NetTCR: sequence-based prediction of TCR binding to peptide-MHC complexes using convolutional neural networks

Jurtz, V. I.; Jessen, L. E.; Bentzen, A. K.; Jespersen, M. C.; MAHAJAN, S.; Vita, R.; Jensen, K. K.; Marcatili, P.; Hadrup, S. R.; Peters, B.; Nielsen, M.

doi:10.1101/433706

bioRxiv · 10.1101/433706

NetTCR: sequence-based prediction of TCR binding to peptide-MHC complexes using convolutional neural networks

Jurtz, V. I.·Jessen, L. E.·Bentzen, A. K.·Jespersen, M. C.·MAHAJAN, S.·Vita, R.·Jensen, K. K.·Marcatili, P.·Hadrup, S. R.·Peters, B.·Nielsen, M.

View original source ↗Export metadata ↓

Abstract

Predicting epitopes recognized by cytotoxic T cells has been a long standing challenge within the field of immuno- and bioinformatics. While reliable predictions of peptide binding are available for most Major Histocompatibility Complex class I (MHCI) alleles, prediction models of T cell receptor (TCR) interactions with MHC class I-peptide complexes remain poor due to the limited amount of available training data. Recent next generation sequencing projects have however generated a considerable amount of data relating TCR sequences with their cognate HLA-peptide complex target. Here, we utilize such data to train a sequence-based predictor of the interaction between TCRs and peptides presented by the most common human MHCI allele, HLA-A*02:01. Our model is based on convolutional neural networks, which are especially designed to meet the challenges posed by the large length variations of TCRs. We show that such a sequence-based model allows for the identification of TCRs binding a given cognate peptide-MHC target out of a large pool of non-binding TCRs.

Explore related subjects

bioinformatics Preprint

Keep this discovery

Explore connections, maps & timelines

BibTeX RIS

Jurtz, V. I., Jessen, L. E., Bentzen, A. K., Jespersen, M. C., MAHAJAN, S., Vita, R., Jensen, K. K., Marcatili, P., Hadrup, S. R., Peters, B., Nielsen, M.. 2018-10-03. NetTCR: sequence-based prediction of TCR binding to peptide-MHC complexes using convolutional neural networks. https://doi.org/10.1101/433706

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related preprints

bioRxiv2026

Transcript quantification using context-aware MAPQ modeling

Long-read RNA-sequencing enables accurate transcript-level quantification, but prior static filtering of MAPping Quality (MAPQ) can ignore transcriptome-intrinsic ambiguity, and discarding reads on this basis distorts abundance estimates. We present CALMeR (Context-Aware Long-read MAPQ-based Rescuer), which conditions MAPQ filtering on transcript properties rather than a fixed cutoff. CALMeR fits a zero-one inflated beta regression on various transcript and read-level features to model the expected MAPQ distribution for each alignment. Reads whose MAPQ fits this context, such as low-MAPQ reads from multi-isoform genes and paralog families are retained; reads with anomalously low MAPQ are removed as likely artifacts. Retained reads can be summed per transcript or passed to any downstream quantifier. Compared with a MAPQ [≥] 1 filter and expectation-maximization-based transcript quantification methods, CALMeR recovers more spike-in transcripts and better matches known spike-in concentrations. CALMeR is available at https://github.com/sparthib/CALMeR.

bioinformatics↗

bioRxiv2026

Structure-resolved fold annotation of the medicinal-leech salivary-EV cargo: a Ly-6/uPAR three-finger scaffold and a structurally dark fraction

Many proteins that medicinal leeches package into salivary extracellular vesicles (EVs) carry no interpretable sequence annotation, so the least-characterised part of a pharmacologically important secretome lacks even a fold assignment. A companion Astral-DIA study of Hirudinaria manillensis nominated 17 ultra-enriched candidate EV cargo proteins. Here we ask whether structure prediction and structural search can assign folds where sequence methods fail, and which candidates remain unresolved under prespecified criteria. With an offline pipeline (ESMFold, TM-align, Foldseek against AFDB-SwissProt and PDB100, UniRef90 search, protein-to-genome alignment), the candidates separate into intracellular/membrane machinery and a secreted, cysteine-dense fraction of three proteins with no Pfam domain. The predicted structure of the leech Protein-sleepless homolog g21166 supports its assignment to the Ly-6/uPAR three-finger fold (TM 0.58 vs CD59; best significant PDB match Lynx1); no function is inferred. g436 and g20791 returned no match meeting the criterion (alnTM [≥] 0.5, E [≤] 0.05) in either database, g436 even after homolog-informed modelling. Gene structure and splice junctions support a conserved locus for the two-Trypsin-segment model g8114 (catalytic unit not established). A Boltz-2 thrombin pilot did not support target engagement, and a pre-registered nicotinic test was terminated as undecidable. The results yield a structure-informed shortlist for experimental follow-up.

bioinformatics↗

bioRxiv2026

EpiRank: partner-agnostic surface-patch retrieval for epitope prediction of protein binding interfaces

Structure-based interface predictors are normally reported as a single pooled ROC AUC over a residue test set. For a protein designer the operative question is different and much harder: given one target chain, will the top-ranked surface patch be a real binding site? We built EpiRank, a partner-agnostic epitope predictor, on HumanPPI 29,246 predicted human dimers, 12,296 target chains, 4.18 M residues using only single-chain geometry and physicochemistry (170 features, no partner, no sequence search, no structural alphabet), and evaluated it as a retrieval problem with homology- and partner-graph-controlled splits, on the 3,529 chains that survive an explicit partner-contamination filter. The model places a real epitope in its top-ranked patch 64 % of the time. That number is far less informative than it looks. Compact patches placed at random hit a real epitope 41 % of the time under the same rule, so the model's lift is 1.57 x, not 64 %; 54 % of all predicted patches land on nothing; and when the same chains are scored by models trained on a 2,000-chain and an 11,669-chain superset, a substantial fraction change answer while the aggregate barely moves. We document four specific ways this task invites overstatement partner-graph percolation across splits, single-linkage homology percolation at the conventional 30 % identity threshold, a count-only hit rule that rewards spatial smoothing rather than signal, and hotspot precision conditioned on success, which ranks a random baseline above the trained model. Each was hit during development and each inverted a conclusion we had already drawn. We argue that interface-prediction papers should report lift over a geometry-matched random baseline, patch precision, and per-target stability alongside AUC, and we release the evaluation harness that computes them.

bioinformatics↗