We will discuss the “Lab-in-the-loop” system, a collaboration between Prescient Design and Antibody Engineering at Genentech, to build and integrate state-of-the-art machine learning methods with large molecule design and discovery capabilities. Lab-in-the-loop encompasses generative models, pseudo-oracles, physics-based modeling, large language models, wet-lab assays, and active learning to fundamentally change early-stage drug discovery.

Most therapeutic antibodies are simple antagonists. However, like all proteins, antibodies can be sophisticated nano-machines. Biolojic Design uses AI to program antibodies to become dynamic functional switches affecting biology in new ways. I will describe our AI-design process, and share clinical data from the first AI-designed therapeutic antibody. I will also show preclinical data on multi-specific antibodies illustrating their potential to improve outcome in cancer and autoimmune diseases.

Incorporating predictive modeling into experimental workflows holds great promise for accelerating discovery, guiding optimization, and prioritizing leads. Here we discuss application of ML to design of synthetic libraries for discovery, hybrid experimental-modeling approaches for selection of functionally diverse antibodies, and developability predictions that decrease resource-intensive experiments. Our integration of ML into a larger informatics/data platform enables predictions to inform candidate selection throughout the discovery and lead optimization process.

AI applications have become increasingly powerful, and play an increasing role in today's research and development. At the same time, there is no consensus yet regarding AI methodologies, and how to best integrate them in the discovery and development process. Building internal capabilities and establishing external collaborations can help navigate through this exciting new augmentation of biologics drug development.

In 2015, we introduced the CamSol method, which enables accurate prediction of protein solubility solely by analysing the physico-chemical properties of their sequences. We recently extended the method to enable accurate prediction of non-natural amino acids such as chemically or posttranslationally modified ones. In this talk, I will highlight how CamSol can be used in drug development pipelines and explore the possibilities to extend CamSol using machine learning.

We will discuss how Seismic Therapeutic is using its IMPACT platform to integrate machine learning, structural biology, protein engineering and translational Immunology to accelerate the discovery and development of therapeutics for autoimmune diseases, caused by a dysregulated adaptive immune system.

Engineering new biological molecules with desired activity profiles requires time consuming and expensive cycles of design-make-test-analyse (DMTA) work. Even for short protein sequences, the available exploration space is intractable for traditional methods of experimental biology. In this talk, I will discuss how machine learning can be used to augment the design of peptide therapeutics. Specifically, I will present our recent studies focussed on de novo design and optimisation of GPCR ligands.

Generative language models trained on protein sequences have proven incredibly powerful for protein sequence design. In this talk, we will demonstrate how protein language models enable discovery of diverse proteins, which often function on par with natural counterparts- despite significant deviation in sequence space. Beyond generation of sequences, protein language models are effective zero-shot predictors of fitness, enabling direct optimization of function.