Journal of Chemical Information and Modeling 最新文献

Predicting protein–ligand binding is a centralchallengein computational drug discovery, and while machine learning (ML) andcofolding methods have advanced rapidly, their ability to generalizebeyond training or parametrization regimes remains insufficientlyunderstood. DNA-encoded libraries (DELs) enable ultralarge screeningof billions of molecules simultaneously, providing a useful testbedfor evaluating these approaches at scale. A recent NeurIPS competitionrevealed that even top-performing ML models trained on DEL data failedat generalizing to out-of-distribution (OOD) chemical space. We investigatedwhether integrating structural modeling could bridge this generalizationgap. We systematically assessed state-of-the-art ML, docking, andcofolding methods, including Schrödinger Glide, Rosetta GALigandDock,and Boltz-2 with three biologically diverse protein targets screenedagainst libraries containing multiple DEL synthesis formats. WhileML excels in-distribution, OOD hit discrimination is dependent onboth the target and ligand context, with no single method consistentlydominating. These findings demonstrate that benchmark performancealone is insufficient to predict OOD performance, highlighting theneed for system-dependent evaluation of binding prediction methods.We provide an open-source package for assessing protein–ligandprediction methods and analyzing high-throughput screening data: DEL-iver.

Peptide therapeutics occupy a critical niche betweensmall moleculesand biologics, yet the sequence design of peptide binders remainschallenging because their bioactive conformations and interfacialfitness are dictated by the receptor microenvironment. Current computationalapproaches largely treat this as a generic inverse folding problemthat relies on backbone geometry alone, overlooking the detailed physicochemicalconstraints of the receptor pocket. To bridge this gap, we introducethe REceptor-Aware Peptide Sequence Designer (REAPS), a geometricgraph neural network that reframes sequence design for peptide bindersby treating the receptor as a fully observable all-atom context. Multidimensionalevaluations show that REAPS outperforms the widely used inverse foldingmodel ProteinMPNN in sequence recovery, structural fidelity, and interface-levelbiophysical metrics across both linear and macrocyclic peptide binders.We further integrate REAPS with structural hallucination in a closed-loop de novo discovery pipeline, yielding peptide candidateswith experimentally verified functional activity at the neurokinin-3receptor (NK3R). The source code, model checkpoints, processed datasets, preprocessing scripts, and an example design workflow are publiclyavailable at https://github.com/Mistletoe-git/REAPS.

Enantioselective liquid–liquid extraction (ELLE)based oncyclodextrin host–guest recognition is a promising green strategyfor chiral resolution. However, its industrial application is oftenhindered by restricted chiral recognition efficiency and ill-definedinterfacial mass transfer mechanisms. In this work, a multiscale approach─integratingmacro-equilibrium experiments, molecular dynamics simulations, andquantum chemical calculations─was employed to decode the interfacialmolecular interactions and their decisive roles in the ELLE of ibuprofen.The interface was regulated and reconstructed through the introductionof surfactants. The anionic surfactant SDS shortened extraction equilibriumtime by approximately 30%, while the facial amphiphilic NaDC extendedequilibration but improved the enantiomeric excess. MD simulationsrevealed that the addition of SDS and NaDC increased the interfacialthickness from 0.41 nm by 1.23 and 1.51 times, respectively. Via umbrellasampling, the chiral migration process exhibited characteristic free-energychanges of approximately −11.48 kJ·mol^–1 and +35.58 kJ·mol^–1 when transferring fromthe organic and aqueous phases to the interface, respectively. Furthermore,quantum chemical analysis quantified the interaction energy differences,demonstrating that steric hindrance and hydrogen-bonding intensityvariations drove the observed changes in equilibrium parameters. Thesefindings establish a conceptual microscopic model for interface-mediatedchiral partitioning, providing a robust theoretical framework foroptimizing separation efficiency and process intensification in ELLEsystems.

Per- and polyfluoroalkyl substances (PFAS) comprise morethan 21,000chemicals, yet fewer than 1.3% have any experimental bioactivity measurement,making structure-based hazard prediction the gating problem for prioritizingthe untested majority. Existing PFAS models are single-task and rarelyevaluated under leakage-aware cross-validation. We present the firstmultitask chemical foundation-model benchmark for PFAS bioactivity,fine-tuning off-the-shelf MolFormer-XL with 156 ToxCast classificationheads under Tanimoto-Butina cluster-out cross-validation, which isappropriate because 89% of PFAS have empty Murcko scaffolds. The modelreaches macro AUROC 0.764, beating a Morgan-fingerprint XGBoost baselineby +0.18 and winning on 155 of 156 end points. That baseline is, however,partly degenerate: binary Morgan fingerprints map a quarter of thecompounds onto identical vectors, and sweeping eight 1D/2D/3D representations on the same folds raises the best classicalmodel to 0.688, so the margin properly attributable to the foundationmodel is +0.076. A matched factorial ablation shows that backbonefine-tuning and multitask learning are synergistic: multitask learninghelps only once the backbone is fine-tuned (+0.08). This interactionis backbone- and task-dependent: it reproduces on Tox21 for two backbonesbut is absent for ChemBERTa-2 on PFAS, and label-free encoder diagnosticsshow this is not because ChemBERTa-2 represents PFAS poorly, so staticrepresentation quality does not predict postfine-tuning performance.PFAS-specific domain-adaptation pretraining gives no detectable improvement(Δ = – 0.01 to – 0.03). Mondrian-by-class conformalcalibration improves rare-active-class coverage from 0.36 to 0.78,and an applicability-domain analysis shows only 26% of untested PFASlie close enough to the training set for the model to be applied withoutcaveat. The representation also transfers to soil-water partitioning(K_d) regression, where it holds at Spearmanρ = 0.75 even on PFAS chemistries absent from training whilethe fingerprint baselines collapse.

