NaviDiv: A Comprehensive Tool for Monitoring Chemical Diversity in Generative Molecular Design

Mohammed Azzouzi, Thanapat Worakul, Clémence Corminboeuf

Keywords: Chemical diversity analysis, Generative molecular design, Reinforcement learning.

TL;DR: NaviDiv provides a comprehensive framework with multiple complementary metrics to monitor chemical diversity evolution during reinforcement learning optimization in generative molecular design.

Abstract: The rapid progress in generative models for molecular design has led to extensive libraries of candidate molecules for biological and chemical targets. However, ensuring these molecules are diverse and representative of the broader chemical space remains challenging. Without proper tools, researchers may over-explore limited regions or miss promising candidates. This work presents NaviDiv, a comprehensive set of tools for analyzing and steering chemical diversity in generative molecular design, introducing multiple complementary metrics that capture different aspects of molecular diversity through representation distance-based, string-based, fragment-based, and scaffold-based approaches. Our package not only monitors diversity evolution but also provides adaptive diversity constraints that can be integrated into the optimization process to guide generative models toward maintaining desired levels of chemical space exploration. Through a case study on singlet fission material discovery using REINVENT4, we demonstrate how different diversity metrics evolve during reinforcement learning optimization and show that our diversity constraints can prevent model collapse while preserving property optimization performance. The package is freely available in \href{https://github.com/lcmd-epfl/NaviDiv}{NaviDiv GitHub repository}. This initial implementation serves as a foundation for future extensions to additional molecular representations and generative architectures, addressing a critical bottleneck in automated molecular discovery.

Submission Track: Findings, Tools & Open Challenges

Submission Category: AI-Guided Design

Institution Location: Lausanne, Switzerland

AI4Mat RLSF: Yes

Submission Number: 22