LF3DRoPE

Antibody-specific epitope prediction asks a narrower question than conventional B-cell epitope prediction: not which patches of an antigen are broadly immunogenic, but which antigen residues a particular antibody contacts. Antibody-agnostic predictors such as EpitopeTransfer and the patch-centric conformational predictor score an antigen in isolation, so they cannot distinguish two antibodies that bind opposite faces of the same target. The paired task depends on three-dimensional complementarity between the antibody's CDR loops and the antigen surface, and most existing methods supply that geometry through bolt-on graph, surface, or point-cloud encoders while the attention mechanism itself still positions residues by one-dimensional sequence order.

LF3DRoPE, from Chuanliu Fan, Nan Yu, Junjie Wu, and Guohong Fu at Soochow University, takes the geometry into the positional mechanism instead. Rotary position encoding rotates query and key vectors by an angle proportional to a token offset; LF3DRoPE — Local-Frame 3D Rotary Position Encoding — replaces that scalar offset with the displacement vector between two residues, expressed in the backbone-defined local frame of the query residue. Because each displacement is written in a frame built from the residue's own N, Cα, and C atoms, the resulting attention bias is invariant to any global SE(3) transformation of the input coordinates by construction, rather than by data augmentation.

The model was posted as an arXiv preprint on 2 August 2026 and has not yet been peer reviewed. The paper cites a code repository, but that repository is not public, and no trained weights have been released.

#Key Features

• Geometry as the positional mechanism: Inter-residue displacement vectors are injected directly into rotary attention, so folded geometry shapes every attention score instead of arriving through a separate structural encoder.
• Local frames, not global coordinates: Expressing displacements in per-residue backbone frames makes predictions exactly invariant to rigid motions of the complex.
• Frozen protein language model inputs: Residue representations come from AntiBERTy for the antibody and ESM-2 for the antigen, both frozen, with one-hot amino acid features and an optional CDR-type embedding.
• Paired antibody–antigen reasoning: Cross-attention layers couple the two encoded chains, producing per-residue epitope scores conditioned on the specific antibody.
• Generalization beyond the training distribution: The single trained checkpoint ranks mutations of therapeutic antibody variants it never saw during training.

#Technical Details

The antibody and antigen are each encoded by four pre-normalized Transformer layers with hidden dimension 256, eight attention heads, feed-forward dimension 512, and dropout 0.1; five rotary frequencies along each of the three local-frame axes give 30 two-dimensional rotation channels. Two cross-attention layers then pair the encoders. Training used the AsEP benchmark — 1,723 non-redundant antibody–antigen complexes curated from AbDb, split 1,383 / 170 / 170 — optimized with AdamW at a learning rate of 1×10⁻⁴ for at most 100 epochs.

On AsEP, LF3DRoPE reaches a Matthews correlation coefficient of 0.410 ± 0.008 on the ratio split and 0.171 ± 0.010 on the harder epitope-group split, where test epitope groups are absent from training. Reported baselines include PEPNet (0.401 / 0.139), WALLE (0.210 ± 0.020 / 0.077 ± 0.015), and ESMBind (0.016 ± 0.008 / 0.002 ± 0.008). Two controls isolate the contribution of local-frame geometry: substituting standard 1D sequence-offset RoPE drops the group-split MCC to 0.159 ± 0.004, and a global-frame variant (GF3DRoPE) falls to 0.324 ± 0.011 on the ratio split. Under rigid transformations of the whole complex, LF3DRoPE's MCC change is exactly zero, while GF3DRoPE loses 0.037 ± 0.013.

#Applications

The immediate use is epitope mapping for a known antibody–antigen pair: identifying the contact residues that define a binding site, which informs escape-mutation analysis, epitope binning, and structure-guided affinity maturation. The paper's case study applies the trained checkpoint to six therapeutic targets from the IgDesign mutation benchmark and finds positive Spearman correlations between epitope scores and measured binding affinity on all six, reaching significance on the largest cohort (TSLP, ρ = 0.30, p < 10⁻⁴) — evidence that the scores carry usable signal for ranking antibody variants, not just for labeling contacts.

#Impact

LF3DRoPE's contribution is architectural: it shows that folded-structure information can enter a Transformer through the positional encoding itself, with SE(3) invariance guaranteed rather than learned, and that doing so beats both a sequence-offset baseline and a global-frame variant on the same backbone. The idea is not specific to epitopes and transfers to any paired-chain protein task. The evidence, however, rests on a single benchmark: all quantitative results come from AsEP splits, with the mutation-ranking study as the only out-of-distribution check. With code and weights not yet released and the work awaiting peer review, independent reproduction and evaluation on other antibody–antigen datasets remain open.