Génération de nouvelles molécules et réactions par intelligence artificielle guidée par la chemographie
Bort, William · HAL (Le Centre pour la Communication Scientifique Directe) · 2023
This thesis is dedicated to the exploration and understanding of neural network latent spaces, to allow the creation of a link between the latter and classical structural descriptors to perform inverse QSAR. The generative potential of seq2seq architectures often comes with a blurry understanding of the rules governing its chemical spaces. A study of an Autoencoder’s chemical space construction showed its ability to recreate existing property distributions and molecular structures with varying degrees of success depending on complexity and density factors. The model was even successfully modified to generate feasible and novel reactions.However, the sequential interpretation of chemical structures through SMILES strings tend to create weaknesses in the resulting chemical spaces. As such, structural descriptors like ISIDA, which are more robust, are usually preferred to map and identify zones of interest when searching for active compounds. Several methods to harness the efficiency of ISIDA descriptors and combine it with the generative power of an Autoencoder latent space resulted in the development of a new architecture based on Conditional Variational Autoencoders and the Attention Mechanism to generate potent molecules against biological targets.