The unavailability of large-scale datasets hinders the ground-truth-based benchmarking of antibody-antigen binding prediction. We developed the Absolut! software that enables generation of synthetic lattice-based 3D-antibody-antigen binding structures with ground-truth access to conformational paratope, epitope, and affinity. We confirmed that accuracy-based rankings of ML methods trained on experimental data hold for ML methods trained on Absolut!-generated data. The Absolut! framework enables real-world relevant benchmarking of ML strategies for biotherapeutics design.

Specificity, chemical stability, and manufacturability are important facets for an antibody therapeutic. We combine auto-regressive modeling on sequences from NGS repertoires and functional output across diverse targets, with models for biophysical property prediction to generate novel synthetic libraries with improved developability. We discuss the application of this approach to the development of Adimab’s HCab platform generating high-affinity developable output across therapeutically relevant antigens, in a traditionally difficult modality.

Computational approaches are transforming therapeutic antibody discovery on speed and success. Among them, knowledge-based machine learning becomes a powerful tool enabling virtual screening of thousands of sequences for desired properties. This talk will discuss the development of prediction models, as well as the effective utilization of prediction algorithms to speed up quality hits discovery.

Current antibody development methods are limited by the availability of large datasets of antibody-antigen binding data. In this talk, we demonstrate how large datasets of multi-dimensional antibody-antigen data and associated machine learning models enable the discovery and optimization of antibodies with desired affinity, cross-reactivity, epitope, and bio-developability properties. We introduce A-Alpha Bio’s protein-protein interaction database with over 100M protein-protein interactions and describe initial work towards developing a generalizable antibody-antigen binding model.

Exploring the vast sequence space is a major challenge in designing de novo proteins. Optimizing for one trait often comes at the expense of another. Utilizing pre-trained language models and graph neural networks, we show that de novo cytokine mimetics that express well, fold into the desired structure, are thermostable and bind to their desired targets can be generated at a much higher success rate than with traditional physics-based simulation.

In this talk, we will present our recent development by combining a high-throughput molecular dynamics simulation platform and neural networks to accelerate the prediction of antibody biophysical properties, including solvent accessible surface area, charge, and hydrophobicity. These models, requiring only protein sequences, can accelerate the prediction of antibody biophysical properties from hours using supercomputers to seconds using laptops, an essential advantage for screening antibody drug candidates.

Antibody sequencing data has grown exponentially concomitant with the discovery and engineering of next-generation sequencing technologies. Generative machine learning models have the potential to learn complex relationships between sequences and benefit from large, heterogeneous sequence datasets. Here we describe advances in generative modeling of antibody sequences towards antibody design and property prediction, with a focus on incorporating structural information via structure prediction software.

Miniproteins are a powerful yet underutilized therapeutic modality. They are only 30-90 amino acids in length, yet they adopt a folded tertiary structure like a much larger protein; this structure enables miniproteins to bind with high affinity and specificity to their targets. Miniproteins found in nature are very challenging to engineer to bind new targets. We solved this problem using computational de novo design, finally unlocking miniproteins for therapeutic development.

In this talk, I shall introduce my strategic vision of Biopharmaceutical Informatics and how it can be used to discover developable biotherapeutics.

We constructed two antibody display libraries using a Generative Adversarial Network (GAN) machine learning algorithm. ML pattern recognition of IgG human repertoire has enabled application of V(D)J-gene diversity of CDRs and frameworks into our Just Humanoid Antibody Library (J.HAL). The phage Fab library exhibited success in multiple discovery campaigns and is currently undergoing diversity expansion to improve success rates. The yeast VHH library has demonstrated value for novel therapeutic discovery.

We developed a deep learning-based protein-protein interface (PPI) design pipeline that leverages a generative model (Ig-VAE) for 3D protein generation. This novel strategy uses neural networks' capabilities in capturing dynamic structures to create a fully flexible structural docking process for PPI design. The approach can sample protein conformations at an unprecedented speed and optimize structures for predefined functions. I will describe our current results on designing epitope-specific protein binders.

AlphaFold has revolutionized the structure prediction of proteins alone or in the complex. The need for co-evolutionary sequence constraints for structure prediction limits its use against antibody-antigen complexes. We predicted the structure of antibody-antigen complexes using traditional physics-based protein-protein docking tools. We evaluated the ability of AlphaFold in the quality assessment of models. Our results highlight that AlphaFold can rescue poorly-ranked models and better discriminate good-quality models from decoys.

We show that deep learning algorithms known as protein language models can evolve human antibodies with high efficiency, despite providing the models with no information about the target antigen, binding specificity, or protein structure, and also requiring no additional task-specific finetuning or supervision. We performed language-model-guided affinity maturation of seven diverse antibodies, screening 20 or fewer variants of each antibody across only two rounds of evolution. Our evolutionary campaigns improved the binding affinities of four clinically relevant antibodies up to 7-fold and three unmatured antibodies up to 160-fold across diverse viral antigens.