Thus, if the neural network falsely classifies a gene as a positive, this is not necessarily a failure of the learning scheme,īut often represents a gene participating in biological processes closely associated with the original class to be learned. This of course reflects the fact that cellular processes do not represent isolated, modules.

The MIPS catalog, which has been compiled based on the extensive biological knowledge in the literature, They are (1) class size, (2) heterogeneity of the classes, and (3) the high degree of intersectionĪmong functional classes. In addition to this, we identified three reasonsįor the poor learning performance that are purely related to the biology underlying the data rather than to the technicalĪspects of machine learning. Ratio, and poor learning can be a result of the noise eclipsing the signal. First, the output of DNA array technology can have a poor signal to noise Of reasons that can affect learning performance. A priori, oneĬould suspect that this is caused by a poor performance of the neural network learning method. Note that a typical expression data set usually contains several thousands of genes but perhaps 60% false negatives for 92% of the functional classes. 2000) have been applied to the molecular classification of different cancer tissues. 2001), and pattern discovery methods ( Califano et al. Thus, support vector machines ( Furey et al. Supervised methods can be used if there is some prior knowledge about the classes to be analysed. The molecular classification of leukemia ( Golub et al. For instance, aggregative hierarchical clustering has been used extensively for Has been investigated in several studies. Of genes with correlated expression patterns, to explore shared functions or regulation. Phenotypes, etc.) in terms of expression values, regarded as their molecular signatures, or (2) conversely, the classification The data can be clustered along these two dimensions for two purposes: either (1) the classification of conditions (tissues, 1997).ĭNA array experiments primarily involve the measurement of thousands of gene expression levels under different conditions. 2000), whereas in higher organisms, like humans, tens of thousands genes can be monitored ( Zhang et al. 1998) and Caenorhabditis elegans( Hill et al. 2000) or in simple eukaryotes like yeast ( Eisen et al. It is now possible to study genomic patterns of gene expression in prokaryotes ( Arfin et al. 1996) allows for the simultaneous recording of thousands of gene expression levels and has opened new ways of looking at organisms We exemplify this methodology using the well-studied tricarboxylic acid cycle.ĭNA array technology ( Schena et al. Related to the original class, allowing for a coarse reconstruction of the interactions between associated biological pathways. We show that in a few iterations this process converges toĪ gene set that is learnable with considerably low rates of false positives and negatives and contains genes that are biologically Procedure for combining false positives with the original class.

Our analysis indicates thatĬlassification systems with a lower Borges effect are better suitable for machine learning. We term this the “Borges effect” and introduce two new numerical indices for its quantification. High degree of interconnections among functional classes confounds the signatures that ought to be learned for a unique class. Positives (and negatives) in a machine-learning context are not necessarily “false” in a biological sense. We found that only ∼10% of these are learnable (based on the rate of false negatives). Report systematic results for ∼100 classes in the Munich Information Center for Protein Sequences (MIPS) functional catalog. The comprehensive application of supervised neural networks (SNNs) for functional annotation. Methods have been applied to a limited number of functional classes for this purpose. Previously, support vector machines (SVMs) and other machine-learning

Using prior knowledge, these algorithms can assign functional annotations based on more complexĮxpression signatures found in existing functional classes. This has been demonstrated by unsupervised clustering of co-expressed genes and, more importantly, by Recent advances in microarray technology have opened new ways for functional annotation of previously uncharacterised genes

Mps itslearning