By demonstrating in this study that three additional tripeptides are novel non-canonical PTS1 tripeptides, we show that novel tripeptides, even if positioned close to the prediction threshold, are correctly predicted as containing non-canonical PTS1 tripeptides.
As the TFBS prediction threshold increases, fewer TFBSs are predicted leading to less certainty about a gene's true operonic status.
Figure 4a shows that at a low TFBS prediction threshold (where more TFBS are predicted) it is more likely that sequences without a predicted TFBS are operonic (although only reaching a positive predictive value peak of ~0%).
One potential application to ML-based predistors is dynamically setting the prediction threshold such that the proportion of predicted active residues matches the estimated prior distribution for the query protein.
Indeed, only ~28% of genes predicted to be operonic are actually operonic at the highest TFBS prediction threshold.
To further assess the accuracy of the PWM models when the prediction threshold is varied, we have plotted them as receiver operator characteristic (ROC) curves (Figures 5,6).
We postulate that this finding is indicative for the assumption that based on the full current literature a meaningful percentage of the 'unknowns' that pass the prediction threshold will be actual pairs worth studying in more detail.
In a ROC diagram, the TP rate, or sensitivity, is plotted on the y-axis against FP rate (or false discovery rate) on the x-axis as the prediction threshold is varied over all possible values.
However, in many of these instances the misalignment affected only part of a repeat, or a minor fraction of the sequences in the alignment, and the scores obtained with the raw alignment were often still better than prediction threshold (data not shown).
The prediction threshold of 9% is not a commonly accepted threshold for distinguishing asthma from COPD.
