Sustainability-Guided Small Molecule Generation with Generative Flow Networks

Summary

Generative Artificial Intelligence (AI) has emerged as a promising tool for accelerating scientific discovery in chemistry, particularly in the field of molecular design. In response to the environmental pollution crisis, generative models can be leveraged to propose environmentally friendly candidate molecules, thereby accelerating the design process and minimizing harmful effects on the environment. To that end, we investigate the use of Generative Flow Networks (GFNs), a family of generative models that allow sampling combinatorial objects, such as molecules, by optimizing a property of interest.

This is done by learning to sample proportionally to an externally defined reward function, which is typically a model (e.g. a neural net) functioning as a proxy to the aforementioned property. In this thesis, we design and train GFNs to propose new molecules with low water solubility, a crucial physicochemical property of molecules that is linked to environmental leaching and pollution. Experimentally, we evaluate the capacity of GFNs to simulate the desired distribution (via a battery of metrics), assessing their performance across various training objectives. Finally, we further screen the generated environmentally friendly molecules to identify potential agrochemicals, such as pesticides, herbicides, and insecticides. The code is available at https://github.com/johnsaveus/gflownet .