Offre en lien avec l’Action/le Réseau : – — –/Doctorants

Laboratoire/Entreprise : LERIA (Université d’Angers)
Durée : 18 months
Contact : nicolas.gutowski@univ-angers.fr
Date limite de publication : 2024-03-31

Contexte :
In chemistry, the discovery of new molecules often results from the refinement of an already known effective compound through chemical reactions to enhance its properties. The emergence of a truly new molecule is a rarer phenomenon. It is around this objective that a theme has developed focusing on the de novo generation of molecules with desired properties. Among the challenges in this research area are the size of the search space and the difficulty of generating synthesizable molecules.
Molecules can be represented as graphs, where vertices are labeled according to the type of atom, and edges are labeled according to the type of bond. This is a problem of generating a graph structure, where the goal is the combination of one or more functions to optimize and constraints to satisfy. Thus, the G-GENOCOD project (Graph-Generation for Novel Compound Discovery), although applied to chemistry, addresses a much broader problem of generating complex graph structures with a very large space of composed actions.

Sujet :
The G-GENOCOD project follows EvoMol, an evolutionary algorithm for molecule generation developed by an interdisciplinary team from LERIA and MOLTECH. While EvoMol achieves benchmark results, significant challenges that G-GENOCOD aims to address remain:
1. The first will be conditioned by the goal of realism (generating synthesizable molecules). This objective is crucial for real-world applications. Methods derived from the goal-conditioned RL approach will enable the attainment of strong synthesizability properties while being explainable.
2. The second will be conditioned by the selection of optimal actions to achieve the desired chemical properties (“properties conditioned”), i.e., the choice of actions on the graph is currently random. One would expect an intelligent method (here, reinforcement learning: RL) to apply a policy of selecting actions that has worked in the past, as a chemist would add a known chemical function to enhance a target property.
3. The third objective will finally be to evolve towards the ability to generate molecules according to several defined objectives (multi-objective optimization).

Profil du candidat :
Knowledge :
– Reinforcement learning
– Graph theory
– Theorems and proofs of convergence
– Probabilistic reasoning

Know-how:
– Python development
– Development using scikit-learn, PyTorch
– Writing scientific articles in LaTeX

Soft skills:
– Efficient and responsive
– Autonomous
– Proactive
– Rigorous

Formation et compétences requises :
PHD degree of less than 3 years
Specialty : Computer Science, Artificial Intelligence

Adresse d’emploi :
UFR Sciences, 2 Bd de Lavoisier, 49000 Angers, FRANCE

Document attaché : 202401082139_FP_POST-DOC_G-GENOCOD-ENG.pdf