Fig. 5: AlphaFold2 prediction of Gp38-OmpA binding.

To compare MG1655 OmpA variant infection data with all adhesin-receptor predictions using AlphaFold2-multimer, the directionality of receptor/adhesin pairs was determined by defining the principal axis of inertia for each subunit, serving as a proxy for potential binding pairs. A correct interaction model must exhibit acceptable directionality. The percentage of predicted models with acceptable directionality deviates from random placement across different angle (α) thresholds (a). Proportion values for each adhesin/receptor pair were calculated, as well as the proportion for a thousand generated random datasets, each containing 54 samples with random vector orientations. The distribution of the AlphaFold2 pair models was significantly different from all random distributions (Mann–Whitney U test), with a mean p value of 0.007 (b). We tested (two-sided t test) if each individual proportion was different from 0.25. * Represents a p value < 0.05; green represents the true binding partners (confirmed by experiments in Fig. 4c). The dashed line represents the expected proportion for a randomly generated dataset at α = 90° (0.25) (c). A confusion matrix shows the reliability of using such criteria to select true positives (d).