Seonglae ChoESM Fold, EvolutionaryScale
ESM 3
www.nature.com
https://www.nature.com/articles/d41586-024-02214-x
New glowing molecule, invented by AI, would have taken 500 million years to evolve in nature, scientists say
An artificial intelligence model has created a new protein that researchers say would have taken 500 million years to evolve in nature — if nature were capable of producing such a thing.
https://www.livescience.com/technology/artificial-intelligence/new-glowing-molecule-invented-by-ai-would-have-taken-500-million-years-to-evolve-in-nature-scientists-say
Protein folding prediction proceeds in two stages inside the model → AI protein folding models first create chemical relationships, then calculate spatial structure. Direct manipulation of internal representations (activation patching, steering) → enables intentional modification of protein structure. In other words, the model's internal computations are interpretable and causally manipulable
Stage 1 (early blocks) → transfers chemical information (charge, type, etc.) from amino acid sequences → to pairwise residue representations → i.e., preparation stage for "who can interact with whom"
Stage 2 (later blocks) → forms spatial information such as inter-residue distances and contacts → determines actual 3D structure
arxiv.org
https://arxiv.org/pdf/2602.06020