Diffusion-based generative model for antibody CDR design and antigen conditioning
DiffAb is an SE(3)-equivariant diffusion model that jointly generates the 3D structure and amino acid sequence of antibody CDR loops conditioned on target antigen epitopes.
Joint structure-sequence diffusion on antibody CDR-H3 and light chain loopsAntigen-conditioned generation with physical clash penaltiesBenchmark evaluations showing higher affinity and diversity than RosettaAntibodyOpen source codebase with PyTorch and PyG implementations
DiffAb is an SE(3)-equivariant diffusion model that jointly generates the 3D structure and amino acid sequence of antibody CDR loops conditioned on target antigen epitopes.
Information checked against an official source; not a hands-on test. Source · Last reviewed: 20/09/2026, 11:10:15
Key Features
Joint structure-sequence diffusion on antibody CDR-H3 and light chain loops
Antigen-conditioned generation with physical clash penalties
Benchmark evaluations showing higher affinity and diversity than RosettaAntibody
Open source codebase with PyTorch and PyG implementations
Interactive 3D Structure
DiffAb Predicted Complex
Streams real 3D atomic coordinates from RCSB Protein Data Bank
⇄ Drag to rotate · Scroll to zoom
pLDDT / B-Factor:
>90 Very high 70-90 Confident 50-70 Low <50 Very low
Academic Context & Research Evidence
Biological & Workflow Fit
Biological Application
Antibody CDR design, therapeutic optimization, and paratope engineering
Research Workflow
Target antigen PDB -> generate complementary CDR conformations and sequences
Use this citation format when referencing DiffAb in scientific publications and benchmark papers.
@software{diffab_2026,
title = {{DiffAb}},
author = {{Tsinghua University (AIR)}},
year = {2026},
url = {https://github.com/luost26/diffab},
note = {Indexed on aibioatlas - AI for Biology and Drug Discovery}
}
Peer-Reviewed Literature & Preprints
Live scientific citations streamed from Europe PMC and PubMed for DiffAb.
⏳ Fetching real-time literature from Europe PMC & PubMed...
Technical / Product Information
Missing values mean the catalog has no recorded information. They do not mean a feature is absent.
Entry typeAI Model
Access modeOpen Source
AI roleEquivariant Generative Diffusion
Input dataNot recorded
Output dataNot recorded
Licence conditionsMIT License
Commercial eligibilityOpen source release
Compute requirements1x GPU (>= 16GB VRAM)
ValidationNot recorded
TypeEquivariant diffusion model
Intended useNot recorded
CompatibilityNot recorded
ManufacturerTsinghua University
Biological applicationAntibody CDR design, therapeutic optimization, and paratope engineering
Research workflowTarget antigen PDB -> generate complementary CDR conformations and sequences