They mention AI used to weed out molecules but don't mention how it was used.
I'd be interested in know did they use pre-trained models or train one from the ground up.
I would expect some custom training was needed in order to address this highly specialized use case.
There are 2 primary uses cases emerging where LLMs are proving advantageous.
1) Summarization --- where the results can be nebulous and it don't really much matter. For example, casual web search.
2) Research based on trial and error where the nebulous results will be subjected to thorough verification --- i.e. this case.
In any case, the statistical results can be nebulous and using them without verification is a recipe for disaster. For example:
I don't think it's a matter of firing up an LLM and saying "Hey LLM, please find me a molecule that acts on human GLP-1 receptors but does not have the negative side effects of semaglutide." And it's all like "Certainly! Here are some candidate GLP-1 agonists that blah blah blah..."
It's probably more like AlphaFold -- a statistical model of molecular structure and action, not a language model.
