RAAD

RAAD — Relation-Aware Antibody Design — co-designs the sequence and the three-dimensional structure of an antibody's complementarity-determining regions against a given antigen. It was developed in the AI Lab of the Research Center for Industries of the Future at Westlake University with Zhejiang University, released as a preprint in December 2024 and published at AAAI 2025.

Two assumptions in prior CDR designers limit them in practice. The first is that the antigenic epitope is known and supplied as input, which is exactly the information a designer usually lacks; the second is that antigen–antibody contacts can be treated as a static graph fixed at construction time. RAAD drops both. It takes the complete antigen, builds an attributed heterogeneous graph over antigen and antibody residues with several typed edge relations, and updates those relations layer by layer, so the interaction pattern is learned rather than imposed.

The paper also argues that antibody optimization has been measured wrongly. Iterative affinity optimization tends to produce "universal antibodies" that bind the target and unrelated antigens with high affinity — a gain in the reported metric and a loss in the property that actually matters. RAAD introduces an SP-score that contrasts binding energy against the target with binding energy against off-target antigens, plus a contrastive constraint that trains for it directly.

#Key Features

• Epitope-unknown design: The full antigen is the input, so no epitope needs to be identified in advance; the model attends to the binding region itself.
• Relation-aware message passing: Node and edge features are E(3)-invariant and messages are passed separately per edge relation type, giving the network more contextual and geometric information than a single homogeneous graph provides.
• Dynamic interaction modelling: Antigen–antibody edge relations are re-estimated at every layer rather than fixed at graph construction, which the ablation identifies as one of the two largest contributors to performance.
• One-shot co-design: Sequence and structure are produced in a single forward pass, without autoregressive decoding or iterative structure refinement.
• Specificity-aware optimization: A contrastive constraint based on mutual information between the CDR embedding and the target antigen embedding steers optimization toward specific, not merely tight, binders.

#Technical Details

RAAD is trained on the Structural Antibody Database, using 3,127 antigen–antibody complexes after removing structures lacking light chains or antigens. CDR sequences are clustered with MMseqs2 at 40% identity — yielding 765, 1,093, and 1,659 clusters for CDR-H1, H2, and H3 — and split 8:1:1 for training, validation, and test. On ten-fold cross-validation, RAAD reaches amino-acid recovery of 67.64% at RMSD 0.77 Å for CDR-H1, 58.87% at 0.72 Å for CDR-H2, and 37.71% at 1.91 Å for CDR-H3, improving on dyMEAN by 6.49%, 6.24%, and 1.67% in recovery and reducing RMSD by 8.33%, 12.20%, and 8.61%.

On the RAbD benchmark of 60 complexes, RAAD designs antigen-binding CDR-H3s with 41.26% amino-acid recovery, TM-score 0.9874, and RMSD 1.46 Å, against 39.14%, 0.9825, and 1.66 Å for dyMEAN. For optimization, the generator is fine-tuned on 53 antigen–antibody complexes from SKEMPI v2.0 with iterative target augmentation: without the specificity constraint it reaches ΔΔG of -12.163 kcal/mol against the target but -1.748 against off-target antigens, for an SP-score of 10.415; with the constraint, off-target binding falls to -0.238 while on-target binding holds at -11.787, raising the SP-score to 11.549 and cutting the number of mutated residues from 6.50 to 5.98. Supplying only a pre-given epitope instead of the full antigen costs 1.65% in modelling recovery and 4.67% in generation recovery.

#Applications

RAAD is aimed at therapeutic antibody engineering, where the two tasks it covers — proposing CDRs against a target of interest and maturing an existing antibody — are the routine computational steps. Because it does not require an epitope, it can be pointed at a newly characterized antigen structure directly, and its advantage over baselines widens as CDR-H3 length grows, which is where design is hardest and most valuable. The specificity metric and constraint are reusable independently of the generator, giving groups working on affinity maturation a way to check that optimization has not traded away selectivity.

#Impact

RAAD's contribution is less a new backbone than a reframing: it shows that the contextual information discarded by epitope-conditioned, statically-graphed designers is worth real accuracy, and that specificity needs its own objective. The authors are candid about the remaining limits — optimization depends on the quality of an external ΔΔG predictor, and the study of pretraining strategies is preliminary. Code is released under an MIT licence; no pretrained checkpoint accompanies it, so reproducing results requires training from the SAbDab data pipeline in the repository.