Multi-armed bandit algorithm for sequential experiments of molecular properties with dynamic feature selection.

Area of Science:

• Computational Chemistry
• Machine Learning
• Drug Discovery

Background:

• Sequential optimization aims to find optimal candidates with desired properties while minimizing experiments.
• High dimensionality in feature spaces (e.g., molecular descriptors) complicates feature utilization in candidate selection.
• Existing methods like Bayesian Optimization (BO) often use fixed feature sets, limiting adaptability.

Purpose of the Study:

• To develop a novel sequential optimization algorithm for molecular problems.
• To address the challenge of high dimensionality and dynamic feature relevance in optimization.
• To improve the efficiency and reliability of identifying optimal molecular candidates.

Main Methods:

• Developed a new algorithm integrating reinforcement learning, multi-armed linear bandits, and online dynamic feature selection.
• Implemented a stopping condition to ensure the reliability of the selected candidate.
• Compared the algorithm against Bayesian Optimization (BO) using synthetic and real-world molecular datasets (hydration free energy, enantiomer product free energy difference).

Main Results:

• The developed algorithm demonstrated superior performance, particularly in reducing the time to find the best candidate and stop experiments.
• Dynamic feature selection, adapting molecular descriptors with experiments, significantly improved optimization efficiency.
• The multi-armed linear bandit approach with dynamic feature selection outperformed standard BO with fixed features.

Conclusions:

• Dynamic feature selection is crucial for efficient sequential optimization in high-dimensional molecular spaces.
• The proposed algorithm offers a more effective and reliable approach for identifying optimal molecular candidates compared to traditional methods.
• This work advances the application of machine learning in computational chemistry and drug discovery.