TriStageHLA splits epitope discovery into three stages, lifting immunogenicity AUROC from 60.1% to 81.4%

Epitope discovery is split into three questions asked in sequence: whether a peptide looks like something that will be displayed, which HLA allele it binds, and whether it actually triggers a T-cell response, with the three modules trained separately and chained. Earlier designs folded binding and presentation prediction into one model and often treated immunogenicity as the final endpoint, so binding and T-cell responsiveness were conflated. On the held-out immunogenicity test set the immunogenicity module reached AUROC 81.4% and AUPRC 73.8%, against immunogenicity-specific baselines ImmuneApp at 70.7%/63.6% and BigMHC-IM at 60.1%/51.5%.

The binding module reached AUROC 96.2% and AUPRC 95.2% on combined BA/EL evaluation, and AUROC 94.54% with AUPRC 89.74% at the allele level. On the same evaluation BigMHC reached AUROC 95.5% and MixMHCpred3.0 92.7%; at the allele level BigMHC reached 94.21%/88.77% and MixMHCpred3.0 91.94%/87.46%. Stratified by difficulty, on rare alleles with fewer than 200 training instances (n=34) the binding module averaged AUROC 0.899 versus BigMHC's 0.857 (+0.042, p=0.019); on the rare-and-distant intersection (n=10) the advantage was largest, at +0.080 AUROC and +0.172 AUPRC.

The screening module updates only low-rank adapters on a frozen protein language model, leaving about 35 million trainable parameters, roughly an order of magnitude below TransHLA's about 0.7 billion. Earlier peptide-presentation front ends ran at hundreds of millions of parameters and did not support parameter-efficient adaptation of the learned representation to binding or immunogenicity tasks. Under the same protocol with metrics computed from raw logits, the screening module reached 84.85% accuracy and AUROC 92.05%, versus TransHLA's 84.70% and 91.90%.

The binding module uses a dual-branch design that models global context and local motifs together, with a mutual-information objective that stabilizes training under class imbalance. Removing the local convolutional branch dropped mean AUROC across lengths from 95.33% to 90.44% and mean AUPRC from 90.85% to 81.65%, indicating local motif extraction is not replaceable. In a simulated screen over 300,000 human-proteome background peptides plus 14,490 true candidate peptide-allele pairs, screening retained 82.4% of true candidate peptides (10,747/13,044) while discarding 87% of background peptides, and the binding stage then retained 74.4% of true candidate pairs (10,779/14,490) while removing 90% of background pairs.

The demonstrated scope is HLA class I with 8- to 14-mer peptides, evaluated on IEDB-derived training data, an external CEDAR set and one simulated screen over 300,000 background peptides; no prospective validation in patient-derived cohorts is reported. The authors propose next steps including HLA class II extension, deployment-oriented calibration and thresholding, broader multi-laboratory external validation, and semi-supervised or contrastive objectives to use unlabeled ligandomes.

Immunogenicity labels are operationally defined by measurable T-cell assay responses in IEDB and may reflect cohort-specific TCR repertoires and prior exposure rather than an intrinsic peptide property. The authors note that although identical peptide-allele pairs and overlapping peptide sequences between CEDAR and IEDB were excluded, close-homologue peptides could still partially inflate apparent generalization, and that multi-center datasets would strengthen external validity; batch effects and prevalence shifts can also affect thresholds.