Shiliang Li

Unfavorable drug metabolism drives clinical failure, limited by the low accuracy of existing AI predictors. To address this, we constructed a database of 11,665 human-specific reactions and developed Mettle, a novel AI model integrating chemical feature interaction with contrastive learning. By explicitly training the model to distinguish true metabolic transformations from structurally similar decoys, Mettle achieves state-of-the-art performance with ∼80% top-5 accuracy. We demonstrate Mettle’s utility by tackling poor oral bioavailability in RSK4 inhibitors. This metabolite-aware design strategy yielded R636, which maintains high potency while exhibiting a remarkable 63-fold increase in absolute bioavailability (to 63%). R636 showed a favorable safety profile and significant antitumor efficacy in two ESCC PDX models. Mettle thus emerges as a powerful tool for metabolite-aware oral drug design.

Despite breakthroughs in predicting acute organ toxicities, advanced computational models continue to struggle with complex, insidious end points like drug-induced bone toxicity. A fundamental limitation of most supervised pipelines is their failure to explicitly model intermolecular similarities during task-specific training. Neglecting this macroscopic topology of the chemical space leads to fragmented latent representations, poor generalization, and high false negative rates. To address this, we propose MacroTox, a macroscopic topology-driven multimodal deep learning framework. Beyond intramolecular fusion, its core innovation is the dynamic construction of an intrabatch drug–drug similarity graph. Guided by a topology-aware synergistic optimization, this mechanism captures latent network correlations, maximizing the utilization of the chemical space to resolve representational fragmentation and enhance the predictive capacity under data scarcity. Evaluated on a rigorously curated bone toxicity data set, MacroTox achieved an area under the receiver operating characteristic curve of 0.93 and a Matthews correlation coefficient of 0.73. Crucially, it effectively balances the sensitivity-specificity trade-off (SEN: 0.88, SPE: 0.86), substantially mitigating the underreporting risks in bone toxicity screening. Notably, ablation studies confirm that these predictive enhancements stem from the synergistic effect of dynamic graph topology and edge loss regularization. Extensive benchmarking of MoleculeNet further confirms its robust transferability. Furthermore, via multilevel feature attribution, MacroTox elucidates chemically intuitive structure–activity relationships. By pinpointing specific toxicophores and resolving complex activity cliffs, the framework proves that it captures authentic toxicological mechanisms rather than memorizing superficial data set biases. Ultimately, MacroTox offers an reliable, generalizable, and interpretable virtual screening engine for early stage drug discovery.

RNA-based technologies have demonstrated significant potential for diverse applications, ranging from vaccination to gene editing. However, their widespread adoption is limited by the critical challenge of efficient delivery. Lipid nanoparticles (LNPs) have emerged as a widely utilized RNA delivery system, yet their formulation design and optimization primarily rely on empirical trial-and-error, which is labor-intensive, time-consuming, and cost-prohibitive, thus hindering the rapid development of RNA therapeutics. To facilitate the early-stage design and optimization of LNPs for enhanced delivery efficiency, in this study, we construct LNPs-TE, a benchmark dataset comprising over 10 000 experimentally measured transfection efficiency (TE) values, and introduce LNPs integrated feature fusion Transformer (LIFT), a deep learning framework for LNPs TE prediction. Comprehensive experiments demonstrate that LIFT effectively integrates multidimensional molecular representations of ionizable lipids, the key component in LNPs formulation, achieving superior predictive performance, with an average Pearson correlation coefficient of 0.845 for regression and an area under the receiver operating characteristic curve (AUC-ROC) of 0.818 for multi-class classification across multiple datasets. Through scaffold-based splitting and activity cliff tasks, we further validated the exceptional generalization ability and robustness of LIFT, which achieved over a 10% improvement in the coefficient of determination (R²) compared with state-of-the-art baseline models, highlighting its potential as a practical and stable approach for the virtual screening of efficient LNPs formulation. The relevant data, model and code are made publicly available at https://github.com/U12458/LIFT.

