Assessment and development of a preexisting pipeline for computational design of multistate proteins

Tobias Johannes Dorer

Kopenhagen, Denmark

A proteins function relies on its dynamic behavior and the ability to program specific dynamics into de novo designed proteins is hence a momentous goal. Allostery, substrate binding and even enzymatic catalysis often depends on the specific dynamic behavior of proteins built by evolution.
While recent method development led to elaborate tools such as ProteinMPNN or RF-Diffusion for the design of proteins of arbitrary shape, the tools for the precise design of dynamic behvavior are yet to be created. Design pipelines for dynamic proteins lack elaborate inverse folding methods as well as computational methods to predict the populations of different conformations.
This thesis builds on improving the general design pipeline for the purpose of dynamic proteins. I show that the elaborate combination and modification of state-of-the-art deep learning tools promises higher success rates. I found that, while many new deep learning models are designed, the ingenious usage and the deliberate choice of existing models can have a strong influence on success. The comparison of probability distributions generated by inverse folding models holds information on the compatibility of conformations. With tied logits sampling in deep learning models, the resulting probability distribution given by the inverse folding tool can be stronger influenced by one state depending on the confidence of the model. Such principles can be used to screen and balance the sampling process for dynamic de novo proteins in the future.

Go back