DeepPROTECTNeo

DeepPROTECTNeo is a deep learning framework for cancer neoepitope prioritization built at the Indian Institute of Technology Kharagpur by Debraj Das, Soumyadeep Bhaduri, Avik Pramanick, and Pralay Mitra, first posted as a preprint in January 2025. It addresses a persistent gap in personalized cancer vaccine design: most pipelines rank candidate neoepitopes by predicted peptide-MHC binding alone, even though presentation on MHC is necessary but not sufficient for an anti-tumor response. Whether a presented peptide activates T cells depends on recognition by a T-cell receptor, and that step is routinely omitted from candidate selection.

The system couples a conventional variant-to-peptide workflow with a learned TCR-epitope binding model. Whole-exome or whole-genome reads go through somatic variant calling and annotation, HLA alleles are inferred with arcasHLA, mutant peptides are filtered by NetMHCpan-4.1 peptide-MHC binding, and patient-matched CDR3-beta sequences are reconstructed in parallel with TRUST4. Surviving peptides are then scored against the patient's own receptor repertoire by a dual-branch, context-aware transformer-CNN with explicit cross-attention.

DeepPROTECTNeo sits alongside single-chain TCR-epitope predictors already in the catalog, including TULIP and DecoderTCR, but packages its trained model inside an end-to-end clinical pipeline rather than shipping it standalone. The workflow is hosted as a public web server, and trained weights plus processed benchmark data are archived on Zenodo under CC BY 4.0.

#Key Features

• Dual-branch sequence and descriptor encoding: A sequence branch applies bidirectional LSTMs with learned positional encodings to peptide and CDR3-beta sequences, while a convolutional branch encodes 102 physicochemical and BLOSUM-derived descriptors through residual networks with channel-wise self-attention.
• Gated fusion with rotary cross-attention: A gated fusion module combines the two representations, and bidirectional cross-attention with rotary embeddings models residue-level contacts, recovering known anchor positions and additional contact motifs.
• Stringent TCR-split benchmarking: Five disjoint folds are split by receptor sequence rather than at random, so each fold is scored on TCRs never seen in training — the setting that matches clinical use on a new patient's repertoire.
• End-to-end variant-to-neoepitope workflow: Variant calling, HLA typing, peptide-MHC filtering, and repertoire mining are wired into one hosted pipeline, removing the manual handoffs that normally separate raw sequencing from a ranked candidate list.

#Technical Details

The binding model was trained on a corpus assembled from McPAS-TCR, VDJdb, and IEDB comprising 140,992 unique CDR3-beta-epitope pairs drawn from 130,303 receptors and 1,903 epitopes. Because verified non-binders are scarce, negatives were generated by epitope shuffling that preserves the underlying sequence distribution. On TCR-split five-fold cross-validation the model reaches a mean AUROC of 0.7856 (s.d. 0.0023) and AUPRC of 0.7932 (s.d. 0.0024), roughly four percentage points above six single-chain baselines — ATM-TCR, ERGO-AE, ERGO-LSTM, TEINet, epiTCR, and NetTCR-2.0 — each retrained on the same corpus. Inter-fold variability is far lower than the baselines, with epiTCR fluctuating by more than 0.12 AUROC across folds. Ablations identify skip connections and cross-attention as the components contributing most.

Generalization was tested on ImmuneCODE, a repertoire of validated interactions from more than 1,000 SARS-CoV-2-infected individuals; after removing every pair overlapping the training corpus, 18,811 interactions remained. In this held-out setting the model achieved a balanced accuracy of 0.6237, macro F1 of 0.6223, AUROC of 0.6752, and AUPRC of 0.6766, each a significant gain over the nearest baseline, alongside the lowest Brier score in the comparison. Per-epitope AUROCs across fourteen 9-mer viral peptides span 0.53 to 0.74. The parameter count is not disclosed.

#Applications

The intended use is patient-specific target selection for cancer vaccines and adoptive T-cell therapy, where a clinical team needs a short, defensible candidate list from a tumor-normal sequencing pair. The authors demonstrate this on five TESLA consortium subjects — three with melanoma, two with non-small-cell lung cancer — whose validated neoepitopes provide ground truth. Peptide-MHC filtering on SNVs and indels retained 350 of the 532 complexes TESLA evaluated, including 24 of the 34 confirmed immunogenic epitopes. A TCR-specificity cutoff of 0.48 on the 95th-percentile binding score per epitope left 252 candidates, capturing 18 of the 34 validated neoepitopes at 52.9% sensitivity and 53.0% specificity.

#Impact

DeepPROTECTNeo argues that TCR recognition belongs inside neoantigen prioritization rather than downstream of it, and supplies both a trained model and the pipeline needed to test that claim on clinical data. Its clearest contribution is methodological: TCR-split evaluation, shuffling-based negatives, and calibration analysis form a more honest assessment protocol than the random splits common in this literature, and the low inter-fold variance is as notable as the headline accuracy. Performance on the TESLA cohort is modest, with balanced accuracy near chance on that small validated set, so the tool narrows a candidate list rather than settling it. The work remains a preprint under a CC BY-NC-ND license, and no source repository accompanies it — the model is reachable through the hosted web server and the Zenodo archive rather than as installable code.