Guido UguzzoniInvited by Serge Picaud, Guido Uguzzoni (Italian Institute for Genomic Medicine) will give a talk on Monday November 20th, 12.00 PM, in the conference room of the UCL, 13 Rue Moreau.

This talk will be dealing with "Machine learning approach leveraging protein screening experiments".

Over the last years, Artificial intelligence (AI) applications in molecular biology have made significant strides, exemplified by the breakthroughs achieved by Alpha Fold 2 in protein structure prediction from genomic data.

However, a general approach to manipulating protein function remains elusive, presenting significant technical and fundamental challenges while holding enormous potential for diverse biomedical applications.

In this presentation, I will introduce an innovative machine-learning procedure to train accurate computational models from the screening of mutational libraries of proteins. Screening experiments, such as Directed Evolution, are used to select protein variants with high fitness for specific tasks. Despite significant advances, these methods have inherent limitations, including the combinatorial explosion of possible variants, local random-fashion exploration, and experimental noise and biases. On the other hand, the massive data generated by screening experiments can be used to train in-silico models. Which can provide an accurate statistical description of the evolution of the variant population through the panning rounds, predict the biophysical activity of protein variants, and finally computationally design new improved variants.
Practical applications of this approach will be discussed, ranging from antibody specificity design based on in-vitro phage display experiments to the optimization of adeno-associated viral capsids for tissue-specific delivery using in-vivo Directed Evolution data.

Guido Uguzzoni is a researcher at the Italian Institute for Genomic Medicine with a background in statistical physics and computational biology. Currently, his research focuses on developing and applying machine learning methods to protein design and optimization. These methods aim to generate protein sequences optimized for biochemical activity leveraging high-throughput screening experiments, such as Directed Evolution experiments. In particular, some ongoing collaborations concern the design of multispecific antibodies and adeno-associated viral capsids, both of which have significant therapeutic implications.