AlphaFold 2 wasn't Q-learning based. It was supervised SGD and the "evoformer" they introduced is very close to a transformer. So it's not exactly an LLM, but it's a pretty close equivalent for protein data.