Policy-Based Active Learning for Efficient Molecular Identification.

Area of Science:

• Computational Chemistry
• Machine Learning in Drug Discovery
• Chemical Informatics

Background:

• Automated chemical design faces challenges in identifying valuable molecules efficiently within budget constraints.
• Pool-based active learning is used in automated settings to guide experimental selection, but often optimizes model accuracy over search efficiency.
• Existing methods may not directly align with the goal of finding the best molecule within a specific candidate pool.

Purpose of the Study:

• To develop a policy-based active learning framework for optimizing molecular pool selection in automated chemical design.
• To reformulate molecular pool selection as a sequential decision-making problem solvable with Markov decision processes.
• To directly optimize search efficiency for identifying molecules with the most favorable properties under budget constraints.

Main Methods:

• A policy network is trained to optimize cumulative search performance within limited evaluation budgets.
• Property prediction models are used as contextual signals, not direct optimization targets.
• The framework is evaluated on 1409 tasks from MoleculeNet and ChEMBL, plus six literature-curated in vivo tasks.

Main Results:

• The proposed policy-driven active learning framework demonstrates improved efficiency in identifying optimal molecules.
• The method successfully identified top-performing molecules within fewer evaluation cycles compared to conventional approaches.
• Performance was validated across both benchmark datasets and in vivo experimental settings.

Conclusions:

• Policy-driven active learning offers a generalizable strategy for enhancing molecular identification efficiency in budget-constrained chemical discovery.
• The framework effectively optimizes the search for high-value molecules within predefined candidate pools.
• This approach shows significant potential for accelerating the discovery of novel chemical entities.