Structure-based virtual screening (SBVS) is a widely used approach in in silico drug discovery, requiring a receptor structure or binding site to predict a ligand's binding pose and affinity. Consequently, the performance of SBVS depends on the protein's conformation. The most common SBVS method is protein-ligand docking, which relies on physics-based, empirical, or knowledge-based scoring functions to estimate binding affinity. However, these methods are highly sensitive to structural variations and conformational changes of the receptor structure. The performance of SBVS often drops substantially when the apo form of the receptor used for screening is different from its holo form. To address this issue, we previously introduced an SBVS method called PL-PatchSurfer. This method uses a surface representation to describe a binding site of receptors and drug molecules, which is more robust to structural changes, offering more stable screening performance. The surface of the molecules is segmented into patches, and their shapes and physicochemical properties, such as the electrostatic potential, are represented by the 3D Zernike descriptor, a compact and rotationally invariant mathematical representation. Here, we introduce PL-PatchSurfer3, an improved version of PL-PatchSurfer, which incorporates two key enhancements: a refined definition of hydrogen bond complementarity and the integration of visibility, which captures curvature information of a patch. Our evaluation demonstrates that the new program outperforms its predecessor and other SBVS methods while maintaining its robustness against receptor structure variations.Scientific contributionThis study presents PL-PatchSurfer3, an improved surface patch-based virtual screening method that incorporates molecular surface shape, physicochemical properties, hydrogen-bonding compatibility, and visibility, capturing local curvature using rotationally invariant 3D Zernike descriptors. PL-PatchSurfer3 demonstrates robust performance across holo, apo, modeled, and AlphaFold-predicted receptor structures and is superior to or competitive with conventional approaches and recent deep learning-based methods.

Modern biochemistry is producing vast amounts of chemical knowledge. Ontologies, such as the Chemical Entities of Biological Interest (ChEBI) ontology, can help organising this knowledge. With manual classification alone however, ontologies cannot keep up with the growth of their domain. In this work, we propose a novel taxonomy of 67 classes related to peptides, a large branch in ChEBI with nearly 15,000 compounds. The existing natural language definitions in ChEBI have been expanded and specified more precisely. These natural language definitions are accompanied by a logical axiomatisation in monadic second-order logic (MSOL). To use the axiomatisation for automated classification, a methodology has been developed that translates monadic second-order definitions first into partial first-order definitions and finally into an algorithmic classification. This connects three aspects important to ontological definitions: They reflect the opinions of experts, they are unambiguous, and they can be checked automatically. In our evaluation, we compare the results of our classification to the current taxonomy of ChEBI . This reveals potential inconsistencies in ChEBI as well as areas that might benefit from automated extensions. We also evaluate our natural-language definitions in an expert survey.Scientific contribution: This work provides precise natural-language definitions of 14 current ChEBI classes as well as 53 new peptide-related classes. These definitions are formalised in MSOL and come with an efficient implementation that allows for large-scale molecule classification, including a full classification of ChEBI and PubChem.

Limited experimental data remains a key challenge in applying machine learning to drug discovery, particularly for cancer-related targets. In this study, we present a data-efficient active meta-deep learning framework to predict mitogen-activated protein kinase 1 (MAPK1) inhibitors, which are promising candidates for cancer-related therapies. Our approach integrates active learning (AL) with a meta-model that combines four deep architectures: a convolutional neural network, an attention, a graph convolutional network, and a graph neural network-attention, trained on molecular descriptors and graph-based representations. These models generate four probability-based features that feed into an attention-based meta-learner, improving predictive performance by 5.12% in the area under the precision-recall curve (AUPRC) and 5.48% in the Matthews correlation coefficient (MCC) using only 10% of the training data. Among the AL sampling strategies evaluated, entropy sampling showed competitive performance in selecting informative molecules for model improvement. Overall, our framework achieves an AUPRC of 0.835 ± 0.017 and MCC of 0.817 ± 0.017, on par with a traditional training method despite using only 26.7% of the training data. Compared to a conventional random forest model trained on brute-force, a 100% full training set, our approach shows a 10.6% improvement in AUPRC and modest gains in MCC, confirming the effectiveness of the proposed framework. Under severe class imbalance, balanced accuracy steadily increased across AL iterations, reaching values greater than 0.85 at the final iteration for all uncertainty-driven strategies. Molecular docking confirmed successful prioritization of the top four predicted compounds. Evaluation on an external MAPK1 data set demonstrated generalizability, with our approach achieving an AUPRC of 0.818 and an MCC of 0.403, comparable to the independent test set. These results highlight the potential of combining intelligent data selection with deep learning architectures through the meta-model to accelerate predictive performance in data-scarce drug discovery. Scientific contribution: This study contributes a novel, data-efficient active meta-deep learning framework for predicting MAPK1 inhibitors, addressing the challenge of limited experimental data in a cancer-specific target. By integrating AL with a meta-model composed of four deep architectures, the approach significantly enhances the predictive performance using only a fraction of the training data. The framework achieves superior metrics compared to traditional training methods, highlighting its potential to accelerate drug discovery in data-scarce settings.

