Data-driven approaches for identifying hyperparameters in multi-step retrosynthesis

Area of Science:

• Computational Chemistry
• Artificial Intelligence in Chemistry
• Retrosynthesis Planning

Background:

• Multistep retrosynthesis is crucial for drug discovery and chemical synthesis.
• The efficiency of retrosynthesis planning algorithms, like Monte Carlo Tree Search (MCTS), is highly dependent on hyperparameter tuning.
• Balancing search time and route solvability is a key challenge in computational retrosynthesis.

Purpose of the Study:

• To investigate the impact of MCTS hyperparameters on multistep retrosynthesis performance.
• To identify optimal MCTS hyperparameter settings for improved speed and accuracy in retrosynthesis.
• To compare data-driven hyperparameter optimization strategies against default settings.

Main Methods:

• Evaluated MCTS hyperparameters (iterations, depth, width) using systematic grid search, Bayesian optimization, and a machine learning approach.
• Assessed performance using Linear Integrated Speed-Accuracy Score (LISAS) and Inverse Efficiency Score.
• Tested optimized hyperparameters on proprietary and public datasets (ChEMBL).

Main Results:

• Identified a hyperparameter set that significantly outperforms the default AiZynthFinder settings.
• Achieved 93% solvability with a 151s median search time on an in-house dataset.
• Reached 74% solvability with a 114s median search time on the ChEMBL dataset, surpassing default performance.

Conclusions:

• Data-driven hyperparameter optimization effectively enhances MCTS performance for multistep retrosynthesis.
• The identified optimal settings provide a superior balance of speed and solvability for retrosynthesis planning.
• Dynamic hyperparameter prediction shows promise for real-time optimization in chemical synthesis.