Class prediction and discovery using gene microarray and proteomics mass spectroscopy data: curses, caveats, cautions

Abstract

Using very simple classifiers, we show for several publicly available microarray and proteomics datasets how these curses influence classification outcomes. In particular, even if the sample per feature ratio is increased to the recommended 5-10 by feature extraction/reduction methods, dataset sparsity can render any classification result statistically suspect. In addition, several 'optimal' feature sets are typically identifiable for sparse datasets, all producing perfect classification results, both for the training and independent validation sets. This non-uniqueness leads to interpretational difficulties and casts doubt on the biological relevance of any of these 'optimal' feature sets. We suggest an approach to assess the relative quality of apparently equally good classifiers.

Protein expression in experimental malignant glioma varies over time and is altered by radiotherapy treatment

“…For mass spectra in a biomedical context SFRs are typically in the range 1 20 -1 500 . In machine-learning approaches such as neural networks the conventional solution is to reduce the feature space dimensionality by variable selection (Somorjai et al, 2003). The multivariate methods described and used herein rely on linear algebraic operations and are transparent in contrast to neural nets (for example), which are often seen as black boxes.…”

Pattern Recognition Analysis of MR Spectra

“…6 For this reason, the features containing little discriminative power must be discarded, which means that we will need a large number of cases (about three to five times more cases than the number of features to extract) from which to extract those features. 7,8 However, in real life, the number of available cases will be limited by epidemiology and budget. This problem is called the curse of dimensionality.…”

“…This problem is called the curse of dimensionality. 7 It is common to try to compensate for this restriction with large multicenter studies 8 (see Clinical Trials of MRS Methods), which brings an added problem: data compatibility in the face of slightly different acquisition conditions (field strength, localization pulse sequence, echo time (TE), and recycling time, among others). These factors will introduce variability or noise into the classifier training process.…”

Prediction of carcinogenic potential by a toxicogenomic approach using rat hepatoma cells

“…Recently, use of this approach for predictive toxicology using gene expression profiles has been reported by several groups. (6)(7)(8)(9)(10)(11)(12)(13)(14)(15)(16)(17) In the present study, we aimed to provide a basis for a rapid and easy method to predict carcinogenicity of chemicals based on microarray technology with cultured MH1C1 rat hepatoma cells, selected as a model system to minimize complicating factors such as cell type heterogeneity and interindividual differences of animals. For this purpose, 39 chemicals that have been well characterized for carcinogenicity were first tested for cytotoxicity in cells using reductase activity at 3 days and non-toxic doses were determined for measurement of gene expression.…”