搜索结果：npj drug discovery

Drug-drug interaction (DDI) poses a major challenge in clinical pharmacology, often compromising therapeutic efficacy or causing serious adverse events. Traditional detection methods, heavily dependent on experimental assays and expert knowledge, are constrained by high costs and limited scalability. This work explores emerging machine learning (ML)-based strategies for predicting DDIs by leveraging the rapidly expanding biomedical data landscape. Recent advances in deep learning architectures, graph neural networks and sophisticated feature engineering have markedly improved predictive performance, offering scalable and data-efficient alternatives to conventional approaches. We further highlight real-world clinical applications where ML-based models have enhanced drug safety monitoring and informed therapeutic decision-making. Finally, we discuss critical challenges like model interpretability, generalizability and integration with clinical workflows, and outline future directions toward building robust, explainable and clinically actionable DDI prediction systems. This work provides a comprehensive perspective on how AI-driven methodologies are reshaping pharmacovigilance and precision therapeutics.

Artificial intelligence (AI) is being used in oncological drug development to address the high costs, low success rates, and long timelines that characterize traditional drug development pipelines. The use of machine learning (ML) and deep learning (DL) models in computer-aided drug design is constantly growing owing to their capacity to analyze large, heterogeneous datasets, their ability to capture nonlinear biological trends, and their integration of various molecular and clinical characteristics. AI applications accelerate target discovery by predicting protein structures, ranking disease-relevant genes, and assessing target drugability. AI can be used to conduct rapid searches of multiplexed chemical libraries, predict drug-target interactions, and optimize the pharmacological and physicochemical properties of drugs in virtual screening. Advanced neural network designs also aid in de novo drug design, which involves developing new molecular structures with therapeutic properties of interest. This review outlines how AI has been used for target identification, virtual screening, de novo molecular design, and, specifically, in cancer applications. It further discusses the major issues in AI-based drug development, such as data quality, model interpretation, computational constraints, and ethical and regulatory considerations, which remain essential obstacles to broader clinical translation.

Phenotypic drug discovery (PDD) identifies new drugs by observing the effects of compounds on living systems without prior knowledge of their targets. Advances in biological data and machine learning have made PDD more systematic and data-driven. This review outlines a computational framework, including phenotype representations, key tools, and public datasets. It also discusses major challenges and strategies to improve PDD's efficiency and translational potential, offering a practical guide for researchers in the field.

Rhinoviruses (RV) comprise three species, RV-A, RV-B, and RV-C, with approximately 170 types. RV-C is associated with severe respiratory illness, particularly in children and individuals with asthma or chronic obstructive pulmonary disease, underscoring the need for effective antiviral strategies. Progress in RV-C research and drug discovery has been limited by the lack of robust, scalable cell-based infection models that recapitulate the complete RV-C replication cycle. Here, we describe a high-content imaging (HCI)-based high-throughput infection system for RV-C. Rather than relying solely on receptor overexpression, we used a genetically stable fluorescent reporter virus (RV-C15a-mGL) to screen ~300 monoclonal cell lines expressing the RV-C receptor variant CDHR3-Tyr529. This approach identified a clone that efficiently supports RV-C replication and revealed that productive infection depends on determinants beyond receptor abundance alone. Using this clone, we established and validated a robust, scalable screening platform with Z' > 0.75 in both 96- and 384-well formats. The system was readily adapted to additional RV-C types (C11 and C41), as well as RV-A and RV-B. A pilot screen of approximately 10,000 small molecules identified both known and novel RV-C inhibitors, supporting the utility of this platform for antiviral discovery and for advancing the study of RV-C biology.

Hepatocellular carcinoma (HCC) remains a major global health challenge due to its molecular heterogeneity, late diagnosis, and limited therapeutic options. Recent studies have identified isonicotinylation (Kinic), a novel lysine acylation, as a regulatory modification influencing carcinogenic protein activity and liver cancer progression. In this study, we established the Kinic Index (KinicI), an artificial intelligence (AI)-driven predictive model that integrates multi-omics data and consensus clustering to classify HCC patients into two distinct Kinic subgroups. Patients in the high-Kinic subgroup exhibited significantly worse overall survival, demonstrating the value of KinicI for risk stratification and outcome prediction. Machine learning approaches (LASSO, RSF) coupled with Shapley additive explanation (SHAP) analysis identified CYP2C9 and G6PD as the most influential prognostic variables associated with HCC progression. Single-cell and spatial transcriptomic analyses confirmed that CYP2C9 and G6PD are primarily localized in malignant hepatocytes with high metastatic potential, underscoring their clinical relevance. Importantly, using the GraphBAN deep learning framework and ADMET-AI screening, we prioritized candidate compounds targeting CYP2C9 and G6PD, followed by molecular docking that validated strong binding affinities, suggesting their potential as novel therapeutics. Together, our study demonstrates that KinicI is a powerful AI-enabled platform for prognostic modeling, molecular stratification, and multitarget drug discovery, providing a foundation for precision oncology and resistance-aware treatment strategies in HCC patients.

