Investigations into the Efficiency of Computer-Aided Synthesis Planning.

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Machine learning for synthesis planning

Background:

• Machine learning (ML) models are crucial for efficient chemical synthesis route generation.
• Current retrosynthesis models are often too slow for practical application.
• Reducing inference time and carbon footprint of ML models is a significant challenge.

Purpose of the Study:

• To decrease inference times of the Chemformer model for retrosynthesis.
• To investigate alternative transformer architectures, knowledge distillation (KD), and hyper-parameter optimization.
• To assess the impact of these optimizations on single-step and multi-step search efficiency.

Main Methods:

• Evaluated closely related transformer architectures with knowledge distillation (KD).
• Investigated feature-based and response-based KD.
• Optimized hyper-parameters based on inference time and model accuracy.
• Assessed single-step prediction speed and multi-step search efficiency.

Main Results:

• Alternative transformer architectures under-performed when using KD.
• Reducing model size and improving single-step speed are important but not sufficient.
• Multi-step search efficiency is more influenced by the diversity and confidence of single-step predictions.
• Knowledge distillation alone did not significantly improve multi-step efficiency.

Conclusions:

• Further research should combine KD with other techniques for improved synthesis planning.
• Multi-step search efficiency in retrosynthesis is complex and influenced by factors beyond single-step model speed.
• Balancing exploration and exploitation in Monte Carlo-based retrosynthesis is critical.