With the disclosure of the important role of substructural alerts (SA) in drug development and toxicity evaluation, many automatic substructure extraction tools based on different theoretical knowledge have been reported in recent years. To compare the emphasis of various substructure extraction methods and the reliability of their results, we were encouraged to conduct a comprehensive analysis of seven representative tools to find the best one. In this paper, we introduced a well-designed evaluation of seven popular tools (Bioalerts, KRFP, MoSS, PySmash_circular, PySmash_group, PySmash_path, and SARpy) based on 43 toxicity datasets, consisting of four components: comparison of substructures derived by different methods, comparison of predictive models based on substructural rules, comparison of the efficiency of extracting toxic substructures, and the effect of SAs on quantitative structure-activity relationship (QSAR) predictive models. The results demonstrated that PySmash_circular performed best overall, with satisfactory results in substructure information carrying and the rule-based predictive models. PySmash_path and Bioalerts were also recommended for their similar performance to PySmash_circular, but the main problem was that they took too much time and generated too many substructures. Specifically, Bioalerts and PySmash_circular could obtain substructures carrying richer information, while SARpy had the best predictive rule-based models, but it only focuses on precision (PR) value in the evaluation of individual SA. More than that, the substructures obtained by all 7 methods can enhance the recognition ability of the QSAR models for toxic compounds and make them interpretable. Finally, we have also made a baseline substructure set of 43 toxicity endpoints available to the public to facilitate further development of drug research and environmental safety assessment in a rapid and accurate direction.Scientific Contribution: Based on 43 toxicity datasets, we conducted a comprehensive evaluation of 7 representative substructure extraction tools from both the perspective of individual substructure and substructure-based models. This work not only enables users to make more autonomous choices of the optimal substructure extraction tool, but also provides the public with a benchmark substructure set of 43 toxicity endpoints, promoting the further development of computational toxicology.

Accurate prediction of compound-protein interactions (CPIs) is crucial for chemical biology and drug discovery. Despite recent advancements, existing deep learning (DL)-based CPI models often struggle to simultaneously achieve high generalization performance, quantify prediction confidence, and ensure explainability. Here, we propose ChemGLaM, a chemical genomics language model designed to address these three crucial challenges, thereby enabling reliable and explainable CPI predictions. ChemGLaM integrates independently pre-trained chemical and protein language models through an interaction block with a cross-attention mechanism, achieving near state-of-the-art performance in predicting novel CPIs at a low computational cost. Incorporating uncertainty estimation and attention visualization enables ChemGLaM to enhance the success rate of virtual screening and to provide molecular insights into CPIs. To demonstrate the practical impact of ChemGLaM, we constructed a publicly available database containing large-scale CPI predictions for every possible pairing between all 20,434 human proteins and all 11,455 drugs and validated its practical applicability in a case study on amyotrophic lateral sclerosis. ChemGLaM marks an important step forward in addressing the challenges of AI-driven CPI exploration and drug discovery.Scientific ContributionThis study established a unified CPI prediction framework that simultaneously achieves high generalization performance, confidence quantification, and explainability. We leveraged this framework to create a community resource by constructing a comprehensive CPI database and demonstrated its practical utility by successfully prioritizing hit compounds and deconvoluting their targets in a phenotypic screening for amyotrophic lateral sclerosis.

