MOTIVATION: Multilayer perceptrons (MLP) represent one of the widely used and effective machine learning methods currently applied to diagnostic classification based on high-dimensional genomic data. Since the dimensionalities of the existing genomic data often exceed the available sample sizes by orders of magnitude, the MLP performance may degrade owing to the curse of dimensionality and over-fitting, and may not provide acceptable prediction accuracy. RESULTS: Based on Fisher linear discriminant analysis, we designed and implemented an MLP optimization scheme for a two-layer MLP that effectively optimizes the initialization of MLP parameters and MLP architecture. The optimized MLP consistently demonstrated its ability in easing the curse of dimensionality in large microarray datasets. In comparison with a conventional MLP using random initialization, we obtained significant improvements in major performance measures including Bayes classification accuracy, convergence properties and area under the receiver operating characteristic curve (A(z)). SUPPLEMENTARY INFORMATION: The Supplementary information is available on http://www.cbil.ece.vt.edu/publications.htm

MOTIVATION: The most commonly utilized microarrays for mRNA profiling (Affymetrix) include 'probe sets' of a series of perfect match and mismatch probes (typically 22 oligonucleotides per probe set). There are an increasing number of reported 'probe set algorithms' that differ in their interpretation of a probe set to derive a single normalized 'signal' representative of expression of each mRNA. These algorithms are known to differ in accuracy and sensitivity, and optimization has been done using a small set of standardized control microarray data. We hypothesized that different mRNA profiling projects have varying sources and degrees of confounding noise, and that these should alter the choice of a specific probe set algorithm. Also, we hypothesized that use of the Microarray Suite (MAS) 5.0 probe set detection p-value as a weighting function would improve the performance of all probe set algorithms. RESULTS: We built an interactive visual analysis software tool (HCE2W) to test and define parameters in Affymetrix analyses that optimize the ratio of signal (desired biological variable) versus noise (confounding uncontrolled variables). Five probe set algorithms were studied with and without statistical weighting of probe sets using the MAS 5.0 probe set detection p-values. The signal-to-noise ratio optimization method was tested in two large novel microarray datasets with different levels of confounding noise, a 105 sample U133A human muscle biopsy dataset (11 groups: mutation-defined, extensive noise), and a 40 sample U74A inbred mouse lung dataset (8 groups: little noise). Performance was measured by the ability of the specific probe set algorithm, with and without detection p-value weighting, to cluster samples into the appropriate biological groups (unsupervised agglomerative clustering with F-measure values). Of the total random sampling analyses, 50% showed a highly statistically significant difference between probe set algorithms by ANOVA [F(4,10) > 14, p < 0.0001], with weighting by MAS 5.0 detection p-value showing significance in the mouse data by ANOVA [F(1,10) > 9, p < 0.013] and paired t-test [t(9) = -3.675, p = 0.005]. Probe set detection p-value weighting had the greatest positive effect on performance of dChip difference model, ProbeProfiler and RMA algorithms. Importantly, probe set algorithms did indeed perform differently depending on the specific project, most probably due to the degree of confounding noise. Our data indicate that significantly improved data analysis of mRNA profile projects can be achieved by optimizing the choice of probe set algorithm with the noise levels intrinsic to a project, with dChip difference model with MAS 5.0 detection p-value continuous weighting showing the best overall performance in both projects. Furthermore, both existing and newly developed probe set algorithms should incorporate a detection p-value weighting to improve performance. AVAILABILITY: The Hierarchical Clustering Explorer 2.0 is available at http://www.cs.umd.edu/hcil/hce/ Murine arrays (40 samples) are publicly available at the PEPR resource (http://microarray.cnmcresearch.org/pgadatatable.asp http://pepr.cnmcresearch.org Chen et al., 2004).

Data analysis and visualization is strongly influenced by noise and noise filters. There are multiple sources of "noise" in microarray data analysis, but signal/noise ratios are rarely optimized, or even considered. Here, we report a noise analysis of a novel 13 million oligonucleotide dataset - 25 human U133A (/spl sim/500,000 features) profiles of patient muscle biopsies. We use our recently described interactive visualization tool, the hierarchical clustering explorer (HCE) to systemically address the effect of different noise filters on resolution of arrays into "correct" biological groups (unsupervised clustering into three patient groups of known diagnosis). We varied probe set interpretation methods (MAS 5.0, RMA), "present call" filters, and clustering linkage methods, and investigated the results in HCE. HCE's interactive features enabled us to quickly see the impact of these three variables. Dendrogram displays showed the clustering results systematically, and color mosaic displays provided a visual support for the results. We show that each of these three variables has a strong effect on unsupervised clustering. For this dataset, the strength of the biological variable was maximized, and noise minimized, using MAS 5.0, 10% present call filter, and average group linkage. We propose a general method of using interactive tools to identify the optimal signal/noise balance or the optimal combination of these three variables to maximize the effect of the desired biological variable on data interpretation.

BACKGROUND: We provide a systematic study of the sources of variability in expression profiling data using 56 RNAs isolated from human muscle biopsies (34 Affymetrix MuscleChip arrays), and 36 murine cell culture and tissue RNAs (42 Affymetrix U74Av2 arrays). RESULTS: We studied muscle biopsies from 28 human subjects as well as murine myogenic cell cultures, muscle, and spleens. Human MuscleChip arrays (4,601 probe sets) and murine U74Av2 Affymetrix microarrays were used for expression profiling. RNAs were profiled both singly, and as mixed groups. Variables studied included tissue heterogeneity, cRNA probe production, patient diagnosis, and GeneChip hybridizations. We found that the greatest source of variability was often different regions of the same patient muscle biopsy, reflecting variation in cell type content even in a relatively homogeneous tissue such as muscle. Inter-patient variation was also very high (SNP noise). Experimental variation (RNA, cDNA, cRNA, or GeneChip) was minor. Pre-profile mixing of patient cRNA samples effectively normalized both intra- and inter-patient sources of variation, while retaining a high degree of specificity of the individual profiles (86% of statistically significant differences detected by absolute analysis; and 85% by a 4-pairwise comparison survival method). CONCLUSIONS: Using unsupervised cluster analysis and correlation coefficients of 92 RNA samples on 76 oligonucleotide microarrays, we found that experimental error was not a significant source of unwanted variability in expression profiling experiments. Major sources of variability were from use of small tissue biopsies, particularly in humans where there is substantial inter-patient variability (SNP noise).