Predicting the Timing of Sequential Cancer Therapy from Sparse Early Tumor Dynamics

This repository contains the simulation code, frozen analysis inputs, publication source data, and final figures for the predictive-timing extension of the two-strike evolutionary-rescue cancer model.

Overview

This study asks whether the future tumor-population nadir can be predicted from a small number of early Drug-1 measurements, and whether such prediction can recover the treatment benefit obtained from much denser reactive longitudinal monitoring.

The final evaluation cohort contains 80 heterogeneous in-silico patients. For each patient, the pretreatment and Drug-1 evolutionary history is fixed across candidate second-strike times. Stochasticity is reintroduced only after the second strike, allowing alternative treatment decisions to be evaluated against the same underlying patient history.

The ABC population-nadir predictor is trained and tuned using an independent reference library and does not use post-switch extinction outcomes during training or hyperparameter selection.

Repository structure

extinctionTherapyTminPred/
├── README.md
├── requirements.txt
├── .gitignore
├── sims/                  Final SLiM simulation models
├── scripts/               Cohort generation, ABC, regret, and plotting code
├── hpc/                   Longleaf SLURM submission scripts
├── data/
│   ├── cohort/            Frozen 80-patient analysis inputs
│   ├── reference/         ABC reference features and sampled reference paths
│   └── source_data/       Data underlying manuscript figures and tables
└── figures/               Frozen manuscript figures

Main simulation models

• sims/hybrid_regen80_expand.slim
• sims/hybrid_regen80_prefix.slim
• sims/hybrid_regen80_outcome_extend300.slim

Main analysis scripts

• scripts/generate_regen80_manifest.py
• scripts/run_regen80_pipeline.py
• scripts/sanity_check_regen80.py
• scripts/build_hybrid_abc_reference_library.py
• scripts/tune_and_run_hybrid_abc_80.py
• scripts/decompose_abc_oracle_regret.py
• scripts/make_final_results_figures.py

Final cohort design

• 80 in-silico patients
• 20 candidate second-strike lags
• 10 independent post-switch replicates per patient and lag
• 16,000 counterfactual outcome simulations
• extinction scored at cycle 300
• quantitative Drug-1 and Drug-2 resistance
• partial cross-resistance
• no resistance fitness cost

ABC reference library

A total of 10,000 reference simulations were attempted. After excluding unusable simulations and trajectories with boundary nadirs, the final nadir-prediction reference set contains 9,983 trajectories.

The final predictor uses four early Drug-1 population measurements from treatment cycles 0--3 to predict the future population nadir.

Reproducing the manuscript figures

From the repository root, run:

python scripts/make_final_results_figures.py

Reproduced figures and source data are written to:

reproduced_results/
├── figures/
└── source_data/

The reproduced_results/ directory is not tracked by Git. The committed figures/ and data/source_data/ directories contain the frozen manuscript versions for comparison.

Software

The simulations were performed with SLiM 5.1.

Python dependencies are listed in requirements.txt.

Large intermediate files

Large SLiM population snapshots, scheduler logs, scratch analyses, and the complete set of raw reference-library trajectories are intentionally not tracked. The frozen tables required for the reported analyses are included. A sampled subset of reference trajectories used for the supplementary reference-support figure is also included.