New preprint uses AlphaFold3 ensembles to predict TCR–pMHC binding

quaidmorris · x · 2026-07-25

A recent preprint introduces enFoldX, a structure-ensemble method for predicting whether a TCR and a peptide-MHC pair are cognate binders.

Instead of trusting a single AlphaFold3 co-folded structure, the method featurizes many sampled complex structures.
The key insight is that AF3 can produce plausible-looking false positives for non-cognate pairs, especially when sequences differ only subtly.
enFoldX turns that weakness into a signal by analyzing how confidence and interface geometry vary across the ensemble.
The approach focuses on features at the binding interface rather than on one “best” predicted structure.

This is a concrete AI-for-biology method paper rather than a product announcement.

More from Apps