[Paper Review] De novo antibody design with SE(3) diffusion

Daniel Cutting, Frédéric A. Dreyer+3 more|arXiv (Cornell University)|May 13, 2024

Monoclonal and Polyclonal Antibodies Research10 citations

TL;DR

IgDiff is an antibody variable domain diffusion model that generates designable, novel backbone structures and demonstrates experimental expression, outperforming state-of-the-art backbone diffusion models on antibody design tasks.

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.

Motivation & Objective

• Motivate de novo protein design via structure-based generative modeling for antibodies.
• Extend SE(3) diffusion to handle paired heavy and light chain variable regions.
• Assess designability, novelty, and experimental expressibility of generated antibodies.
• Compare IgDiff to state-of-the-art backbone diffusion models on antibody engineering tasks.
• Provide design tasks (CDR design, light chain pairing, CDRH3 length changes) and evaluate performance.

Proposed method

• Adopt SE(3) diffusion framework on antibody backbones with paired heavy/light chains (SE(3)^N).
• Represent backbones as frames T_i in SE(3) built from N*, C_alpha*, C, O coordinates and torsion psi in SO(2).
• Use a Riemannian score-based model with a score network s_theta trained by denoising score matching loss (L_DSM).
• Decompose loss into translation and rotation components (L_DSM = L_DSM^x + L_DSM^r) and add auxiliary losses L_bb and L_2D near t≈0 to improve fine-grained backbone details.
• Train on synthetic antibody structures from OAS (ABB2-predicted) and fine-tune FrameDiff weights; include chain-break encoding and chain-type masking.
• Perform sampling via Euler–Maruyama on SE(3)^N with geodesic walk, reconstructing O atoms from predicted torsion psi; use AbMPNN for inverse folding to predict sequences.

Figure 1: Left: Schematic representation of an antibody. Centre: Backbone atoms of the variable region, showing both a heavy (green) and light (blue) chain domain. Right: The parametrisation of residues into frames used by the diffusion model, with each frame consisting of four heavy atoms connected

Experimental results

Research questions

• RQ1 Can IgDiff generate plausible, designable, and novel paired heavy/light chain antibody variable regions?
• RQ2 How does IgDiff compare to existing backbone diffusion models (e.g., RFDiffusion) on antibody design tasks?
• RQ3 What is the designability and novelty (dihedral distributions, CDR loop canonical forms) of generated antibodies?
• RQ4 Are IgDiff-generated antibodies expressible with high yield and do they maintain structural integrity under sequence recovery via antibody-specific inverse folding?
• RQ5 How effective is conditioning for design tasks like all CDR design, light-chain design, and CDRH3-length changes?

Key findings

• IgDiff produces highly designable antibodies with novel binding regions and backbone dihedral angles in good agreement with reference antibody distributions.
• Generated antibodies show Ramachandran distributions matching OAS-derived structures and achieve scRMSD < 2 Å across regions for most samples.
• Approximately 88% of generated antibodies pass region-wise designability tests; 79.1% pass RMSPE-based consistency check across full variable domains.
• Unconditioned IgDiff designs show notable novelty in CDR H3 and diversity among CDR loops, with a mean CDR H3 RMSD to closest training match around 1.39–1.50 Å.
• Experimental validation: 28 selected antibody sequences expressed with high yield, confirming practical viability of IgDiff designs.
• In conditional design tasks, IgDiff outperforms RFDiffusion on self-consistency (scRMSD) and canonical cluster matching across most regions; light-chain design and CDRH3-length changes show substantial improvements.

Figure 2: Examples of IgDiff generated antibody structures. Light chains are highlighted in blue, heavy chains in green. (A-B) Unconditionally generated antibodies. (A) IgDiff generated antibody (dark green/dark blue) compared to ABodyBuilder2 prediction on the lowest self-consistency RMSD sequence