Abstract: We introduce IgDiff, an antibody variable domain diffusion model based on a general protein backbone diffusion framework which was extended to handle multiple chains. Assessing the designability and novelty of the structures generated with our model, we find that IgDiff produces highly designable antibodies that can contain novel binding regions. The backbone dihedral angles of sampled structures show good agreement with a reference antibody distribution. We verify these designed antibodies experimentally and find that all express with high yield. Finally, we compare our model with a state-of-the-art generative backbone diffusion model on a range of antibody design tasks, such as the design of the complementarity determining regions or the pairing of a light chain to an existing heavy chain, and show improved properties and designability.
Introduction
Antibodies are essential components of the immune system and represent a premier class of therapeutics for treating cancer, autoimmune disorders, and infectious diseases. Designing these molecules de novo—from scratch—remains a significant challenge in computational biology. Traditional methods often struggle with the immense conformational space of hypervariable loops (CDRs) and the complex orientation between heavy and light chains.
Researchers have introduced IgDiff, a generative diffusion model specifically tailored for the variable domains of antibodies. By adapting the SE(3) diffusion framework to handle multiple protein chains and fine-tuning it on a massive synthetic dataset, IgDiff generates novel, structurally sound antibody backbones that outperform general-purpose models.
Methods: SE(3) Diffusion for Multiple Chains
IgDiff is built upon the FrameDiff architecture, a diffusion model that operates on the SE(3) group (the group of all rotations and translations in 3D space). Unlike models that predict raw atom coordinates, IgDiff models the "frames"—the position and orientation—of each amino acid residue's backbone.
To make this framework suitable for antibodies, the authors introduced several key innovations:
- Multi-chain Support: The framework was extended to handle the interaction between heavy and light chains by incorporating "jumps" in residue indexing and specific chain labels, allowing the model to learn the relative orientation of the two domains.
- Design-Sequence-Fold Pipeline: The workflow follows a rigorous validation loop. First, IgDiff generates a 3D backbone. Then, AbMPNN (an inverse folding model) predicts an amino acid sequence for that backbone. Finally, ABodyBuilder2 folds the sequence back into a 3D structure to check for "self-consistency." If the folded structure matches the original design (scRMSD < 2Å), the design is considered "designable."
Specialized Training on Synthetic Data
The model was trained on a subset of the Observed Antibody Space (OAS). Due to the relative scarcity of experimental crystal structures, the authors utilized 148,832 synthetic structures predicted by ABodyBuilder2. This approach allowed the model to learn from a significantly wider variety of antibody configurations than would be possible using only the Protein Data Bank (PDB). Fine-tuning on this high-quality synthetic data enabled the model to capture antibody-specific distributions, such as canonical loop clusters and Ramachandran dihedral angles, more accurately than general protein models.
Key Results and Experimental Validation
IgDiff demonstrates superior performance across both unconditional and conditional design tasks. In unconditional generation, 88% of IgDiff’s designs passed the self-consistency threshold, indicating a high likelihood of successful folding.
Beyond digital metrics, the team synthesized 28 AI-designed antibodies in the lab. Remarkably, 100% of these designs expressed with high yield (80–130 mg/L), proving that IgDiff produces biologically viable proteins rather than "digital hallucinations."
In comparative benchmarks against the state-of-the-art RFDiffusion, IgDiff showed a significant advantage in specialized tasks. For example, in a "CDR-H3 length change" task, IgDiff achieved a 74% success rate, whereas RFDiffusion managed only 6%. This highlights the importance of antibody-specific fine-tuning for maintaining the geometry of critical binding regions.
Limitations and Future Directions
While IgDiff represents a major advancement, it currently focuses solely on the protein backbone, relying on external tools for sequence prediction and side-chain packing. Future iterations may move toward "full-atom" diffusion for increased precision. Additionally, while this work focuses on de novo design and inpainting, the next frontier is antigen-conditioned design, where the model is tasked with generating an antibody to bind a specific target of interest, such as a viral protein.
Reference: Cutting, D., et al. (2024). De novo antibody design with SE(3) diffusion. arXiv:2405.07622v1 [q-bio.BM].
Summary updated