The RNA cytosine-5 methyltransferase NSUN2 is an emerging therapeutic target in precision oncology, with aberrant overexpression driving tumor progression, metastasis, and therapy resistance across multiple malignancies. Despite its critical role in cancer biology, selective small-molecule inhibitors remain limited. We employed an AI-accelerated workflow to screen approximately 101 million compounds from the ZINC database using structure-based virtual screening. The AlphaFold2-predicted human NSUN2 structure was aligned with the experimentally determined M. jannaschii TRM4 homolog (PDB: 3A4T, 34.2% sequence identity, 1.82 Å RMSD). A CatBoost ensemble classifier trained on Morgan fingerprint descriptors with AutoDock Vina-derived labels achieved robust performance (training: recall 0.87, ROC-AUC 0.89; test: recall 0.71, ROC-AUC 0.85), with low test precision reflecting extreme class imbalance inherent to virtual screening. Multi-stage filtering identified 12,000 high-scoring compounds with binding affinities of -9.933 to -8.375 kcal/mol. ADMET profiling yielded 34 drug-like candidates with favorable pharmacokinetic and toxicological profiles. Molecular dynamics simulations over 50 nanoseconds validated binding stability of lead compounds ZINC-1000507789 and ZINC-1000507824. These structurally diverse non-covalent reversible inhibitors targeting the SAM cofactor binding pocket warrant experimental validation through biochemical assays and cellular studies to overcome therapeutic resistance in NSUN2-driven malignancies.

Structure-based drug design is rapidly evolving, driven by advances in both physics-based and knowledge-based methods. These computational approaches are increasingly integrated across all stages of drug discovery. Despite remarkable progress, challenges remain in achieving accuracy, generalizability, computational efficiency, and chemical synthesizability. In this review, we provide a critical overview of advances, strengths, and limitations of recent methods. We also discuss synergies between the two concepts that hold promises for future advancements towards their practical applicability.

Genetics-informed drug repurposing holds promise in translating genetic findings into therapeutics. In this study, we developed a Genetics Informed Network-based Drug Repurposing via in silico Perturbation (GIN-DRIP) framework and applied the framework to repurpose drugs for type-2 diabetes (T2D). In GIN-DRIP for T2D, it integrates multi-level omics data to translate T2D Genome-Wide Association Study (GWAS) signals into a genetics-informed network that simultaneously encodes gene importance scores and a directional effect (up/down) of risk genes for T2D; it then bases on the GIN to perform signature matching with drug perturbation experiments to identify drugs that can counteract the effect of T2D risk alleles. We identified telmisartan that belongs to the Angiotensin Receptor Blockers (ARB) drug category as a candidate, and validated the ARB antihypertensive drugs' potential therapeutic effects on T2D in Vanderbilt University Medical Center (VUMC)'s Electronic Health Records (EHR) data with over 3 million patients.

Precision oncology faces critical challenges in interpreting complex cellular signals and predicting drug responses across heterogeneous cancer environments. Here, we present BioGDR, a multimodal interpretable deep learning framework that integrates structure-based predicted biological features, including differential gene expression and kinase inhibition profiles, eliminating the need for experimental measurements. By modeling tumor transcriptomic states through pathway-informed graph neural networks and employing a drug-guided attention strategy, BioGDR enables mechanistic insights into drug sensitivity across compound and cellular contexts. Comprehensive evaluations demonstrate that BioGDR outperforms existing methods in compound screening relevant to early-stage drug discovery and in predicting cell line sensitivity across heterogeneous cellular states characteristic of precision oncology, while analyses on clinical patient cohorts further confirm its practical utility and generalization capability. Experimental validation with a novel ALDH1B1 inhibitor confirms its ability to identify sensitive cell populations and reveal underlying mechanisms. This work establishes a robust, biologically informed framework that bridges preclinical drug development and clinical applications, advancing precision oncology through integrative, multimodal learning and interpretable mechanism analysis.

