Bridging CGR representations and language models for reaction property prediction

Abstract

Accurate prediction of reaction properties, such as activation energies, enthalpies, rate constants and yields, is key to the design of efficient and sustainable chemical processes. While graph-based models using Condensed Graphs of Reaction (CGR) have shown strong performance, text-based transformer models trained on reaction SMILES have gained popularity due to their scalability and performance. In this work, we analyze the performance of an existing CGR-based string representation for reaction property prediction, the SMILES/CGR. In contrast to conventional reaction SMILES, this representation explicitly encodes atom and bond changes within a unified SMILES-like sequence. A key limitation of SMILES/CGR is that it does not encode stereochemistry. To overcome this, we introduce the superimposed reaction SMILES (sr-SMILES), a compact CGR-based representation that preserves full stereochemical information. To evaluate both representations, we pretrained a BERT transformer on two million USPTO reactions via masked language modeling, and then finetuned separate models on six benchmark datasets. We compared against models trained on standard reaction SMILES, naïve baselines, and results from the literature. Our results show that sr-SMILES consistently matches or outperforms reaction SMILES, particularly on benchmarks where mechanistic information is critical, as well as in low-data regimes. In contrast, SMILES/CGR tends to underperform, highlighting that the details of CGR encoding matter. While sr-SMILES did not close the performance gap with graph-based CGR methods, it shows that language models can benefit from explicit mechanistic information when appropriately encoded. Moreover, by comparing SMILES/CGR and sr-SMILES models with different atom mappings, we reinforce the notion that CGR-based approaches are highly sensitive to the mapping quality. A plausible explanation for the similar performance observed across string-based representations is that language models implicitly learn the atom mapping during (pre-)training, rendering explicit mapping unnecessary for many prediction tasks.