202409212321
Status: #lecture
Tags: #ai #drug_discovery #alphafold #proteomics #protein_folding
# AlphaFold Multimer for Interaction Prediction
Presented by Ernst Schmid
Website: https://predictomes.org/
- Created method to measure confidence of protein-protein interactions by:
- Finding interacting residue pairs between the two proteins (where "interacting" is residues that are close together)
- Looking at confidence (pLDDT) of the individual interacting residues
- Finding the estimated relative distance error of the pair of interacting residues (i.e. how confident is AlphaFold in the distance between these two residues)
We can make predictions of protein-protein interactions by looking at an overall score of AlphaFold's "confidence" in the residue interactions using the above methods. Sometimes AlphaFold will just place residues of two different chains in a multimer near each other arbitrarily, so it is important to distinguish between arbitrary placements and asserted interactions.
Below is a selection from their matrix looking at proteins involved in human genome maintenance. Each cell in the matrix represents a score of the interactions between proteins $i$ and $j$ for some indices $i$ and $j$ in the set. Colored squares in the matrix represent predictions in which at least one pair of interfacial residues is closer than 8Å and meets minimum confidence criteria (PAE < 15 and pLDDT > 50).
![[Pasted image 20240305110428.png]]
Also interesting part of the talk - he created an easy-to-use interface for AlphaFold structure prediction. Allows users to submit jobs using a UI which get sent to private cloud GPUs. After the job finishes, it runs through an analysis pipeline that checks the protein-protein interaction confidence, and then finally saves the results to Dropbox.
---
# References