Generation of Synthetic Transcriptome Data with Defined Statistical Properties for the Development and Testing of New Analysis Methods

Abstract

We have previously developed a combined signal/variance distribution model that accounts for the particular statistical properties of datasets generated on the Applied Biosystems AB1700 transcriptome system. Here we show that this model can be efficiently used to generate synthetic datasets with statistical properties virtually identical to those of the actual data by aid of the JAVA application ace.map creator 1.0 that we have developed. The fundamentally different structure of AB1700 transcriptome profiles requires re-evaluation, adaptation, or even redevelopment of many of the standard microarray analysis methods in order to avoid misinterpretation of the data on the one hand, and to draw full benefit from their increased specificity and sensitivity on the other hand. Our composite data model and the ace.map creator 1.0 application thereby not only present proof of the correctness of our parameter estimation, but also provide a tool for the generation of synthetic test data that will be useful for further development and testing of analysis methods.

Quality Assessment of Transcriptome Data Using Intrinsic Statistical Properties

“…It is independent of cross-hybridizing monkey RNA to human arrays ( Figure 1B ). Importantly, synthetic, random pseudo-data generated using our mathematical model display the same distribution properties ( 16 ) ( Figure 1C ). When analyzing more carefully the parameter value distributions of the entire set of 750 arrays ( Figure S1 and Data File S1 ), these initial findings are confirmed.…”

“…We have already shown previously how this invariance of the data structure of AB1700 experiments can be efficiently used to draw random pseudo-data from the parameterized model we had previously presented ( 16 ) . In order to even better capture the statistical properties of individual experiments, and after having observed a certain variability in the signal and signal-variance ranges that is only poorly captured by the parameters of the two lognormal distributions, we have decided to add the natural logarithm of the signal range and the signal-variance range, as estimated on the 99% quantile in both dimensions, as independent characteristic parameters to our model.…”

“…Theoretical studies had suggested that transcriptome signal distributions are always lognormal distributed ( 15 ) , while the signal distribution of AB1700 data clearly is composed of two independent lognormal distributions as evidenced by expectation maximization and likelihood analysis ( 14 ) . Based on the characterization of the signal and signal-variance distribution of this initial set of 50 experiments, we had developed a mathematical model that can be efficiently used to describe the statistical properties of AB1700 data by estimating the model parameters on individual datasets 14 , 16 (Figure S1) .…”

“…The estimation of parameters for our signal and signal-variance model is described in detail in previous studies 14 , 16 . Briefly, the ensemble of parameters is estimated for every data file individually using a combination of techniques in the following (also compare Figure S1 ):…”

“… ( 14 ) ], it is of no importance to the fact that it can be exploited in characterizing this particular type of microarray experiments. In fact, we have shown that a model created to capture the main characteristics of this invariant property of the AB1700 data can be used to estimate a parameter set from individual experiments, which in turn can serve to generate synthetic, random pseudo-data with indistinguishable statistical properties ( 16 ) .…”