EMBER predicts TCR-pMHC binding without a docked complex
TL;DR
- EMBER folds the TCR and pMHC separately rather than predicting a docked complex, removing the placement error the authors say dominates TCR-pMHC binding predictions.
- Against six retrained baselines, EMBER reports 18.5 to 25.3 percent gains in micro-AUPRC and 10.6 to 24.1 percent in MCC under epitope-disjoint evaluation.
- EMBER's ranking is up to 85% more stable across folding tools, and preprocessing scales linearly in monomers instead of quadratically in receptor-antigen pairs.
A new method for predicting which T-cell receptors bind which cancer peptides claims to beat six retrained baselines on every metric, in part by refusing to dock them together first.
The paper, posted to arXiv, calls the method EMBER. It 'eliminates the need for a predicted docked complex, and with it the placement error that dominates such predictions,' the authors write. Instead of co-folding the receptor and antigen into a shared coordinate frame, EMBER folds each independently, encodes them as SE(3)-invariant residue graphs, and compares the learned representations.
The reported margins, under epitope-disjoint evaluation against six retrained baselines across four folding tools, are 0.9 to 14.4 percent in micro-AUROC, 18.5 to 25.3 percent in micro-AUPRC, and 10.6 to 24.1 percent in MCC. EMBER's ranking is also 'up to 85% more stable across folding tools', the authors report, meaning its scores shift less when the underlying structure predictor is swapped out. Preprocessing scales linearly in the number of monomer folds rather than quadratically in their product.
The abstract reports no absolute AUROC or AUPRC values, no dataset size, and no separate evaluation on unseen TCR families.
Originally reported by paper
Read the original article →Original headline: EMBER Predicts T-Cell/Cancer-Peptide Binding +25% Better by Eliminating the Docking Step