Biomarker development from functional precision medicine datasets via explainable machine learning.

Abstract

10061 Background: Genomics precision medicine, deployed via tumor panel sequencing, now assists in deploying targeted therapies to cancer patients. Numerous clinical trials have investigated the utility and benefit of genomics precision medicine in multiple cancer indications. Current large-scale studies report actionability rates from ~35% to ~60%, although clinical benefit rates have been shown to be closer to 10%. While this has positively impacted patients in need, the gap between actionability and benefit remains a clinical challenge attributed to multiple factors including the complex, multi-factorial relationship between molecular status and response to therapy. These differences go beyond simple disease states and may be reflective of multiple clinically relevant features including age, sex, and race/ethnicity. Methods: We implemented a functional precision medicine (FPM) program where patients with advanced pediatric cancers were prospectively profiled via high-throughput drug sensitivity testing (DST) of FDA-approved agents on patient-derived tumor cells as well as genomics testing. The objective was to investigate the clinical utility and benefit of FPM guidance in the treatment of pediatric cancer and elucidate the relationship between molecular status of patients’ diseases and treatment responses. We generated DST data (n = 21 patients) and genomic profiling data (n = 20 patients) on pediatric cancer patients in Miami, FL, as well as post-hoc whole exome and transcriptome sequencing data (n = 13 patients) and investigated three specific relationships. Results: We analyzed the relationship between racial/ethnic background and functional response to anti-cancer agents, determining potential differences in response to therapeutic classes. Next, we examined relationships between functional response and cancer type, identifying an unanticipated lack of clustering between disease indications in patients with advanced pediatric cancers. Finally, we applied an explainable machine learning (xML) framework to the functional genomic dataset to develop multi-omics biomarker hypotheses for the chemotherapy agent idarubicin, pinpointing a potential multi-cancer relationship between response to idarubicin and known disease mechanisms in acute myeloid leukemia (AML), the sole indication where idarubicin is approved. We further present additional proof-of-concept studies generating biomarker hypotheses via xML, demonstrating a framework for development of multi-omics biomarkers. Conclusions: We are now expanding our pan-pediatric cancer functional genomics dataset through an NIMHD-funded expansion cohort (NCT05857969, n = 65 patients) to further investigate multi-omics relationships between functional and molecular characteristics and understand the role of race/ethnicity in the complex relationship.