Accurate recognition and reconstruction of chemical structures from molecular images is critical in cheminformatics, with applications spanning drug discovery, materials science, and patent analysis. However, much of this chemical matter exists in implicit forms that are not directly machine-readable, which limits its integration into structured databases. Traditional rule-based and machine learning-based methods often perform poorly when images contain overlapping atoms, ambiguous bonds, or low visual quality. To address these challenges, we introduce I2M, a framework integrates Vision Transformer (ViT) encoders enhanced with adapter modules and utilizes pre-trained Detection Transformer (DETR) weights for improved accuracy and efficiency. I2M identifies atoms, charges, and bonds by predicting their locations and categories, then assembles these elements into chemically interpretable molecular graphs. The model employs robust query selection mechanisms proposed in DETR, improving decoding accuracy, and advanced validation routines to effectively handle complex cases. Compared with existing models, including MolVec, DECIMER-Image-to-SMILES, DECIMER Image Transformer, MolGrapher and MolScribe, I2M achieves good performance in both accuracy and computational efficiency. Additionally, to facilitate the automatic and convenient testing of other chemical structure recognition methods in future studies, we have developed a chemical structure image benchmark and published the datasets. This work highlights I2M’s transformative potential in cheminformatics, offering a robust solution to bridge the gap between molecular imagery and digital molecular representations. I2M introduces the first box-based detection framework for molecular structure recognition, enabling direct prediction of atoms, bonds, and charges with spatial localization while preserving molecular 2D layouts. Compared with existing image-to-SMILES and graph-based recognition approaches, I2M provides an interpretable detection-based paradigm with improved accuracy and computational efficiency across diverse chemical image benchmarks. Furthermore, we establish publicly available chemical structure image benchmarks to facilitate reproducible evaluation and future development of molecular image recognition methods.
The development of materials science is undergoing a shift from empirical approaches to data-driven and algorithm-oriented research paradigm. The state-of-the-art platforms are confined to inorganic crystals, with limited chemical space, sparse experimental data and a lack of integrated online computation for rapid validation. Here, we introduce the Clean Energy Materials Platform (CEMP, https://cleanenergymaterials.cn), an open-access platform that integrates high-throughput computing workflows, multi-scale machine learning (ML) models and a comprehensive materials database tailored for clean energy applications. A key feature of CEMP is the online calculation module, which enables fully automatic quantum and molecular dynamics simulations via structured table uploads. CEMP harmonizes heterogeneous data from experimental measurements, theoretical calculation and AI-based predictions for four material classes, including small molecules, polymers, ionic liquids, and crystals. The platform hosts 380,000 entries, including 20,000 experimental records, 50,000 quantum-chemical calculations and 310,000 AI-predicted properties. The database covers 12 critical properties and the corresponding ML models demonstrate robust predictive power with R2 ranging from 0.64 to 0.94, thus ensures rapid material screening, structure-property relationship analysis and multi-objective optimization for clean energy applications. CEMP aims to establish a digital ecosystem for clean energy materials, enabling a closed-loop workflow from data acquisition to material discovery and real-time online validation. CEMP establishes a unified, source-attributed data and computation infrastructure for clean-energy materials by integrating curated records with automated molecular simulations and machine-learning-based property prediction. Its central contribution is the direct connection of searchable materials data to standardized quantum-chemical and molecular dynamics workflows, enabling users to move from data retrieval to reproducible computation and property screening within a single web-based platform. By bridging materials data, physics-based simulation, and data-driven prediction, CEMP provides an accessible framework for discovery of clean energy materials.
