Abstract:
Advances in analytical instrumentation have provided the possibility of examining thousands of genes, peptides, or metabolites in parallel. However, the cost and time-consuming data acquisition process causes a generalized lack of samples. From a data analysis perspective, omics data are characterized by high dimensionality and small sample counts. In many scenarios, the analytical aim is to differentiate between two different conditions or classes combining an analytical method plus a tailored qualitative predictive model using available examples collected in a dataset. For this purpose, partial least squares-discriminant analysis (PLS-DA) is frequently employed in omics research. Recently, there has been growing concern about the uncritical use of this method, since it is prone to overfitting and may aggravate problems of false discoveries. In many applications involving a small number of subjects or samples, predictive model performance estimation is only based on cross-validation (CV) results with a strong preference for reporting results using leave one out (LOO). The combination of PLS-DA for high dimensionality data and small sample conditions, together with a weak validation methodology is a recipe for unreliable estimations of model performance. In this work, we present a systematic study about the impact of the dataset size, the dimensionality, and the CV technique used on PLS-DA overoptimism when performance estimation is done in cross-validation. Firstly, by using synthetic data generated from a same probability distribution and with assigned random binary labels, we have obtained a dataset where the true classification rate (CR) is 50%. As expected, our results confirm that internal validation provides overoptimistic estimations of the classification accuracy (i.e., overfitting). We have characterized the CR estimator in terms of bias and variance depending on the internal CV technique used and sample to dimensionality ratio. In small sample conditions, due to the large bias and variance of the estimator, the occurrence of extremely good CRs is common. We have found that overfitting peaks when the sample size in the training subset approaches the feature vector dimensionality minus one. In these conditions, the models are neither under- or overdetermined with a unique solution. This effect is particularly intense for LOO and peaks higher in small sample conditions. Overoptimism is decreased beyond this point where the abundance of noisy produces a regularization effect leading to less complex models. In terms of overfitting, our study ranks CV methods as follows: Bootstrap produces the most accurate estimator of the CR, followed by bootstrapped Latin partitions, random subsampling, K-Fold, and finally, the very popular LOO provides the worst results. Simulation results are further confirmed in real datasets from mass spectrometry and microarrays.