Getting low interaction scores with existing complexes

[outlierd29]

Hello there,

We are currently running different experiments with somewhat confusing results. We selected two proteins, STING (homodimer) and NPC1, to test the af2complex model. According to this paper -> (https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8859990/), a physical interaction between them is expected. However, our predicted "Interface Score" for “STING:2+NPC1” is only 0.0065, which is lower than the scores for some random proteins we tested (ranging from 0.01 to 0.24).

We generated the features using the Alphafold2 collar notebook from the deepmind repository (https://github.com/deepmind/alphafold/blob/main/notebooks/AlphaFold.ipynb) and then used them as input for Af2complex, both with default settings.

I was wondering if someone could help us understand if we are using the model in the right way and interpreting the results correctly.

[FreshAirTonight]

Your are correct that the interface score you observed is insignificant. There are several possibilities: First, the deep learning modeling missed the expected interactions, i.e., a false negative. For example, hypothetically, if the N-term transmembrane helix of NPC1 moves dramatically to grab the STING dimer, it will be very difficult to model such conformational changes correctly. Second, the complex model might be a physically correct, but the confidence metric is not sure about it. If the inter-protein contacts make physical and biological senses, the complex model may still be useful even with a low score. Third, other protein(s) might be involved in the complexation but not considered. The STING dimer might not directly interacts with NPC1. The evidence presented in the reference regarding the direct physical interactions is based on immunoprecipitation from cell extracts, which could be noisy and one cannot rule out the involvement of other unknown players. If you have a short list of other possible players, you may try to screen the interactions with either STING dimer or NPC1, and see if they help to improve complex modeling.

One last note, you will get a high score with "SING:2/NPC1", but the high value is largely due to the SING dimer, rather than the STING and NPC1 interactions. It is important to confirm that the interactions between the dimer and NPC1 are confident according to the AF2 models.