Journal of Chemical Information and Modeling 最新文献

Candidate structure prioritization in non-target screening(NTS) can beperformed using collision cross-sections (CCSs) predicted by a machine-learning (ML) model, witha threshold on the relative difference between measured and in silico CCSs. Such thresholds generally are heuristicand lack a statistical foundation. In this context, we have constructeda conformal predictive system to quantify the uncertainty of the ML-predictedCCSs and provide guarantees on the minimum percentage of prioritizedtrue structures among the candidate lists of the NTS features detectedin a sample. The system is designed as a wrapper around the open-sourceML model GraphCCS and can, in principle, be combined with any CCSmodel. Additionally, we propose the detection of the applicabilitydomain of the conformal predictive system based on maximum common-edge subgraph (MCES) molecular distances. We showcase theapplication of our tool to two independent in silico NTS datasets. Results on the two tested external datasets suggestthat applicability domain detection based on the MCES distance allowsto maintain the guarantees on the minimum percentage of true structuresprioritized. In terms of candidate structure prioritization performance,on average ∼16% of incorrect candidates could be deprioritizedwhile retaining 90% of all true structures. Although this constitutesa modest 8% improvement over random prioritization, our tool additionallyoffers adaptivity of removal efficiency to the difficulty of the NTScandidates based on quantified uncertainty. Furthermore, the frameworkcan potentially be combined with other analytical properties to improvecandidate structure prioritization in NTS.

pH plays a central role in many biological events. Toexplore theunderlying molecular mechanisms, it is of importance to determinepK_a values of ionizable residues in proteins.In this work, DeepKa, a deep learning-based protein pK_a predictor, was employed to create the pK_a database DeepKa DB (http://computbiophys.com/DeepKa/database), which contains pK_a data of over 30million residues in about 200 thousand soluble proteins. The databasehas been integrated into the existing DeepKa web server, from whichusers can obtain the data of interest freely. In addition to the datapresentation, four representative case studies were set up to demonstratehow pK_a’s extracted from the databasecould be applied to the investigation of pH-dependent processes inproteins.

Buildinga working mental model of a protein typically requiresweeks of reading, cross-referencing crystal and predicted structures,and inspecting ligand complexes, an effort that is slow, unevenlyaccessible, and often requires specialized computational skills. Weintroduce Speak to a Protein, a new capability thatturns protein analysis into an interactive, multimodal dialogue withan expert coscientist. The AI system retrieves and synthesizes relevantliterature, structures, and ligand data; grounds answers in a live3D scene; and can highlight, annotate, manipulate, and see the visualization.It also generates and runs code when needed, explaining results inboth text and graphics. We demonstrate these capabilities on relevantproteins, posing questions about binding pockets, conformational changes,or structure-activity relationships to test ideas in real time. Speak to a Protein reduces the time from question to evidence,lowers the barrier to advanced structural analysis, and enables hypothesisgeneration by tightly coupling language, code, and 3D structures. Speak to a Protein is freely accessible at https://open.playmolecule.org.

Moleculegeneration has emerged as a powerful computational toolfor de novo drug design, enabling the explorationof the chemical space beyond the limits of conventional virtual screening.The field has progressed rapidly, driven by advances in molecularrepresentations, generative architectures, and target-aware modelingstrategies. However, existing reviews typically address specific modelfamilies or application scenarios in isolation rather than offeringan integrated perspective on how these components collectively forma coherent generation workflow. In this review, we present a comprehensiveevaluation of molecule generation models for de novo drug design, covering 82 methods across five deep generative frameworks,including recurrent neural network (RNN)- and transformer-based models,variational autoencoders (VAEs), generative adversarial networks (GANs),flow-based models, and diffusion models. We first summarize widelyused benchmarks and molecular representations and then examine themethodological principles underlying both general and pocket-conditionedgeneration. A central contribution of this work is a systematic synthesisand comparative analysis of the reported performance across commonlyused benchmarks and evaluation metrics. We also summarize representativeexperimentally validated cases. Looking ahead, we discuss future directionsin standardized 3D data, interaction-aware generation, receptor flexibility,and multiobjective molecular design, with the aim of improving thereliability and experimental relevance of molecule generation. Allcollected benchmark resources, evaluation metrics, and model referencesare provided in a publicly accessible repository at https://github.com/JacklinGroup/molecule-generation-review.

