Biologically informed TCR representations improve in-distribution prediction but fail to generalize to unseen epitopes

Abstract

Background. TCR specificity predictors work on epitopes seen in training and fail on unseen ones. One explanation holds that receptor representations ignore the organization created by V(D)J recombination. We tested it. Methods. From 121,467 deduplicated human TCRβ clonotypes in VDJdb we derived reference-validated germline anchors, split each CDR3 into germline termini and a junctional N-region, and compared biologically informed features against raw CDR3 3-mers and a V/J-only control, on seen epitopes, on withheld epitopes, and on the official IMMREP23 benchmark. Results. Junctional regions carry order-dependent structure: conditioning a bigram model on the (V,J) pair improves discrimination of real from order-shuffled N-regions by +0.1013 AUC, not attributable to model capacity (permuted-label control). A boundary-trim control locates most of it at the germline boundaries, leaving +0.0392 in the junctional interior. Binding prediction differs: morphological features beat raw 3-mers on seen epitopes (0.7135 vs 0.6589) but not V/J identity alone (0.7089), and on withheld epitopes every arm sits at chance, the advantage falling to +0.0046 with a CI spanning zero. Validation. Conclusions replicate on IMMREP23, where no method exceeds 0.52 on the seven unseen peptides, including a TCRdist-style baseline that otherwise outperforms our model and NetTCR-2.2 retrained on the identical split (0.4868 unseen). The advantage over k-mers proves contingent on training-set size. Conclusions. Biologically informed segmentation captures genuine germline-conditioned organization that does not extend to unseen epitopes, and most of the within-distribution advantage reflects V/J identity rather than junctional sequence. Receptor representation alone appears insufficient.