T-SCAPE

Predicting whether a peptide will actually provoke a T-cell response is bottlenecked by data scarcity. The functional assays that measure the endpoint — ELISPOT and related cytokine-release readouts — are slow and expensive, so the datasets that matter most are the smallest. Most in-silico tools work around this by modeling one upstream step of the antigen-presentation cascade — peptide-MHC binding or TCR-pMHC recognition — and treating that proxy as a stand-in. That substitution is leaky: strong MHC binders routinely fail to elicit cytokine release.

T-SCAPE (T-cell immunogenicity Scoring via Cross-domain Aided Predictive Engine) takes the opposite approach, pretraining a single shared encoder across five immunologically distinct data domains — MHC class I and II presentation, peptide-MHC binding affinity, TCR-pMHC interaction, source-organism "selfness," and T-cell activation — before fine-tuning on the small activation dataset that defines the real target. It was developed by Chaok Seok's group at Seoul National University with Galux, posted to bioRxiv in May 2025 and published in Science Advances in December 2025.

The headline result is a transfer effect the authors attribute directly to that pretraining: the fine-tuned model scores the anti-drug antibody (ADA) risk of therapeutic antibodies without receiving any MHC input at all, a setting it was never supervised on. Where entries such as mhcSFM and DecoderTCR each sharpen one link in the cascade, T-SCAPE targets the composite outcome and lands nearer the deimmunization problem addressed by PEARL.

#Key Features

• Five-domain multi-task pretraining: A shared integrative encoder is trained simultaneously on presentation, affinity, TCR recognition, and species-of-origin tasks, then fine-tuned on T-cell activation data.
• Adversarial domain adaptation: An adversarial loss trains the shared encoder to confuse a task discriminator, while orthogonality constraints separate shared from task-specific representations — suppressing dataset artifacts such as the peptide-length gap between class I and class II corpora.
• MHC-free antibody ADA scoring: Gated attention lets the model ignore padded inputs, so a full-length antibody can be scored by sliding-window 9-mer segmentation with no HLA allele specified.
• Leakage-controlled evaluation: Training peptides sharing a 9-mer with any benchmark peptide at fewer than two mismatches were removed, closing intra- and inter-task leakage paths through the cascade.
• Broad peptide and allele coverage: Handles peptides up to 20 residues across MHC class I and class II, including HLA-DP, -DQ, and -DR, where competing tools impose 8-11 or 8-14 residue limits.

#Technical Details

T-SCAPE is a convolutional network rather than a transformer, built on ByteNet encoder-decoder blocks with kernel size 3 — a deliberate choice given that immunogenicity tasks operate on short 9-mer fragments where local motifs dominate over long-range dependencies. The integrative encoder embeds the peptide plus MHC class I and class II pseudo-sequences, combined by gated attention; MLP decoders emit per-task predictions and a separate MLP discriminator drives the adversarial objective. Pretraining used 557,151 class I and 206,055 class II eluted-ligand mass-spectrometry entries from IEDB, 129,973 class I and 516,463 class II binding-affinity measurements, 65,030 curated TCR-pMHC pairs, and human versus non-human sequence tasks drawn from OAS and UniProt; fine-tuning used just 12,782 T-cell activation points. The model has 6 layers at dimension 280.

On the infectious-disease vaccine benchmark it reached precision 0.916 and PR-AUC 0.904 against 0.791 and 0.760 for BigMHC-IM, and AUROC 0.784 against 0.569 for BigMHC-ELIM. On a leakage-free class II test set of 777 peptides it scored mean PPVn 0.633, precision 0.611, PR-AUC 0.710, and AUROC 0.678, ahead of NetMHCIIpan-4.1. On 216 FDA-approved antibodies at a 10% ADA threshold it reached PR-AUC 0.743 versus 0.607 for the next-best method. Neoantigen discovery was the weakest setting, with mean PPVn of 0.424 marginally below BigMHC-IM's 0.429. An identical model trained without pretraining collapsed on the ADA task, predicting every antibody non-immunogenic.

#Applications

The intended use is triage: ranking candidate epitopes for neoantigen cancer vaccines and infectious-disease vaccines, and flagging immunogenic hotspots during deimmunization of biologics. Because ADA scoring needs no HLA input, antibody engineers can screen humanized candidates directly from sequence, and the authors demonstrate reduced predicted ADA scores across deimmunized variants. Weights and inference code are public, and a hosted web server accepts sequence input.

#Impact

T-SCAPE makes a concrete case that multi-domain pretraining, not merely more data on the target task, is what buys generalization in immunogenicity prediction — the ablation against an identical non-pretrained model is its strongest evidence. Its leakage-control protocol and the new leakage-free class II benchmark are contributions in their own right, since inflated scores from cascade-induced overlap are endemic in this literature. The authors are explicit that the model is an auxiliary screening tool rather than a replacement for experimental validation: performance remains uneven across HLA alleles, class II activation data is thin, and the code and weights are released under CC BY-NC 4.0, which rules out commercial use without a separate license.