Research: CANCER Folding Project #17650

In drug discovery, particularly that of cancer, maximizing state exploration is a useful novel strategy – providing new protein states and conformations to point drug design methods at increases the likelihood that a potential binder and inhibitor may be found. However, in many cases a new state that is "useful for design" (ie.

ones distinct enough to be worth targeting to identify novel drugs) require a lot of sampling or simulation.

Sometimes, even exascale computers like Folding@home are not enough! Adaptive methods are very powerful here, but have the drawback of requiring system knowledge, or having to guess which protein features are worth adaptively exploring on, which may not always turn out to be true. Identifying druggable states or exploring conformational state space relevant to disease is an existing challenge.

The embarassingly parallel nature of Folding@home allows us to massively scale up our exploration.

However, the underlying methods still rely on luck to a large extent – we must discover the states in work units as the dataset grows in size and more work units are run.

This can be an incredibly inefficient process, wasting work units on regions of state space that are irrelevant or uninteresting to the question at hand.

Adaptive Seeding is a way to tackle this inefficiency.

Rather than applying a "boost" potential that alters the physics of our system, or having to do live-streamed analysis like Adaptive sampling, Adaptive Seeding sets up multiple starting structures across conformational space.

The intention is that having multiple distinct starting structures will increase the rate at which the landscape is traversed and lead to transitions/pathways connecting functional states in less simulation time, while preserving physics.

These projects seek to test different "seeding" approaches that yielded a different spread of starting structures.

Each unique structure starts at a different RUN.

As with other projects, we will be studying MET kinase.

17645: Run using starting structures generated using AI-methods like AlphaFold2
17646: Run using starting structures that are stemmed from "fixing" experimentally derived structures in the PDB
17649: A clone of 17645 but with 20X higher saving frequency for higher resolution dynamics
17650: A clone of 17650 but with 20X higher saving frequency for higher resolution dynamics.