Accurate prediction of small-molecule bioactivity is critical for accelerating drug discovery. Here, we report MoActiPred, a multimodal AI framework that integrates 2D molecular graphs, physicochemical descriptors, and 3D E (3)-equivariant geometric representations. Using a curated ChEMBL-derived influenza virus neuraminidase (NA) inhibitor dataset (1245 compounds; IC50 threshold 20 nM), MoActiPred showed the strongest overall performance among the evaluated methods, including traditional machine-learning models (Random Forest, SVM, and XGBoost), SMILES-based language models (ChemBERTa, ChemFM, and PolyBERT), and graph-based baselines (MolCLR and SEGNN). On the public test set, MoActiPred obtained an accuracy of 0.952, MCC of 0.797, and F1 score of 0.898, while retaining high average precision, supporting the value of combining 2D and 3D representations. To evaluate its practical utility, MoActiPred was applied prospectively to 26 oseltamivir analogues designed, experimentally synthesised and evaluated against influenza A (H3N2) NA. The model correctly predicted 6 of 7 high-activity compounds (IC50 < 20 nM) and 18 of 19 inactive or weakly active derivatives, yielding 92.3
Allosteric regulation, whereby small molecules or mutations modulate protein function from sites distal to the orthosteric active site, is a pervasive mechanism of biological control and an important opportunity for drug discovery. Yet the prospective identification of allosteric binding sites and functionally relevant mutations remains difficult because allosteric effects arise from subtle long-range communication and are not readily captured from heterogeneous sequence, structural, and ligand information. Although recent computational methods have improved allosteric-site prediction, most remain limited to pocket-level classification, are built on a single modality, or do not address the functional consequences of mutations. Here, we present MMAllo, a multimodal deep learning framework that integrates protein language model embeddings, ligand Morgan fingerprints, and graph-based structural representations to enable unified prediction of allosteric binding residues and allosteric mutation effects. MMAllo comprises two modules: MMAlloBind, for ligand-specific allosteric binding-residue prediction, and MMAlloMut, for assessing the functional impact of candidate allosteric mutations. On benchmark datasets, MMAlloBind achieves competitive residue-level performance on AlphaFold3-predicted structures, while MMAlloMut effectively prioritizes curated allosteric mutations over putative neutral mutations. In case studies of PTPN2–FRJ (9C56) and CBLB–compound 9 (8QTG), MMAlloBind recovers experimentally characterized allosteric pockets and identifies representative high-probability candidate communication paths connecting ligand-contact residues to distal functional regions. For the disease-related THRβ V264D mutation (P10828; 6KKB/6KNU), MMAlloMut predicts an above-threshold allosteric effect using only the wild-type AlphaFold3 model; subsequent Gaussian accelerated molecular dynamics, NRIMD network analysis and MM/PBSA calculations provide supporting evidence that V264D perturbs local packing, long-range communication patterns and protein–ligand interaction energetics. Together, these results establish MMAllo as a general framework for linking residue-level allosteric site discovery with mutation-impact prediction and for supporting allosteric drug discovery, particularly in settings where high-resolution experimental structures are unavailable. Scientific contribution The key advance of this work is the development of a unified multimodal framework that moves beyond the current scope of computational allostery prediction. Whereas prior methods have largely focused on identifying allosteric pockets, often from a single source of information, MMAllo performs ligand-specific prediction at residue resolution and explicitly evaluates the functional impact of mutations within the same framework. By integrating protein language model embeddings, graph-based structural representations, and ligand fingerprints, MMAllo captures complementary features of allosteric regulation that are not jointly modeled in existing approaches. Importantly, the framework does not only improve predictive capability, but also provides interpretable computational outputs, including candidate communication paths and mutation-associated changes in inferred long-range coupling patterns. Its ability to operate on AlphaFold3-predicted wild-type structures further extends allostery analysis to proteins lacking experimentally resolved mutant or complex structures. In this way, MMAllo differentiates itself from prior work by unifying site discovery and mutation-effect prediction, while offering a practically deployable strategy for mechanistic analysis, functional variant prioritization, and allosteric drug design.
Deep learning models are data-hungry, and synthetic (artificial) data has been shown to be invaluable when data availability is low. While this has been demonstrated in certain technology areas, adopting such an approach is new in machine learning (ML) applications in chemistry, except for some pre-training tasks. In drug discovery, predicting binding energy between proteins and ligands is crucial. Many ML-based studies have been proposed to predict protein-ligand binding affinity using existing experimental data. However, these models suffer from inherent biases. Recent efforts have produced PLAS-20k, a synthetic dataset of multiple protein-ligand complex (PLC) conformations generated using molecular dynamics (MD) simulations as a viable option to complement existing experimental data and improve binding affinity prediction. For the binding affinity prediction task, we employ Pafnucy, a deep convolutional neural network, and propose using multiple structures for each PLC from PLAS-20k for training. We compare four different statistical and ML-based result-aggregation techniques. This work demonstrates the utility of dynamic datasets in enhancing binding affinity predictions, laying the foundation for future improvements in predicting similar protein properties using synthetic datasets and more sophisticated models and methods. We propose that physics-based synthetic datasets can significantly help develop more accurate data-driven methods. Scientific contribution This study shows that incorporating synthetic molecular dynamics data improves deep learning models for protein–ligand binding affinity prediction beyond static experimental structures. By systematically evaluating frame selection and prediction aggregation strategies, we demonstrate that training on diverse conformational snapshots significantly enhances generalization and accuracy. Our results highlight that dynamic synthetic datasets can enable deep learning models to outperform conventional methods such as MM-PBSA while remaining computationally efficient.
