The Tumor Evolves Faster Than the Plan

Current cancer treatment selects therapies based on a single tumor biopsy taken at diagnosis, then follows a fixed protocol through multiple lines of treatment. But tumors evolve continuously — acquiring resistance mutations, shifting clonal composition, and altering their microenvironment — often within weeks of starting therapy. By the time imaging or clinical markers reveal treatment failure, the tumor has already evolved past the point where the next-line therapy was optimal. No clinical system exists that can dynamically track tumor evolution during treatment and update therapy recommendations in real time.

Metastatic cancers (breast, lung, colon) kill over 600,000 Americans annually. Current oncology relies on a limited set of biomarkers from a single data modality — usually genomic sequencing of the initial biopsy — to guide treatment across months or years of therapy. Treatment response rates for second-line and third-line therapies in metastatic cancer are typically 10–30%, partly because the tumor the physician is treating has diverged significantly from the tumor that was biopsied. An estimated 30–50% of patients with actionable mutations at diagnosis acquire additional mutations during treatment that could redirect therapy, but these changes go undetected until clinical progression.

Liquid biopsy (circulating tumor DNA) can detect some mutations non-invasively, but current assays track a narrow panel of known resistance markers rather than mapping the full clonal architecture of the evolving tumor. Adaptive clinical trial designs (like I-SPY 2 for breast cancer) have shown that dynamically assigning patients to treatments based on biomarkers improves outcomes, but these operate at a population-trial level, not at the individual patient-treatment level. Computational models of tumor evolution exist in research settings but have not been validated prospectively — they can retrospectively explain clonal dynamics but cannot yet predict which clones will dominate under a given treatment pressure. The infrastructure to integrate multi-modal data (genomics, transcriptomics, imaging, liquid biopsy, pathology) into a real-time treatment recommendation system does not exist in clinical practice.

A system that integrates longitudinal multi-modal tumor data (serial liquid biopsies, imaging, pathology) into a computational model of clonal evolution that can predict treatment response and recommend therapy switches before clinical failure would transform oncology from reactive to anticipatory. This requires advances in three areas simultaneously: (1) high-frequency, multi-analyte liquid biopsy platforms that capture tumor heterogeneity beyond point mutations; (2) validated computational models of tumor evolutionary dynamics under therapeutic pressure; (3) clinical trial infrastructure capable of evaluating adaptive, model-driven treatment protocols.

A student team could develop a computational model of clonal evolution using publicly available cancer genomics datasets (e.g., TCGA, GENIE) and test whether serial sampling data can predict which clonal populations will expand under specific drug pressures. A more engineering-focused team could prototype a data integration pipeline that combines simulated liquid biopsy, imaging, and pathology data into a unified tumor state representation. Relevant disciplines: computational biology, machine learning, oncology informatics, biomedical engineering.