Abstract

Multi-target property prediction has the potential to improve generalization by exploiting the positive transfer between targets. Molecular generative models utilize independent single-target property prediction networks to discover novel molecules. We propose using multi-target networks to jointly predict several molecular properties and learn better representations by exploiting auxiliary information. Our multi-target model shows improvement in prediction accuracy on the test set. We additionally present results demonstrating promising performance in property prediction in these generative models does not translate to optimization. More specifically, random exploration is competitive with gradient-based strategies and better methods are needed.

By Anirudh Jain, Markus Heinonen, Heikki Käsnänen, Julius Sipilä and Samuel Kaski