Despite their role as oncogenic drivers and predictive biomarkers, alterations in rearrangement during transfection (RET), a receptor tyrosine kinase (RTK), remain significant challenges due to off-target effects and reduced efficacy against emerging mutations, necessitating more selective and innovative drug design. Herein, we report a multi-layered framework encompassing both reinforcement learning (RL) training and post-processing phases, for the efficient de novo design of RET inhibitors striking a balance between structural novelty and predictive reliability. We evaluated two RL strategies: potency-focused optimization (PFO) and polypharmacological multi-parameter optimization (PMPO), with the latter integrating 131 predictive models for off-target selectivity and phenotypic activity. For the 23 kinase targets, pIC50 classification thresholds were defined using Youden's J statistic (mean J = 0.762, range 0.679–0.918). The 108 NSCLC phenotype models showed limited discriminative performance (mean AUC = 0.550). After applicability domain filtering, the increased mean AUC allowed the use of the phenotype models as a confidence-weighted filter in the generative workflow. Post-hoc similarity distribution analysis confirmed that the generated candidates represent structurally novel yet domain-consistent chemotypes, as evidenced by peak densities situated within the Tanimoto similarity range of 0.20–0.40 (generated vs known). Notably, PMPO significantly enhanced multi-parameter consistency (CV: 5.1
Protein folding remains a fundamental problem in computational biology, where understanding folding trajectories and dynamic behavior under physiological conditions is critical for linking structure to function. Recent advances such as AlphaFold have revolutionized sequence to structure prediction, establishing a powerful and robust foundation for computational protein modeling. Building upon this success, we introduce a novel physics informed computational framework that expands static structure prediction toward dynamic and physically consistent modeling of protein folding processes. The proposed framework unifies a mathematically grounded reaction diffusion model of folding kinetics with deep sequence embeddings through a physics informed neural network (PINN), ensuring consistency with governing physical laws while remaining biologically grounded. We refer to this unified framework as PhyFold. Our results illustrate that the proposed approach effectively captures environment dependent folding dynamics and shows strong agreement with both numerical reference solutions and experimental fluorescence measurements under physiologically relevant conditions. By explicitly integrating environmental factors such as temperature, pressure, and pH, PhyFold enables realistic, physics aware simulation of protein folding dynamics. This approach offers a natural extension of modern structure prediction methods and opens new directions for protein design, structure guided drug discovery, and personalized medicine. The code to reproduce the work is available at https://github.com/jamshaidwarraich/PhyFold.git. Scientific contribution PhyFold introduces a hybrid Bio-Physics Informed Neural Network framework that models protein folding kinetics as a reaction diffusion partial differential equation with environmentally modulated kinetic coefficients, integrated with ProtT5 sequence embeddings for sequence specific spatiotemporal prediction an approach not previously demonstrated in the literature. Unlike purely data driven or static structure prediction methods, PhyFold explicitly encodes temperature, pH, and pressure dependence into governing kinetic coefficients, enabling physically interpretable simulation of cooperative folding transitions across diverse physiological conditions. Dual validation against finite difference reference solutions and experimental GFP fluorescence measurements confirms mathematical consistency and biological relevance, establishing PhyFold as a computationally efficient surrogate for environment aware protein folding dynamics.
A fast pipeline based on the OPLS-AA force field is presented that enables the automated calculation of non-bonding intermolecular (dimer) interaction energies derived from a set of small organic (monomer) molecules. To calculate the non-bonding contributions, optimized geometries of the monomer molecules as well as their OPLS-AA van der Waals and atomic partial charge parameters are required. The key advantage of the new pipeline lies in its fast, highly parallelized in-memory computations without slow I/O operations, which make it possible to thoroughly process comparatively large numbers of (more than a hundred) monomer molecules within acceptable time frames (hours and days): Compared to an analogous, more versatile, and comprehensive approach, the new computational scheme delivers comparable results while being more than two orders of magnitude faster. For locally estimating the required OPLS-AA force field parameters with the LigParGen and BOSS software packages, a user-friendly graphical user interface for the Windows operating system is provided. Scientific contribution Rapid calculation of mutual intermolecular energies for a set of monomer molecules based on the widely used OPLS-AA force field enables a considerable expansion of this type of calculation for practical purposes. The performance improvement can be used to achieve significant improvements of interaction energy averages as well as considerable expansions in the size of the monomer molecule set.
