Leveraging tumor dynamics to discover mutations influencing progression and treatment response for precision oncology

BACKGROUND: Current research in precision oncology focuses on progression-free survival (PFS) and overall survival (OS) as the key outcomes. This stands in stark contrast to the primary outcomes guiding cancer treatment in the clinic- dynamic changes in tumor size in response to treatment. While PFS and OS are meaningful, highly interpretable and easily collected, we hypothesize that revealing associations between somatic mutations and tumor size dynamics can uncover genetic factors that can influence treatment response and generate hypotheses for future clinical investigation, while further improving our understanding of tumor biology. METHODS: We propose and benchmark a new statistical framework - Andersen-Gill based Tumor Dynamics (AG-TD) to model recurrent outcomes and apply it to data from ~ 17 K radiology reports of patients with non-small cell lung cancer undergoing routine treatment at the Dana-Farber Cancer Institute. Our approach enables modeling recurrent outcomes while considering the time-dependent nature of biomarker-treatment interaction, patient mortality, and loss to follow-up. RESULTS: We identify 27 somatic mutations that lead to differential response to common treatments including both protective and hazardous effects. Among these, deletions in MYBL1 and GATA3 were independently validated by replication in an external EGFR-TKI drug sensitivity screen, reaching Bonferroni correction significance. We also identify 13 somatic mutations associated with differential rates of progression and response during the entire treatment course, of which 9 are undetectable using conventional OS or PFS models. We further verify that the results are consistent in a homogeneous sub-cohort of advanced stage (3 and 4) patients only. CONCLUSIONS: Overall, we demonstrate the potential for using longitudinal outcomes in cancer genetic studies to complement traditional survival-based analyses, and to enable systematic identification of candidate gene-treatment interactions for future validation.