Human-in-the-loop active learning for goal-oriented molecule generation.

Area of Science:

• Computational chemistry and cheminformatics.
• Artificial intelligence in drug discovery.
• Machine learning for predictive modeling.

Background:

• Machine learning (ML) models, including quantitative structure-property relationship (QSPR) and structure-activity relationship (QSAR) predictors, are crucial for accelerating drug discovery by predicting molecular properties.
• Current ML predictors often lack generalizability due to limited training data, leading to the generation of molecules with inflated predicted properties that fail experimental validation.
• Generative AI agents guided by these predictors can explore chemical spaces but are susceptible to these limitations, producing molecules with poor real-world performance.

Purpose of the Study:

• To develop an adaptive framework that refines ML property predictors for more effective goal-oriented molecule generation.
• To enhance the accuracy and reliability of ML-guided drug discovery by integrating active learning (AL) and iterative human feedback.
• To improve the generation of drug-like molecules with desirable properties by addressing the generalization issues of current predictors.

Main Methods:

• Integration of active learning (AL) with iterative feedback mechanisms to refine ML property predictors.
• Utilization of the Expected Predictive Information Gain (EPIG) criterion for selecting molecules to reduce predictive uncertainty.
• Leveraging human experts as an oracle for cost-effective feedback to augment limited training data and refine predictors.
• Empirical evaluation through simulated and real human-in-the-loop experiments.

Main Results:

• The proposed approach successfully refines property predictors, aligning them better with oracle assessments.
• Demonstrated improvement in the accuracy of predicted molecular properties.
• Observed enhanced drug-likeness and other practical characteristics (e.g., synthetic accessibility) among the top-ranking generated molecules.
• The framework shows robustness to noise in human feedback.

Conclusions:

• The developed adaptable framework effectively integrates active learning and human expertise to refine property predictors for goal-oriented molecule generation.
• This approach enhances the reliability of generative AI in drug discovery by ensuring generated molecules satisfy predicted profiles and score well on oracle models.
• Prioritizes practical drug development considerations, balancing chemical space exploration with exploitation of existing knowledge and similarity.