Training on experimental function improves prediction of tumor-reactive T-cell receptors

bravo_abad · x · 2026-09-21

A writeup of new immunology-AI work: instead of cheap proxy labels, St. Paul, Hendrikse et al. collected direct functional labels — testing whether patient-derived T-cell receptors actually activate against multiple myeloma cells — then trained PreGame, a random-forest model on gene expression and surface protein data. The model can now score many receptors without the full lab assay. Key lesson: expensive but direct functional labels beat easy proxies for scientific ML.

Original post →

More from Research

Research channel →