DiffAb is a deep generative model that designs antibody complementarity-determining regions (CDRs) using diffusion probabilistic models and equivariant neural networks. Published at NeurIPS 2022, it was the first deep learning method to generate antibodies explicitly targeting specific antigen structures, representing an early application of diffusion models to protein design.
The model jointly generates both sequences and 3D structures of CDR loops while maintaining physical constraints through equivariance—a mathematical property ensuring predictions remain valid under rotations and translations. This antigen-aware approach distinguishes DiffAb from prior methods that lacked structural context about the target.
ProteinIQ hosts DiffAb on GPU infrastructure, providing browser-based access to antibody CDR design without local installation or environment configuration.
| Input | Description |
|---|---|
Antibody-Antigen Complex | PDB file or RCSB PDB ID. Must contain antibody heavy chain, light chain, and antigen. |
Heavy chain ID | Chain identifier for the antibody heavy chain (default H). Use the chain ID as it appears in your PDB file. |
Light chain ID | Chain identifier for the antibody light chain (default L). Use the chain ID as it appears in your PDB file. |
| Setting | Description |
|---|---|
Design mode | Co-design single CDR: Generate sequence and structure for one CDR loop. Co-design multiple CDRs: Simultaneously optimize all heavy chain CDRs (H1, H2, H3). Fixed backbone: Design sequence only while keeping the backbone structure fixed. |
Target CDR | Which CDR loop to redesign. Options: HCDR1, HCDR2, HCDR3 (default), LCDR1, LCDR2, LCDR3. HCDR3 typically has the greatest impact on antigen binding specificity. |
Number of samples | How many design variants to generate (1–10, default 3). More samples provide better exploration of the design space but increase computation time. |
Random seed | Optional seed for reproducible results. Leave empty for random generation. |
DiffAb returns designed antibody structures in PDB format with:
Download the PDB files to analyze structures in PyMOL, ChimeraX, or other molecular viewers.
DiffAb uses a diffusion probabilistic model that learns to denoise antibody CDR structures through iterative refinement. Starting from random noise, the model progressively generates both amino acid sequences and atomic coordinates that fit the binding site.
The neural network architecture is equivariant under SE(3) transformations (rotations and translations), meaning it treats spatial geometry consistently regardless of how the complex is oriented. This ensures physically realistic antibody conformations.
During training, DiffAb learned from antibody-antigen complexes in the Protein Data Bank, capturing patterns in how CDRs bind to different epitopes. The model can operate in three modes:
Co-design generates sequences and structures simultaneously, allowing backbone flexibility to optimize binding geometry.
Fixed backbone designs only the sequence given a predetermined structure, useful for optimizing existing antibody scaffolds.
Multi-CDR co-design jointly optimizes multiple CDR loops, capturing cooperative effects between adjacent loops that single-CDR design misses.
DiffAb generates multiple design candidates. Evaluate them using:
The model outputs are starting points for optimization, not final therapeutics. Most designs require experimental testing and iterative refinement.
DiffAb maintains a rigid antigen structure during design. Real antibody-antigen interfaces often involve conformational changes upon binding that the model cannot capture.
The framework chain IDs must be specified correctly. If your PDB uses non-standard chain naming, the tool may fail to identify antibody regions properly.
Generated sequences may contain developability liabilities. The model optimizes for structural fit to the antigen, not manufacturability, stability, or immunogenicity. Additional filtering and engineering are typically required before therapeutic development.
