This presentation will provide a pharma industry perspective on translating de novo protein design from computational promise to therapeutic reality. We will showcase Sanofi's integrated approach to VHH discovery, featuring our computational de novo design toolbox and its seamless integration with specialized wet-lab workflows. Through concrete proof-of-concept studies on therapeutically relevant targets, we will demonstrate how generative design methods are being applied in real-world drug discovery programs. The talk will address both the current capabilities and practical limitations of de novo approaches, offering insights into how computational VHH design is evolving from experimental tool to established design platform.

Obtaining novel antibodies against specific protein targets is experimentally laborious. Meanwhile, computational design methods have been limited by low success rates. We introduce Germinal, a generative pipeline that designs functional antibodies against specific epitopes while requiring minimal experimental screening. Our method co-optimises structure and sequence by integrating AlphaFold-Multimer with an antibody-specific language model to perform de novo CDR design onto user-specified frameworks. Tested against four diverse protein targets, Germinal successfully designed functional antibodies across all formats, requiring only tens of designs per antigen. Validated designs exhibited high novelty, robust mammalian expression, and broad applicability. 

We present a controlled framework for evaluating binding specificity using 106 experimental nanobody-antigen complexes and 11130 shuffled non-cognate pairings. Benchmarking AlphaFold3, Boltz-2, and Chai-1 showed that although predicted complexes were often geometrically plausible, internal confidence scores (eg, ipTM) failed to distinguish correct from incorrect pairings, and increased sampling improved refinement but not discrimination. We therefore conclude that confidence scores require validation against realistic decoys and release 1.8 million AI-generated complexes.

Directed evolution provides a robust platform to dissect and engineer antibody affinity maturation in vitro, thereby informing the rational design of antibody therapeutics with enhanced protective efficacy and immunogens capable of guiding broadly neutralizing antibody development through vaccination. We applied yeast display-directed evolution to demonstrate that two closely related HIV bnAb lineages, differing only in LCDR3 length, follow distinct evolutionary pathways to acquire optimal neutralising activity. We also show that early- and late-stage LCDR3 maturation do not converge on identical sequence solutions, indicating that light-chain optimisation is shaped through iterative coordination between epitope engagement and intra-paratope structural refinement.

Proteins have promised to solve many grand challenges of modern society. However, the existing processes for discovery, characterization, and engineering of proteins are slow, expensive, and inconsistent. To address this limitation, my lab has been developing an AI-powered biofoundry since 2013. In this presentation, I will highlight a few representative accomplishments and discuss our efforts in making our AI-powered biofoundries accessible to the broader research community.

A Robot Scientist (AI scientist, self-driving lab) is a physically implemented robotic system that applies techniques from artificial intelligence to execute cycles of automated scientific experimentation. The Robot Scientist ‘Adam’ was the first machine to autonomously discover scientific knowledge. I am co-organising the ‘Nobel Turing Challenge’ to develop: AI systems capable of making Nobel-quality scientific discoveries comparable, and possibly superior, to the best human scientists by 2050 or sooner.

While deep learning has transformed protein design, purely data-driven methods struggle to generalise to novel targets and lack physical grounding. We present a computational pipeline that integrates a physics-based force field with evolutionary sequence optimisation for de novo antibody design. Our approach jointly optimises binding affinity, stability, and developability. Benchmarks demonstrate that physics-informed methods match state-of-the-art ML approaches on antibody binding prediction while offering interpretability without target-specific training data.