Figure 2
From: Protein embeddings predict binding residues in disordered regions
Test set performance. Methods to predict binding regions in IDPRs (intrinsically disordered regions): IDBindT5 (here), SOTA methods ANCHOR214 and DeepDISOBind34, and a Random baseline (frequency of 38.8% positives and 61.2% negatives). Data set: test set (Mobi195 with 195 proteins). The performance is captured by per-residue measures (scaled to [0,1]): precision (Eq. 1), recall (Eq. 2), balanced accuracy (Eq. 5), F1-Score (Eq. 6) and MCC (Eq. 7). Error bars reflect the 95% confidence interval (CI) of 1.96*standard errors. Only IDBindT5 is significantly better than the random baseline for all measurements. While it also reaches the highest numerical values for all measures, not all differences are significant at 95% CI. Noteworthy is the superior recall value at top precision (more measures in Supplementary Table SOM_T4).
