Journal of Cheminformatics

A compound's toxicity can be gauged from the toxicities of the compounds it resembles. We showed that in 2003 for rat oral LD50: the compounds in a reference set of 13,645 that most resemble the query form a consortium and vote on its toxicity. This work returns to that method twenty-three years later, on public data, because the reference set used in 2003 is no longer public. Rebuilt on 11,989 curated rat oral LD50 values, the method answers for 67.0% of the held-out compounds carrying an exact value, and 92.1% of those answers land within one log unit; in 2003 the figures were 94% within one log unit on 57.3% of compounds. We then ran it over 28 further endpoints, each recording whether a compound showed a particular toxicity, from cytochrome inhibition to mutagenicity to clinical toxicity: 149,224 measurements on 34,162 compounds. The method answers only when a compound has close enough neighbors, and on those it correctly identifies a median 72.9% of the compounds that showed the toxicity and 96.5% of those that did not. Separately we asked whether a reverse screen carries information about these endpoints beyond molecular size and lipophilicity. We search a compound against a published index of 27,910 ligands solved inside proteins and count the protein families those sites belong to. A compound whose matches reach no protein family inhibits CYP3A4 2.5% of the time; one whose matches span sixteen or more families inhibits 51.2% of the time. Added to heavy atom count, molecular weight and cLogP, the screen raises the mean area under the ROC curve from 0.72 to 0.76 across the 28 endpoints. The OpenADMET consortium has since measured CYP3A4 inhibition on 2,335 compounds, and the same pattern holds. Compounds that resemble known binders of sixteen or more protein families inhibit CYP3A4 24.6% of the time; compounds whose matches reach no protein family inhibit it 7.3% of the time. The more protein families a compound resembles binders of, the more likely it is to inhibit CYP3A4. This association is strong enough to act on before anything is made. Scoring a library on how far its compounds reach across protein families, together with their size and lipophilicity, and setting aside the riskiest fifth removes two thirds of the androgen receptor binders and about half of the CYP3A4 inhibitors. Scoring needs only a two-dimensional structure, with no docking and no conformer generation at query time. Interactive scoring, the data and the pool are at toxpred.ai. Keywords: toxicity prediction; structural similarity; nearest neighbors; reverse screening; target breadth; promiscuity; LD50; cytochrome P450; virtual library triage

Accurately predicting cellular responses to drugs remains a challenge with the potential to reduce experimental screening costs and accelerate drug discovery. Current computational approaches represent compounds through chemical structures, which carry little information regarding their activity within biological systems. We show that gene expression responses, measured in a reference cell line, can instead serve as transferable representations of chemical perturbations. The proposed framework, BioPert, uses biological representations with a small neural network to predict transcriptional delta responses in cell lines of different lineages. Across the Tahoe-100M and LINCS L1000 datasets, BioPert outperforms molecular fingerprints and embeddings learned from chemical structure, which provide limited improvements over random controls. BioPert's prediction correlations reach 0.79 on Tahoe-100M, corresponding to an improvement of 0.34 over the next-best representation. On LINCS, performance varies with the experimental reproducibility of the test conditions, highlighting the impact of batch effects. Notably, BioPert considerably surpasses predictions obtained by copying the reference response, demonstrating that it captures context-dependent effects. Further analysis of a C32-cobimetinib case shows pathway-level accuracy of the predictions, including when the reference and target responses differ. These results open a new path for chemical perturbation prediction and could ultimately reduce the burden of phenotypic screening.