The XII-ZOMES Conference (Shenzhen, 2024) showcased advances in PCI complexes—proteasome, COP9 signalosome, and eIF3—which regulate cellular proteostasis through conserved architectures. Presentations covered translation initiation, proteasome assembly, ubiquitin and ubiquitin-like modification regulation, protein quality control mechanisms, and emerging therapeutic technologies including PROTACs and molecular glues for targeted protein degradation.

To investigate whether antidiabetic drugs have a biological basis to be repurposed in PD prevention, we applied a drug target Mendelian randomization framework to assess associations between genetic variation in antidiabetic drug targets and PD risk or age at onset (AAO). Instrumental variables (IVs) were derived from GWAS summary statistics on fasting glucose (FG), glycated hemoglobin (HbA1c), and gene expression data from GTEx. Apart from SGLT2 inhibitors, all other antidiabetic drugs of interest could be instrumented through our methods. Positive and negative control analyses were carried out to validate 20 IVs in the FG arm and 23 IVs in the HbA1c arm. DPP-4 inhibitors failed the positive control. GWAS summary statistics for PD risk and AAO data were sourced from the IPDGC and COURAGE-PD consortia, resulting in 42 083 cases/457 090 controls for risk and 37 103 PD cases for AAO. MR analyses showed no significant associations across consortia or in meta-analysis. These findings do not support a causal role of genetic variation in antidiabetic drug targets in PD risk or AAO.

暂无摘要（点击查看详情）

A multi-drug resistant emm92-type strain of group A Streptococcus (GAS) has emerged as an important causative agent of invasive infections-particularly affecting people who inject drugs-in the United States. To curtail this developing threat, we aimed to identify and repurpose FDA-investigated compounds as antimicrobials. To identify growth-inhibiting compounds, a machine learning-based model was trained on the emm92-iGAS growth response to 2560 bioactive compounds. The model was used to screen a 6111 FDA-evaluated drug library in silico. Of the 9 validated compounds, GW3965 experimentally exhibited a 99% reduction in iGAS survival at an MIC of 6.25 µM. Treatment with GW3965 aided complete wound closure in a human skin equivalent model, and decreased lesion size and reduced bacterial burden in a mouse model of skin and soft tissue infection. Application of a machine learning model expedited the discovery of GW3965 as a therapeutic for iGAS skin and soft tissue infections.

Antimicrobial resistance (AMR) poses a major global health threat that demands the discovery of new antimicrobial agents. Antimicrobial peptides (AMPs) offer a promising therapeutic alternative due to their broad-spectrum activity and reduced likelihood of resistance development. In the current study, we developed COMPASS, a comprehensive database aggregating 75,381 unique AMP sequences from nine public repositories, and created AmpGPT2, a transformer-based generative model specifically fine-tuned for AMP sequence generation. Unlike directed approaches, which optimize antimicrobial sequences or certain properties, our foundational model learns general AMP sequence patterns through an undirected training strategy. AmpGPT2 generated peptide sequences, of which 95.41% were predicted to be AMPs by AMP Scanner, representing a substantial improvement over existing models. The generated peptides exhibit physicochemical properties consistent with natural AMPs, including appropriate length distributions and molecular characteristics. Experimental validation demonstrated that one of five tested peptides, which shares structural features with dermaseptin-family AMPs, exhibited significant concentration-dependent antimicrobial activity against Klebsiella pneumoniae and Pseudomonas aeruginosa, supporting the model's potential for functional AMP discovery. Highlighting the persistent challenge of translating computational predictions into biological function, this work establishes a foundational framework for AMP discovery that can serve as a basis for subsequent directed optimization strategies, potentially accelerating the development of novel antimicrobial therapeutics.

The complexity of disease-causing signaling networks is indicative of the failure of single-target therapeutics to work, particularly because of feedback, redundancy and activation of compensatory responses. The review describes the recent movement to network pharmacology and purposeful polypharmacology facilitated by the emergence of artificial intelligence (AI) and massive biological knowledge graphs. This review explains how machine learning and graph neural networks can be used to characterize molecular interactions systematically, predict targets that are of disease relevance, as well as priorities on multi-target intervention strategies. Generative models and reinforcement-based learning strategies are addressed to create compounds and combinations of drugs designed to modulate networks, and not individual protein inhibition. It describes the experimental validation processes, such as CETSA, NanoBRET, and Perturb-seq, and patient-derived models and MIDD systems to aid the translational evidence. Data quality, bias, interpretability, and reproducibility are taken into consideration. In sum, this review presents a feasible and combined model of AI-assisted network-mediated drug discovery.