In machine learning for QSAR/QSPR, the choice of train–test splitting algorithm alters both a model’s realized performance and the accuracy with which that performance is estimated from the held-out test set. Prior work has established that structure-aware splits such as Kennard–Stone and SPXY produce optimistically biased internal estimates, but these characterizations typically examine one method family at a time. Here we quantify how fifteen splitting strategies from four method families differ across fifteen drug-discovery datasets on two criteria (realized external benchmark performance and performance-estimation bias), and further assess whether splitting strategy affects model-selection quality when multiple model classes are optimized simultaneously. A single, consistent pattern holds across all fifteen datasets and all eleven evaluation measures (RMSE, MAE, MedAE, R^2 , Spearman ρ , Pearson r, Kendall τ , enrichment factor at 5/10/20
Structure-activity relationship (SAR) modeling of natural products presents a persistent methodological challenge: datasets typically contain fewer than 100 compounds, which restricts the use of data-hungry deep learning models, while conventional QSAR approaches lack mechanistic interpretability and function as black boxes. We present MK-Ensemble, a systematic four-stage optimization framework for interpretable, fragment-based SAR modeling with small-sample natural product datasets. The framework integrates multi-kernel support vector regression, hybrid molecular representations, adversarial domain adaptation, and stacking ensemble learning with strictly nested cross-validation. As a validation case study, we applied the framework to a curated dataset of 91 antioxidant compounds–comprising 24 steroidal saponins from Polygonatum cyrtonema and 67 structurally diverse reference compounds–with 128 activity records across DPPH, ABTS, and FRAP assays. We additionally performed applicability domain characterization via Williams plots and descriptor-space distance analysis, Y-randomization testing (500 permutations), and rigorous statistical model comparison using corrected resampled t-tests and Bayesian correlated t-tests. The Stacking Ensemble achieved R^2 = 0.846 (95
Hypusine modification of eIF5A by Deoxyhypusine Synthase (DHPS) is essential for the replication of a broad range of viruses, making DHPS an attractive host-directed therapy (HDT) target for antiviral drug discovery. AI-based drug design methods spanning binding affinity prediction and de novo molecular generation are rapidly proliferating, yet systematic benchmarking of diverse algorithmic paradigms under identical conditions on a shared target, with experimental validation, remains scarce. Here, we present an integrated AI benchmarking pipeline targeting DHPS and evaluate the full process from computational prediction to experimental validation. We compared 13 MM-GBSA score estimation models and 10 de novo molecular generation methods under 5-seed repeated experiments. For the prediction task, we propose an Out-of-Fold stacking ensemble integrating gradient boosting, graph neural networks, and chemical language models, achieving the highest Spearman rank correlation of 0.861 among all compared models. For the generation task, we propose RL-Design, a reinforcement learning-based framework using the ensemble predictor as a scoring oracle. RL-Design achieved the highest hit rate of 92.02
The size of synthetically accessible small molecule chemical space has rapidly grown to billions, and even trillions, of molecules, unlocking tremendous new opportunities for drug discovery. Fully leveraging these ultralarge chemical libraries, however, will strain existing computational tools, requiring novel approaches to efficiently search and process this vast space. The Chemically Reasonable Mutations (CReM) method is a powerful approach for the generation of novel and synthesizable molecules using precomputed fragment libraries. While expanding the size of these libraries could greatly enhance CReM’s ability to afford higher quality ideas, improving the scalability of the current framework is needed to take advantage of these larger libraries in practice. Here, we introduce an optimized fragment library framework, oCReM, that eliminates redundancies in the fragment database along with modifying corresponding core functions in CReM. oCReM affords a reduction of storage requirements and query times by significantly over 50
