Abstract
Predicting molecular structures with target properties and specific reaction conditions is a critical task in drug discovery and material science. Simplified molecular input line entry system (SMILES) representation, widely used in molecular generation tasks, encodes molecular graph structures as linear sequences of characters. This redundancy introduces ambiguity in feature mapping, potentially confusing generation models. To address this issue, we propose a pretrained contrastive learning-based SMILES Transformer Encoder (Contras-STE) to capture invariant, high-level features. This approach guides the SMILES generation model in learning the inherent structure of SMILES while maintaining a differentiable conversion process. To avoid nondifferentiability, we sample the generated SMILES from the token distribution using Gumbel-Softmax and input them into Contras-STE to compute the InfoNCE loss, which quantifies the similarity between the generated SMILES, target SMILES, and irrelevant SMILES. We test Contras-STE on MoleculeNet benchmarks against existing fingerprint-based methods and RNN-based methods, and the results show that Contras-STE outperforms other methods in most cases. To evaluate the performance of the SMILES generation model with Contras-STE, we test models on the OSDAs prediction task under synthesis conditions and the properties of zeolites production, and the results demonstrate that Contras-STE can highly improve the validity rate and novelty rate of generated SMILES.