Fold and dock proteins using reinforcement learning
||Fold and dock proteins using reinforcement learning|
||Björn Wallner <firstname.lastname@example.org>|
||2021-08-17 – 2022-09-01|
In the last couple of years machine-learning (ML) methods that play various computer and board games has been developed. Despite being fun the practical benefit of having a computer playing computer games is rather limited. Here we want to experiment with a similar technique to train ML methods to fold and dock proteins. We will integrate ML into the Rosetta package for macromolecular modelling, which will enable us to sample conformational space and use ML to assess the sampled conformation to make guided decisions during the sample trajectory. We have already used pre-trained ML methods to guide sampling in Rosetta. Here, we will allow the method train online to adapt and learn during sampling. This means that we will generate the input data on the fly, and finding the right balance between data generation and weight updates (training) will be crucial. We believe that having access to both CPU and GPU nodes will be vital for the success of this project.
In addition, we will also use the allocation to run the AlphaFold2 protein structure prediction method. This method requires GPU at the inference step.