Combination therapy offers a promising strategy for cancer treatment by reducing toxicity and overcoming drug resistance. However, existing substructure-based prediction methods may struggle to effectively capture the scale-specific features of substructure interactions and often overlook cell-line-specific substructure selection, which can limit mechanistic interpretability. To address this issue, we propose HSSynergy, a hierarchical substructure-aware deep learning framework for predicting anticancer drug synergy. It first employs Graph Attention-Convolution Fusion Module to adaptively extract multi-scale substructure features from molecular graphs. Rather than indiscriminately mixing features, it introduces a Scale-Aware Masked Attention mechanism that enforces precise layer-wise alignment, and utilizes hierarchical grouping with mask constraints to achieves same-scale focusing while shielding against cross-scale noise. Furthermore, shifting away from passive cell line representations, a Cell-Active Cross-Attention mechanism models the active selection of specific substructures by heterogeneous cancer cells, capturing precise drug-cell contexts. Rigorous evaluations on two benchmark datasets show that HSSynergy achieves superior performance compared to state-of-the-art methods with robust generalization to unseen drugs and cell lines. Beyond predictive metrics, it provides mechanistic interpretability insights, revealing the hierarchical emergence of substructures and accurately pinpointing literature-validated functional groups driving synergy in specific drug combinations. Notably, several novel synergistic combinations predicted by HSSynergy are supported by existing literature and clinical evidence, suggesting its potential utility in aiding anticancer drug discovery. HSSynergy (i) Scale-Aware Masked Attention restricts substructure interactions within scale groups to reduce cross-scale noise. (ii) Cell-Active Cross-Attention models dynamic, cell-specific substructure selection instead of static cell-line fusion. (iii)Hierarchical attention links synergy predictions to pharmacologically functional groups.
The representation of molecular structure in textual form plays a central role in data-driven cheminformatics, not just for deep learning models that often rely on sequence-based inputs, but also for classic machine learning pipelines which are still relevant in this field. The Simplified Molecular Input Line Entry System (SMILES) representation and its explicit variants differ substantially, yet the implications of this difference for tokenization and downstream property prediction remain insufficiently characterized. This paper systematically investigates the role of explicitness in SMILES representation and its impact on aqueous solubility prediction. Four SMILES variations ranging from basic and canonical to explicit and explicit canonical are reviewed to illustrate how increasing explicit chemical detail alters the available information content. Building on this analysis, a novel tokenization approach termed ExACT (Explicit Atom level Context Tokenization) is introduced, which directly leverages atom level explicitness and local chemical context. In addition, a general strategy for enhancing existing tokenization methods through increased explicitness is proposed and demonstrated by upgrading the current state of the art atom-in-SMILES method to an explicit variant. All approaches are evaluated on the AqSolDB dataset using consistent machine learning pipelines and cross validation protocols. The results show that increased explicitness systematically improves predictive performance across tokenization strategies, yielding higher coefficients of determination, lower mean absolute and root mean squared errors, and reduced variance across folds. The proposed ExACT method achieves the best overall performance, largely outperforming traditional molecular fingerprint representations while remaining statistically comparable to the best competing tokenization methods. Furthermore, the efficiency analysis demonstrated that ExACT offers additional computational advantages, including lower dimensionality and the elimination of the decoding step, resulting in faster feature computation. These findings demonstrate that explicit SMILES representations encode chemically meaningful information that can be effectively exploited through context-aware tokenization, providing an interpretable and computationally efficient alternative representation for molecular property prediction. Scientific contribution This work investigates the role of explicitness in SMILES and its impact on aqueous solubility prediction, proposes a novel ExACT tokenization method and a general strategy for enhancing existing tokenization methods through increased explicitness.
