Comparative Analysis of Template-Based and Neural Network Approaches for Computer-Aided Retrosynthesis: ASKCOS vs. IBM RXN

Keywords: LivChat, retrosynthesis, chemistry

TL;DR: Comparing publicly available retrosynthesis models using an AI virtual assistant

Abstract: Computer-aided retrosynthesis tools are transforming synthetic chemistry by automating the design of synthetic routes. This study compares two leading retrosynthesis models: ASKCOS, which uses a template-based approach, and IBM RXN, which employs a neural network-based methodology. Using three target molecules—ethyl acetate, ε-caprolactone, and 2,4,6-tribromoaniline—we evaluate each model’s predictions, highlighting their strengths and weaknesses. ASKCOS generally excels at proposing literature-supported, interpretable synthetic routes, while IBM RXN demonstrates creativity and flexibility, occasionally at the expense of practicality. We discuss optimal use cases for each model and propose future directions for improving both model performance and evaluation methodologies.

Submission Number: 107