Antibody Co-Design

Unlike traditional methods that generate a sequence and then predict its structure, co-design models produce sequence and 3D coordinates simultaneously. This coupled process uses a shared latent space where mutations in sequence space immediately propagate to structural adjustments, and vice versa. The model learns the biophysical energy landscape of protein folding, ensuring every generated residue has a physically plausible local environment. This eliminates the common failure mode where a high-affinity sequence is predicted but cannot actually fold into the required paratope conformation.

Co-design models leverage SE(3)-equivariant neural networks that respect the rotational and translational symmetries of 3D space. Key architectural components include:

• Tensor field networks that operate on point clouds of atoms
• Equivariant message passing that updates node and edge features while preserving geometric relationships
• Invariant attention mechanisms that focus on geometrically meaningful residue interactions This ensures that rotating or translating the entire antibody-antigen complex does not change the predicted sequence or binding properties, a critical inductive bias for generalization.

The complementarity-determining region H3 (CDR-H3) loop is the most variable and functionally critical part of the antibody. Co-design models must accurately sample its vast conformational space. Modern approaches use:

• Denoising diffusion models that gradually refine random coordinates into valid loop geometries
• Flow matching on the Riemannian manifold of protein backbone torsion angles
• Cyclic coordinate refinement that alternates between sequence optimization and local structure relaxation This enables the generation of novel CDR-H3 conformations not seen in natural repertoires, accessing previously unexplored binding geometries.

Co-design models condition the generation process on the 3D structure of the target antigen epitope. The model learns to sculpt the antibody's paratope surface to achieve shape complementarity and favorable interaction energy with the target. Key conditioning mechanisms include:

• Cross-attention between antigen residue embeddings and antibody generation states
• Interaction fingerprints encoding hydrogen bond patterns, hydrophobic contacts, and salt bridges
• Surface-aware convolutions that operate on the solvent-accessible surface of the antigen This ensures the generated antibody is not just physically plausible but functionally targeted to the desired epitope.

Therapeutic antibodies must satisfy multiple competing objectives beyond binding affinity. Co-design frameworks incorporate multi-objective optimization to balance:

• Affinity (binding free energy ΔG)
• Specificity (discrimination against off-target homologs)
• Developability (aggregation propensity, thermal stability, solubility)
• Immunogenicity (T-cell epitope burden) The model navigates the Pareto frontier, presenting designers with a family of optimal trade-off solutions rather than a single point estimate. This shifts decision-making from post-hoc filtering to integrated design.

Co-design is inherently iterative. Generated candidates are validated through a cascade of computational assays before experimental testing:

• Rosetta energy scoring for physics-based stability assessment
• AlphaFold-Multimer for independent fold and binding verification
• Molecular dynamics simulations for conformational stability under thermal motion
• Aggrescan and CamSol for developability profiling Candidates that fail any checkpoint are fed back as negative examples, enabling active learning loops that progressively sharpen the generative model's understanding of viable design space.