Table 5.

Classification performance of AlphaFold2 models predicted for the physiological and non-physiological homo dimers of the benchmark dataset. This is evaluated on the subset of 1480 targets for which results were produced with all AlphaFold2 versions.

Column 1 lists the version of AlphaFold used for the predictions, and the score used to quantify the similarity between the AlphaFold2 predicted model, and the homodimer structures of the corresponding benchmark entry (see Main text for detail). Column 2 lists the Area Under the Curve (AUC) of the ROCs computed using the listed scores. Columns 3 and 4 list the mean and median values for the computed scores considering only the physiological dimers. Those for the non-physiological dimers are not reported, as AlphaFold2 tends to predict alternative association modes for a significant fraction of these dimers, as expected.

Score	ROC AUC	Model Accuracy Mean (Physiological dimer)	Model Accuracy Median (Physiological dimers)
QS-score (AF-gap)	0.938	0.778	0.914
QS-score (AF-multimer)	0.957	0.833	0.940
QS-score (AF_multimer relaxed)	0.957	0.832	0.938
DockQ (AF-gap)	0.954	0.707	0.833
DockQ (AF-multimer)	0.964	0.786	0.877
DockQ (AF-multimer relaxed)	0.964	0.787	0.876