Journal of Chemical Information and Modeling 最新文献

Publicly accessible compound repositories are essentialfor virtualscreening and hit identification, but their continuous growth bringsinconsistent structure representations, reactive groups, and a computationalburden from standardizing millions of entries. Many repositories alsocontain outdated or out-of-stock compounds that are no longer procurable,weakening the practical value of screening results. We present a 20-yeareffort to build and maintain an open-access compound repository: theMolport in-stock catalog, now a collection of 6.1 M purchasable compoundssourced exclusively from validated suppliers. We describe the cheminformaticsworkflow applied to this catalog, which harmonizes structures fromopenly accessible, heterogeneous sources into a standardized format,maintains dynamic currency by including only compounds confirmed instock and removing outdated or unavailable molecules, and profilesphysicochemical properties, druglikeness, three-dimensional (3D) shape,and structural diversity to produce chemically meaningful, medicinallyrelevant subsets. The result is an organized, up-to-date view of purchasablechemical space that supports faster and more reliable identificationof experimentally accessible hit candidates for downstream screeningcampaigns.

Reliable prediction of hydrogen adsorption free energy(ΔG_H) is essential for acceleratingelectrocatalystdiscovery for the alkaline hydrogen evolution reaction (HER), yetpractical machine-learning workflows remain limited by inconsistentenergetic definitions, heterogeneous density functional theory (DFT)protocols, and poor out-of-distribution (OOD) generalization. Here,we present an energetically anchored machine-learning framework thatintegrates machine-learning interatomic potential (MLIP)-derived energeticdescriptors with pretrained Crystal Hamiltonian Graph Neural Network(CHGNet) latent embeddings to predict DFT-defined hydrogen adsorptionenergetics across chemically diverse catalyst surfaces. The frameworkemploys protocol-consistent single-point MLIP energy evaluation onreference geometries to construct thermodynamically aligned energeticdescriptors while combining them with structural representations throughgradient-boosting regression. Using curated data sets from CatalysisHub and AQCat25 containing single-, bi-, and trimetallic adsorptionsystems, the embedding-augmented energetic model achieved high predictiveaccuracy (R² = 0.976, MAE = 0.054 eV,RMSE = 0.105 eV) under site-level data splitting, substantially outperformingembedding-only and physicochemical-descriptor-only models. Shapleyadditive explanations analysis revealed a hierarchical learning mechanismin which the protocol-consistent energetic descriptor serves as thedominant thermodynamic anchor, whereas structural descriptors providesecondary refinements that improve adsorption-energy discrimination,particularly near the thermoneutral regime relevant to catalyst screening.Explicit evaluation of MLIP-only geometry workflows further demonstratedthat the framework retains meaningful adsorption-energy ranking capabilitydespite degradation introduced by MLIP-relaxed geometries. AdditionalOOD and transfer-learning analyses showed that predictive robustnessdepends strongly on the balance between compositional diversity andreference-protocol consistency. These results establish energeticallyanchored MLIP embeddings as an effective strategy for scalable post-DFTadsorption-energy refinement and data-efficient electrocatalyst screeningwhile clarifying the practical limitations of MLIP-driven workflowsfor heterogeneous catalysis.

This review addressesa critical need in the emerging intersectionof digital chemistry and material properties by critically evaluatingthe current capabilities and future prospects of applying machinelearning (ML) to ionic liquids (ILs) for thermal energy storage (TES)applications. We examine available data sets for key thermophysicalproperties relevant to TES, including thermal conductivity, heat capacity,density, viscosity, surface tension, melting point, enthalpy, CO₂ solubility, and toxicity. Particular emphasis is placed onchemical space coverage, biases, and practical limitations. We comparedifferent modeling approaches, spanning from conventional QSPR/QSARand group-contribution methods to contemporary ML approaches, includingkernel models, ensemble methods, graph neural networks (GNNs), anddeep learning architectures. The impact of molecular representationsand descriptor choices and validation strategies in determining modelrobustness and transferability is examined. To structure this assessment,we introduce a three-tier property-selection framework for IL-basedTES and apply a six-criterion evaluation checklist consistently acrossall nine reviewed properties. The analysis is supported by literature-summarytables that consolidate data sets, chemical space coverage, modelingapproaches, and validation practices across different property domains.Finally, we outline recommendations for data set standardization,open sharing of data and code, and the establishment of benchmarkvalidation protocols to improve reproducibility and accelerate thediscovery of ILs for TES applications.

