Trustworthy Inverse Molecular Design via Alignment with Molecular Dynamics.

Area of Science:

• Computational Chemistry
• Machine Learning
• Molecular Modeling

Background:

• Data-driven inverse molecular design (IMD) is gaining traction.
• Current IMD methods lack trustworthiness due to misalignment with ground-truth molecular dynamics functions.

Purpose of the Study:

• To propose TrustMol, a novel IMD method designed for trustworthiness.
• To invert a reliable molecular property predictor for enhanced IMD.

Main Methods:

• Constructs a latent space using a variational autoencoder (VAE).
• Trains an ensemble of property predictors for latent to property space mapping.
• Employs a new reacquisition method for representative training samples.
• Optimizes latent designs by minimizing predictive error and ensemble-quantified uncertainty.

Main Results:

• Achieves state-of-the-art performance in IMD accuracy.
• Demonstrates alignment with the ground-truth function, ensuring trustworthiness.
• Generates desired molecules with improved reliability.

Conclusions:

• TrustMol offers a trustworthy solution for inverse molecular design.
• The method addresses the critical need for reliable molecular property prediction.
• Future applications in drug discovery and materials science are promising.