Explaining and avoiding failure modes in goal-directed generation of small molecules

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery

Background:

• Automated in-silico molecular design is advancing.
• Concerns exist about goal-directed algorithms exploring novel chemical spaces unbiasedly.
• A phenomenon shows generated molecules scoring high on optimization but low on control models.

Purpose of the Study:

• To investigate the cause of biased exploration in goal-directed molecular design.
• To determine if predictive models or generation algorithms are responsible for scoring discrepancies.
• To demonstrate a method for achieving unbiased exploration and high scores across models.

Main Methods:

• Analysis of goal-directed generation algorithms.
• Evaluation of predictive model performance in molecular design.
• Comparison of scoring across optimization and control models.
• Implementation of improved predictive models.

Main Results:

• The observed phenomenon of biased exploration is attributed to issues within the predictive models.
• The goal-directed generation algorithms themselves are not inherently flawed.
• Utilizing appropriate predictive models resolves the scoring discrepancies.
• Generated molecules achieve high scores on both optimization and control models.

Conclusions:

• Predictive model quality is critical for unbiased in-silico molecular design.
• Goal-directed generation can effectively explore novel chemical spaces when paired with robust predictive models.
• This work resolves a key concern in automated molecular design, enabling more reliable discovery.