DuaDeep-SeqAffinity

An antibody's specificity lives in six short hypervariable loops — three on the heavy chain, three on the light — set into a germline framework nearly identical from one antibody to the next. Concatenate an antigen and both antibody chains into one string for a single encoder, and that invariant framework dominates by sheer length, diluting the complementarity-determining region (CDR) variation that decides how tightly the pair binds. DuaDeep-SeqAffinity is organised around avoiding that dilution: antigen, heavy chain and light chain are embedded and encoded in three independent streams that meet only in the last layer before regression.

Inside each stream two branches read the same embeddings in parallel — a Transformer encoder for long-range evolutionary context, and a 1D convolutional network for local physicochemical motifs — and their pooled outputs are concatenated. The representation layer is a frozen ESM-2 checkpoint, never updated during training; what is fitted is the shared projection, the per-stream branches and the regression head. The result is a structure-free predictor: no docking, no predicted complex, and none of the error that structure-based pipelines inherit from the flexible CDR-H3 loop, which remains hard for AlphaFold 3 and Boltz-2 in the antibody-antigen setting.

The model comes from ENSIA, the National School of Artificial Intelligence in Algiers, with a clinical co-author at CHU Bab El Oued. It was posted to arXiv in December 2025 and revised in September 2026.

#Key Features

• Tri-stream separation: The three chains are never merged before embedding; each gets its own forward pass and feature extractor, preserving chain-specific signal up to late fusion.
• Two branches per stream: A Transformer encoder and a 1D CNN read the same projected embeddings in parallel, so global fold context and local paratope motifs are extracted at their own scales rather than competing inside one backbone.
• Small trainable head on a frozen backbone: About 7.11M trainable parameters sit on a frozen 35M-parameter ESM-2 encoder, and a training run fits in roughly 1.6 GB of GPU memory.
• Sequence-disjoint evaluation: Unique antibody and antigen sequences are split before any pair is assigned, so a test pair counts only when neither partner appeared in training — the strictest Park and Marcotte regime.
• Residue-level interpretability: Averaged self-attention and gradient saliency both localise to CDR1, CDR2 and CDR3 on the heavy chain, while antigen saliency picks out candidate epitope residues.

#Technical Details

Each sequence is embedded by the frozen esm2_t12_35M_UR50D checkpoint into 480-dimensional residue features, projected to 256 dimensions by a shared linear layer, and read in parallel by a two-layer, eight-head Transformer encoder (feed-forward width 1024) and two Conv1d layers with kernel sizes 3 and 5. Average pooling gives a 256-dimensional global vector and a 128-dimensional motif vector per stream, concatenated to 384; the three then concatenate to 1152 and pass through a 1152-1024-512-1 multilayer perceptron. Training uses AdamW at a constant learning rate of 1e-4, batch size 16, smooth L1 loss, FP16 precision and a 512-token cap, for 10 epochs.

Training data is the AbRank benchmark, whose Kd values are filtered to a physically meaningful range and converted to pKd = 9 - log10(Kd). Across four random seeds the model reaches RMSE 0.7366, MAE 0.5147, R-squared 0.460, Pearson 0.683, Spearman 0.683 and pairwise ranking AUC 0.895. Both single-branch ablations trail it on every metric — ESM-2 plus Transformer at Pearson 0.613 and AUC 0.850, ESM-2 plus CNN at 0.636 and 0.856 — with paired t-tests across the matched seeds at p < 0.05 in every comparison. Ranking figures quoted from other publications, not recomputed on this split, include MINT at 0.775 and the structure-aware WALLE-Affinity at 0.866.

#Applications

The target use is pre-screening, not final measurement: ranking antibody repertoires or mutant panels by predicted affinity so that surface plasmon resonance or bio-layer interferometry is spent only on the top of the list. Needing neither a solved complex nor a predicted one, it applies to novel antigens with no structural template. The saliency maps give antibody engineers a second output: a per-residue ranking on the antigen that can seed epitope hypotheses before structural work begins.

#Impact

The contribution that generalises beyond this dataset is architectural: decoupling local and global feature extraction, and keeping the three chains apart until fusion, beats folding them into one monolithic encoder, and the paired seed-level statistics make that systematic rather than a lucky run. An R-squared of 0.460 places the model as a ranking and triage tool rather than a quantitative affinity oracle, which is how its authors position it. The limits are explicit: average pooling discards residue-to-residue coordination, there is no cross-attention between antibody and antigen, the split is sequence-disjoint but not homology-aware, and the interpretability evidence is example-level. No source code or trained weights accompany the preprint — its availability statement releases only the AbRank dataset.