Generative Modelling for Chemical Design

Abstract

The discovery of novel molecules and materials with desirable properties is a fundamental objective of chemical design. This involves searching through a vast space of different possibilities, with an estimated 10^60 plausible molecules of drug-like size. As the demand for increasingly effective drugs and materials never ceases to grow, traditional approaches relying on human intuition and trial-and-error are becoming insufficient for the task. Data-driven solutions offer a paradigm shift, whereby large chemical databases enable this vast space to be efficiently navigated. This thesis demonstrates that generative models, based on deep-learning architectures, are capable of inverse design of novel molecules with targeted physicochemical and quantum-mechanical properties.

In this work, two different conditional generative-modelling frameworks are proposed. Firstly, a generative adversarial network (GAN) is used to generate molecular conformations with selective total energy values. A local inversion algorithm is developed to convert the generated atomic descriptors into interpretable atomic structures. The second generative-modelling framework, MolGPT, employs a generative pre-trained transformer (GPT) model with text-based chemical representations. A new representation called MolBlox is proposed, with a corresponding method to invert the generated text into valid molecules. The model is conditioned using a variety of physicochemical and quantum-mechanical properties. In both frameworks, the generated molecules are verified as being structurally valid, with accurate property matching, as confirmed by density-functional theory (DFT) calculations.

This thesis concludes that these generative frameworks offer scalable solutions to accelerate chemical design. Published work on local inversion and ongoing manuscripts on conformation generation and MolGPT underscore their impact. These tools are considered a promising avenue of further research, demonstrating the viability of automated methods in the discovery of novel pharmaceuticals and materials.