EMBER: Predicting TCR--pMHC Binding Without a Docked Complex

EMBER. Predicting TCR--pMHC Binding Without a Docked Complex. By Lin Ha and five others, from Northwestern University. Published on the nineteenth of October. No code repository is linked to it.

Structure-based predictors of binding between a T cell receptor and a peptide bound to an MHC usually start from a predicted docked complex. They co-fold the receptor and the antigen together and read binding off the contact surface between them. The trouble is what that inherits. Any deviation introduced by the structure-prediction tool is carried forward into every downstream feature, and the authors say this placement error dominates such predictions.

EMBER removes the docked complex, and with it the error. The receptor and the peptide MHC are folded independently rather than together. Each is encoded as a residue graph that is invariant under rotation and translation, and the two learned representations are compared without ever building a shared coordinate frame. A relative pose is therefore not an input to the model at all, so placement error has no path into the prediction.

There is a second benefit, in cost. Because each side is folded on its own, EMBER preprocesses a number of monomer folds that grows with the sum of the two sets. A complex-based predictor has to work through the product instead, one operation per pairing.

Evaluation is epitope-disjoint, against six retrained baselines, under each of four folding tools. EMBER places first on every metric in all four settings. The reported margins are zero point nine to fourteen point four percent in micro AUROC, eighteen point five to twenty-five point three percent in micro AUPRC, and ten point six to twenty-four point one percent in MCC. Its ranking is also up to eighty-five percent more stable from one folding tool to another.

The stability number is the one that ties back to the diagnosis. If the usual predictor's weakness is inherited placement error, then swapping the folding tool underneath it should move the answers around. A model that never builds a joint structure has less to inherit, and the results show it changing less when the tool beneath it changes.