Session:

In recent years, generative design methods have gained significant attention in the field of molecular ideation, offering new possibilities for the discovery and optimization of novel molecules. These methods employ algorithms and machine learning techniques to explore vast chemical spaces, generating diverse molecular structures with desired properties. As these approaches rapidly evolve, it becomes crucial to benchmark and evaluate their performance to ensure their reliability and effectiveness in real-world applications.
In this work we study two classes of generative algorithms from literature, generative adversarial networks and normalizing flow models. Generative Adversarial Networks (GANs) have emerged as a powerful framework for generating ideas in various domains, offering unique advantages over traditional generative models. GANs offer remarkable capabilities for generating ideas by capturing complex data distributions, promoting creativity and exploration, facilitating the generation of diverse ideas, and adapting to specific requirements. For this study we benchmark MolGAN: An implicit generative model for small molecular graphs [1] and A de novo molecular generation method using latent vector based generative adversarial network [2]. The other architecture studied in this benchmark are normalizing flow algorithms, a family of algorithms which promise easy training and fast sampling. For normalizing flows we benchmark GraphAF: a Flow-based Autoregressive Model for Molecular Graph Generation [3], and GraphNVP: An Invertible Flow Model for Generating Molecular Graphs [4].
In this benchmark we build molecular distributions over 250,000 OLED like molecules using these four methods and attempt to minimize hole reorganization energy using three different quality machine learning models. We then analyze the performance characteristics of the methods, the ideas they create, the idea diversity, and the compute cost.

[1] Nicola De Cao, presented at ICML 2018 workshop on Theoretical Foundations and Applications of Deep Generative Models (unpublished)
[2] O. Prykhodko et al., J. Cheminf. 11 (2019) 74.
[3] C. Shi, M. Xu, Z. Zhu, W. Zhang, M. Zhang, J. Tang, arXiv (2020), doi:10.48550/ARXIV.2001.09382.
[4] K. Madhawa, K. Ishiguro, K. Nakago, M. Abe, GraphNVP: An Invertible Flow Model for Generating Molecular Graphs (2019), doi:10.48550/ARXIV.1905.11600.

Co-Author(s): Andrea Browning (Schrodinger), Alex Goldberg (Schrodinger), Mathew Halls (Schrodinger), Tsuguo Morisato (Schrodinger), Nobuyuki Matsuzawa (Panasonic Corporation), Hiroyuki Maeshima (Panasonic Corporation), Tatsuhito Ando (Panasonic Corporation)

Karl Leswing

Appearances