How to Use DiffAb Antibody Design

Overview. DiffAb is a diffusion-based generative model for antigen-specific antibody design. The NeurIPS paper frames antibody generation around the complementarity-determining regions (CDRs), jointly modeling CDR sequence and three-dimensional structure while conditioning the design on the target antigen structure.

Use Cases. This makes DiffAb relevant for early antibody discovery when the antigen surface is known but a high-quality binder has not yet been established. The original model was designed as a multipurpose system capable of CDR sequence-structure co-design, sequence design for fixed backbones, and optimization-style refinement, although the Neurosnap workflow is focused on antigen-conditioned candidate generation from an Input Antigen structure.

Workflow on Neurosnap. Researchers upload the antigen PDB and launch design in a browser-based workflow without local setup. Experiment Seed can be fixed for reproducibility or varied to probe different parts of the generative space. The output is a candidate panel of designed antibody structures and sequences for downstream review.

How It Works. DiffAb uses diffusion probabilistic modeling together with equivariant neural networks so that sequence and structure are generated in a way that respects three-dimensional geometry. That is important in antibody design because antigen recognition depends on the coupled shape and chemistry of the CDR loops rather than on sequence motifs alone.

Scientific Context. In the original study, DiffAb was one of the earliest diffusion-based protein-structure design models and the first deep-learning method explicitly aimed at generating antibodies against specific antigen structures. The paper reports competitive designed candidates by biophysical energy functions and related protein-design metrics, supporting its use as a serious structure-conditioned generator rather than a purely illustrative model.

Workflow. On Neurosnap, DiffAb is best used as the front end of an antibody discovery workflow. Researchers should inspect generated candidates for CDR plausibility, antigen-surface complementarity, heavy- and light-chain sequence diversity, and structural reasonableness before moving shortlisted designs into affinity prediction, developability screening, or experimental validation.

• Antigen-conditioned design: DiffAb starts from target structure, which is the correct regime for epitope-aware antibody generation.
• Joint sequence-structure modeling: CDR residues and geometry are designed together instead of being treated as independent sequence edits.
• Reproducible exploration: Experiment Seed makes it easy to repeat a run exactly or deliberately sample a different candidate set.
• Strong pipeline fit: Generated antibodies can move directly into structural review, affinity scoring, developability analysis, and wet-lab screening.