Dual-stream attention model for peptide-MHC class II binding that runs on ten interchangeable frozen protein representations.
No providers recorded yet. Browse all providers
A peptide-MHC class II predictor's accuracy is usually reported as a single number, but that number confounds two things: what the protein encoder already knows about sequence, and what the downstream head does with it. Papers in this area typically change both at once, so when a new method beats the last one, neither contribution is recoverable. PREpiBind is built to separate them — one architecture, one training protocol, one set of data splits, and the protein representation as the only thing that moves.
Developed at Seoul National University's College of Pharmacy with collaborators at Chonnam National University Medical School and Korea's National Immunotherapy Innovation Center, PREpiBind is a dual-stream, joint-attention head over frozen representations. The MHC alpha/beta chain pair and the epitope enter separate two-layer transformer encoders, their token sequences are concatenated, a joint self-attention interaction block lets the two sides exchange information, and a masked mean pool feeds a two-layer prediction head trained with binary cross-entropy. Nothing in the backbone is fine-tuned, so the head is small and the representation is a selectable input.
Ten encodings were plugged into that slot: the BLOSUM62 substitution matrix, residue and pairwise features distilled from AlphaFold 3, Boltz-1 and Chai-1, and protein language model embeddings from ESMC at 300M/600M/6B and ESM-3 at Small, Medium and Large. These are existing models PREpiBind consumes, not contributions of the work.
Training data came from an April 2025 IEDB export restricted to linear, unmutated 15-mers against HLA-DR, HLA-DQ, HLA-DP and murine H2, assembled into three arms: qualitative binding (161,223 entries), mass-spectrometry ligandomics (111,444) and IC50 measurements (47,154) binarised at 500 nM and 1,000 nM. Each arm was split 70:30 with grouping by peptide and beta chain to block leakage, then five-fold cross-validated across three seeds; reported figures are means over the resulting fifteen runs. On the qualitative test set, ESM-3 Large reached 0.927 ± 0.002 ROC-AUC, Chai-1 0.914, Boltz-1 0.884, AlphaFold 3 0.876 and BLOSUM62 0.871, against 0.838 for NetMHCIIpan-4.3 and 0.815 for MixMHC2pred-2.0; the mass-spectrometry arm topped out at 0.994. On the thresholded IC50 arms the leading language-model configurations clustered at 0.835-0.839, below NetMHCIIpan-4.3's 0.909, which is scored by a head fitted directly to measured affinities. The language-model advantage narrowed under stress: averaged over 38 withheld molecules, ESMC 300M reached 0.796 and Chai-1 0.792, and in the cross-species H2 evaluation the ordering flipped with the aggregation rule — language models led on pooled rows, Chai-1 led when eight H2 molecules were weighted equally, and the panel was too small to support a stable ordering. Scale did not track accuracy either: ESMC 6B scored 0.909, below ESMC 300M's 0.919.
The released checkpoints score peptide-MHC class II binding for 98 human HLA class II chains and 18 mouse H2 chains, the practical input to neoantigen shortlisting, vaccine epitope selection and immunogenicity risk assessment for biologics. The benchmark itself is the other deliverable: groups building pMHC-II tools can read off which representation family suits an allele-rich pooled setting versus transfer to an unseen molecule, without re-running ten encoders themselves.
PREpiBind gives the pMHC-II field a common yardstick where it had none, and its findings cut against two comfortable assumptions — that the largest language model is the best feature extractor, and that a single pooled leaderboard ranks representations. The authors are explicit that structure-derived features from folding models stay competitive exactly where generalization matters most, and that the choice of encoder should follow the intended prediction scenario. The work is a preprint and has not been peer reviewed; its comparisons are confined to frozen features, 15-mer peptides and partially overlapping assay-derived datasets, and the authors call for external validation beyond them.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.