Lactylation, a recently identified histone modification derived from lactate metabolism, has emerged as a critical regulator of epigenetic reprogramming, tumor proliferation, and immune evasion. In ovarian cancer, lactate dehydrogenase A (LDHA) and other metabolic enzymes contribute to lactate accumulation, which supports chemotherapy resistance and disease progression. Although lactylation is increasingly linked to therapy failure, its precise molecular connection with ovarian cancer, as well as its therapeutic potential are unclear. Traditional analytical approaches often fail to integrate the complexity of multi-omics, limiting the discovery of actionable lactylation-associated vulnerabilities. This research aims to develop an AI-driven multi-omics framework to identify lactylation-related genes, stratify patient drug responses, and establish prognostic signatures in ovarian cancer. Transcriptomic, epigenomic, pharmacogenomic, mutation, and clinical outcome data were collected from The Cancer Genome Atlas (TCGA), the Genomics of Drug Sensitivity in Cancer (GDSC), and independent ovarian cancer cohorts. Deep learning models, including variational autoencoders (VAEs), Long Short-Term Memory (LSTM) networks, and Multitask Multilayer Perceptrons (MLPs) (LSTM-MLP), were applied for molecular subtyping, survival analysis, and IC50 prediction. Findings were validated through pathway enrichment, mutation mapping, immune infiltration profiling, and structure-guided drug repurposing, the proposed method achieved precision of (0.955). Key lactylation-related genes, including LDHA and SLC16A3, were associated with immune exhaustion and cisplatin resistance. The Gln-TEx score and lactylation risk signature robustly predicted patient survival and drug response across TCGA and validation cohorts. Perturbation sensitivity and repurposing analyses revealed novel therapeutic vulnerabilities. This study establishes a precision oncology framework that integrates lactylation biology with AI-driven analytics to uncover druggable targets, enhance patient stratification, and inform the design of multi-target therapies in ovarian cancer.

Breast cancer's genomic heterogeneity complicates drug discovery, making repurposing an attractive but challenging strategy. Advances in artificial intelligence now enable integration of multi-omics data to reveal drug-gene-disease relationships and generate subtype-specific repurposing hypotheses. In this Review, we examine AI-driven computational approaches from signature-based to multi-modal frameworks and propose an integrated interpretability-driven framework linking mechanistic validation with clinical translation toward more transparent and actionable precision oncology.

Evidence on whether menopausal hormone therapy (MHT) affects neurological or psychiatric disease is conflicting. As MHT acts by binding to oestrogen receptors (ERα and ERβ), we used drug-target Mendelian randomisation (MR) to test whether perturbing these targets alters the risk of Alzheimer's disease (AD), brain structure, depression, or anxiety. Genetic variants in the genes encoding these oestrogen receptors (ESR1 and ESR2) that were associated with positive controls were leveraged as instrumental variables. In two-sample MR analyses using large genome-wide association studies, genetically proxied ERα and ERβ perturbation showed no evidence of effect on AD or on cortical grey matter, hippocampal volume, or white matter hyperintensities. Genetically proxied ERβ perturbation significantly increased risk for depression (β = -0.66, 95% CI [-0.99, -0.32], p = 0.002), but not anxiety. Our study highlights psychiatric considerations when targeting oestrogen receptors with MHT, but provides no evidence for either harmful or protective effects on AD risk.

Bone metastasis is a major cause of morbidity and mortality in breast cancer, yet effective prognostic models and targeted therapies remain limited. Here, a machine learning (ML)-driven multi-omics framework integrating epithelial-mesenchymal transition (EMT) and nucleotide metabolism (NM) signatures is presented to uncover prognostic biomarkers and guide rational drug discovery. Using gene expression omnibus (GEO) and the cancer genome atlas-breast invasive carcinoma (TCGA-BRCA) bone metastasis datasets, applied the least absolute shrinkage and selection operator (LASSO) ML to identify NM-associated hub genes, revealing peroxiredoxin 4 (PRDX4) as a key risk-associated gene. Multi-level analyses demonstrated that PRDX4 expression correlates with immune cell infiltration, microsatellite instability (MSI), tumor mutational burden (TMB), EMT activation, and poor overall survival. Consensus clustering stratified patients into distinct EMT-NM molecular subgroups with divergent clinical outcomes, immune checkpoint expression, and tumor stemness scores, providing a foundation for precision patient stratification. To accelerate translational impact, we performed drug repurposing and molecular docking, identifying Docetaxel as a high-affinity PRDX4-targeting compound with favorable binding energetics. Together, this work demonstrates how ML-driven multi-omics analysis can bridge biomarker discovery and drug design, guiding multitarget and multi-drug strategies to improve outcomes in bone metastatic breast cancer.