Reliable identification of homodimers using AlphaFold

Protein-protein interactions are central to understanding biological processes. Predicting interaction partners is extremely valuable for avoiding costly, time-consuming experiments. AlphaFold has shown an unsurpassed ability to predict protein-protein interactions accurately. However, a protein can also form homomeric interactions, that is, interact with itself. When identifying homodimers, AlphaFold yielded a significantly higher false-positive rate than when identifying heterodimers. True Positive Rate (TPR) at 1% False Positive Rate (FPR) drops from 81% for heterodimers to 22% for homodimers. When we investigated the high-scoring false positives, that is, non-homodimers with high AlphaFold scores that were predicted as such, we found that their homologs were enriched for homomultimeric proteins, indicating that stoichiometry is not always conserved among homologs. We split our dataset into a training and test set and with a simple logistic regression model that combines AlphaFold scores with stoichiometry information among homologs and interface features from the predicted structures. This logistic regression model was able to increase the performance on the test set from 27% TPR (at 1% FPR) to 36%. Without homology stoichiometry information, no improvement was seen. The higher number of high-scoring false predictions for homodimers indicates that homodimer prediction still has significant room for improvement.