An active learning scheme for optimizing protein sequences

We demonstrated how the wazy package (Yang et al. 2022, developed by members of the White lab at U. Rochester) can be trained on protein sequence-property prediction tasks. For training we ran coarse-grained simulations using HOOMD-blue 2.9.7 extended with azplugins. The simulations were run on the MOGON II computing cluster of JGU Mainz.

Here, we provide the code used for training and the results of the simulations (extracted quantities, e.g. $B_{22}$ or $\Delta G$ for protein sequences), along with the scripts used for generation the simulations and computing aforementioned quantities.

The code presented here was used in the study by Changiarath, Arya, Xenidis, Padeken, Stelzl 2024, under review for Faraday Discussions.

Dependencies

Our code builds on wazy, which performs the featurization using UniRep, and has its own construction for doing bayesian optimization using MLPs as a surrogate model. We use localCIDER for computing descriptors.

Apparently metapredict has a lot of dependencies and it'll download large libraries while installing.

pip install cython metapredict wazy localcider