RADAb

The six complementarity-determining regions of an antibody are loops, and CDR-H3 in particular is the least constrained of them — variable in length, sequence and conformation, and the region that most determines what an antibody binds. Generative models that write those loops from scratch have to learn that variability from SAbDab, a database of fewer than ten thousand antigen-antibody complexes, which is small enough that a model can overfit it and still produce sequences that no natural repertoire would contain.

RADAb takes the opposite route from de novo generation: it looks the loop up first. Before generating anything, it searches the whole Protein Data Bank for CDR-like fragments — short linear motifs from any protein, antibody or not, whose backbone geometry matches the CDR loop being designed within an RMSD threshold. Proteins that fold into the same local shape tend to share sequence consensus, so a set of such fragments carries evolutionary information about which residues that shape tolerates, drawn from a structure space vastly larger than the antibody databases. Those retrieved sequences become conditions on a discrete diffusion process that denoises the CDR sequence in place on the fixed antigen-bound backbone.

The model comes from Shuangjia Zheng's group at the Global Institute of Future Technology, Shanghai Jiao Tong University, with the School of Software and Microelectronics at Peking University, and was published at ICLR 2025. Its denoising architecture follows DiffAb, a structure-conditioned antibody diffusion model from a separate group, but RADAb's dual-branch network and its retrieval machinery are trained as its own weights rather than layered onto a borrowed checkpoint.

#Key Features

• Structural retrieval instead of sequence homology: Fragment search runs through MASTER on backbone coordinates alone, so no sequence information leaks into the retrieved conditions and non-antibody proteins can supply usable motifs.
• Dual-branch denoiser: A global branch encodes the antigen-antibody complex with invariant point attention; a local branch reads the retrieved motifs as a pseudo-MSA. A skip connection fuses the two probability distributions before the softmax.
• Tied row attention over retrieved motifs: The local branch borrows the MSA Transformer's tied row attention, which shares attention scores across rows of structurally similar fragments and keeps memory use low.
• Binding-energy optimization, not just recovery: The same sampling procedure re-writes an existing antibody's CDRs toward lower interface energy, treating a known binder as the starting template rather than a blank slate.

#Technical Details

RADAb uses residue features of dimension 128 and pair features of dimension 64 through six IPA layers, with a frozen ESM-2 650M encoder supplying 1280-dimensional evolutionary embeddings of the partially noised antibody sequence at each timestep. The local branch stacks two axial attention layers (embedding dimension 384, hidden dimension 1536, six heads) over a pseudo-MSA whose first row is the framework sequence and whose remaining rows are the top 15 retrieved fragments — an ablation on the number of retrieved motifs peaks at 15 and degrades beyond it. Training minimizes a KL divergence between the forward posterior and the network's prediction over SAbDab complexes clustered at 50% CDR-H3 identity to keep training and test sets disjoint, run separately for CDR-H3 (100,000 iterations) and jointly for the other five regions (250,000 iterations) on a single RTX 4090.

On the 63-complex SAbDab test set, amino acid recovery reaches 57.0% on CDR-H3 and 76.6% on CDR-H1, against 49.2% and 74.9% for a retrained DiffAb; on CDR-H3 loops longer than 14 residues, where all methods fall off, it recovers 51.4%. Sequences refolded with ABodyBuilder2 and relaxed in PyRosetta give a mean ΔΔG of 109.16 and improve on the native sequence's binding energy in 37.3% of designs.

#Applications

The natural use is affinity maturation of an antibody whose complex structure is already solved: fix the backbone, hand the model the antigen context, and sample CDR variants ranked by predicted interface energy. Because retrieval works on backbone geometry, the same machinery applies to any of the six loops rather than only the hypervariable H3, and to therapeutic leads where the framework must be preserved.

#Impact

RADAb is the first retrieval-augmented generative framework for antibody design, and it demonstrates that the PDB's non-antibody structure space can substitute for the antibody-specific training data that these models otherwise lack. It has since become a standard CDR-H3 baseline for later generative protein models, including LOGOS. The authors state plainly that the work has not been validated in wet-lab experiments — all results are in-silico recovery, self-consistency RMSD, language-model plausibility and PyRosetta energies. The released code covers inference and evaluation; no trained checkpoints accompany it.