[Paper Review] AbDiffuser: Full-Atom Generation of in vitro Functioning Antibodies

Karolis Martinkus, Jan Ludwiczak+9 more|arXiv (Cornell University)|Jul 28, 2023

Innovative Microfluidic and Catalytic Techniques Innovation21 citations

TL;DR

AbDiffuser is an equivariant, physics-informed diffusion model that jointly generates antibody sequences and full-atom 3D structures, validated in silico and in vitro with promising binding results.

ABSTRACT

We introduce AbDiffuser, an equivariant and physics-informed diffusion model for the joint generation of antibody 3D structures and sequences. AbDiffuser is built on top of a new representation of protein structure, relies on a novel architecture for aligned proteins, and utilizes strong diffusion priors to improve the denoising process. Our approach improves protein diffusion by taking advantage of domain knowledge and physics-based constraints; handles sequence-length changes; and reduces memory complexity by an order of magnitude, enabling backbone and side chain generation. We validate AbDiffuser in silico and in vitro. Numerical experiments showcase the ability of AbDiffuser to generate antibodies that closely track the sequence and structural properties of a reference set. Laboratory experiments confirm that all 16 HER2 antibodies discovered were expressed at high levels and that 57.1% of the selected designs were tight binders.

Motivation & Objective

• Motivate and enable generation of full antibodies (heavy and light chains) with both sequence and structure in an end-to-end fashion.
• Incorporate family-specific priors and physics constraints to improve diffusion-based antibody design.
• Develop a memory-efficient architecture that handles variable sequence lengths and produces backbone and side-chain atom placements.
• Demonstrate that generated antibodies closely match reference sequence and structural properties, including experimental expression and binding evidence.

Proposed method

• Introduce AbDiffuser, an SE(3) equivariant diffusion model with a novel aligned protein representation (APMixer).
• Use a fixed-length AHo-aligned antibody representation to handle length changes and facilitate end-to-end generation of full antibodies.
• Employ a projection-based layer to enforce bond lengths and angles while operating in global coordinates during diffusion.
• Incorporate informative priors: (i) position-specific residue frequencies (AHo) and (ii) a learned Gaussian Markov Random Field capturing conditional atom dependencies.
• Frame averaging to achieve SE(3) equivariance for both positions and residue types.
• Show memory-efficiency and scalability advantages over baselines, enabling full-atom generation on a single GPU.

Experimental results

Research questions

• RQ1 Can a diffusion model jointly generate antibody sequences and full-atom 3D structures while respecting SE(3) symmetry and physics-based constraints?
• RQ2 Do family-specific priors and the AHo-aligned representation improve generation quality and efficiency compared to backbone-first or sequence-only approaches?
• RQ3 Is the model capable of producing experimentally expressible antibodies and identifying high-affinity binders without extensive post-selection?
• RQ4 How does AbDiffuser perform against state-of-the-art baselines in paired OAS generation and HER2 binder design across structure- and sequence-based metrics?

Key findings

• AbDiffuser improves generation of antibodies that closely track the sequence and structural properties of a reference set in silico.
• All 16 HER2 antibodies submitted for in vitro validation were expressed; 57.1% of designed binders showed tight binding to HER2.
• In Paired OAS experiments, AbDiffuser often outperformed baselines across multiple metrics, approaching or matching the distribution of natural antibodies on several structure- and sequence-based scores.
• Incorporating AHo priors and learned atom dependencies improves diffusion performance; using a side-chain-aware setup modestly enhances sequence naturalness and packing scores (Rosetta) for generated antibodies.
• AbDiffuser achieves a memory footprint and generation speed advantage (169M parameters, 3GB during generation, 2.3 minutes for 10 samples) versus baselines, enabling full-atom generation on a single GPU.