Poster #53: Johnna Berryhill MS

OncoProfilingTools: a human-in-the-loop AI framework for pharmacogenomic drug-response analysis

Johnna Berryhill, MS¹; Jason LaPierre, MS²; Arjun Reddy Banda, MS¹; Alexandra Chittams-Miles, PhD³; Adam Khader, MD⁴; Katarzyna M. Tyc, PhD¹ ¹ Department of Biostatistics, Virginia Commonwealth University, Richmond, Virginia, USA ² School of Life Sciences and Sustainability, Virginia Commonwealth University, Richmond, Virginia, USA ³ Richmond Institute for Veterans Research, Richmond, Virginia, USA ⁴ Department of Surgery, VCU Health, Virginia Commonwealth University, Richmond, Virginia, USA

Large pharmacogenomic screens can reveal differences in drug response across molecularly or clinically defined groups, but translating statistical findings into biological hypotheses often requires multiple tools, custom code, and manual review. We developed OncoProfilingTools, an open-source R package that combines reproducible drug-response analysis with human-in-the-loop, agent-assisted interpretation. Methods: Users can provide their own molecular and drug-response data or load pre-integrated PRISM/DepMap datasets. OncoProfilingTools stores these data in standard Bioconductor structures and compares drug sensitivity between user-defined molecular or clinical groups. A system of large language model (LLM)-based agents coordinates downstream analyses, including pathway enrichment, transcription-factor analysis, immune deconvolution, and functional annotation. The agents synthesize results into biological explanations, while a report-writing agent presents these explanations alongside supporting statistical evidence for review. We applied this workflow to baseline gene-expression profiles and secondary PRISM drug-response data from 27 colorectal cancer cell lines, including 11 consensus molecular subtype 4 (CMS4) and 16 non-CMS4 cell lines. CMS4 is an aggressive, mesenchymal subtype associated with treatment resistance and poor prognosis. Results: Of 1,405 compounds, 43 were nominally more active in CMS4 cell lines (Wilcoxon p < 0.05), and 10 also met the activity criterion of median CMS4 AUC < 0.70, where lower AUC indicates greater sensitivity. None remained significant after Benjamini–Hochberg correction. Leading active candidates included YM-155, romidepsin, NMS-E973, and elesclomol. Pathway analyses showed higher mesenchymal, extracellular-matrix, morphogenesis, and selected PI3K-related activity in CMS4 cell lines, along with lower metabolic and mitochondrial activity. The generated report linked each LLM-proposed explanation to the relevant drug-response and pathway results, allowing researchers to prioritize hypotheses for further study. Conclusions: OncoProfilingTools provides a reproducible approach for moving from drug-response results to biological hypotheses that can be reviewed and tested. The CMS4 analysis demonstrates how agent-assisted interpretation can bring together statistical and biological evidence while maintaining human oversight.