Predicting a protein’s structure from its primary sequence has been a grand challenge in biology for the past 50 years. In this talk, we will describe work at DeepMind to develop AlphaFold2, a new deep learning-based system for structure prediction that achieves high accuracy across a wide range of targets. The talk will cover both the underlying machine learning ideas and the implications for biological research.

Describing an antibody’s binding site using only one single static structure limits the understanding of the antibody’s function. This limitation is even more pronounced when no experimentally determined structure is available or the crystal structure is distorted by packing effects, which can result in misleading antibody paratope structures. To improve antibody structure prediction and to take the strongly correlated loop and interface movements into account, antibody paratopes should be described as interconverting states in the solution. Therefore, the definition of kinetically and functionally relevant states can be successfully used to improve the accuracy and enhance the understanding of antibody-antigen recognition.

Determining which protein regions selectively affect function and biophysical properties is crucial in development of diagnostics and therapeutics. The identification of key regions of a protein structure can be achieved through antibody scanning, which involves combining a phenotypic assay with a library of antibodies targeting different epitopes on a protein’s surface. Beta-2 microglobulin (B2M) is a well characterised protein that is associated with a wide range of diseases. Understanding the structural or functional role of B2M's surface epitopes can be clinically beneficial. Using an in silico design strategy we recently developed, nanobodies with high stability and affinity against B2M were generated to perform antibody scanning. 

Lipopolysaccharide (LPS, endotoxin) is a cell wall component of gram-negative bacteria and highly toxic upon entering the human body. Reliable endotoxin removal or detection assays are of high demand. Many assays face difficulties in providing animal-free, LPS-specific, and low-priced approaches. This study focuses on the creation of an artificial antibody via Phage Display that specifically binds Lipid A and paves the way towards a bead-based detection and removal tool.

Early and accurate disease detection is crucial for preventing disease development and defining therapy strategies. Autoimmune diseases are notoriously challenging to diagnose for clinicians. CDSS are a promising technology to enhance precise diagnostics. However, due to the difficulties of integrating omics data in conjunction with clinical characteristics, such systems are often limited to certain data types. We were able to develop an integration pipeline that reliably diagnosed patients based on multi-omics data combined with clinical and laboratory data. Our results uncover insights in the field of autoimmune diseases and can be adapted for applications across disease conditions.

Doctors struggle to diagnose autoimmune diseases because of overlapping symptoms and a multitude of laboratory test results. Together with physicians and researchers we have developed a Swiss and European prototype of a clinical decision support system, Personalis. The software uses artificial intelligence to detect patterns in the genetic, biomedical, and clinical data. These predictions help physicians make an early and correct diagnosis of autoimmune diseases.

Computational methods are emerging techniques to design in silico, new therapeutic single-domain antibodies (nanobodies) targeting a wide range of biological molecules. A humanness tool has been here developed using a deep machine learning vector-quantised variational auto-encoder (VQ-VAE) with unsupervised learning to assess the similarity of de novo-designed nanobodies with antibodies produced by human immune systems. This humanness assessment even shows significant correlation with immunogenicity antibody data.

Our novel, automated high-throughput engineering platform enables the fast generation of large panels of multi-specific variants (up to 10.000) giving rise to large data sets (more than 100.000 data points). By mining our data sets we were able to extract engineering patters and to develop AI-based virtual screening workflows to guide the exploration of huge design spaces for multi-specific biologics drug discovery.

The multiple biophysical properties that overall define the developability of biologics depend not only on protein sequence but also on buffer composition. Here we show how machine learning algorithms can accelerate the design of biopharmaceutical formulations that simultaneously optimize multiple biophysical properties.

At Just – Evotec Biologics, we’re developing cutting-edge machine learning technologies to accelerate antibody drug development. In this talk, I’ll focus on the Antibody-GAN – an ML framework that allows us to generate limitless antibody drug candidates with desirable properties. I’ll describe the discovery platform we’ve built upon this technology (J.HAL) and how we’re using data from this platform to develop even more powerful tools for discovery and design.

The development of biologics with suitable functionality into practically useful molecules is often impeded by developability issues. Conformational stability and solubility are arguably the most important biophysical properties underpinning developability potential, as they determine colloidal stability and aggregation, and correlate with yield and poly-reactivity. I will present a computational pipeline, and corresponding experimental validation, for the automated design of antibody variants with improved stability and solubility.

An individual’s B cell receptor (BCR) repertoire encodes information about past immune responses and potential for disease protection. Deciphering the information in BCR sequence datasets will transform our understanding of disease and enable discovery of novel antibody therapeutics. Here, we present an antibody-specific language model, AntiBERTa; it learns a rich, biologically relevant representation of BCR sequences, and the model is generalizable to a number of applications, such as paratope prediction.