Hong Kong-based AI drug discovery company Insilico Medicine, in collaboration with NVIDIA, has unveiled a new language processing tool for biomedical and chemical tasks, named nach0. This AI model is designed to solve chemical and biological problems, generate new molecules, and answer biomedical questions.

The creators of nach0 claim that the model can bridge the gap between biomedical natural language texts and chemical structure descriptions. It consists of abstract texts extracted from PubMed and chemistry-related patent descriptions taken from the US Patent and Trademark Office (USPTO) as well as molecular structures by the simplified molecular-input line-entry system (SMILES). The model reportedly includes up to 100 million documents and 2.9 billion patent descriptions.

The system is built on the NVIDIA BioNeMo generative AI platform for drug discovery applications. The gathered data has been converted into tokens and given unique annotations by the researchers in order to train the system. The training was done using NVIDIA NeMo end-to-end custom generative AI platform, and the datasets were managed by NVIDIA’s memory-mapped data loader modules.