-
Artificial intelligence: currently mainly focused at the intersection of deep learning and discrete probabilistic graphical models in the context of neuro-symbolic AI and protein design. Deep learning is used to learn the structure and parameters of graphical models from raw data, while our graphical model solver (toulbar2) is used to reason on the infered knowledge and sample optimal constrained solutions. Alltogether, this defines a generative neuro-symbolic architecture that can sample a constrained (conditional) learned probabilistic model without having to retrain the model from scratch. This architecture is quite versatile, with applications in protein design, combinatorial optimization (decision focused learning), and other domains. Eventually, the aim is to use several (learned or hand built) Graphical Models to combine learned information (intuition) with knowledge (Logic), especially in the context of Designing new physical (molecular) objects.
I'm also interested in the use of generative AI technology (espcially Denoising Diffusion Probabilistic Models) to sample solutions of NP-hard problems, especially to sample good quality solutions of the Maximum a Posteriori problem (in Markov Random Fields) or the Weightd Constraint Satisfaction Problem (in the context of Constraint Programming and Cost Function Networks).
A good fraction of this relies on the efficiency of our graphical model solver toulbar2, and I'm therefore also interested in the development of efficient algorithms for answering complex queries (optimization, counting,....) on discrete graphical models, with a bias toward algorithms offering guarantees and/or provable complexity bounds.
This forms the core of the topic of the chair in the Artificial and Natural Intelligence Toulouse Institute (ANITI). - Bioinformatics: the application of Discrete Graphical Models and Machine/Deep learning technologies mostly to constrained optimisation problems arising in computational biology. My current application domain of interest is Computational Protein Design, with applications in health and green chemistry (and a startup: Amineo). In the past, I have contributed to genetic markers ordering, genetic map joining, RNA secondary structure prediction and also RNA/protein gene finding and prediction (with frameshift detection, using Conditional Random Field like models) both for prokaryotic and eukaryotic organisms, then biological network inference (by learning parameters and structures of probabilistic graphical models) and recently a bit of highly-duplicated genomes assembly assistance.
RESEARCH THEMES
EXTERNAL COLLABORATIONS
- With Sophie Barbe of the CIMES team in Toulouse Biotech Institute for designing and applying Computational Protein Design algorithms and tools.
- With Juan Cortés in LAAS/CNRS for the use of robotics-inspired flexible protein modeling.
- With the Combinatorial optimization group of CERT (Centre d'Études et de Recherches de Toulouse de l'ONERA) on algorithms for CSP.
- With Hélène Fargier, Jérôme Lang, Martin Cooper from the Institut de Recherches en Informatique de Toulouse for extensions of the CSP formalism and valued constraint network properties and algorithms.
- With the team of F. Rossi, University of Padova, Italy, for the comparison of Valued CSP and Semi-ring CSP.
- With Pedro Meseguer and Javier Larrosa (Polytechnic University of Catalunya, Spain), for the algorithmic of Valued and weighted CSP.
- With P. Rouzé (University of Ghent, Belgium) for gene finding in Arabidopsis thaliana. and other plants.
- With Marie-France Sagot on the algorithmic of gene finding.