TCR–pMHC binding prediction skips the docked complex: two monomers alone top every metric and steady rankings by 26% to 85%
Synopsis
TCR–pMHC binding prediction no longer has to start from a docked complex. Structure-based predictors co-fold the pair and read binding off the interface, so placement error changes which residues and contacts reach the model. EMBER folds each molecule as an unbound monomer, encodes both as rigid-motion-invariant residue graphs, and compares them through bidirectional attention over CDR3 and peptide residues. Across four folding tools and five epitope-disjoint folds it leads every metric, with 26% to 85% higher cross-tool ranking agreement than structure-based baselines.
Interpretation
TCR–pMHC binding can be scored without assembling a complex: receptor and antigen are folded separately as monomers, and the score comes from comparing two residue graphs, with no relative pose entering the computation. Earlier structure-based predictors co-fold or dock the two molecules into one assembly and infer binding from the predicted interface, typically selecting residues within 8 to 14 Å of the partner. On 49 class-I TCR–pMHC crystal structures and six co-folding tools, median 5 Å contact Jaccard runs from 0.188 to 0.459 and median receptor Cα RMSD from 5.87 to 8.09 Å, setting Boltz-2 aside; Boltz-2's 0.833 Jaccard falls to 0.155 on the 10 targets released after its training cutoff.
The two molecules exchange information only between CDR3 loops and peptide, and those residue sets come from sequence annotation rather than predicted geometry, so placement error cannot change which residues enter the model. Under distance-cutoff interfaces, a placement error changes not only coordinates but which residues and intermolecular contacts appear in the input graph. Each residue carries a Cα coordinate, a 37-dimensional local geometric descriptor and a frozen ESM-C embedding; edges link 30 nearest neighbors with distance and residue-identity chemistry terms, a shared encoder rebuilds neighborhoods each layer, and the contact head takes no coordinate argument.
Monomer folds can be reused, so preprocessing scales with monomers rather than with candidate pairs. Complex-based predictors generate one complex structure per candidate pair, so cost grows with the product of receptor and target counts. On an evaluation set of 21,477 TCRs, 353 pMHCs and 45,909 pairs, folding all monomers with tFold-TCR costs 4.2 GPU-hours against 80.7 GPU-hours as complexes, a reported saving of up to 19.4×.
Perspective
Evaluation covers four public datasets, 343 epitopes and 45,909 candidate pairs, using five epitope-disjoint folds and three seeds, with every method retrained on structures from tFold-TCR, ESMFold2, TCRmodel2 and Boltz-2. The released encoder receives only the TCR β chain; the α chain sits in the cached graph but does not reach the encoder, and restoring it costs 8.6% of the micro-AUPRC margin. Preprocessing timings ran on one A100-SXM4-80GB, each cell a single measurement rather than a mean over repeats.
The authors report that co-folding interface confidence separates true binders from decoys at Cohen's d of 0.20 to 0.60, yet no complex-consuming method converts that evidence into discrimination. Feeding the complex to EMBER instead costs 12.3% of the micro-AUPRC margin, more than removing structure entirely at 11.1%. Sampling does not average the docking bias away: directional coherence γ across 20 folds per target is 0.84 to 0.992, and equal-weight averaging moves the displacement by −2.5% to +6.7%. The preprocessing saving is a constant factor on a fixed candidate set rather than a change of regime; TCRmodel2 spends 45% to 48% of wall clock outside inference, and Boltz-2's timing excludes multiple sequence alignment.