Generalizability of molecular docking predictions acrossnovelprotein targets and distinct small-molecule chemotypes is key to thesuccessful application of deep learning co-folding approaches in drugdiscovery, as highlighted by recent benchmarks, including Runs-n-Poses[Škrinjar, P.; et al.Nat. Struct. Mol. Biol.2026, 33, 782–794.]. To comparegeneralizability across physics-based, hybrid rescoring models, andco-folding approaches, we systematically benchmarked several moleculardocking tools representing these paradigms. Our results support previousobservations that co-folding approaches show superior performanceon targets resembling their training datasets. However, they deterioratesharply to unacceptably low 20–40% success rates on novel dissimilarsystems, consistent with memorization. In contrast, physics-basedand hybrid methods exhibit greater out-of-distribution robustness,maintaining more than 60% success rates for docking protein–ligandcomplexes with minimal similarity to previously known complexes. Ourresults highlight the complementarity of AI-based approaches and methodsbased on physical sampling in all-atom models, in terms of their applicabilityrange, and argue for the benefits of tighter integration.

Bayesian flow networks (BFNs) jointly model atomic coordinatesand atom types for structure-based drug design, but their generativedynamics remain poorly understood. We show that the two channels resolveon different terms: atom types pass through a sharp commitment transitionat t_c1, whereas coordinateprecision improves smoothly with no critical point of its own, reachingsub-Ångstrom resolution only at a stated threshold t_c2. Extreme-value theory predicts thetype channel’s posterior half-maximum in closed form, and thepredicted $1/\sqrt{{\beta }_1}$ scaling holds across seven β₁ values (R² = 0.996) and over13 points spanning three BFN architectures and two domains, thoughmeasured times run 11%–20% below the predicted absolute values.Exploiting that structure, phase-aware iterative refinement alignsthree rounds to these boundaries and trains no auxiliary regressor,reward function, or property predictor. On 100 CrossDocked2020 testpockets at K = 200, it reaches a mean Vina Dock of−8.72 kcal/mol. Gradient-guided methods that need such trainingstay ahead by a small margin: over the 100-pocket intersection, thepaired gap to MolJO (−8.98) is 0.25 kcal/mol (95% CI [−0.13,+0.60], paired Wilcoxon p = 0.022); equivalence testingat a preregistered ±0.5 kcal/mol margin does not establish equivalencewithin it. CByG (−9.16) leads by 0.44 kcal/mol on unpairedmeans. Phase-aware iteration improves the Round-1 to Round-3 meanby 1.34 kcal/mol (paired Wilcoxon p = 4.80 ×10^–18, Cohen’s d_z = −1.59) and shifts the median by −1.44kcal/mol, a shift Best-of-K resampling cannot produce.Paired on the pocket, an audited rerun passes every PoseBusters check4.7 percentage points more often than MolJO.

The extreme persistence of per- andpolyfluoroalkyl substances(PFAS) in the environment is rooted in their three-dimensional molecularconformation. The helical twist adopted by perfluoroalkyl chains dueto hyperconjugation involving their recalcitrant C–F bondsgoverns their resistance to degradation, yet a quantitative, continuousmetric for helicity has remained absent. Here we introduce a data-drivenframework that quantifies backbone helicity using three statisticaldescriptors: void-state (binary occupancy) spatial autocorrelations,local backbone principal component analysis, and persistent homology.We further validate each statistical descriptor against geometricdihedral benchmarks and DFT-calculated Vibrational Circular Dichroism(VCD) spectra. The framework is initially established on a homologousseries of perfluorocarboxylic acids (^FC₂ through ^FC₁₆) and their hydrogenated analogues, then extendedacross perfluorosulfonic acids, fluorotelomer alcohols, polyfluoroalkylhexanoic acid analogues, and longer-chain PFOA and PFOS analogues.Agreement between statistical, geometric, and spectroscopic definitionsof helicity is established for PFCAs and extended to structurallydistinct subfamilies, including PFOA analogues of varying fluorinecontent and perfluorosulfonic acids, demonstrating that the descriptorsare robust to changes in headgroup chemistry and fluorine substitutionpattern. The framework provides a chemistry-agnostic toolset for incorporatingbackbone conformation into predictive models of PFAS environmentalfate and degradation reactivity.