The inhibition of Aβ₄₂ (β-amyloid)fibrilformation is a key therapeutic strategy in Alzheimer’s diseaseresearch. Paddlewheel diruthenium complexes have shown promising activityagainst Aβ₄₂ aggregation and preformed fibril disaggregation,yet their molecular mode of action remains poorly understood. In thiswork, we perform atomistic simulations to explore how charge modulationinfluences the interactions of three analogous paddlewheel dirutheniumcomplexes, the parent neutral complex [Ru₂Cl(D-p-FPhF)(O₂CCH₃)₃], andits anionic [Ru₂Cl₂(D-p-FPhF)(O₂CCH₃)₃]⁻ and cationic[Ru₂(D-p-FPhF)(O₂CCH₃)₃]⁺ counterparts (D-p-FPhF^– is the N,N′-bis(4-fluorophenyl)formamidinatoligand) with Aβ₄₂. Our results indicate that electrostatictuning governs binding affinity and the extent of interaction acrossthe Aβ₄₂ fibril surface. As the complexes’charge changes from −1 to +1, the interaction pattern shiftsfrom localized contacts to widespread, multi-site engagement encompassingkey charged, aromatic, and hydrophobic regions of Aβ₄₂. This enhanced binding correlates with longer-lived, thermodynamicallystable interactions at the fibril interface, which effectively lowerthe free energy penalty for fibril disassembly. Overall, our findingspropose a mechanism in which charge-dependent activation through ligandexchange enhances fibril recognition and promotes disruptive bindingmodes, demonstrating the potential of charge-tunable diruthenium complexesas therapeutic modulators of Aβ₄₂ fibril stability.

Alzheimer’s disease (AD) remains a major neurodegenerativedisorder with limited therapeutic options, while currently approvedacetylcholinesterase inhibitors (AChEIs), such as donepezil, are associatedwith adverse effects including cardiotoxicity. Here, we integrateda deep learning-based framework to design novel AChEIs candidateswith improved predicted cardiac safety. A reinforcement learning-guidedGraphVAE (RL-GraphVAE) was employed for target-biased molecular generation.The generated compounds were prioritized through CardiotoxPred-basedcardiotoxicity screening, molecular docking, triplicate moleculardynamics simulations, and chemical synthesis. Experimental evaluationidentified D0209 as a lead molecule exhibiting a noncompetitive inhibitionmechanism. Notably, compared with donepezil, D0209 showed ∼43-foldlower hERG channel inhibition, indicating an improved cardiac safetyprofile. Overall, this integrated computational and experimental workflowdemonstrates the utility of generative modeling for the discoveryof novel leads with improved predicted safety profiles, providingpromising starting points for further optimization and biologicalevaluation.

Accurate prediction of drug selectivity is critical forprioritizingcandidate compounds with reduced off-target effects in AI-based drugdiscovery. Although selectivity can be inferred indirectly from drug–targetaffinity (DTA) prediction, cumulative errors from affinity predictionsacross multiple targets can reduce reliability, motivating the developmentof methods that directly predict compound-level selectivity. We proposea Chemical Uncertainty-aware Embedding (CUE) framework that integratesmolecular fingerprints and 2D molecular structure image embeddings.The two modalities are combined via Loss Trajectory Analysis for Uncertainty(LTAU)-based weighted fusion, which adaptively reweights featuresaccording to their predictive reliability. Across eight benchmarkdata sets, CUE achieved RMSE values ranging from 0.163 to 1.691 andoutperformed affinity-based and quantitative structure–activityrelationship (QSAR) baseline models. Furthermore, a virtual screeningcase study for EGFR(T790M/C797S) inhibitors further demonstrated itsutility for identifying selective hits. The source code is availableat https://github.com/jjjabcd/CUE.

Microscopic deformation stage recognitionfrom moleculardynamics(MD) trajectories is crucial for understanding the evolution of materialdamage; however, traditional empirical analysis and black-box singledeep-learning models lack both high-throughput spatiotemporal modelingand transparent physical interpretability. This work develops a multiscaleinterpretable deep-learning framework to automatically classify elastic,plastic, and fracture stages from uniaxial tensile MD trajectoriesand physically decode model decision logic. First, a 3D atomic trajectoryis converted to 2D gray-scale image sequences; a CNN-LSTM hybrid architectureis built to jointly extract spatial atomic textures and long-timedeformation dynamics, reaching 98.8% test accuracy and far surpassingspatial-only CNN baselines in both supervised classification and unsupervisedclustering. More importantly, a hierarchical multiscale interpretabilitytoolkit, including multilayer feature heatmaps, smoothed gradientsaliency maps, gradient-weighted class activation maps, and regularizedSHAP attribution, is integrated to quantify positive/negative featurecontributions and localize model focus regions. The visualized attentionzones perfectly match core physical fields (von Mises strain, atomicdisplacement, and nonaffine deformation), resolving the explainabilitygap of conventional MD data mining pipelines. This work establishesa generalizable, trustworthy AI paradigm that connects data-drivenprediction to intrinsic microscale mechanical mechanisms for computationalmaterial research.

