Part III: AlphaFold2

Question

For our exercises, we have been looking at CASP13. Now we will take a look at CASP14 to

learn about a particularly important new result.

(https://predictioncenter.org/casp14/zscores_final.cgi). A new method called AlphaFold2

shattered the competition by predicting the most accurate protein tertiary structures ever

reported. Use the various links on https://predictioncenter.org/casp14/index.cgi to answer the

following questions:

1. Who came in second? What was their average z-score? How did this compare to

Alpha Fold?/nUse https://predictioncenter.org/casp14/results.cgi?view=tb-sel to compare the runner-up to

AlphaFold2. Select AlphaFold2 and the runner-up on the left, all targets on the right, and

click show results. Click the little arrows in the RMSD column to sort by RMSD (smallest to

largest).

2. What is the smallest RMSD achieved by AlphaFold2 and what was the smallest

RMSD for the runner-up?

3. How many of the first entries sorted by RMSD belong to AlphaFold2?

AlphaFold2 used a new artificial intelligence algorithm recently developed for learning

human language. It is the first application of this type of AI to a bioinformatics problem that

has demonstrated this extraordinary level of success. In terms of our ability to predict the

structure of proteins, AlphaFold2 is indeed a breakthrough.

Use the following papers as starting points to learn more about AlphaFold2 and answer the

questions that follow.

https://www.nature.com/articles/d41586-021-02265-4

• https://www.nature.com/articles/s41586-021-03819-2

• https://www.nature.com/articles/s41592-021-01365-3

4. AlphaFold2 invents a neural network block they call an evoformer. What type of

neural network is the evoformer based on? What are some of the well-known

applications for this network?

5. What types of information are used to train AlphaFold2?

6. AlphaFold2 has an average GDT_ TS score of 92. Why is this so impressive?

7. What are some of the key limitations of AlphaFold2?

8. Some have said that AlphaFold2 represents such a tremendous breakthrough that "the

protein folding problem is solved". Is this true? Why or why not?

9. How does WSME-Linker improve on AlphaFold2?