Curing Cancer
Cancer as Evolution — Adaptive Therapy · Oncology
Treat the tumor as an evolving population — schedule therapy to keep it controllable forever instead of selecting for resistance.
📋 The problem
Tumors are evolving populations. Maximum-tolerated dosing selects for resistant clones and relapse. Adaptive therapy treats cancer as a control problem — dosing just enough to keep a drug-sensitive population that suppresses resistant cells, aiming for indefinite control.
🧗 Why it's a grand challenge
Clonal composition is only partially observable, dynamics are noisy and patient-specific, and the optimal schedule depends on competition parameters that must be inferred online.
🧮 Governing model
dx_s/dt = r_s x_s(1−N/K) − d·u(t)·x_s; dx_r/dt = r_r x_r(1−N/K)Competing sensitive/resistant clones under drug schedule u(t): dx_s/dt = r_s x_s(1−N/K) − d·u·x_s, dx_r/dt = r_r x_r(1−N/K); control to delay competitive release.
Current best: Adaptive therapy (Gatenby/Zhang prostate-cancer trial)
🧭 Possible approaches
- Online inference of clonal dynamics from ctDNA / imaging
- Reinforcement-learning dosing policies that delay competitive release
- Game-theoretic schedules that exploit the fitness cost of resistance
🎯 Build the benchmark
From longitudinal tumor-burden / clonal-fraction series, predict and maximize time-to-progression under a toxicity budget, vs continuous MTD.
Metric: TTP_months — months time-to-progression
Datasets to start from: Adaptive-therapy prostate PSA trajectories, ctDNA clonal-dynamics longitudinal panel, PDX drug-response time series
🤖 Build an AI agent to solve it
A per-patient controller that ingests ctDNA/imaging and outputs the next dose to maximize progression-free time.
Once a benchmark exists, an AI4Science agent can iterate solutions against it — every verified solution earns PWM.
This is a frontier framing page — an open problem, not yet benchmarked or verified, unlike PWM's mature computational-imaging benchmarks.