EDGE-II

EDGE-II predicts which peptides a person's HLA class II molecules will present, and how likely those peptides are to elicit a CD4+ T-cell response. It was built at Gritstone bio, the Emeryville biotechnology company whose earlier class I algorithm, EDGE, was used to select neoantigens for its individualized and off-the-shelf cancer vaccine programs. EDGE-II is the class II counterpart to that model — a separate algorithm for a harder problem, and not a newer version of the class I predictor.

Class II presentation has lagged class I for two reasons. Training data are scarce, because class II immunopeptidomics is generated on far fewer samples, and the binding groove is open at both ends, so peptides are longer and the sequence rules governing which register is presented are looser. The practical consequence is that class II predictors have been noticeably less accurate than class I ones, even though CD4+ T-cell help is understood to be necessary for a broad, durable anti-tumor response.

EDGE-II attacks both problems at once. It transfers representations from ESM-2, a pretrained protein language model, rather than learning peptide chemistry from immunopeptidomics alone, and it replaces the heuristics prior methods use to attribute a multi-allelic mass-spectrometry hit to one of several alleles with a network that learns the attribution directly. The work was posted as a preprint in December 2024.

#Key Features

• Protein language model backbone: Peptide and allele sequences are embedded with ESM-2 and fine-tuned for presentation, so the model starts from learned protein chemistry rather than random initialization.
• Learned Allele Network: An auxiliary attention network computes per-allele weights from the sample's complete HLA genotype at inference time, operating on latent embeddings rather than on final output probabilities.
• Locus-resolved training data: The authors generated new mass-spectrometry data using HLA-DR-, DP-, and DQ-specific antibodies, which both enlarges the training set and sharpens which peptides are attributed to which locus.
• Sequence-only inputs: Predictions require only the peptide and the HLA genotype, with no gene expression or cleavage covariates, which keeps the model applicable where transcriptomic data are unavailable.
• Immunogenicity signal: Presentation scores are calibrated against ELISpot T-cell response data, so the output is usable directly for ranking candidate epitopes.

#Technical Details

The sequence encoder is initialized from ESM-2; in the head-to-head that motivated this choice, the authors compared TAPE (38M parameters, 512 hidden dimensions) against the closest ESM-2 checkpoint (35M parameters, 480 hidden dimensions) and found ESM-2 significantly better. On the Reynisson et al. class II presentation data used to train both NetMHCIIpan-4.1 and BERTMHC, adding the Learned Allele Network raised average precision to 0.848 ± 0.001 and ROC-AUC to 0.963 ± 0.003, against 0.807 ± 0.001 and 0.954 ± 0.0004 for BERTMHC's maximal-output-probability deconvolution. Adding the locus-specific immunoprecipitation data as an intermediate training stage brought final validation performance to AP 0.925 ± 0.005 and ROC-AUC 0.981 ± 0.004.

On the independent mass-spectrometry test set curated by Cheng et al., EDGE-II trained on exactly BERTMHC's data reached AP 0.684 ± 0.007 and ROC-AUC 0.898 ± 0.004, versus 0.648 and 0.889 for BERTMHC, 0.577 and 0.821 for NetMHCIIpan-4.3, and 0.442 and 0.790 for MixMHC2pred-2.0; with the locus-specific data included, AP rose to 0.775 ± 0.004 and ROC-AUC to 0.934 ± 0.002. Scoring the ELISpot peptides of Ott et al. with a sliding 15-amino-acid window, the 9 of 97 peptides scoring above 0.99 presentation probability had an expected CD4 immunogenicity of 0.41 ± 0.12, against 0.05 ± 0.04 for the lowest-scoring decile — roughly an eightfold enrichment.

#Applications

The model is aimed at neoantigen selection for cancer immunotherapy, where a vaccine designer must choose a handful of peptides from hundreds of candidate tumor mutations. In the paper it was applied to a patient with KRAS G12C-positive non-small cell lung cancer enrolled in a Phase 1/2 study of a KRAS G12C-directed vaccine. EDGE-II predicted that the patient's HLA-DQ alleles were the strongest presenters of G12C mutant peptides and identified a presentation motif; post-vaccination samples contained clonally expanded CD4+ T cells with cytotoxic transcriptional profiles recognizing those epitopes.

#Impact

EDGE-II is evidence that the transfer-learning recipe that reshaped protein modelling generally also lifts class II presentation prediction, and that learning multi-allelic deconvolution beats the hand-designed heuristics that preceded it. Its practical reach is limited by availability: no weights, code, package, or public API has been released, the locus-specific immunopeptidomics data underpinning the best results are proprietary, and the preprint is licensed CC BY-NC-ND and has not been peer reviewed. Gritstone bio filed for Chapter 11 bankruptcy protection in 2024, so the model should be read as a published research result rather than as an accessible tool.