Protein-cofactor selectivity prediction is important for enzyme annotation, metabolic engineering, and biocatalyst discovery, but it remains challenging because experimentally supported associations are limited and unevenly distributed, and because biologically relevant differences can be subtler than those captured by general protein-ligand interaction models. Here, we present MINT (Multimodal Interaction Network with Teacher supervision), a multimodal framework for cofactor specificity prediction. MINT combines protein sequence features with cofactor sequence and graph representations and is trained with a gradual adaptation strategy that transfers useful information from a pretrained protein-ligand interaction model to the cofactor prediction task. On a fixed balanced held-out test set, MINT achieved the best overall performance among all compared methods (AUC = 0.932), and this ranking was reproduced in an independent family-stratified evaluation at the natural positive rate (pooled AUC = 0.939). Further analyses suggest that the pretrained model provides transferable structural information, while the gradual training scheme helps adapt that information to cofactor specificity. In a case study of the chemically near-identical NAD/NADP pair, MINT recovered 95.0-98.6% of true binders and resolved a single-phosphate-level distinction with direct relevance to engineering enzyme cofactor specificity. These results show that gradual transfer learning is an effective approach to cofactor specificity prediction when annotation is limited.

Predicting ligand activity across G protein-coupled receptors (GPCRs) requires models that capture both molecular structure and receptor-specific information while remaining robust to chemical and target-domain shifts. We developed a multimodal graph neural network that combines explicit ligand molecular graphs with frozen protein language model representations of GPCR sequences and jointly predicts quantitative pActivity and binary activity. The model was trained on 271,739 curated ligand-GPCR pairs spanning 183,694 ligands and 216 human GPCRs and evaluated using random, Bemis-Murcko scaffold, and strict cold-ligand partitions. With ESM-2 650M receptor embeddings, the selected model achieved mean absolute errors of 0.513 {+/-} 0.006, 0.540 {+/-} 0.006, and 0.641 {+/-} 0.005 pActivity units under random, cold-ligand, and scaffold evaluation, respectively, substantially outperforming a protein-aware fixed-feature multilayer perceptron and consistently improving quantitative prediction over a matched GINE reference. Removing receptor embeddings markedly degraded both regression and classification, whereas differences among ESM-2 35M, ESM-2 650M, and ProtT5 were comparatively small. Independent evaluation on 6,319 ChEMBL 37/BindingDB pairs revealed a substantial external domain shift, with mean absolute error increasing to approximately 0.94-0.95 despite chemically stringent internal validation. Nevertheless, receptor-specific information remained highly informative in a DRD2-DRD3 selectivity analysis: the cold-ligand ensemble achieved R^2 = 0.853, Spearman {rho} = 0.916, and ROC-AUC = 0.966 for strong DRD3 selectivity, whereas the receptor-independent ablation approached chance-level discrimination. These results establish protein-language-model-conditioned molecular graph learning as a scalable strategy for GPCR bioactivity prediction while identifying chemotype novelty, external-domain transfer, and richer ligand-receptor interaction representations as key remaining challenges.

Large language model (LLM) agents are increasingly capable of carrying out autonomous computational research, but it remains unclear whether they can develop molecular modeling methods that compete with strong human-developed approaches. Here, we evaluate autonomous method development across four settings: Therapeutics Data Commons (TDC) ADMET tasks, the OpenADMET ExpansionRx Challenge, the activity prediction track of the OpenADMET PXR Induction Challenge, and the Practical Molecular Optimization (PMO) benchmark. Agents followed the prescribed data splits, metrics, and evaluation protocols for each benchmark. On TDC, the best agent-developed score across configurations improved on the leaderboard reference on 18 of 21 tasks and on the combined leaderboard and peer-reviewed reference on 17, while individual Codex configurations improved on the combined reference on 13 tasks. On ExpansionRx, five-agent and single-agent Codex configurations achieved overall scores corresponding to fourth and sixth place relative to the published final leaderboard, with the five-agent configuration reaching top-five performance on seven of nine endpoints. On PXR, five-agent Codex achieved a mean absolute error slightly outperforming the best published challenge entry. On PMO, five-agent Codex developed a molecular optimizer that outperformed existing methods evaluated under the benchmark's standard setup; other methods obtained higher reported scores under modified conditions, including additional molecular pretraining data or a different oracle-call budget. Together, these results show that current LLM agents can autonomously develop high-performing molecular prediction and optimization methods across substantially different drug-discovery settings, in several cases matching or exceeding leading human-developed approaches.

