Friday, May 10, 2019 – 15:00 to 16:30
Salle SC 23.01 – Université de Montpellier – Campus Triolet
The recent spectacular advances of Artificial Intelligence are often exclusively associated with DeepLearning. In this talk, I will show how the more discrete but still remarquable progresses of automatedreasoning in Artificial Intelligence can be leveraged to contribute to the computational design of newproteins (a challenging but important problem with many applications in health, green chemistry and biotechnologies). Starting from a description of the usual rigid backbone Computational Protein Design problem with a discrete rotamer library and a decomposable force field (an NP-hard problem that is usually tackled using Monte Carlo methods), we will see how automated resoning technology derived from Constraint Programming in AI are able to solve challenging examples of computational protein design problems, including formulations accounting for some backbone flexibility. The computational efficiency of these approaches is sufficient to bring to light for the first time the increasing limitations of a specialized Monte Carlo approach in exploring regions of very low energy (high probability/stability). To bridge automated reasoning-built designs with reality, we recently contributed to the design of an amino acid sequence which, once synthesized, was shown to fold and self assembles into the expected hyper- stable protein, a symmetric protein called Ika.
Thomas Schiex is a member of the Statistics and Algorithmics for Biology team at INRA MIA Toulouse. The team develops, adapts and applies Statistical and Artificial Intelligence to problems in Genetic, Molecular and Structural Biology. In AI, Thomas is more specifically interested in solving discrete optimization problems on Graphical Models (Markov Random Fields, Bayesian Networks, Cost Function and Constraint Networks) with a recent focus on applications in Protein Design.