Hundreds of computational methods for predicting ligand binding pockets exist, but the problem of finding druggable pockets throughout the human proteome persists. Different strategies for pocket-finding excel in different use cases. Ensemble models that leverage multiple different pocket-finding strategies can best capture diverse pockets at scale. Despite this, no publicly available human-proteome-wide datasets of pocket predictions from multiple pocket-finding methods exist. We present the Human Omnibus of Targetable Pockets (HOTPocket), a dataset of over 2.4 million predicted pockets over the entire human proteome that utilizes both experimentally-determined and computationally-predicted protein structures. We assembled this dataset by running seven diverse, established pocket-finding methods over all PDB and AlphaFold2 structures of the canonical human proteome. We created a novel pocket scoring method, hotpocketNN, which we used to filter candidate pockets and assemble the final proteome-wide dataset. Our hotpocketNN method is able to recover known ligand binding pockets, including those which are dissimilar from any pocket seen in its training set. The hotpocketNN method outperforms all constituent methods, including P2Rank and Fpocket, when assessing the precision withDCAcriterion on the Astex Diverse Set and PoseBusters dataset. Additionally, hotpocketNN was able to identify recently-discovered druggable pockets on KRAS and the mu opioid receptor. We make both the HOTPocket dataset and the hotpocketNN method freely available.

In recent years, machine learning models have shown substantial progress in predicting molecular properties. However, integrating molecular graph structures with sequence information continues to present a significant challenge. In this paper, we introduce Multi-MoleScale, a novel multi-scale framework designed to address this challenge. By combining Graph Contrastive Learning (GCL) with sequence-based models like BERT, Multi-MoleScale enhances the prediction of molecular properties by capturing both structural and contextual representations of molecules. Specifically, the model leverages GCL to effectively capture the intrinsic graph-based features of molecules while utilizing BERT’s pretraining capabilities to learn the contextual relationships within molecular sequences. The contrastive learning component enables Multi-MoleScale to distinguish between relevant and irrelevant molecular features, thereby enhancing its predictive accuracy across diverse molecular types. To assess the performance of our method, we conducted experiments on several widely used public datasets, including 12 molecular property datasets, the ADMET dataset, and 14 breast cancer cell line datasets. The results show that Multi-MoleScale consistently outperforms existing deep learning and self-supervised learning approaches. Notably, the model does not require handcrafted features, making it highly adaptable and versatile for a variety of molecular discovery tasks. This makes Multi-MoleScale a promising tool for applications in drug discovery, materials science, and other molecular research fields. Our data and code are available at https://github.com/pdssunny/Multi-MoleScale. Supplementary InformationThe online version contains supplementary material available at 10.1186/s13321-025-01126-w.

Surfactant mixtures play a critical role in industries such as drug delivery, cosmetics, firefighting foams, and lubrication, serving as foundational components of the global economy. Their performance hinges on micelle formation, a self-assembly process governed by the critical micelle concentration (CMC), which enables key functions like solubilization, emulsification, and targeted molecular delivery. However, rapidly and accurately predicting the CMC of mixtures remains a significant challenge due to the chemical diversity and nonlinear interactions between surfactants. Here, we introduce an artificial neural network (ANN)-based machine learning framework to predict the CMC of binary surfactant mixtures. Our workflow leverages cheminformatics-derived molecular descriptors for each surfactant component, which are then aggregated using strategies such as concatenation, arithmetic mean, and harmonic mean. We find that pairing the arithmetic mean strategy with ANN yields the best performance, effectively capturing complex molecular interactions and enabling dual predictive capabilities: (1) precise interpolation of CMC values at untested mole fractions within known mixtures, and (2) accurate prediction of complete CMC-composition profiles for entirely novel surfactant combinations. SHAP-based interpretability analysis highlights that features such as hydrophobic surface area, electronic topological descriptors, and headgroup basicity drive model predictions, aligning with core principles of surfactant chemistry and reinforcing the mechanistic validity of our model. Overall, this framework accelerates data-driven surfactant design by reducing experimental burden and enabling rapid, rational optimization of formulations across pharmaceuticals, personal care, environmental remediation, and enhanced oil recovery.Scientific contributionThis study presents a novel machine learning framework that, for the first time, predicts full critical micelle concentration (CMC)-composition profiles for binary surfactant mixtures, including untrained systems. By strategically combining the features of individual components of mixtures using arithmeticmean, our artificial neural network model deciphers nonlinear interactions between chemically distinct surfactants, enabling accurate and generalizable CMC predictions. Beyond performance gains, this framework facilitates rapid and systematic exploration of formulation space via inverse design and high-throughput screening, establishing a powerful foundation for the rational development of next-generation surfactants with applications in energy, environmental remediation, pharmaceuticals, and biomedical science.

