Transformers in drug discovery: fine-tuning ChemBERTa for high-accuracy prediction of solubility, toxicity and...

Drug discovery is still an expensive and time-consuming enterprise, a majority of clinical trial failures being traced to inaccurate predictions of molecular properties. To address this challenge, a new technique: fine-tuning transformer-based model ChemBERTa, is presented to achieve an accurate molecular property prediction. The hybrid architecture outperforms traditional networks, such as MolBERT, by combining ChemBERTa and graph neural networks. This superiority is supported by statistical tests, and the effectiveness of this model on systems of complex chemical structures is proved by the systematic error analysis. The technique proves effective on pharmaceutical and non-pharmaceutical compounds alike in the pesticide sector. This approach underlines the ability of AI to help and improve efficacy in drug discovery, lower the cost and improve clinical success rates.