Uncertainty quantification with graph neural networks for efficient molecular design.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Materials science

Background:

• Optimizing molecular design in large chemical spaces is challenging due to domain shifts and maintaining predictive accuracy.
• Existing methods often struggle with broad, open-ended chemical exploration.
• Computational-aided molecular design (CAMD) requires robust optimization strategies.

Purpose of the Study:

• To evaluate the effectiveness of uncertainty quantification (UQ) integrated with directed message passing neural networks (D-MPNNs) for optimizing expansive chemical spaces.
• To identify optimal implementation strategies for UQ-enhanced D-MPNNs in molecular design.
• To assess the performance of UQ-guided optimization in both single- and multi-objective tasks.

Main Methods:

• Integration of uncertainty quantification (UQ) with directed message passing neural networks (D-MPNNs).
• Utilization of genetic algorithms (GAs) for optimization.
• Evaluation using benchmarks from the Tartarus and GuacaMol platforms.
• Application of probabilistic improvement optimization (PIO) for UQ integration.

Main Results:

• UQ integration via probabilistic improvement optimization (PIO) generally enhances optimization success across diverse chemical spaces.
• UQ-enhanced D-MPNNs demonstrate improved reliability in exploring chemically diverse regions.
• PIO is particularly advantageous in multi-objective optimization tasks, effectively balancing competing goals.
• The proposed methods outperform uncertainty-agnostic approaches in several benchmarks.

Conclusions:

• Uncertainty quantification significantly improves the performance of D-MPNNs for computational-aided molecular design (CAMD).
• Probabilistic improvement optimization (PIO) offers a robust strategy for navigating complex chemical landscapes and balancing multiple objectives.
• This study provides practical guidelines for implementing UQ in CAMD workflows, enabling more reliable and efficient molecular discovery.