Proteins are crucial in biological processes and are important substances that mediate biochemical reactions, regulate cellular processes, and facilitate drug binding through their active sites and surface residues. The pK_a values of proteins determine the protonation state of ionizable amino acids under specific pH conditions, profoundly impacting protein structure, function, and drug design. However, experimental determination of pK_a values is normally laborious and complex. Moreover, existing prediction methods are limited by the data quantity and quality, as well as their inability to address the intricate structural and physicochemical attributes of proteins, thereby hindering accuracy and generalization, especially in predicting pK_a values for buried residues. In this study, we developed a multimodal protein pK_a prediction model named ME-pK_a (Multimodal ESM pK_a), which leverages the multimodal information and employs a multifidelity learning strategy to speedily and accurately predict molecular pK_a values. The ME-pK_a method facilitates data augmentation by integrating the local environmental attributes of amino acids with the FASTA sequence characteristics of proteins. Furthermore, the incorporation of multifidelity learning has addressed the challenge of limited data availability to some extent. Our ME-pK_a model outperforms several state-of-the-art models in predicting protein pK_a values, achieving impressive results with a low RMSE of 0.845 ± 0.09 and MAE of 0.641 ± 0.07, a high R² of 0.921 ± 0.02, and R of 0.959 ± 0.01 on the PE-pK_a data set. Notably, ME-pK_a demonstrated balanced and robust performance across major ionizable residue types (ASP, GLU, HIS, LYS). It demonstrates superior accuracy in predicting pK_a values for buried residues (RSA < 0.2), achieving the lowest MAE values of 0.921 ± 0.05 on the PE-pK_a data set and 0.911 ± 0.06 on the Small Set data set, which collectively excel in capturing complex environmental influences on pK_a. Moreover, our method confirmed pH-dependent binding of PD-L1 antibodies mediated by the protonation state of His-69 in PD-L1, emphasizing the critical role of amino acid protonation states in drug design. The source code of ME-pK_a can be found at https://github.com/yzjyg215/ME-pKa.

The transient receptor potential vanilloid 4 (TRPV4) channel, a polymodal calcium-permeable cation channel, is activated by diverse stimuli, including hyposmolarity, moderate heat, mechanical stress, and acidic pH. TRPV4 plays a critical role in the onset and progression of pain and is implicated in multiple pathological conditions, positioning it as an attractive therapeutic target. However, the development of TRPV4-targeted therapeutics has been hindered by limited mechanistic understanding of its inhibition. Here, the natural product (R)-1-(3-ethylphenyl)ethane-1,2-diol (AH001) is identified as a potent, subtype-selective TRPV4 inhibitor. A potential gating mechanism of TRPV4 inhibition is proposed through a combination of structural analyses and site-directed mutagenesis. Electrophysiological studies and cryo-electron microscopy (cryo-EM) further reveal that AH001's glucuronide metabolite, AHP, retains robust TRPV4 inhibitory activity, providing a critical structural basis for rational optimization of AH001 as a lead compound. Notably, AH001 demonstrated significant efficacy in reducing neuronal hyperexcitability in dorsal root ganglia (DRGs) and produced pronounced analgesic effects across multiple pain models, highlighting its therapeutic potential in pain relief. This study offers critical structural and mechanistic insights into TRPV4 inhibition and lays the groundwork for the structure-based design of TRPV4-targeted therapeutics.

Biomedical knowledge graphs (KGs) provide a structured and traceable framework for target-oriented, AI-assisted drug discovery. By linking genes, proteins, diseases, phenotypes, compounds, pathways, assays, structures, adverse events, and clinical evidence, biomedical KGs can partially mitigate evidence fragmentation and support hypothesis generation across molecular, cellular, and clinical levels. This review summarizes how biomedical KGs are constructed, reasoned over, evaluated, and translated into practical drug-discovery tasks. We first outline major data resources, entity and relation extraction, knowledge fusion, quality control, and large language model-assisted KG construction, with emphasis on ontology grounding, evidence tracing and hallucination control. We then compare key KG reasoning paradigms, including symbolic rule-based reasoning, representation learning, path- and subgraph-aware reasoning, graph Transformer models, and KG-LLM hybrid reasoning. Their strengths and limitations are discussed in relation to interpretability, scalability, leakage risk, benchmark bias, and prospective validation. We further highlight the need to incorporate three-dimensional structural evidence, including protein structures, binding pockets, variants, conformational states, and ligand compatibility, so that graph-based relevance can be connected with chemical tractability. Finally, we discuss applications of biomedical KGs in target prioritization, druggability assessment, mechanism interpretation, drug-target interaction prediction, molecular generation, lead optimization, ADMET and safety assessment, drug repurposing, combination therapy, biomarker discovery, precision medicine, and drug-resistance modeling. Current limitations include heterogeneous data quality, incomplete causal evidence, static graph representations, privacy constraints, and insufficient experimental or clinical validation. Overall, biomedical KGs should be viewed as evidence-integration and prioritization engines rather than automatic proof-generating systems for drug discovery, guiding rigorous, iterative, and experimentally grounded translational decisions.