Within the ADMET continuum, metabolism and excretion (ME) form a critical bridge between drug exposure established by absorption and distribution and downstream efficacy and toxicity. However, no existing framework for drug ME prediction has simultaneously achieved broad coverage of endpoints and robust data recency and completeness. Here, we developed MetExPred, a multi-view prediction framework that comprises 17 classification endpoints and two regression endpoints, covering the overall drug ME process. The framework combines sequence-based and graph-based molecular representations, while optionally incorporating ESM-2 protein representations when experimentally annotated targets are available. A protein-aware masking strategy enables the same architecture to operate in both molecular-only and target-enhanced settings, and multi-view attention adaptively integrates the available representations. Across the classification tasks, MetExPred achieved mean AUROC, AUPRC and F1 scores of 0.844, 0.734 and 0.697, respectively. For clearance and half-life prediction, the model achieved RMSE values of 0.686 and 0.668. MetExPred showed the strongest average performance among the evaluated baselines, while ablation studies confirmed the complementary contributions of molecular sequence, graph and protein information. These results provide a unified and flexible modeling framework for systematic ME prediction of drug compounds, enabling early-stage virtual screening and pharmacokinetic assessment during lead optimization.

Accurate molecular property prediction requires effective molecular representations that can describe a molecule from multiple complementary perspectives. Existing deep learning approaches typically use SMILES strings, two-dimensional molecular graphs, or three-dimensional conformations as inputs. These representations capture different aspects of molecular information: SMILES encodes a sequential description, molecular graphs describe atom-bond connectivity, and 3D conformations provide spatial information from atomic coordinates. However, these modalities are often learned separately or merged with low effective fusion operations, which may not sufficiently capture interactions among sequential, topological, and geometric features. In this work, we propose Tri-Modality Cross-Attention (TMCA), a multimodal framework that integrates 1D SMILES, 2D molecular graphs, and 3D molecular conformations for molecular property prediction. TMCA uses pretrained encoders for the 1D and 3D branches, with a SMILES-based Transformer for the 1D branch and a recent conformation-aware pretrained model for the 3D branch. At the same time, a trainable 2D graph encoder is designed to support modality fusion. All the parameters from the pretrained models are frozen to make overall training more efficient. We evaluate TMCA on four datasets (i.e., BBBP, BACE, ClinTox, and HIV) from the MoleculeNet database for classification tasks and compare its performance with state-of-the-art competing approaches. Experimental results show that TMCA achieves the best overall performance, demonstrating the value of multi-modality integration for drug property prediction and the effectiveness of our cross-attention-based fusion strategy. The source code of TMCA is available at: https://github.com/Ay-Zhao/TMCA

Recent advances in deep learning-based structure prediction, including AlphaFold3 and the open-source Boltz model family, have extended biomolecular modeling to joint prediction of protein-ligand, protein-nucleic acid, and multi-chain complexes with binding-affinity estimation. Boltz-2 is among the most feature-complete of these open models, but its practical use requires a local CUDA-capable GPU, command-line execution, and manually authored YAML configuration files, limiting accessibility for researchers without dedicated computational infrastructure. We developed Boltz2-Notebook, a Colab-native interface comprising four integrated stages - automated environment setup, interactive parameter-to-YAML generation, execution management, and automated confidence and affinity visualization - together with a manifest-driven batch mode for multi-target screening. All modelling capabilities are inherited unmodified from Boltz-2; Boltz2-Notebook's contributions are limited to accessibility, input construction, and workflow automation. Independent of the software, we curated a benchmark of 317 protein-ligand pairs (122 proteins, 277 ligands) from BindingDB and predicted binding affinity in triplicate using the Boltz-2 command-line engine on high-performance computing infrastructure. Predicted and experimental pIC50 values showed moderate correlation (Pearson r = 0.609 [95% CI 0.540-0.675]; Spearman {rho} = 0.625; R2 = 0.371; MAE = 0.968 pIC50 units), with high triplicate reproducibility (pairwise r = 0.97) but a systematic compression of the predicted affinity range and no measurable relationship between Boltz-2's self-reported confidence metrics and prediction accuracy. Boltz2-Notebook is freely available as open-source software and provides external, reproducible evidence - including a specific confidence-calibration limitation - relevant to interpreting Boltz-2 affinity predictions responsibly.

