AI-Based Modeling and Drug Development

National Center For Advancing Translational Sciences

Investigators

Abstract

Structure-based virtual screening for ADCK3 inhibitors. ADCK3, an atypical protein kinase of the UbiB family, is crucial for coenzyme Q10 (CoQ10) biosynthesis, a molecule vital for ATP production and antioxidant defense. However, ADCK3 mutations lead to CoQ10 deficiency that are linked to severe neurological, muscular, and renal disorders like encephalomyopathy and cerebellar ataxia, yet its precise mechanisms remain unclear. To elucidate the underline mechanisms and identify potential therapeutic leads, we undertook structure-based virtual screening for ADCK3 inhibitors, beginning with an extensive structural analysis of ADCK3's active site to generate a pharmacophore model. This model facilitated a quick screening of a large virtual compound library, yielding 800 initial candidates. Subsequently, 129 confirmed ADCK3 inhibitors were found through confirmation assays, including 6 compounds exhibiting selectivity against p38 kinase. Further molecular dynamics simulations predicted the binding modes of the most potent compounds, and metadynamics analysis pinpointed key amino acid residues involved in intermolecular interactions. Our results are detailed in the research paper (Gao P. et al. 2024. J Chem Inf Model. 64, 6072-80. PMID: 39025788) that provides a strong foundation for future ADCK3-targeted drug development. AI-based lead compound discovery for Glioblastoma (GBM). GBM is a rare brain cancer with a devastatingly high mortality rate, highlighting the urgent need for effective new therapeutics. Despite significant research efforts, the underlying biological mechanisms of GBM remain unclear, and there is currently no targeted therapy for this disease. We developed a novel computational pipeline that leverages gene expression data analysis for GBM and conducts virtual compound screening for drug development. By constructing the GBM Gene Expression Profile (GGEP) using multi-omics data, we identified drugs that could reverse gene expression patterns from the Integrated Network-Based Cellular Signatures (iLINCS) database. Candidates were prioritized through hierarchical clustering of their expression signatures and quantification of their reversal strength using two self-defined indices based on the log2 fold change (LFC) of GGEP genes induced by the drug candidates. Among the five prioritized candidate compounds identified computationally, Clofarabine and Ciclopirox were validated through in vitro experiments as highly effective in selectively targeting GBM cancer cells. This study demonstrates a promising approach to accelerating drug development by uncovering the underlying gene expression effects between drugs and diseases, applicable to both rare and common diseases. (Sun S. Et al. 2025. J Transl Med 23: 25. PMID: 39773231). Integrative analysis of large-scale RNA-Seq and proteomics profiling data for cancer drug discovery. The pathogenesis of cancer is complex, with different types often exhibiting distinct gene mutations, resulting in varied omics profiles. We conducted a systematic analysis of large-scale RNA-Seq and proteomics profiling data from 16 human cancer types to identify cancer-specific biological pathways and potential cancer-targeting drugs. Statistical approaches were employed to identify significant molecular targets and pathways associated with each type of cancer. Subsequently, potential anti-cancer drugs that can target these pathways were identified. The number of significant pathways identified for each cancer type ranged from four (stomach cancer) to 112 (acute myeloid leukemia), and the number of therapeutic drugs targeting these cancer-related pathways ranged from one (ovarian cancer) to 97 (acute myeloid leukemia and non-small-cell lung carcinoma). Notably, some of these drugs are FDA-approved therapies for their corresponding cancer types, that validates our method. Our findings offer a rich source of testable hypotheses to unravel the complex mechanisms underlying human cancers and provide a basis for prioritizing and repurposing drugs as anti-cancer therapies (Xu T. Et al. 2025. Pharmacogenomics J 25: 2. PMID: 39971899). A novel machine leaning model for identification antiviral agents. This study uses advanced techniques in genome sequencing and machine learning to improve the discovery of antiviral drugs. By combining the genetic information of viruses with the structural details of existing antiviral drugs, we created models that can effectively identify drugs tailored to specific viruses and those that work against multiple viruses. These models proved to be highly accurate, with strong performance metrics. We applied these models to a large collection of around 360,000 compounds, aiming to find potential drugs against SARS-CoV-2, the virus responsible for COVID-19. Our innovative approach, which merges several machine learning methods and integrates detailed compound structures with viral genetic data, enabled us to pinpoint both specific and broad-spectrum antiviral agents. To ensure the best candidates, we used a thorough five-step filtering process. This process selected compounds that not only resembled known antiviral drugs but also minimized toxicity and ensured a variety of structural types. We tested 346 of these predicted compounds using two lab-based experiments: one that checks if the virus can enter cells (PP entry assay) and another that examines the virus's ability to replicate (RdRp assay). The tests revealed promising results, with 24 compounds showing effectiveness in the PP entry assay and 47 in the RdRp assay. The most effective compounds showed strong antiviral activity at very low concentrations. This study highlights the power of combining viral genetic information with drug structure data to quickly find potential new treatments for emerging viral threats (Xu T. et al. 2025. Commun Chem 8, 189. PMID: 40542184).