGR with hierarchical windowattention under a reproducible protocol: five-seed repeats, confidenceintervals, and paired tests quantify comparison uncertainty, and patch-embeddingheatmaps with candidate motif screens expose the patterns prioritizedby the model. The unified implementation, traceable sequence processing,and compact 4.25-million-parameter model provide a fixed-dimensional,auditable basis for RNA localization prediction and pattern prioritization.

Enzymes are extensively utilized for catalyzing the productionof value-added compounds in diverse sectors with significant economicimpact. However, challenges such as poor protein expression, low catalyticactivity, substrate/co-factor limitations, and toxicity of final productslimit the efficiency of these biocatalysts. Protein engineering offersa solution by redesigning enzyme catalytic properties to enhance biosyntheticpathways. In this study, we present an automated platform for genediscovery and enzyme engineering, aimed at overcoming these bottlenecks.The GDEE platform finds and optimizes enzymes responsible for rate-limitingsteps in biosynthetic pathways, with objectives ranging from improvingcatalytic efficiency to enabling novel transformations. The platformoperates through four key steps: an initial sequence step that eithersources natural enzyme sequences or generates mutant variants, followedby an atomistic protein structure prediction step, a docking step,and a binding energy evaluator, obtaining a set of variant enzymesequences that are prioritized for experimental validation. Additionally,machine learning-based re-scoring of binding poses to improve bindingaffinities significantly enhances the accuracy of predicting the effectof mutations on binding, highlighting its potential in metabolic engineeringapplications.

Enzymes play a central role in green chemistry and biomanufacturing.However, precise enzyme selection via Enzyme Commission (EC) numberprediction remains a critical bottleneck between the retrosyntheticpathway planning and experimental implementation. Existing computationalmethods predominantly rely on single-modality representations thateither lose spatial topological structures or treat reactions as staticsnapshots, failing to capture the explicit atom-level state transitionsin chemical bond evolution. To address these limitations, we proposeGMSF, a dual-path multimodal deep learning framework. GMSF featurestwo core innovations: Difference Graph Encoding, which leverages atom–atommapping (AAM) to compute node feature differences and extract structuraldifference signatures of reaction centers; and Multiscale SequenceEncoding, which uses AAM-aligned node mappings to hierarchically capturechemical semantics. Experimental results on the ECREACT dataset demonstratethat GMSF achieves an accuracy of 93.17% in the highly challenginglevel 3 (subsubclass) enzyme function prediction, outperforming thestate-of-the-art (SOTA) baseline by a margin of 1.36%. Furthermore,its successful application to real-world metabolic networks in theBioCyc database validates its immense potential for deciphering complexbiochemical reactions, serving as a powerful computational screeningtool to narrow down catalytic mechanisms and coenzyme dependenciesprior to the exact enzyme selection.

Antioxidant peptidesare emerging as promising therapeuticandnutraceutical agents because of their capacity to neutralize freeradicals, modulate oxidative stress, and contribute to disease preventionin fields ranging from oncology to dermatology and food science. However,traditional computational approaches often fail to capture the biochemicalsemantics and evolutionary context of peptide sequences, limitingthe predictive accuracy and mechanistic interpretability. Here, wepresent an explainable deep learning framework integrating evolutionaryscale modeling (ESM) embeddings with a temporal convolutional network(TCN) and long short-term memory (LSTM) architecture to classify antioxidantpeptides and uncover key sequence determinants of activity. ESM embeddingsprovide rich evolutionary and structural priors, enabling the modelto capture both local k-mer motifs and global sequencedependencies. To ensure transparency, we employed a multimethod interpretabilitystrategy using Anchor, LIME, and SHAP, which identified robust, potentiallybiologically meaningful hypotheses containing residues rich in cysteine(C), arginine (R), lysine (K), histidine (H), aspartic acid (D), glutamicacid (E), threonine (T), glutamine (Q), proline (P), methionine (M),tyrosine (Y), and tryptophan (W) as central to antioxidant activity.Anchor offered deterministic, high-precision motif rules, while LIMEand SHAP highlighted redundant and context-dependent contributions,together forming a layered interpretability framework. This studynot only achieves state-of-the-art classification performance butalso bridges black-box modeling and biochemical understanding, offeringa generalizable, hypothesis-generating platform for rational peptidedesign and discovery that is important in human health.

Dynamins, crucial GTPases involved in membrane remodeling,playkey roles in processes like endocytosis and membrane fission. Whiletheir importance is well-established, the precise dynamics of theirconformational changes remain unclear. This study investigates thebacterial dynamin-like protein (BDLP) from Nostoc punctiforme, which undergoes a dramatic closed-to-open transition upon GTP binding.Using standard and enhanced all-atom molecular dynamics simulations,we map this transition, revealing a complex energetic landscape. Theinitial barrier arises from disruption of salt bridges formed by theGTPase domain with the trunk and paddle, setting off the main hingemotion. Subsequent conformational changes involve further hinge motionsand lateral shearing to minimize electrostatic forces. Notably, GTPbinding accelerates the transition by 5 orders of magnitude throughan allosteric mechanism. Our multiscale simulations of 37.1 μsaccumulated simulation time show that BDLP dimerization and membranebinding are necessary to initiate the stabilization of the open conformation.These findings provide crucial insights into BDLP’s membraneremodeling mechanism, advancing our understanding of dynamin dynamics.

Preparing ChEMBL bioactivity data for quantitative structure–activityrelationship (QSAR) modeling requires navigating curation decisionsthat often lack documentation and are hard to reproduce. We presentCAPRICHO (ChEMBL Aggregation Package with Robust Inspection and CurationHandling Options), a Python command-line tool addressing two gapsin current workflows: transparency and flexibility. Unlike approachesthat remove problematic data, CAPRICHO flags quality issues whilepreserving all data for inspection, letting users assess how eachcuration decision affects cross-assay comparability. Aggregation isconfigured via customizable grouping columns, supporting both traditionalQSAR and emerging assay-aware modeling paradigms, with all parametersrecorded in recipe files for reproducibility. Three case studies demonstratethe tool: (i) programmatic support of ChEMBL max-curation standards,(ii) CYP inhibition multitask curation, and (iii) Caco-2 permeabilitywith unit standardization. Together, these show that preserving ratherthan removing flagged data lets practitioners identify which qualityissues most affect comparability and balance data quality with availabilityfor QSAR modeling. CAPRICHO is freely available at https://github.com/CDDLeiden/Capricho

Kinetic processesplay a central role in structural transformationsacross natural, engineered, and biological systems, ultimately determiningtheir final properties. Traditional kinetic characterization methodsare often time-consuming and resource-intensive, and require priorassumptions about the reaction mechanism. In this work, we presenta physics-informed machine learning framework to extract closed-formkinetic representations from time-series data without preselectinga single reaction mechanism. The framework combines neural networks,physics-informed features, and sparse regularization. The frameworkwas applied to a range of processes, including thermally driven reactionssuch as polymer curing and crystallization, and nonthermal processessuch as enzymatic browning. The method identified sparse, closed-formkinetic representations while reducing reliance on a single predefinedkinetic model. R² values exceeded 98% for the syntheticcases and ranged from 76% to 92% for the experimental cases. In addition,a systematic ablation study evaluated robustness to noise, data availability,sampling density, incomplete measurements, and candidate functionlibrary variation, and an experimental case study was included toassess the performance under realistic conditions. These results demonstratethe potential of this approach to accelerate kinetic characterizationacross polymer, food, and biomolecular processing applications, whereaccurate kinetic models are essential.