Cyclic peptides with backbone N-methylated amino acids (BNMeAAs) and D-amino acids (D-AAs) have gained increasing attention for their stability, membrane permeability, and other therapeutic potentials. Currently, Rosetta simple_cycpep_predict (SCP) can predict their structures through energy-based calculations, but this approach is computationally intensive and time-consuming. Moreover, the available crystal structures of such cyclic peptides remain highly limited, hindering the development of data-driven structure prediction models. To address these challenges, we propose HighFold-MeD, a deep learning-based framework that distills knowledge from Rosetta SCP by fine-tuning the AlphaFold model. First, a cyclic peptide structure dataset is constructed using Rosetta SCP by sampling massive conformations for cyclic peptides with BNMeAAs and D-AAs and evaluating their energy scores. The AlphaFold model is then fine-tuned to incorporate the extended 56 BNMeAAs and D-AAs. Besides, a relative position cyclic matrix is introduced to explicitly model head-to-tail cyclization. Finally, a force field is employed to minimize steric clashes in the predicted structures. Empirical experiments demonstrate that HighFold-MeD achieves accuracy comparable to that of Rosetta based on the sampled datasets by the SCP module of Rosetta, with the key parameters that nstruct = 500 and cyclic_peptide: genkic_closure_attempts = 1000, while accelerating structure prediction by 50-fold, thereby significantly expediting the development of cyclic peptide-based therapeutics.Scientific contributionWe propose HighFold-MeD, which provides a rapid and relatively accurate approach for predicting the structures of cyclic peptides containing backbone N-methylated amino acids and D-amino acids—key building blocks in peptide drug design. By distilling the knowledge of Rosetta SCP under specific parameters into a fine-tuned AlphaFold framework, our method achieves a 50-fold acceleration while maintaining relatively high accuracy, thereby enabling large-scale cyclic peptide drug design.Supplementary InformationThe online version contains supplementary material available at 10.1186/s13321-025-01111-3.

Accurately classifying chemical structures is essential for cheminformatics and bioinformatics, including tasks such as identifying bioactive compounds of interest, screening molecules for toxicity to humans, finding non-organic compounds with desirable material properties, or organizing large chemical libraries for drug discovery or environmental monitoring. However, manual classification is labor-intensive and difficult to scale to large chemical databases. Existing automated approaches either rely on manually constructed classification rules, or are deep learning methods that lack explainability. This work presents an approach that uses generative artificial intelligence to automatically write chemical classifier programs for classes in the Chemical Entities of Biological Interest (ChEBI) database. These programs can be used for efficient deterministic run-time classification of SMILES structures, with natural language explanations. The programs themselves constitute an explainable computable ontological model of chemical class nomenclature, which we call the ChEBI Chemical Class Program Ontology (C3PO). We validated our approach against the ChEBI database, and compared our results against deep learning models and a naive SMARTS pattern based classifier. C3PO outperforms the naive classifier, but does not reach the performance of state of the art deep learning methods. However, C3PO has a number of strengths that complement deep learning methods, including explainability and reduced data dependence. C3PO can be used alongside deep learning classifiers to provide an explanation of the classification, where both methods agree. The programs can be used as part of the ontology development process, and iteratively refined by expert human curators. Supplementary InformationThe online version contains supplementary material available at 10.1186/s13321-025-01092-3.