Hit-to-lead optimization is a multi-objective molecular-design problem in which binding-related scores must be considered together with drug-likeness and synthetic accessibility. We developed a closed-loop language-model medicinal-chemistry agent that combines target and starting-hit context, PubMed retrieval, molecular-property tools, docking, and feedback from previously scored analogues. In a matched computational benchmark comprising six kinase targets, six methods, a 40-candidate budget, and six independent campaigns per method, the agent achieved the highest mean aggregate score under a pre-specified composite objective on four targets. Re-analysis of the same candidate sets with drug-likeness-gated docking and docking-drug-likeness hypervolume produced different target-level winners, showing that evaluation rules can change the comparative interpretation of an optimization campaign. Candidate distributions and component ablations indicated that the agent's advantage was associated primarily with retention of scored-candidate feedback and higher-QED regions rather than with docking alone. An EGFR case study illustrates a scaffold-preserving in-silico optimization trajectory from erlotinib. The study is a computational benchmark and evaluation framework; the prioritized molecules are hypotheses for subsequent experimental testing, not experimentally validated leads.

Protein complex structure prediction depends on both the breadth of candidate coverage and the ability to select accurate models within a limited output budget. Most existing approaches emphasize only one stage. End-to-end models and molecular docking workflows generate candidate structures, whereas quality-assessment models primarily rerank a predefined candidate pool. A complete strategy must account for complementary sampling across generators, differences among scoring scales and the allocation of final candidate quotas. Here, EnsPlex is presented as a multi-source framework that couples structural sampling with candidate selection for protein-protein interaction complex prediction. EnsPlex combines conformation-expanded docking, AlphaFold-Multimer, AlphaFold3 and Boltz-1 to expand the sampled conformational space. Conformation-expanded docking comprises monomer conformational expansion followed by flexible HADDOCK docking. FACET is trained to predict candidate quality using DockQ and its component metrics as supervision. Across 102 antigen-antibody systems, EnsPlex achieved up to a 27.8% relative improvement in target success rate over AlphaFold3 with its built-in ranking under matched output budgets. FACET also improved within-source ranking in the internal candidate pools and in homology-filtered external data. These findings support the utility of EnsPlex for structure prediction in the evaluated antigen-antibody systems.

To address widespread data leakage and inconsistent evaluation in protein-ligand affinity prediction, we introduce PLABench, a leakage-controlled and target-centric benchmarking framework that enables standardized comparison across sequence- and structure-based models. We benchmark nine deep learning methods across blind CASP16 targets, leakage-controlled ChEMBL35 sets, and established Davis and KIBA datasets under rigorous data split settings, standardizing structural input via AlphaFold3 to ensure fair comparison. Although pretrained structure-based models achieve the highest overall accuracy, they show severe target-dependent variability, and increasing structural fidelity from predicted to experimental conformations yields no consistent gains. Meanwhile, sequence-based approaches surpass some structure-based methods on select targets, and protein family-level evaluations reveal uneven performance across families and substantial inter-model complementarity obscured by aggregate metrics. These findings demonstrate that training scale and structural input alone cannot guarantee cross-target generalization, highlighting the need for context-aware interaction modeling. PLABench provides an extensible open-source platform to facilitate these developments.

