Predicting reaction conditions: a data-driven perspective

Abstract

The selection of optimal reaction conditions is a critical challenge in synthetic chemistry, influencing the efficiency, sustainability, and scalability of chemical processes. While machine learning (ML) has emerged as a promising tool for predicting reaction conditions in computer-aided synthesis planning (CASP), existing approaches face many significant challenges, including data quality, sparsity, choice of reaction representation and method evaluation. Recent studies have suggested that these models may fail to surpass literature-derived popularity baselines, underscoring these problems. In this work, we provide a critical review of state-of-the-art ML techniques, identifying innovations which have addressed the key challenges facing researchers when modelling conditions. To illustrate how relevant reaction representations can improve existing models, we perform a case study of heteroaromatic Suzuki–Miyaura reactions, derived from US patent data (USPTO). Using Condensed Graph of Reaction-based inputs, we demonstrate how this alternative representation can enhance the predictive power of a model beyond popularity baselines. Finally, we propose future directions for the field beyond improving data quality, suggesting potential options to mitigate data issues prevalent in existing literature data. This perspective aims to guide researchers in understanding and overcoming current limitations in computational reaction condition prediction.