Determining precise drug concentration combinations required to inhibit cancer cell growth remains a critical yet resource-intensive challenge in oncology, particularly in combination therapies where exhaustive experimental screening of dose pairs is impractical. While most existing computational approaches focus on predicting drug synergy or classifying interaction types given predefined doses, they rarely address the inverse problem of estimating the specific drug concentration pairs needed to achieve a predefined inhibitory effect. In this study, we formally define and address the problem of drug combination dose estimation for a target level of growth inhibition, with a particular focus on achieving approximately 50
Mass spectrometry-based metabolomics generates complex, high-dimensional data that holds vast potential for biological discovery but remains difficult to integrate and interpret. Knowledge graphs (KGs) unify this heterogeneous information by representing spectra, annotations, taxa, chemical classes, and biological activities as a single interoperable network; however, their practical use is limited by the steep learning curve of corresponding specialized representation and query languages. Here we introduce MetaboT, an open-source multi-agent Large Language Model (LLM) framework that translates natural-language questions into executable SPARQL queries over metabolomics knowledge graphs. MetaboT is designed to mitigate the hallucination and schema-compliance limitations of single-model approaches through a modular architecture in which specialised agents handle scope validation, entity resolution against authoritative resources, schema-aware query generation, iterative refinement, and result interpretation. We validated MetaboT on the Experimental Natural Products Knowledge Graph (ENPKG), using an expert-authored benchmark of natural-language questions paired with reference SPARQL queries, and demonstrate its ability to answer complex questions about plant–metabolite relationships and biological activities, achieving 83.67
Targeted protein degradation has emerged as a promising therapeutic strategy, yet rational degrader design remains challenged by the dynamic nature of protein of interest (POI)–E3 ligase interactions. While X-ray crystallography and cryo-EM provide valuable structural snapshots, they are insufficient for capturing the conformational heterogeneity underpinning efficient ubiquitination and degradation. Here, we present a unified computational workflow to systematically generate and evaluate POI–E3 ligase conformational states for CRBN- and VHL-mediated proteolysis-targeting chimeras (PROTACs). The workflow integrates warhead connectivity analysis, conformational clustering, ubiquitination accessibility assessment and molecular dynamics simulations to identify productive POI–E3 ligase geometries. Analysis of experimental structures revealed that PROTAC linkers do not exceed 15 Å, providing a practical attachment-atom distance based filter for docking-derived models. Furthermore, POI–E3 ligase conformations differing by more than 7.5 Å Cα RMSD exhibited distinct ubiquitination profiles, offering quantitative guidance for defining structurally and functionally divergent states. Experimental ternary complexes consistently positioned one or more solvent-exposed POI lysine residues within 50 Å of the E2 catalytic Cys111, establishing a mechanistically grounded criterion for ubiquitination competence. Validation against 34 experimentally determined PROTAC ternary complexes achieved a 97
Accurate prediction of drug–target binding affinity plays an important role in structure-based drug discovery, yet existing approaches are constrained by their reliance on scarce experimentally determined protein–ligand complex structures. In this context, recent advances in biomolecular structure prediction models, such as AlphaFold3, have emerged as a promising approach to alleviate the scarcity of experimentally determined protein–ligand complex structures. In this study, we introduce AlphaDTA, a framework that integrates AlphaFold3-predicted structures and embeddings for drug–target binding affinity prediction. AlphaDTA processes two types of AlphaFold3 embeddings, single and pair embeddings. We utilized the single embeddings at both fine-grained and coarse-grained levels to capture local interaction patterns and global binding context, and used the pair embeddings to encode cross-molecular relational features. The AlphaFold3-predicted structures are further processed by a three-dimensional structure-based geometric encoder to produce corresponding structural embedding representations. The resulting single, pair, and structural embeddings are then integrated through adaptive fusion for affinity prediction. When we evaluated AlphaDTA using recently proposed PDBbind data splits designed to reduce train–test structural overlap, AlphaDTA achieved state-of-the-art or competitive performance across multiple independent benchmarks. In a case study on the cystic fibrosis transmembrane conductance regulator, AlphaDTA correctly identifies a clinically approved potentiator and suggests a potential repurposing candidate among FDA-approved drugs. Scientific contribution AlphaDTA enables accurate drug–target binding affinity prediction without relying on experimentally determined protein–ligand complex structures. By integrating AlphaFold3-derived single, pair, and structure-based embeddings through adaptive fusion, AlphaDTA improves generalization on structurally nonredundant benchmarks and demonstrates utility for target-specific drug repurposing.
Cystathionine β-synthase (CBS) has emerged as an important therapeutic target implicated in cancer and Down syndrome, yet the discovery of selective CBS inhibitors remains challenging due to limited structural diversity of known ligands and the scarcity of target-focused virtual screening (VS) benchmarks. In this study, we present the first comprehensive evaluation of CBS-specific artificial intelligence (AI) models for structure-based VS, supported by a carefully curated and up-to-date data set of experimentally validated CBS inhibitors, true inactives and decoys. Using this data set, we developed CBS-specific binary classification models trained on docking-derived features and evaluated them in a rigorous five-fold cross-validation framework that employed similarity-controlled splits, ensuring structural independence between training and test sets while minimizing intra-fold class bias. Predictive performance was assessed primarily using the normalized enrichment factor of true actives at 1
