We will discuss our approach and general considerations for implementing active learning and design of experiments to iteratively optimise therapeutic antibody candidates. Our active learning framework is underpinned by both algorithmic innovations and robust data pipelines. We achieve improvements across binding affinity, expression yield, and developability properties via orthogonal optimisation approaches, analogous to the multitude of affinity maturation pathways observed in immune responses.

The OpenFold Consortium brings together academic and industrial teams to build state-of-the-art protein structure and co-folding prediction models optimised for use on commercial computational hardware. We develop fully open-sourced models and support creation of new experimental datasets, aiming to build more powerful models that can accurately predict complex systems of significance to life sciences. In my presentation, I will present the latest modelling and software developments from the consortium.

Boehringer Ingelheim’s digital transformation journey began in 2023, focusing on three key areas. Firstly, empower scientists by liberating them from routine tasks, allowing them to concentrate on high-value work. Second, create a hub for innovation by building an in-house digital portal, a scientist-driven mechanism to formalize and standardize in silico protocols. Lastly, emphasize digital data capture FAIR and APIs, setting the stage for leveraging AI/ML in future.

AI's potential to create antibodies from scratch is promising but hampered by poor hit rates and binding strengths, rooted in insufficient training data. We have addressed this issue by using computational simulations to determine data requirements such as modality, amount, and diversity. Simulations have been guiding our ongoing experimental data generation work, marking a shift towards a data-centric strategy that complements recent algorithmic progress, aiming to overcome current challenges.

Antibody design is a multi-parameter optimisation problem that integrates data from multiple sources, such as high-resolution structures and sequence libraries. Here we show that predictions from multiple independently-trained machine learning models (ProteinMPNN, ESM, AbLang) can be easily and effectively combined when redesigning antibodies, and that doing so retains the strengths but not the weaknesses of each ML method in isolation.

Inverse folding (IF) and protein large language models (pLLMs) have become useful tools for antibody variant generation, with generally good performance, but limited ability to find mutations that enhance the binding to the antigen. Here, we show how IF models can be used to predict B cell epitopes, how to extend this approach to estimate antibody-antigen interaction energy and find mutations that increase affinity, and to fine tune this model to increase its predictive power.

The high binding affinity of antibodies towards their cognate targets is key to eliciting effective immune responses. We propose ANTIPASTI, a Machine Learning approach that achieves state-of-the-art performance in the prediction of antibody binding affinity using as input Normal Mode correlation maps. The learnt representations are interpretable: they reveal similarities of binding patterns among antibodies and can be used to quantify the importance of antibody regions contributing to binding affinity.

We describe the development of ABS-101, a potential best-in-class anti-TL1A antibody for the treatment of inflammatory bowel disease. Generative AI models were leveraged to design antibodies against desired epitopes to achieve reduced immunogenicity risk and binding to both monomeric and trimeric TL1A. AI-guided lead optimization then produced several high-affinity candidates with high cellular potency, desirable developability properties, and cross-reactivity to non-human primate (NHP) species.

Bispecific immune cell engagers have emerged as a highly effective new therapeutic modality for oncology. Current engagers are limited by their reliance on immunoglobulin proteins, which restricts the valency, geometry of binding, developability, and speed of engineering. We solved these challenges by leveraging de novo-designed miniproteins, which enabled us to rapidly create and optimise highly potent NK cell engagers for AML that control tumour growth using xenograft models.

We develop and apply machine learning models to optimise in parallel multiple drug-like properties of the bacterial enzyme IdeS, to design a therapeutic for chronic autoantibody-mediated diseases, while minimising its immunogenicity and other liabilities. The success of this approach is demonstrated via in vivo and in vitro assays, and we illustrate its generalizability by engineering non-immunogenic bacterial cysteine proteases with a variety of immunoglobulin isotype specificities. Finally, we present a general mathematical framework enabling this process to be applied to virtually any naturally occurring protein in silico, using only the protein sequence as input.