AIDA (Aligned Integrated Design for Antibodies)

AIDA — Aligned Integrated Design for Antibodies — generates antibody sequences conditioned on the structure of a target antigen. It comes from Benyamin Jamialahmadi and colleagues at the University of Waterloo, first posted to bioRxiv in July 2024 and revised in December 2024. The acronym is heavily reused in computational biology, and this model is unrelated to the similarly named AIDO family of foundation models.

Antibody design methods split roughly into two camps. Sequence-only models predict complementarity-determining regions from the rest of the antibody but ignore the antigen's shape. Co-design methods generate sequence and structure jointly, forcing the model to commit to coordinates for the CDR loops — precisely the regions whose conformations are least determined by their sequence, and where paired structural training data is scarcest. AIDA takes a middle route: it encodes the antigen structurally but decodes only sequence, sidestepping flexible loop geometry while still conditioning on the surface the antibody must bind.

The other distinguishing choice is what the model does not require. Competing conditional methods expect either a list of epitope residues or a docked antibody framework positioned against the antigen; AIDA takes the antigen alone, which matters because both of those inputs are themselves difficult predictions in a real design campaign.

#Key Features

• Antigen encoded as sequence plus conformation: A pretrained GearNet structural encoder builds a relational graph over the antigen's residues and coordinates, capturing binding-surface geometry rather than sequence alone.
• Cross-attention adapters into an antibody language model: An adapter projects GearNet's output to 256 dimensions, and a cross-attention block inserted after each of the decoder's 30 self-attention layers injects antigen context into generation.
• No supplementary inputs at inference: Epitope residues and docked frameworks are not required, unlike the co-design baselines it is compared against.
• Non-autoregressive masked generation: A Causal Masked Language Modeling objective with focal loss lets any subset of positions — one CDR, all six, or the whole variable region — be masked and filled in a single pass.
• Affinity optimization by partial remasking: Masking a few H3 positions and regenerating them turns the same model into an affinity-maturation tool with no separate optimization loop.

#Technical Details

The decoder is BALM, a bio-inspired antibody language model with 30 transformer layers, rotary positional encoding, and IMGT-numbered positions; the authors added a secondary pretraining phase on roughly 100,000 non-redundant natively paired heavy–light sequences from the Observed Antibody Space, clustered at 50% identity. The encoder is GearNet pretrained on AlphaFold structures. Training ran on SAbDab's roughly 5,000 paired antibody and nanobody structures, clustered by CDR-H3 at 40% identity, with 60 RAbD benchmark samples and their clusters held out. Pretrained weights were frozen for ten epochs then unfrozen at a tenth of the new layers' learning rate, and the cross-attention feed-forward weights were initialized to zero so the adapters start as a no-op.

On single-CDR design against the RAbD benchmark, AIDA reaches 46.41% amino acid recovery and 33.04% contact amino acid recovery on H3, against 43.20% / 27.83% for dyMEAN and 37.47% / 20.88% for DiffAb, and leads on H1 (80.56% AAR) and H2 (71.25%). Designing all six CDRs at once, it reaches 63% overall AAR, 53.67% CAAR, and 1.01 Å RMSD, versus 60.36% / 50.30% / 1.35 Å for dyMEAN. Predicting the entire variable region, it reaches 72.48% AAR and 45.15% CAAR against dyMEAN's 70.35% and 40.02%. On SKEMPI v2.0 affinity optimization, unrestricted AIDA achieves ΔΔG of −5.48 while modifying an average of 4.10 residues, against −3.27 over 4.20 residues for dyMEAN. Structural metrics come from folding the designed sequences with dyMEAN, so RMSD inherits that predictor's error.

#Applications

The intended workflow is therapeutic antibody engineering against a target whose structure is known or predictable: propose CDR sequences for a new antigen without first identifying its epitope, redesign a single loop while holding a validated antibody fixed, or improve an existing binder's affinity by regenerating a few H3 positions. An ablation study shows each component — focal loss, paired pretraining, the CMLM objective, and the antigen adapter — contributes to accuracy in at least one setting.

#Impact

AIDA's practical significance is the removal of epitope and docked-framework inputs from conditional antibody design, which lowers the assumptions a design campaign has to satisfy before the model is usable. Its limitations are equally concrete and the authors name them: paired antigen–antibody structural data remains scarce, the field lacks computational metrics that reliably track real binding, and precision depends on the structural encoder's representation of the antigen. Availability is the largest constraint — no trained AIDA checkpoint has been released. The repository ships full training and evaluation configuration plus download instructions for the two pretrained component backbones, so the model can be trained from scratch, but there are no weights to download and no licence stated.