Research tool — not medical advice. This tool generates structural predictions from a mathematical formula. It has not been clinically validated and must not be used to guide treatment decisions. Novel combinations shown here are formula outputs, not clinical recommendations. Any clinical application requires independent preclinical and clinical testing. Consult a qualified oncologist for treatment decisions.
Select cancer type
Coming
Drug Combination Response Predictor
Every tumour is a dissipative system — it maintains itself by spending energy. Generative Geometry predicts that all dissipative systems resist intervention the same way: through structural depth, temporal adaptation, and coverage gaps.
This tool derives drug combination response rates from a single formula. No curve-fitting per trial. No machine learning. One structural equation, calibrated once per cancer type, tested against published clinical results.
M = agent potency · k = sub-phases covered · n = depth · τ = tumour age · p = prior lines · σ = cascade bonus
How it works: Select a cancer type, choose a clinical stage (which sets depth, tumour age, and prior treatment), then pick 1–4 agents. The formula computes the predicted objective response rate. Where published trial data exists, the prediction is shown alongside the observed result.
What the parameters mean: Each agent blocks a sub-phase of the tumour's maintenance cycle. More sub-phases covered = exponentially higher blockade (the σ cascade). But deeper tumours (higher n) and older tumours (higher τ) resist more. Prior treatment lines (p) further reduce effectiveness through acquired resistance.
Predictions vs published trials
35 comparisons across 7 cancer types. Same formula, same parameters. Sorted by absolute gap.
CANCER
TRIAL
OBSERVED
PREDICTED
GAP
Mean absolute error
Select a cancer type from the left panel to start building combinations.
Where the formula predicts the biggest gains
Novel combinations vs current standard of care, ranked by predicted clinical impact (response gain × patient volume). Late-stage cancers first.
What these predictions assume: Agents perform as calibrated, act independently, and are tolerable in combination. What they do not account for: cumulative toxicity, dose-limiting side effects, drug-drug interactions, pharmacokinetic interference, or patient-specific factors. Real clinical outcomes will be lower. The purpose is to identify where the structural opportunity is largest — which combinations are worth testing — not to predict patient outcomes.
Select a cancer type from the left panel to start building combinations.
Read the paper →The Geometry of Intervention — full derivation of the blockade formula
Known limitations & next steps
The current model predicts 24/35 published trials within 3 percentage points. The remaining outliers have identified causes. We are actively working on:
Joint parameter optimisation — Recalibrating the system-level parameters (β, γ, ε) per cancer type to minimise MAE across all evidence entries simultaneously, rather than agent by agent.
Expanding the evidence base — Adding published single-agent ORR data for the ~70 agents currently marked as "estimated," converting them to calibrated values backed by trial data.
Biomarker-selected population flags — Some trials (e.g. KEYNOTE-024, PD-L1 ≥50%) select for biomarker-enriched populations that inflate observed ORR relative to what the formula predicts for unselected populations. These will be flagged in the evidence table.
Toxicity modelling — At one fractal level deeper (sub-positions within sub-phases), the geometry should predict where toxicity concentrates when agents overlap. This is the next structural extension of the formula.
If you are a researcher interested in contributing calibration data or testing predictions against unpublished trial results, please contact raimo@generativegeometry.science.
How to use
1 Select a cancer type from the left
2 Choose a clinical stage
3 Pick agents — or click a standard regimen or novel combination to auto-fill