Antigen-Specific Antibody Design and Optimization with Diffusion-Based Generative Models for Protein Structures

Abstract

Antibodies are immune system proteins that protect the host by binding to specific antigens such as viruses and bacteria. The binding between antibodies and antigens are mainly determined by the complementarity-determining regions (CDR) on the antibodies. In this work, we develop a deep generative model that jointly models sequences and structures of CDRs based on diffusion processes and equivariant neural networks. Our method is the first deep learning-based method that can explicitly target specific antigen structures and generate antibodies at atomic resolution. The model is a ''Swiss Army Knife'' which is capable of sequence-structure co-design, sequence design for given backbone structures, and antibody optimization. For antibody optimization, we propose a special sampling scheme that first perturbs the given antibody and then denoises it. As the number of available antibody structures is relatively scarce, we curate a new dataset that contains antibody-like proteins as a complement to the original antibody dataset for training. We conduct extensive experiments to evaluate the quality of both sequences and structures of designed antibodies. We find that our model could yield highly competitive results in terms of binding affinity measured by biophysical energy functions and other protein design metrics.

Protein Sequence and Structure Co-Design with Equivariant Translation

“…Note that context features vary from setting to setting. For example, in antibody CDR design (Jin et al, 2021;Luo et al, 2022), they are derived from antibody framework and binding antigen structures with CDR region masked, while in full protein design (Anand & Achim, 2022), they can be secondary structure annotations and residue-residue contact features.…”

“…We then adopt a variant of Invariant Point Attention (IPA) (Jumper et al, 2021) called SeqIPA to capture the interplay of residue types, residue structures and context features, integrating them all altogether into updated context features. Such a practice is often favored in literature (Anand & Achim, 2022;Luo et al, 2022;Tubiana et al, 2022) as it is aware of the orientation of each residue frame while being roto-translation invariant to input and output features. Distinct from vanilla IPA, our SeqIPA takes residue types as the additional input to bias the attention map and steer the representation of the whole protein generated so far:…”

“…Therefore, efforts have been made to develop models that co-design the sequence and structure of proteins . As a pioneering work, Jin et al (2021) propose an autoregressive model that co-designs the Complementarity Determining Regions (CDRs) sequence and structure of antibodies based on iterative refinement of protein structures, which spurs a lot of follow-up works (Luo et al, 2022;Kong et al, 2022). Nevertheless, these approaches are restricted to proteins with specific domain topologies, i.e., antibodies, and often suffer from high inference costs due to autoregressive sampling or annealed diffusion sampling (Song & Ermon, 2019;Luo et al, 2022).…”

“…As a pioneering work, Jin et al (2021) propose an autoregressive model that co-designs the Complementarity Determining Regions (CDRs) sequence and structure of antibodies based on iterative refinement of protein structures, which spurs a lot of follow-up works (Luo et al, 2022;Kong et al, 2022). Nevertheless, these approaches are restricted to proteins with specific domain topologies, i.e., antibodies, and often suffer from high inference costs due to autoregressive sampling or annealed diffusion sampling (Song & Ermon, 2019;Luo et al, 2022). Very recently, Anand & Achim (2022) propose another diffusionbased generative model (Ho et al, 2020) for general protein sequence-structure co-design, where they adopt three diffusion models to generate structures, sequences, and rotamers of proteins in sequential order.…”

Antibody optimization enabled by artificial intelligence predictions of binding affinity and naturalness

“…The general approach involves training a model on experimental data and applying it to predict which sequences are most likely to improve the measured trait. Several promising approaches have been proposed [8][9][10][11][12][13][14], but only two studies have had in silico predictions validated in the lab [15,16]. While being valuable demonstrations, previous models are limited by throughput and the use of binary (rather than continuous) readouts, which can compromise their accuracy at high mutational loads.…”