Etudiant : LABBOUJ MOHAMMED

Filière : Master Machine Learning Avancé et Intelligence Multimédia (MLAIM)

Encadrant : Pr. ABOUSSALEH ILYASSE

Annèe : 2026

Résumé : Chemical reaction prediction has emerged as one of the most promising applications of artificial intelligence in chemistry, offering the potential to accelerate molecular synthesis while reducing the time, cost, and resources associated with laboratory experimentation. In particular, forward reaction prediction aims to predict the product generated from a given set of reactants and reagents, providing valuable assistance for synthesis planning and drug discovery. This internship investigates forward reaction prediction through a comparative study of three molecular representations: SMILES strings, molecular graphs, and molecular fingerprints. Three deep learning architectures, namely ReactionT5, RXNGraphormer, and an adapted MolBERT model, were evaluated using reaction datasets derived from the United States Patent and Trademark Office (USPTO). A benchmark dataset based on USPTO-MECH was constructed to evaluate model generalization on a reaction distribution different from those used during the original training of the models. In addition, MolBERT, originally developed for molecular property prediction, was adapted into an encoder-decoder architecture capable of performing reaction product prediction from fingerprint representations. The experimental results showed a noticeable decrease in performance when the evaluated models were transferred to USPTO-MECH, highlighting the challenges associated with generalization across reaction datasets. Fine-tuning on the constructed benchmark substantially improved predictive performance for all models. Overall, the results emphasize the importance of molecular representation in reaction prediction and suggest that current representations still present limitations when applied to new reaction distributions. This work contributes to the development of more robust and generalizable AI-driven tools for molecular synthesis prediction.