Patient-level multimodal foundation models for treatment selection · Idea

Pathology and radiology foundation models exist separately; combining them with genomics and clinical data (as in CanSim-style efforts, Tempus, Owkin) is the next step. Prediction of ADC or IO response from routine data would be transformative.

Confidence

35%–60%likelyOnCo editors (initial estimate), 2026-09-07 · Single-modality predictive AI is already cleared; multimodal treatment-effect prediction is unproven.

Hypothesis

A multimodal model trained on trial cohorts predicts benefit from a specific therapy (e.g., TROP2 ADC vs chemotherapy) with clinically useful discrimination beyond current biomarkers.

Rationale

ArteraAI showed single-modality models can be predictive; adding modalities and outcomes from randomised trials allows causal treatment-effect estimation.

What would test it

Train on completed phase 3 datasets (with sponsors), validate on held-out trials; prospective biomarker-stratified trial.

Maturity

preclinical evidence

Key papers

1top

methodsNature 2021

AlphaFold 2: predicting protein structures to near-experimental accuracy

The shape of nearly every protein is now available to any researcher in seconds instead of years, which shortens the path from a cancer target to a designed molecule. It does not by itself produce drugs: binding pockets, dynamics and cellular context still need experiment.