Experimental assessment of developmental neurotoxicity (DNT) is costly, time-consuming, and difficult to scale across large chemical inventories. New approach methodologies (NAMs), including computational approaches based on machine learning, deep learning, and large language models (LLMs), offer new ways to screen and prioritize chemicals. However, these emerging computational approaches have largely focused on predicting toxicity endpoints, with less attention to the mechanisms underlying those predictions. For DNT assessment, understanding this mechanistic basis is equally important. In this study, we integrated LLM inference with the adverse outcome pathway (AOP) framework to connect DNT predictions with plausible mechanisms. We constructed an AOP-informed benchmark linking ChEMBL target-activity measurements for 5,861 compounds to 14 AOPs representing five neurodevelopmental outcomes and evaluated ten LLMs for the prediction of molecular initiating events (MIEs), key events (KEs), key event relationships (KERs), and adverse outcomes (AOs), as well as the generation of mechanistic explanations. Explanation quality was evaluated using both a naive LLM judge and a rubric-based LLM judge. Across the zero-shot models, the highest observed macro-F1 scores were 0.818 for MIEs and 0.463 for AOs, but only 0.056 for KEs and 0.034 for KERs, identifying downstream pathway reconstruction as the principal challenge. Few-shot prompting substantially improved performance, with Gemini 3.8 Flash under four-shot prompting achieving macro-F1 scores of 0.956 for AOs, 0.813 for MIEs, 0.492 for KEs, and 0.457 for KERs. Both LLM-judge configurations indicated improved mechanistic explanation quality with few-shot examples, although strong outcome prediction did not consistently ensure recovery of the recorded pathway. Our results demonstrate the potential of AOP-informed LLMs to support DNT assessment by linking predicted neurodevelopmental outcomes to plausible biological mechanisms and enabling predictions to be evaluated at both the outcome and pathway levels.

Agentic systems increasingly coordinate molecular-design tools, but it is unclear which layer of the stack limits outcomes on real projects. We developed MAGI, an open modular agent that authors objectives, launches and monitors optimization, interprets structure activity relationships, and revises its strategy accordingly. MAGI generates molecules either directly through the LLM or by delegating to REINVENT4, with scoring services interchangeable behind a common contract. We tested it across nine retrospective lead-optimization campaigns from three pharmaceutical companies, replayed under fixed temporal cutoffs. Both routes produced valid structures: LLM proposals stayed closer to local chemistry and reached comparable or higher primary activity in fewer operations, whereas REINVENT explored broader chemical space. Whether a campaign met its objective depended on the predictive models, not on the generation route: attainment followed model accuracy on the chemistry proposed, dropping once that chemistry moved outside the model applicability domain. Separately, a blinded evaluation asked whether the MAGI output could pass as expert work: chemists were not able to discriminate agentic proposals from held-out compounds, and judged the SAR reasoning broadly plausible yet incomplete. Together, these results position MAGI as a coordination layer pluggable into existing computational chemistry workflows. The ceiling on real projects, however, remains currently set by scorer applicability rather than by tool orchestration.

Protein fitness prediction is essential for protein engineering. Recent fitness-landscape benchmarks such as FLIP2 have shown that methods performing well under conventional train-test splits can fail when the test set is shifted out of distribution (OOD), including mutations at unseen positions and variants from unseen protein backbones. In particular, current protein language models (pLMs) often suffer substantial performance degradation under such distribution shifts, limiting their practical use in protein engineering. We therefore develop a framework that augments sequence representations with structural and co-evolutionary information for fitness prediction. Specifically, we pair structure-aware pLMs with features pooled from an AlphaFold-style trunk applied to each sequence and its multiple sequence alignment (MSA), and use a lightweight prediction probe to map these frozen representations to fitness scores. Our framework consistently outperforms the best previously reported FLIP2 scores across several types of distribution shift. We then systematically study the contribution of each component in our framework to generalization. Structural priors yield the largest performance gains on held-out protein backbones, whereas trunk features improve generalization to held-out mutation positions. Rather than proposing a single best method, we characterize which modalities are effective for different protein families and distribution shifts, providing practical guidance for choosing representations according to the generalization regime. These findings support the development of more generalizable fitness predictors that can serve as effective reward functions or oracles for protein design toward desired functions.