Solvent effects playa critical role in determining reactionthermodynamicsand kinetics in liquid-phase heterogeneous catalysis, particularlyin confined nanoporous environments such as zeolite pores. However,accurate prediction of solvation thermodynamics remains computationallyprohibitive due to the need for extensive first-principles calculationson diverse solvent configurations. In this work, we present an attention-enhancedthree-dimensional convolutional neural network (3D-CNN) frameworkto predict solvation interaction energies (ΔE_int) of adsorbates in Ti-faujasite (Ti-FAU) zeolite pores.Molecular dynamics (MD) snapshots of C₁–C₃ oxygenates in aqueous and water–methanol solvent environmentsare transformed into multichannel voxel grid representations thatencode spatially resolved atomistic information. The 3D-CNN architectureincorporates separate adsorbate and solvent branches and employs convolutionalblock attention modules to dynamically prioritize spatial regionsand feature channels relevant to adsorbate–solvent interactions.Trained on density functional theory (DFT)-derived interaction energiesfrom a multiscale modeling workflow, the model achieves a mean absoluteerror of 0.08 eV for individual configurations and 0.04 eV for ensemble-averagedsolvation energies, approaching DFT accuracy while substantially reducingcomputational cost. Cross-validation (CV) using simulation-system,solvent-composition, and pore-type splits demonstrates predictiveperformance on held-out systems within the represented chemical domainand across the solvent and pore conditions considered in this study.Analysis of spatial attention maps and feature attributions revealsthat the model emphasizes adsorbate-proximal solvent structures andhydrogen-bonding interactions, consistent with physical intuition.This framework enables efficient and interpretable prediction of solvationinteraction energies in confined systems and provides a scalable pathwayfor mapping structural information sampled from force-field-MD toDFT targets.

Theinteraction of the SARS-CoV-2 Nsp1 protein with the host NXF1-RRMdomain plays an important role in host gene expression, which influencesvirus replication and host immune surveillance. Thus, understandingthe key factors that determine their binding affinity, such as mutations,is of great interest. In this work, we investigated the effects ofthree Nsp1 mutants─D33K/E36K/E37K/E41K, L27D/V28D, and R124E/K125E─onits binding to the NXF1-RRM domain using structural and energeticanalysis. To quantify the affinity changes, we developed a computationalframework for calculating the free energy landscape of protein–proteinbinding based on the Markov state model (MSM). Our approach is basedon fluctuations of the radius of gyration and the root mean squaredeviation (RMSD) from a reference state. We found that the D33K/E36K/E37K/E41Kmutant exerts the strongest destabilizing effect, resulting in greaterstructural deviation, reduced compactness, and a significant lossof intermolecular hydrogen bonds. In contrast, the L27D/V28D and R124E/K125Emutants retain structural and dynamical properties closer to thoseof wild-type Nsp1 (WT), indicating that these mutations have a morelimited impact on the stability of the Nsp1–NXF1-RRM complex.These results are in good agreement with the previous experimentalwork. We also used the molecular mechanics Poisson–Boltzmannsurface area (MM-PBSA) method to estimate the binding free energyof these complexes, yielding results consistent with the MSM analysisand experimental data. However, the binding free energy values obtainedfrom the MSM method are more realistic compared to the MM-PBSA method,which tends to overestimate binding free energy values or produceunphysically negative results. Therefore, the MSM approach providesa useful framework for characterizing mutation-induced changes inthe structural stability and binding energetics of Nsp1–NXF1-RRMcomplexes.

Surfactant mixturesexhibit non-ideal interactions that governtheir macroscopic behavior, yet their prediction is limited by theneed for experimentally derived interaction parameters. Within Rubingh’sRegular Solution Approximation (RST), these interactions are describedby the parameter β, which is typically obtained by fitting experimentalmixture data, restricting its applicability in large-scale formulationdesign. In this work, a computational hybrid approach combining machinelearning with thermodynamic modeling is introduced to enable end-to-endprediction and optimization of surfactant mixtures. A dataset of interactionparameters derived from experimental binary surfactant mixtures wasconstructed to train predictive models for β, together withmodels for pure-component properties. The predicted, composition-dependentβ values are incorporated into the RST framework to estimatemixture behavior without requiring experimental mixture data, whiletreating the resulting activity coefficients as effective quantities.The overall framework was further validated against a comprehensivedataset of binary surfactant mixtures and further evaluated on ternarysystems, showing good agreement with experimental values across diverseinteraction regimes and compositions. In addition, a multi-objectiveoptimization strategy is introduced to identify formulations thatbalance performance and molecular complexity, enabling efficient explorationof chemical space. A case study on representative systems furtherdemonstrates the applicability of the framework for computationalformulation design. Overall, the proposed approach provides a scalablecomputational tool for surfactant mixture design, reducing relianceon experimental parameterization while enabling rapid virtual screeningof formulation space.