Winter Science Chat

Started by agcohn821

agcohn821 Staff Lv 1

Hi everyone!

Just a friendly reminder that our science chat will be happening tomorrow at 12:30 PM PST; 8:30 GMT in #veterans chat. Hope to see you there!

bkoep Staff Lv 1

I think we answered most of the questions that were posted in advance, but there were a couple more from LociOiling that we didn't get to (about new LUA functions and performance issues).

We answered those in a comment here.

01010011111 Lv 1

What is the implementation of deep learning in protein structure predictions?

open AI algorithm Mohammed AlQuraishi Predict protein combinations in seconds. will there be news of collaboration in deep learning and protein computing?

Does the rosetta use genetic algorithms, or is it trying to create an optimization model with a more efficient algorithm?

In past Rosettа news, it was said that a project can accurately model small peptides, what is the problem in modeling large peptides? - news about rarely comes out, the project should please civil scientists with news) Thank you

bkoep Staff Lv 1

Yes, I think you're probably referencing something like this article about AlQuraishi's method (full reference here)? This is a very cool approach that is super-fast, but it didn't perform so well in the most recent CASP contest (the Nature article suggests unrelated technical difficulties).

The biggest impact for AI structure prediction so far has come from AlphaFold (from Google's DeepMind group), which swept the CASP contest in 2018. We are still awaiting a paper that describes the method in detail, but they have shared their major insights. Andrew Senior from DeepMind gave a talk here at UW a few months ago—you can watch that talk on YouTube here.

Rosetta does not make much use of genetic algorithms, but tends to rely more on Markov-chain Monte-Carlo sampling, with local optimization by simple gradient descent. If you're interested in Rosetta specifically, I recommend this YouTube lecture series by Jeff Gray at Johns Hopkins University.

The problem with larger peptides is that they have more bonds, which mean more degrees-of-freedom (DOFs)! To correctly predict the structure of a protein, we have to correctly predict the value of each DOF (i.e. the rotation of each bond). The sampling space increases exponentially with the number of DOFs (imagine that each DOF adds a new dimension to your sampling space). The problem with modeling large peptides is simply that the sampling space is vast (the number of possible folds is astronomically large), so optimization is very difficult to do with limited time/resources.