3 ms·
AlphaFold2 has learned a new trick: How to do chemistry
- ajuc 4y agoThe training set had shapes that the proteins usually take - which is a result of interactions with the things that usually are in their environment (including copies of itself). So to correctly predict that - the algorithm had to learn to simulate these interactions, right?
- jleyank 4y agoTo be pedantic, the ai/ml approach has no concept of the protein folding process, merely the structural motifs that result when it is complete. These are the X-ray or cryoEM environments used in the training process. The 50-year hunt has been to be able to understand and simulate the dynamics involved in folding itself. This seems so complex to simulate yet nature finds it almost trivial. Protein homology, deriving structures from similar proteins, is the “parent” of this, not molecular dynamics.