With all the advances happening in artificial intelligence (AI), it’s natural to ask if one could develop an algorithm that predicts protein dynamics with far less computer time than running simulations takes.
One of the challenges is having enough training data. One would really need lots of atomically-detailed models of protein dynamics. Such models are hard to derive from experiments though, and most simulations data sets are pretty limited given how much computational investment it takes to generate them.
But Folding@home has LOTS data!
A recent study shows how this data can be used to take a first step towards AI-generated protein dynamics. The algorithm, called BioEmu for biomolecular emulator, takes a protein sequence as input and outputs an ensemble of protein structures. A large portion of the training data came from Folding@home.
In our internal tests, BioEmu recapitulates some of the features we see in simulations. However, its early days and it misses other things.
With your help, we’ll keep expanding the training data need to drive new advances. It will be exciting to see how things progress!
