Efficient Machine Learning for Yield Prediction in Transition Metal-Catalyzed <i>meta</i> -C(sp ² )–H Bond Activation and Other Reactions

Developing machine learning (ML) models for yield prediction of chemical reactions has emerged as an important use case scenario in very recent years. In this space, reaction datasets present a range of challenges mostly stemming from imbalance and sparsity. Herein, we consider chemical language representations for reactions to tap into the potential of natural language processing models such as the ULMFiT (Universal Language Model Fine-Tuning) for yield prediction, which is customized to work across such distribution settings. We contribute a reaction dataset with more than 860 manually curated reactions collected from literature spanning over a decade, belonging to a family of catalytic meta-C(sp2)–H bond activation reactions of high contemporary importance. Taking cognizance of the dataset size, skewness toward the higher yields, and the sparse distribution characteristics, we developed (i) a time- and resource-efficient pretraining strategy for downstream transfer learning, and (ii) the CFR (classification followed by regression) model, which provided more accurate yield predictions than the traditional direct regression (DR) methods. Instead of the prevailing pretraining practice of using a large number of unlabeled molecules (1.4 million) from the ChEMBL dataset, we first created a pretraining dataset SSP1 (0.11 million) by using substructure-based mining from the PubChem database, which is found to be equally effective and more time-efficient in offering enhanced performance. The CFR model with the ULMFiT-SSP1 regressor achieved a notable RMSE of 8.40 ± 0.12 for the CFR-major and 6.48 ± 0.29 for the CFR-minor class in yield prediction on the title reaction, with a class boundary of yield at 53%. Furthermore, the CFR model is highly generalizable as evidenced by the significant improvement over the previous benchmarks on Buchwald–Hartwig coupling, Suzuki coupling, nickel-catalyzed C–O coupling, and USPTO reaction datasets. Our model provides a scalable and efficient framework that can assist in reaction discovery.