Some of the lower-hanging fruit in chemistry data was addressed by earlier versions of deep learning, like AlphaFold. It seems the nature of the domain is such that the language is less ambiguous than most natural language. Does anyone have a perspective on the apparent advantages of mapping chemistry interactions to latent space models for LLM training?
Any arbitrary but meaningful seq2seq data can possibly be modelled by a transformer.
Language models can be trained to generate novel functioning protein structures (by training on protein functions and their corresponding sequences), bypassing any sort of folding process entirely.
