EpiTune

A B-cell epitope is the patch of an antigen that an antibody actually contacts, and locating those patches shortens the path from a pathogen sequence to a vaccine immunogen or a diagnostic antibody. Most recent predictors reach for a protein language model, and nearly all use it the same way: run ESM-2 once, freeze it, and train a separate classifier — a multilayer perceptron, a graph network, a structure-derived feature set — on the embeddings it hands back. The language model never learns anything about epitopes, and the classifier never learns anything about proteins.

EpiTune, from Pfizer's Research & Development Creation Center, deletes the second model. Low-rank adaptation (LoRA) adapters are attached to ESM-2's own query, key and value matrices and trained on per-residue epitope labels, so the network that represents the protein is also the network that emits the call. The paper's control experiment carries the argument: an identical 35M architecture trained from random initialization reaches 0.08 PR-AUC on the linear test set, against 0.33 for the same architecture fine-tuned from ESM-2 weights. The pretrained protein representation, not the architecture, is what makes the task learnable.

A second choice follows from the backbone. ESM-2 encodes position with rotary embeddings, which are relative and so not bound to the 1,024-residue window it was trained in. EpiTune trains and infers on antigens up to 1,724 residues, where the tools it is compared against truncate near 1,024 — a cutoff discarding roughly a thousand IEDB antigens.

#Key Features

• Tuned language model, not a frozen encoder: LoRA at rank 8 with lora_alpha=32 on the attention projections leaves the ESM-2 weights on disk untouched and ships the learned behaviour as a 565 KB adapter.
• Long antigens: raising the input limit from 1,024 to 1,724 residues recovers the long tail of the IEDB antigen-length distribution; variants trained only on short sequences score those antigens poorly.
• Linear and conformational variants trained separately: transferring between the two hurt both, so EpiTune-Linear and EpiTune-Conf are distinct fine-tunes of one backbone.
• No structure required: inference takes a raw amino acid sequence, skipping the crystal-structure or predicted-model lookup that structure-based predictors need first.
• Attribution as a first-class output: with one model doing the whole job, attention matrices and Captum integrated-gradient attributions map onto input residues directly.

#Technical Details

Linear and conformational versions were tuned at three ESM-2 sizes — 35M, 650M and 15B. Training data is every human B-cell epitope in IEDB as of July 2025 — roughly 180,000 linear epitopes over 8,208 antigens and 4,000 conformational over 379 — overlaid onto their parent UniProt sequences as 0/1 per-residue labels. Epitopes with more negative than positive assay records were dropped, then CD-HIT at 80% identity and MMseqs2 at 50% cut the set to 6,284 linear and 321 conformational antigens, split 90-10 into 566 linear and 35 conformational test sequences, the conformational set reaching the 32 used for evaluation after further filtering. Fine-tuning ran on 4×80GB A100s for 40 epochs at a 1e-5 learning rate, batch size 1, using PEFT-LoRA with Accelerate and DeepSpeed ZeRO-3.

On the authors' linear test set EpiTune-Linear-15B reaches 0.39 PR-AUC and 0.83 ROC-AUC, and the 35M variant 0.33 and 0.80, against 0.14/0.64 for BepiPred 3.0, 0.09/0.54 for SEMA-1D, 0.10/0.58 for the structure-based DiscoTope 3.0 and 0.06/0.50 for ScanNet. Two qualifications belong with those numbers: a 1% validation split was carved out but unused — the paper states hyperparameter tuning was not used — and because the 650M and 15B runs overfit before training completed, the compared checkpoints were selected by best PR-AUC on the test set itself. The peer models also score below their published values on this filtered set, recovering them on an unfiltered version in the supplement. Conformational prediction is the weaker half: EpiTune-Conf-15B reaches 0.08 PR-AUC and 0.74 ROC-AUC, behind SEMA-1D's 0.18 PR-AUC on the same 32 sequences.

#Applications

Epitope maps drive immunogen selection, antibody discovery and deimmunization, and are normally obtained by co-crystallography, HDX or scanning mutagenesis — slow enough that a computational shortlist changes what gets attempted. A sequence-only predictor is usable on antigens with no solved structure, and on the long multi-domain surface proteins truncating tools cannot read end to end.

#Impact

EpiTune's transferable result is methodological: on linear epitopes, tuning the protein language model itself outperforms every compared pipeline that treats it as a frozen embedding source, and the randomly initialized control shows that margin is inherited from pretraining rather than from architecture. The Apache-2.0 release includes fine-tuning, inference and checkpoint-selection scripts with the training and test splits; the complete 35M linear and conformational adapters ship in the repository, while the 650M and 15B weights come from the authors by request. The work is a preprint and has not been peer reviewed, and conformational epitopes remain open — EpiTune-Conf trails a purpose-built conformational predictor, and the authors argue the two epitope classes should be treated as separate tasks.