In 1998 we predicted percent human intestinal absorption (HIA) for 86 drug and drug-like compounds from six molecular descriptors chosen by a genetic algorithm with a small neural network serving as the fitness function. That result rested on one division of the compounds into three sets: training, cross-validation and external prediction. Repeating the division shows it was a rather fortunate one. An extraction is one random draw from a wide distribution, so we repeat every experiment 250 times and report the result as a distribution. Across 250 stratified extractions, 14 match the published cross-validation error and only one matches the published external prediction error. Reconstructed this way, the published six descriptors give a mean extraction error roughly twice the external prediction value we reported in 1998 (Table 1). We then applied the same six descriptors to a larger set of 780 compounds from a public compilation. Under the 1998 network they give 22.50 %HIA units on training compounds against 25.77 on extracted compounds, whereas on the 86 they give 13.43 against 30.39. More compounds close the gap between fitting and prediction. Four datasets are modeled here: 780 small molecules with HIA measured in man, 3,771 small molecules and 1,281 cyclic peptides measured across Caco-2 cell monolayers, and 7,298 cyclic peptides measured across artificial membranes. Each is its own prediction problem, and one procedure is applied to all four. A genetic algorithm chooses which descriptors to use and how many, and a random forest grades each choice on compounds its own trees never saw. The descriptors chosen most often across the 250 extractions form one consensus set for each dataset. For absorption, 25 consensus descriptors under a random forest reach an extraction rms of 20.47 %HIA units at R 0.76. On the same compounds and the same extractions, the six 1998 descriptors under the 1998 network reach 25.77 at R 0.59. The four models are served at predicthia.ai, where a structure can be drawn or pasted, scored against its training set for similarity, and the training data with fitted models downloaded. Each model served is its fixed consensus set fitted on every compound in its dataset; the accuracy quoted beside a prediction comes from evaluating that same set across 250 extractions (Table 3). The fixed sets hold 25 descriptors for absorption, 55 for Caco-2 small-molecule permeability, 50 for Caco-2 peptide permeability and 25 for artificial-membrane permeability. Keywords: human intestinal absorption; QSPR; genetic algorithm; descriptor selection; cyclic peptides; membrane permeability; oral bioavailability; repeated extraction

Membrane permeability remains a major challenge in cyclic-peptide drug discovery, where optimization often depends on estimating changes relative to measured reference compounds. Existing predictors mainly model absolute permeability, while heterogeneity across literature sources complicates pooled training and comparative evaluation. We present CycDelta, a multimodal residue-graph model trained directly on within-source permeability differences. CycDelta combines directed message passing with pretrained Uni-Mol monomer embeddings, physicochemical descriptors and assay information in a shared encoder for reference and query peptides. To evaluate this comparative formulation under controlled conditions, we establish CycDeltaBench, comprising in-distribution, scaffold-based out-of-distribution, external and zero-shot unseen-assay tests. On the out-of-distribution task, CycDelta achieved Pearson 0.69 versus 0.46 for the strongest baseline. It also led all evaluated methods on two external data sets and achieved zero-shot Spearman 0.70 versus 0.32 for the strongest baseline on an unseen assay. Residue-level attribution relates predictions to monomer contributions, while gradient-guided screening supports prioritization of permeability-enhancing analogue hypotheses. These results position CycDelta as a transferable and interpretable framework for comparative permeability prediction and cyclic-peptide optimization.

Protein inverse folding, which infers compatible amino acid sequences from a given backbone structure, is central to rational protein design. Current methods rely largely on sequence recovery rate as the evaluation metric, emphasizing structural compatibility but seldom addressing function. Yet functional activity depends on the proper network of non-covalent interactions-protein-ligand contacts being a prime example. We argue that function-oriented protein design is governed by three complementary constraints: structural, interaction, and evolutionary, none of which are handled in a unified manner by existing pretrained models. To bridge this gap, we introduce zPolaron, a graph neural network built on protein-ligand complexes for functional protein design. zPolaron represents residues and ligand atoms as a heterogeneous graph and jointly encodes structural, interaction, and evolutionary information by fusing physical-interaction-enhanced graph representations with ESM-2 embeddings. Across multiple fitness prediction benchmarks, zPolaron outperforms all baselines, with gains evident across structural, interaction, and evolutionary dimensions. Wet-lab experiments further confirm its utility in functional enzyme screening and design: the optimized CalA6-D205A-S160G mutant achieves a 17.84-fold improvement in activity, while the CalB-I121M mutant yields a 1.91-fold increase. Together, these results show that integrating interaction and evolutionary information fills the missing puzzle of inverse folding, enhancing protein fitness prediction beyond structural compatibility. zPolaron is available at https://github.com/zelixirSH/zPolaron.

