Modeling macromolecular structures and ensembles using AI

AlphaFold (Jumper et al., 2021) has revolutionized computational structural biology by setting new standards for what can be achieved with computational methods.

We are proud to have played a leading role in this development not only by introducing new and improved capabilities that enable predictions with even greater accuracy than the original version of AlphaFold. In this proposal, we suggest several improvements that will make AlphaFold more capable without explicit retraining.

These improvements are also transferable to other frameworks such as RoseTTAFold-All-Atom, which is not limited to protein-only systems but can also model DNA/RNA-protein as well as PTMs.

We also like to explore intermediate interactions and structural ensembles by retraining a network with a similar architecture as the AlphaFold network but with the added ability to reason over structural ensembles rather than single structures.

This can be achieved by using the AlphaFold network as a denoising model that transforms noisy structures into clean structures, like how text is used to generate images from noisy images using UNetIn the applied part of the project, we aim to use the developed tools to analyze specific dynamic protein systems for which we have first-hand access to experimental data: The oncoprotein MYC and its role in cancer, the Kv7.1 potassium channel, which repolarizes the cardiac action potential in response to voltage, and viral protein structure evolution.