AlphaFold-Multimer (AF-M) uses the same deep learning principles as AlphaFold to predict the structure of protein complexes. We installed the Colabfold version of AF-M locally and rented cloud-based GPUs to predict all binary protein-protein interactions (PPIs) among the core genome maintenance machinery. Each protein pair was folded in 3 out of the five uniquely-trained AF-M models with templates enabled. This pipeline generated an "all-by-all" matrix of potential PPIs. To save computing time, AF-M structures were not relaxed. Protein pairs that caused AF-M to exceed our available GPU capacity (longer than ~3600 residues total) were not folded (white squares in matrix).

To help assess whether an interaction is likely to be true, we trained a classifier called SPOC (see next section). In addition, we provide standard AF-M confidence metrics (PAE, pLDDT, pDOCKQ), as well as another metric, the average number of AF-M models that agree on a prediction ("avg_models").

We are not uploading datasets that were not generated in-house, but we welcome suggestions for new proteins to fold.

For more detailed information about some of the in-house scripts we use to analyze AlphaFold multimer data please visit our lab's GitHub.

Assessing Predictions

To assess AF-M predictions, we created SPOC (Structure Prediction and Omics Classifier; 0-1 scale), an algorithm that was trained to tell the difference between true and false AF-M predictions Schmid and Walter. We define a good SPOC score as a value above which no more than 5% of "hits" are likely to be incorrect (5% False Discovery Rate, FDR). This value is context-dependent. For example, when screening a limited group of proteins that should be enriched for real interactors (e.g. IP-mass spectrometry data), a 5% FDR is achieved at a SPOC score of >0.75 (16:1 curve in Figure 3F in Schmid and Walter). On the other hand, when performing a proteome-wide screen with a lower proportion of true interactors, a SPOC score of 0.95 is needed to achieve 5% FDR (128:1 curve).

AF-M predictions inevitably yield 'false positives' and 'false negatives.' Indeed, a low SPOC score is not evidence of non-interaction (as seen from some known complexes achieving only low scores; Figure 5B in Schmid and Walter) and a high score does not provide definitive evidence of interaction (as seen from the good scores for protein paralogs like MCM2 and MCM7 that do not normally interact). In general, we recommend prioritizing PPIs with the highest SPOC scores first, and/or ones that are supported by independent evidence or help explain a biological phenomenon. In all cases, experimental evidence is essential to validate all predictions.


This website and SPOC were created by Ernst Schmid in consultation with Johannes Walter.

For SPOC and PPIs found on, please cite: Schmid and Walter.

For AF-M, please cite


Date Comment
12-04-2023 We found an error in our original calculation of pDOCKQ. The calculation has been corrected to match, and all values on the site have been updated. The same error occurred in Lim et al. 2023, which will be corrected. We sincerely apologize for any inconvenience this may have caused.
09-05-2023 The site is officially released to the public.