Predicting drug-target affinity (DTA)is essential forscreeningcandidate compounds in early-stage drug discovery. However, most existingcomputational methods formulate DTA prediction as a regression task,leading to an objective mismatch with the practical need to rank andprioritize candidate compounds based on their relative binding strengths.Moreover, existing ranking-based methods often adopt a fixed rankingobjective or scoring function, which may limit their flexibility whenthey are applied to diverse drug-target interaction scenarios. Toaddress these challenges, we propose MultiRankDTA, a novel frameworkthat formulates DTA prediction as a multilevel ranking task to betteralign with the practical goal of drug screening. By integrating pointwise,pairwise, and listwise ranking strategies through a mixture-of-experts(MoE) module, MultiRankDTA enables dynamic and adaptive compound prioritization.Furthermore, we introduce a fusion network to reduce data set-specificbiases by leveraging auxiliary supervision from multiple sources.Extensive experiments demonstrate that MultiRankDTA achieves superiorperformance compared to state-of-the-art computational methods acrossmultiple benchmark data sets and delivers strong results on an independentbenchmark for virtual screening, indicating its potential for generalizationand practical utility. The source code can be obtained at https://github.com/CSUBioGroup/MultiRankDTA.

We introduce ChemLint, an open-source Model Context Protocol(MCP)server that connects any MCP-compatible large language model to acurated suite of local cheminformatics and machine learning tools,enabling rigorous molecular data handling through a conversationalinterface. Molecular machine learning studies are frequently underminedby inconsistent data preprocessing, including invalid SMILES, unresolvedduplicates, and train-test leakage, yet existing large language model(LLM)-based chemistry tools do not address these data-centric challenges.ChemLint provides tools for data exploration and diagnostics, moleculestandardization, and machine learning modeling. All operations areexecuted deterministically by established libraries and recorded ina project manifest that traces every action, supporting reproducibilityand making curation choices explicit. We demonstrate through severalexamples how ChemLint can be used to identify common data qualityissues, evaluate splitting strategies, and execute complete modelingpipelines from raw data to evaluation. ChemLint is freely availableat https://github.com/molML/ChemLint.

This study employed an integrated ligand-based virtualscreeningpipeline to identify potential CASP4 inhibitors from the DrugBankdatabase, leveraging docking-score prioritization, SMILES-derivedChemBERTa embeddings, and key physicochemical descriptors. Buildingon this foundation, the workflow incorporated virtual screening, cheminformaticsmodeling, PK–PD evaluation, molecular docking, molecular dynamicssimulations, and MM/PBSA free-energy analysis to systematically prioritizerepurposed DrugBank compounds for CASP4-targeted Alzheimer’sdisease therapy. A Random Forest classifier trained on the hybridChemBERTa physicochemical feature set distinguished active from inactivecompounds with ∼95% accuracy (ROC–AUC = 0.73) and achieveda ∼3.5-fold enrichment of active compounds among the top-rankedhits, while a companion Random Forest regressor, trained on the experimentalpIC₅₀ values of active compounds, ranked candidates bypredicted potency. In addition, integrated cheminformatics modeling,PK–PD analysis, and molecular docking further narrowed theselection to the top five candidate compounds: DB00519, DB01068, DB06202,DB08882, and DB05316. Finally, MD simulations and MM/PBSA calculationsestablished clear thermodynamic support for DB05316 and DB00519, whosebinding free energies (−21.4 and −20.9 kcal/mol, respectively)exceeded even the reference compound donepezil, highlighting themas the most promising CASP4 inhibitors. Overall, this workflow providesan efficient and robust strategy for prioritizing repositioned DrugBankcompounds as potential CASP4 inhibitors against Alzheimer’sdisease.

The capabilities of large language models (LLMs) haveexpandedbeyond natural language processing to scientific prediction tasks,including molecular property prediction. However, their effectivenessin in-context learning remains ambiguous, particularly given the potentialfor training data contamination in widely used benchmarks. This paperinvestigates whether LLMs perform genuine in-context regression onmolecular properties or instead rely on verbatim retrieval of memorizedtarget values. Furthermore, we analyze the interplay between pretrainedknowledge and in-context information through a series of progressivelyblinded experiments. We evaluate nine LLM variants across three families(GPT-4.1, GPT-5, Gemini 2.5) on three MoleculeNet data sets (Delaneysolubility, Lipophilicity, QM7 atomization energy) using a systematicblinding approach that iteratively reduces available information,complemented by 0-, 60-, and 1000-shot in-context sample sizes asan additional control for information access. To validate the memorizationanalysis and the blinding experiments, we add a positive and a negativecontrol for the memorization experiments and structural referencebaselines for the multishot experiments, as well as bootstrap confidenceintervals for all results. We find no evidence of verbatim retrievalon the legacy benchmarks and show that blinding exposes conflictsbetween pretrained knowledge and in-context information. This workprovides a principled framework for evaluating molecular propertyprediction under controlled information access.

Sequence models that translate NMR spectra into molecularstructuresreport up to 96% top-1 accuracy, but they are pretrained on simulatedspectra and tested on public databases that can overlap those corpora.We build a leakage-controlled benchmark on nmrshiftdb2 and find that26.9% of public molecules are exact training-corpus matches undera stated identity rule, with substantial training-set proximity remainingafter exact-match removal. On the wider exact-match-removed real-spectrumcohort, released-model top-1 accuracy is 24.4% (95% CI 23.3–25.6%).A training-free rerank using molecular formula and ¹³C-peakcount consistency raises it to 31.1% (29.9–32.4%), supportingleakage-controlled evaluation and decode-time consistency checks beforeclaims of experimental generalization for simulation-pretrained inversemodels.