GuacaMol: Benchmarking Models for de Novo Molecular Design

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Molecular modeling

Background:

• De novo molecular design aims to create molecules with specific properties through iterative virtual design-make-test cycles.
• Deep learning and neural generative models show promise for molecular design but lack standardized evaluation.
• Comparative studies between new neural models and established algorithms are scarce.

Purpose of the Study:

• To propose a standardized evaluation framework for de novo molecular design models.
• To enable consistent assessment of both classical and emerging neural network-based approaches.
• To facilitate benchmarking of molecular generation algorithms.

Main Methods:

• Development of GuacaMol, an open-source Python framework for de novo molecular design evaluation.
• Implementation of a suite of standardized benchmark tasks.
• Inclusion of metrics for property distribution fidelity, novelty, chemical space exploration, and optimization tasks.

Main Results:

• GuacaMol provides a consistent platform for profiling and comparing diverse molecular design models.
• The framework supports assessment of model performance across various chemical tasks.
• An accompanying leaderboard allows tracking of algorithm progress.

Conclusions:

• Standardized benchmarking is crucial for advancing de novo molecular design.
• GuacaMol offers a robust solution for evaluating and comparing molecular generation algorithms.
• The framework promotes reproducible research and accelerates the development of novel molecular design tools.