Ionization sources convert neutral analytes into ions,enablingtheir subsequent separation and detection by mass spectrometry. Dependingon the technique, this mechanism occurs within a solution or in thegas phase. The two most common methods to generate ions are electrosprayionization (ESI) and atmospheric-pressure chemical ionization (APCI),respectively. Both ESI and APCI can generate the same resultant ionspecies; however, it has been shown that preferential charge sitescan change in response to an ionization source environment. Therefore,an additional quality control is warranted a priori to delineate erroneous analyte identification and/or unequivocalstructural assignment. For protonation, protomer fidelity can be assessedby comparing the ion-mobility collision cross section (CCS) valuesobtained for both ESI and APCI. In this work, we utilized proxy standards(i.e., controls) derived from the available literature to developa proof-of-concept multimodel inference application called AGENTito probe differences in ionization for high-throughput protomer shiftexploratory analysis. AGENTi uses in sequence the TensorFlow Autoencoderanomaly detection and the YDF Gradient Boosted Tree machine learningframeworks to generate more balanced and robust anomaly scores. Wefound that using a multimodel screening approach can reduce falsepositive capture at an acceptable loss of true positives. Moreover,we expect to achieve a better recovery of true positives with an improvementon the current Gradient Boosted Tree regression model for ΔCCSprediction as more ESI-APCI dual data are obtained for training. Inall, we demonstrated in this pilot study that our foundation modelfor a priori probing of ionization pathway fidelityis feasible and efficient with AGENTi. The AGENTi prototype applicationand documentation are accessible at https://github.com/mitkeng/AGENTi.

Genetic variants in untranslated regions (UTRs) can alterpost-transcriptionalregulation, yet functional interpretation remains difficult becauseregulatory mechanisms differ between 5′UTR and 3′UTRcontexts and are not fully captured by generic noncoding-variant scores.We developed UTR-FNex, an interpretable machine-learning frameworkfor prioritizing candidate functional variants in both UTR regions.We first constructed MetaUTR, a harmonized resource of approximately200,000 UTR variant loci, and derived model-development sets containing14,627 3′UTR and 3533 5′UTR variants after confidence-basedfiltering. UTR-FNex integrates region-specific sequence, RNA-structure,regulatory-motif, miRNA, conservation, and DNABERT-derived allelerepresentations, with a neural feature enhancement (NFE) module providingadditional interaction summaries. Across 10 repeated random hold-outs,mean AUROCs were 0.876 for 3′UTR and 0.974 for 5′UTR.Under gene-grouped evaluation, AUROCs decreased to 0.769 and 0.672,respectively, indicating substantial context dependence, especiallyfor 5′UTR. In an external prostate-cancer 3′UTR cohort,UTR-FNex achieved an AUROC of 0.720 (95% CI 0.690–0.750), comparedwith 0.696 (95% CI 0.660–0.733) for FunUV; however, FunUV achievedhigher AUPRC and stronger top-k enrichment. A TP53rs78378222 case study connected model sensitivity to a known polyadenylation-signalcontext, whereas RNAfold-derived differences were window-dependentand were treated as exploratory. These results support UTR-FNex asa multimodal prioritization framework while defining important limitsin cross-gene and cross-source generalization. Source code is availableat https://github.com/liyaning233/UTR-FNex, and MetaUTR is deposited at 10.5281/zenodo.19719670.

Quantitative structure–activity relationships(QSARs) remainwidely used in medicinal chemistry, yet their physical interpretationis often limited and the meaning of their applicability domain (AD)is commonly treated in statistical or descriptor-space terms. Herewe propose a conceptual framework in which QSAR is interpreted asa small-deformation theory of effective free-energy landscapes. Ligandbinding or chemical substitution is represented as a perturbationof an effective free-energy landscape defined over reduced state variables,and the resulting free-energy response is expressed as a contractedobservable over the equilibrium ensemble of the reference system viaa Zwanzig-type perturbation formula. At the phenomenological level,this response reduces to a low-order expansion in descriptor space,from which conventional linear QSAR emerges as the leading-order approximation.Within the assumed effective description and under stated regularityconditions, the corresponding first-order coefficients are given exactlyby the ensemble-averaged sensitivities of the free-energy landscapeto descriptor perturbations, without requiring a cumulant truncation.The AD acquires a physical meaning as the region in which a locallow-order expansion of the free-energy response remains valid. Inthe small-deformation regime, this same region further correspondsto ligand-induced landscape perturbations that remain local, smooth,and sufficiently weak over the reference ensemble. Conversely, insystems for which an effective-landscape description is appropriate,QSAR breakdown may be associated with phenomena such as activity cliffs,induced fit, and state-dependent binding, which can involve qualitativereorganization or rapid reweighting of the effective landscape. Fora specified thermodynamic system, the second-cumulant term providesa non-positive variance correction to the leading-order ensemble-averagefree-energy response, arising from the variance of the landscape perturbationover the reference ensemble. This correction reflects a fluctuation-assistedeffect in which configurational heterogeneity within the pre-existingensemble allows Boltzmann reweighting to redistribute statisticalweight toward microstates more strongly influenced by the perturbation.Independently, the exact free-energy response is no greater than theleading-order ensemble-average estimate, consistent with Jensen’sinequality. The present framework is conceptual. Its interpretationof QSAR coefficients as ensemble-averaged sensitivities of the free-energylandscape requires additional assumptions and state-resolved informationbeyond conventional compound-level QSAR data, and it does not by itselfprovide a prospective AD method.