Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AbDiffuser: Full-Atom Generation of in vitro Functioning Antibodies

Karolis Martinkus, Jan Ludwiczak, Kyunghyun Cho, Wei-Ching Liang, Julien Lafrance-Vanasse, Isidro Hotzel, Arvind Rajpal, Yan Wu, Richard Bonneau, Vladimir Gligorijevic, Andreas Loukas

TL;DR

AbDiffuser presents a diffusion-based framework for joint antibody sequence and full-atom structure generation, anchored by an SE(3)-equivariant Aligned Protein Mixer (APMixer) and physics-inspired residue projections. It introduces informative priors—position-specific residue frequencies and learned conditional atom dependencies—to reduce denoising complexity and improve generation quality. The approach achieves competitive in silico metrics and demonstrates practical, in vitro validation with HER2 binders, including expression and binding measurements on 16 designs and a best binder with pKD near 9.5. This work advances end-to-end antibody design by delivering full backbones and side chains under a single diffusion model and offers a path to applying the framework to other protein families with alignment-based representations.

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.

AbDiffuser: Full-Atom Generation of in vitro Functioning Antibodies

TL;DR

AbDiffuser presents a diffusion-based framework for joint antibody sequence and full-atom structure generation, anchored by an SE(3)-equivariant Aligned Protein Mixer (APMixer) and physics-inspired residue projections. It introduces informative priors—position-specific residue frequencies and learned conditional atom dependencies—to reduce denoising complexity and improve generation quality. The approach achieves competitive in silico metrics and demonstrates practical, in vitro validation with HER2 binders, including expression and binding measurements on 16 designs and a best binder with pKD near 9.5. This work advances end-to-end antibody design by delivering full backbones and side chains under a single diffusion model and offers a path to applying the framework to other protein families with alignment-based representations.

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.
Paper Structure (42 sections, 1 theorem, 38 equations, 5 figures, 9 tables, 2 algorithms)

This paper contains 42 sections, 1 theorem, 38 equations, 5 figures, 9 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Denoising Diffusion for Protein Generation
  3. Aligned Protein Mixer
  4. SE(3) Equivariance by Frame Averaging
  5. Handling Length Changes by Multiple Sequence Alignment
  6. The APMixer Architecture
  7. Physics-informed Residue Representation by Projection
  8. Informative Diffusion Priors
  9. AHo-specific Residue Frequencies
  10. Encoding Conditional Atom Dependencies
  11. Experiments
  12. Paired OAS Generation
  13. Generating HER2 Binders
  14. In Vitro Validation
  15. Conclusions
  16. ...and 27 more sections

Key Result

Theorem E.1

For any $f_\theta$ that is an invertible equivariant function w.r.t. a subgroup $G$ of the general-linear group GL(d,$\mathbb{R})$ the following must hold: where $c_q(f) := \left(\min_{g \in \mathrm{G}} \mathbf{E}_{Z \sim q}[\| f(g\, Z) - Z\|_t^t]\right)^{1/t}$ quantifies the expected complexity of the learned model under $q$, $W_t(p, p_\theta)$ is the Wasserstein $t$-distance of our generative m

Figures (5)

  • Figure 1: The proposed internal generic side chain representation. The dihedral-defining atoms (orange) from the full-atom representation (top) are used to construct a generic four-atom representation (bottom). If the side chain has fewer than four angles, additional atoms (gray) are placed in the generic side chain to correspond to a $180^{\circ}$ angle. The full atom representation is recovered by applying matching rotations to an appropriate side chain template.
  • Figure 2: In vitro validation of AbDiffuser designs in terms of their ability to express (left), binding affinity (center), and binding rate (right). The 'raw' column corresponds to randomly selected generated antibodies, whereas 'filtered' designs were additionally filtered by in silico screening.
  • Figure 3: HER2 structure and the 3D structures of generated binders highlighted in different colors. AbDiffuser has learned to redesign part of the binding interface while maintaining affinity. Our tightest binder achieved a pKD of 9.5 which, accounting for experimental variability, is akin to Trastuzumab whose average measured pKD was 9.21.
  • Figure 4: We define a Gaussian prior on the atom positions by learning an adjacency (conditional dependence) matrix for antibody backbone atom positions from all of the folded paired heavy and light chains in the Observed Antibody Space (OAS) olsen2022observed. Dependencies between framework residue atom positions and even correlations between heavy (top-left) and light (bottom-right) chain atom positions are distinctly captured.
  • Figure 5: Logo plots for relative amino acid frequencies in the mutated CDR H3 positions for binders (top) and non-binders (bottom) in the dataset by mason2021optimization.

Theorems & Definitions (2)

  • Theorem E.1
  • proof