AlphaFold2 predicts interactions amidst confounding structural compatibility
This article has been Reviewed by the following groups
Listed in
- Evaluated articles (Arcadia Science)
Abstract
Predicting physical interactions is one of the holy grails of computational biology, galvanized by rapid advancements in deep learning. AlphaFold2, although not developed with this goal, seems promising in this respect. Here, I test the prediction capability of AlphaFold2 on a very challenging data set, where proteins are structurally compatible, even when they do not interact. AlphaFold2 achieves high discrimination between interacting and non-interacting proteins, and the cases of misclassifications can either be rescued by revisiting the input sequences or can suggest false positives and negatives in the data set. Alphafold2 is thus not impaired by the compatibility between protein structures and has the potential to be applied at large scale.
Article activity feed
-
close to the interaction cutoff,
I'm sort of interested in the pairs that are close to the interaction cutoff but on the expected side. For example, would a pair with an interaction score of like 0.56 or so have the structure and interaction that you would expect? I guess, how accurate is the 0.5 cutoff?
-
The fact that no recycling is required opens the possibility to apply this procedure at large scale
This is really exciting! I can think of tons of really interesting hypotheses that could be tested by doing this kind of analysis at a larger scale, and I think your analysis here does a great job opening that up.
-
meaningless because of the disordered parts
Do you see other cases of disorder influencing the predictions of binding? Maybe not to this degree, but are there other cases where disordered regions cause a lower or higher ipTM than expected?
-
There is evidence of physical interaction between these proteins, as detected by affinity purification. However, there is no evidence of direct physical interaction by two-hybrid assay.
Might consider adding a citation here
-
The 5 models are drastically different from each other
Is this similar to what you see with non-interacting pairs? Are the protein structures themselves quite different or just the contact between the two models?
-
I submitted to AF2 prediction a particularly challenging data set from a previous study
I really appreciate the use of this previous dataset for this purpose! I went back to the previous paper, and this dataset seems like the perfect fit for this kind of analysis. I think this is a great test of the limits of this particular feature of AlphaFold2 and provides some great insight into what it can be useful for in the future.
-
