Multi-Omics and AI Integration Reshapes Precision Drug Discovery: From Target Identification to Clinical Translation
核心洞察
The systematic integration of multiomics data with artificial intelligence is transforming drug R&D, enabling holistic systems biology interrogation and therapeutic target identification beyond traditional single-omics approaches.
AI-driven multiomics frameworks accelerate target discovery across oncology, neurology, and cardiovascular diseases, with CRISPR-based functional genomics and single-cell technologies revealing clinically actionable vulnerabilities such as EGFR (搜索), KRAS (搜索), and BCL2 (搜索).
Generative AI models like GENTRL have demonstrated the ability to design potent small-molecule inhibitors in as little as 21 days, while AI-powered drug repurposing platforms have identified baricitinib and other candidates for Alzheimer's disease (搜索).
The convergence of artificial intelligence and multi-omics analysis is fundamentally reshaping the pharmaceutical R&D landscape, moving drug discovery from traditional trial-and-error approaches toward rational, computationally driven design. Multi-omics technologies—including genomics, transcriptomics, proteomics, metabolomics, and single-cell or spatial omics—have generated unprecedented volumes of complex biological data. However, efficiently mining these data to uncover actionable insights for drug design, mechanism elucidation, and safety assessment remains a significant challenge that AI is uniquely positioned to address.
Multi-Omics Integration for Target Discovery
The rapid advancement of multiomics profiling technologies over the past decade has profoundly redefined the methodological paradigm for therapeutic target discovery. By systematically integrating orthogonal molecular layers—encompassing whole-genome variation, single-cell transcriptomics, posttranslational proteomics, dynamic metabolomics, and epigenomic regulatory signatures—researchers can now resolve disease-altered interactomes with unprecedented granularity. This integrative systems biology strategy enables computational reconstruction of pathological signaling cascades, identification of master regulatory nodes via network centrality metrics, and quantification of target tractability through assessments of druggability and functional essentiality.
CRISPR-based gene-editing platforms have further catalyzed this transition. The application of CRISPR-Cas9 screening to dissect drug resistance in non-small cell lung carcinoma (搜索) has elucidated core functional dependencies within oncogenic networks, implicating targets such as EGFR (搜索), KRAS (搜索), TP53 (搜索), and antiapoptotic regulators such as BCL2 (搜索) as clinically actionable vulnerabilities. In a landmark study, Pacini et al. developed the second-generation cancer dependency map (DepMap 2.0) by concurrently profiling the genomic landscape, transcriptome, proteome, and clinical annotations of 930 cancer cell lines.
In glioblastoma (搜索) research, functional CRISPR ablation of carboxypeptidase D significantly attenuated IGF1R-mediated PI3K/AKT/mTOR pathway activation, suppressed tumor-initiating cell expansion, and impaired xenograft tumorigenicity—establishing this protease as both a mechanistic effector of oncogenic signaling and a promising therapeutic target in solid tumors.
AI-Driven Drug Design and Molecular Generation
AI has emerged as a pivotal enabler across the drug development lifecycle. Ensemble algorithms such as random forest and XGBoost support robust feature selection, while deep learning models including autoencoders and transformers uncover latent structures across omics layers, revealing nonlinear dependencies and emergent properties often inaccessible through reductionist methodologies.
In a striking demonstration of AI's potential, Ren et al. employed predictive AI methods to identify TRAF2- and NCK-interacting kinase (TNIK (搜索)) as a promising antifibrotic target for idiopathic pulmonary fibrosis (搜索). This approach led to the rapid development of INS018_055, a small-molecule TNIK inhibitor with favorable drug-like properties and broad antifibrotic activity, achieved in just 18 months.
Generative models have further expanded the frontier. GENTRL (generative tensorial reinforcement learning) was used to identify a potent inhibitor of discoidin domain receptor 1 (DDR1 (搜索)), a kinase involved in fibrosis and other diseases, within 21 days. Several compounds showed activity in biochemical assays, with one leading candidate exhibiting favorable pharmacokinetics in mice. Additionally, DSP-0038—an AI-designed dual-target 5-HT1A agonist and 5-HT2A antagonist developed by Exscientia—has entered phase I trials as a potential treatment for Alzheimer's disease (搜索)-related psychosis.
Drug Repurposing Through Multi-Omics and AI
Multiomics technology has revolutionized drug repurposing by enabling systematic identification of hidden drug-target-pathway associations. Li et al. developed a computational drug repositioning method based on proteomic and transcriptomic profiles, identifying voltage-gated sodium channel blockers and monoamine oxidase inhibitors as promising candidates for Alzheimer's disease (搜索) treatment.
In the context of COVID-19, transcriptomic screening revealed abnormal activation of the IL-6/JAK-STAT pathway, while protein interaction network predictions suggested the efficacy of anti-inflammatory drugs such as tocilizumab and the JAK inhibitor baricitinib—predictions that were swiftly validated in clinical trials.
The machine-learning framework DRIAD (Drug Repurposing in AD) was developed to quantify potential connections between AD-related biological processes and integrated genetic datasets. DRIAD prioritized baricitinib as a leading Alzheimer's candidate, which is now being evaluated in an open-label, biomarker-driven basket trial that includes patients with both AD and amyotrophic lateral sclerosis (ClinicalTrials.gov: NCT05189106).
Analysis of real-world insurance claims from 7.2 million patients in the IBM MarketScan Medicare Supplemental Database revealed that two FDA-approved p300/CBP inhibitors—salsalate and diflunisal—are associated with a reduced incidence of Alzheimer's disease (搜索), with neuroprotective efficacy validated in mouse models.
Predicting Drug Safety and Interactions
AI has demonstrated substantial capabilities in predicting drug safety across preclinical, clinical, and postmarketing phases. Mamoshina et al. developed a model capable of predicting cardiotoxicity by analyzing drug properties from publicly available datasets, achieving an area under the curve (AUC) of 79% for validation data and 66% for unseen data. AI-based methodologies have also been applied to predict drug-induced liver injury, achieving a classification accuracy of 89%.
For drug-drug interaction prediction, Wang et al. developed high-performance predictive models using machine learning methods such as random forest and XGBoost, identifying 54,013 possible drug pairs among FDA-approved drugs that may exhibit DDIs. Joshi designed a customized deep neural network called KGDNN for adverse drug reaction prediction, achieving an AUROC of 0.917.
Challenges and Future Directions
Despite these advancements, several methodological challenges remain unresolved. The integration of cross-omics data requires unified algorithmic frameworks to address issues related to data heterogeneity and batch effects. Complex deep learning models often function as black-box systems with opaque decision-making processes, hindering mechanistic interpretation of predictions—a critical limitation in pharmaceutical contexts where understanding the rationale behind models is essential for analyzing mechanism of action and safety profiles.
The field faces what researchers describe as three critical paradoxes: the data heterogeneity dilemma, where disparities in spatiotemporal resolution between single-cell transcriptomics and population pharmacokinetic data introduce biases; the interpretability gap, where explaining attention-based mechanisms in alignment with traditional pharmacological concepts proves challenging; and the dynamic system modeling deficiency, where current models predominantly focus on static binding affinities while overlooking time-varying drug concentration curves.
Future breakthroughs will likely hinge on the seamless integration of cross-scale simulation infrastructures—spanning from molecular dynamics to digital twin patients—alongside federated learning paradigms, to establish a closed-loop ecosystem for drug prediction, validation, and optimization. As the field progresses, the collaboration between AI and multiomics is poised to transform our understanding of complex diseases and enhance patient care through more personalized approaches.