Drug-drug interactions (DDIs) represent a critical challenge in drug development and clinical practice, as they can lead to severe adverse effects, including toxicity and reduced therapeutic efficacy. Deep learning methods have shown promise in large-scale, rapid DDI prediction; however, current approaches suffer from significant limitations in providing mechanistic insights into these interactions. Here, we propose DualTopoDDI, a dual-topology-enhanced interpretable deep learning model for DDI prediction. We applied DualTopoDDI to predict 9.2 billion potential interactions among approved drugs, achieving 97.99% high-confidence predictions. The model demonstrates molecular-level interpretability, identifying key substructures responsible for drug actions in terms of both atom-centric and bond-centric views. DualTopoDDI also excels in elucidating particular DDI toxicity mechanisms, which is validated by its successful explanation of the controversial cardiac toxicity in two COVID-19 drug combination regimens at the time. Evaluations across 11 benchmark datasets demonstrates that DualTopoDDI not only achieves state-of-the-art performance but also showcases robust generalizability and good interpretability. Overall, DualTopoDDI offers a powerful, interpretable tool for understanding and predicting drug-drug interactions, providing critical insights for drug safety and design.

Structure-guided mutation design requires relating a proposed substitution to its molecular environment to anticipate its effects. Existing evaluations of large language models (LLMs) in protein engineering emphasize sequence-based prediction or tool-integrated workflows, leaving unclear whether LLMs themselves can infer mutation effects from structural evidence without dedicated predictors. We evaluate physical intuition-qualitative reasoning that connects structural observations with molecular interaction principles-using a benchmark that tests the same substitution in the same protein across different binding partners, with predictions based on local atomic coordinates and no dedicated mutation-effect predictors. In the tested comparisons, the best-performing LLM achieved group accuracy comparable to the evaluated energy predictors and Biomni, an LLM-driven tool workflow, while the methods differed in their prediction patterns. Most evaluated LLMs achieved higher group accuracy for larger affinity improvements. Structural interventions support selective dependence on mutation-site and partner information, while assessable structural claims largely agree with the supplied coordinates. These findings support direct LLM structural reasoning as a basis for developing rational mutation-design tools that connect molecular observations with partner-specific effects.

Accurate modelling of biomolecular interactions is fundamental to drug discovery, yet current artificial intelligence (AI) workflows remain fragmented across structure prediction, affinity estimation, molecular design, and experimental decision-making. We introduce AnewDDE, an agentic Drug Discovery Engine that connects these capabilities into a closed-loop system for biomolecular interaction modelling and design. We demonstrate several components of this system: AnewFold delivers superior performance on challenging targets, including antibody-antigen complexes and molecular glues, providing more reliable structural hypotheses together with pocket identification and conformational analysis; AnewDesign achieves a 10.7% success rate in identifying binders with single-digit-nanomolar affinities measured by surface plasmon resonance (SPR) in a representative nanobody discovery campaign, substantially reducing the experimental search burden; AnewAffinity combines accuracy approaching that of free-energy perturbation (FEP) methods with the speed required for high-throughput screening and prioritizes promising transformations; and AnewMind, a large language model (LLM) at the hundred-billion-parameter scale that underwent in-house full-parameter post-training, provides scientific reasoning and absorption, distribution, metabolism, excretion, and toxicity (ADMET) prediction for pharmaceutical research and development (R&D). To systematically evaluate AnewMind, we introduce PharmBench, an internal benchmark designed by pharmaceutical R&D experts to evaluate long-horizon decision-making in drug discovery. Evaluations on ADMET tasks and PharmBench show that AnewMind performs competitively with leading frontier models. Together, these capabilities establish AnewDDE as an integrated engine for addressing challenging therapeutic targets and translating molecular insights into experimentally validated candidates at scale.