Predicting whether a T-cell receptor binds a peptide-MHC complex usually starts from a predicted 3D structure of the whole complex. Any error in that docking step then flows into every downstream feature. EMBER, a new framework from Lin Ha and colleagues, sidesteps the problem by never constructing a shared complex at all.
Instead, EMBER folds the TCR and the pMHC independently and encodes each as an SE(3)-invariant residue graph. The model compares these learned representations without a common coordinate frame, so a relative pose is never an input. This also changes the computational cost: preprocessing requires O(N+M) monomer folds rather than O(NM) operations for complex-based predictors.
The authors evaluated EMBER under epitope-disjoint conditions with six retrained baselines and four folding tools. It ranked first on every metric, with improvements of 0.9–14.4% in micro-AUROC, 18.5–25.3% in micro-AUPRC, and 10.6–24.1% in MCC. Its ranking was also up to 85% more stable across folding tools.
Because this is a single arXiv preprint, the results have not yet been independently verified or peer reviewed. Still, the core idea—removing the docked complex from the prediction pipeline—directly targets a known source of error in structure-based binding models.