Fig. 5: Identifying AlphaFold3 metrics for accurate prediction of flg22 recognition.
From: Unlocking expanded flagellin perception through rational receptor engineering
a, ROC analysis of AlphaFold3. Four confidence metrics were separately analysed for 219 FLS2–flg22–SERK3 complexes: chain-pair average ipTMflg22–FLS2, ipTMflg22–SERK3, Minimal (Min) PAEflg22–FLS2 and Min PAEflg22–SERK3 from three independent modelling approaches. The diagonal dashed line represents a random classifier, and AUC values are annotated for each metric, indicating the model’s performance. b, ROC analyses using average ipTMflg22–FLS2 as the predictive metric. Three groups were analysed separately: all data (blue, n = 219), AtFLS2 only (cyan, n = 97) and non-AtFLS2 (green, n = 122). c, Average ipTMflg22–FLS2 for each FLS2–flg22–SERK3 complex whose perception has been validated experimentally (n = 219 combinations). Each dot represents the average ipTMflg22–FLS2 of three independent predictions extracted from each FLS2–flg22–SERK3 combination. Two-sided Wilcoxon rank-sum test was used to determine significance. The box in the boxplot represents the interquartile range (IQR), with the lower and upper hinges corresponding to the first and third quartiles, respectively. The whiskers extend to the smallest and largest values within 1.5× the IQR from the hinges, while the central line indicates the median. d, Accuracy with different ipTMflg22–FLS2 thresholds. Dashed vertical lines indicate the optimal threshold points for each group where accuracy is maximized with the ipTM threshold = 0.82 for all data and non-AtFLS2 data, and 0.83 for only AtFLS2 data. All data accuracy, 86.8%; AtFLS2 accuracy, 96.9%; non-AtFLS2 accuracy, 79.5%. e, Accuracy of AlphaFold3 for predicting immunogenicity of flg22 variants selected in this study (n = 10 receptors × 10 flg22 variants). Dots indicate the percentage similarity of each flg22 variant to the canonical Pae flg22. f, AlphaFold3 accuracy using a 0.82 ipTMflg22–FLS2 threshold for perception of flg22 ligands. g, AlphaFold3 models of AtFLS2 (Arabidopsis thaliana), AtBAK1 and flg22 from Pae that is perceived, the flg22 variant from Atu that is not perceived, and the impossible ligand csp22 from cold shock protein. High pLDDT indicates higher confidence and more accurate prediction.
