Abstract
BACKGROUND: An important and yet rather neglected question related to bioinformatics predictions is the estimation of the amount of data that is needed to allow reliable predictions. Bioinformatics predictions are usually validated through a series of figures of merit, like for example sensitivity and precision, and little attention is paid to the fact that their performance may depend on the amount of data used to make the predictions themselves. RESULTS: Here I describe a tool, named Fragmented Prediction Performance Plot (FPPP), which monitors the relationship between the prediction reliability and the amount of information underling the prediction themselves. Three examples of FPPPs are presented to illustrate their principal features. In one example, the reliability becomes independent, over a certain threshold, of the amount of data used to predict protein features and the intrinsic reliability of the predictor can be estimated. In the other two cases, on the contrary, the reliability strongly depends on the amount of data used to make the predictions and, thus, the intrinsic reliability of the two predictors cannot be determined. Only in the first example it is thus possible to fully quantify the prediction performance. CONCLUSION: It is thus highly advisable to use FPPPs to determine the performance of any new bioinformatics prediction protocol, in order to fully quantify its prediction power and to allow comparisons between two or more predictors based on different types of data.