We compiled a set of experimentally-derived PWMs representing the binding specificities of 158 C2H2-ZF TFs from human, mouse, rat, chicken, worm, fly, yeast, sea pineapple, African frog, and corn, as in Persikov and Singh (Nucleic Acids Res, 2014). These experimentally-derived PWMs are considered to be our gold standard; computationally-predicted PWMs are evaluated as correct if they are in agreement with these PWMs. For each of these 158 TFs, we  predicted their binding specificity using the three methods SVM, ML, and RF. We then aligned the SVM-predicted and experimentally-derived PWMs for each TF.  Each column of the SVM-predicted PWM is marked as ''correct'' if the PCC between it and the aligned column from the experimentally-derived PWM is greater than a particular score threshold. We tested four score thresholds: 0, 0.25, 0.5, and 0.75.  For each of these four score thresholds, therefore, we can thus calculate the total proportion of SVM-predicted columns across all TFs that are correct.  We then looked at how the proportion of correctly predicted columns changes when we limit the SVM-predicted columns to those that are in agreement with one or more of the two other prediction methods. We look at the subsets of SVM-predicted columns that have a PCC > 0, 0.25, 0.5, and 0.75 to the corresponding column from either the ML-predicted or RF-predicted specificities. At each correctness score threshold, as described above, the proportion of correctly-predicted columns increases as we require stricter consensus cutoffs. We see this same trend when considering columns from all 158 TFs (Table S2A), and when considering columns from the subset of 60 D. melanogaster TFs (Table S2B).