Transcript robust fit
About genewise regulation analysis with regression Janne Nikkilä Contents of the talk • • • • • • Biological background of gene expression regulation Previous work in statistical modeling of gene regulation Our motivation Our analysis approach Some results Discussion Biological background for gene expression regulation • Gene expression is regulated at several stages: – – – – – – • DNA unpacking (demethylation, histone acetylation) Transcription Alternative RNA splicing mRNA degradation Translation initiation Protein processing and degradation Transcription is believed to be the most important one Biological background for gene transcription regulation • • • • In the transcription, mRNA corresponding the coding DNA sequence is formed Transcription initiation is mainly controlled by binding of specific protein complexes, transcription factors (tf), to gene promoter region Tfs may enhance, suppress, or do both As tfs are composed of proteins, which are coded by genes, tf activities can be analyzed by studying the expressions of the genes that code tfs Analysis methods used in the literature • Modelling of gene interactions by – – – – – Boolean networks Differential equations Linear regression Clustering Probabilistic models (e.g. Bayesian networks) Our motivation • • ➔ None of the previous methods seem to work adequately This may be due to methods, due to the quality of the data, or perhaps due to the cumulative effect of these two factors We wish to find some evidence that gene regulation mechanisms can be inferred from gene expression data A simple approach • Study one gene expression at time and try to explain it with the sum of the transcription factor component activities – • Regression as model – • Intuitive interpretation of the set up and the results Easy to interpret, computationally feasible Evaluate the results statistically – Somewhat quantitative interpretation of the results Data • Expression data – – • 300 different knockout mutations of the yeast (300 arrays) over 6000 yeast genes on each cDNA-array Binding data – – – Binding activity of 147 transcription factors to all yeast genes (147 arrays) About same genes on array as above Used to choose a set of candidate tfs for each gene Preprocessing of the data • Normal quantity provided by cDNA-arrays is the log-ratio of the sample and the control intensities from each spot – • Plain log-intensities separately? • ➔ May hinder the discovery of normal regulation mechanisms Not possible because of spotwise variation Only the arraywise and genewise averages were removed and the normal log-ratios were used in the analysis The regression model • • • • The expression of a gene, y, is modelled as a weighted sum of x, the expressions of a set of transcription factor genes The error e is assumed to be normally distributed As a result each transcription factor gene is assigned a coefficient , which denotes its role in gene regulation Fitted with robust fit-method Statistical analysis of the results • • A subset of nine genes: some confirmed transcription factors, the binding activities of the 50 tfs and significances of the same 50 tfs in regression model Tests: – – Test whether binding activity and regression model produce same kind of information about the roles of the tfs for each gene -> no statistical significance Test whether the confirmed tfs are found among the most significant ones in either binding or regression -> no stat signif Discussion • • • Clearly, there is no linear association between the regulator genes and the regulated genes in this data set The biggest problem is perhaps the type of the data: cDNA-data without time dimension -> the change of data to Affymetrix and/or timeseries data might help Another problem may be oversimplified model, but with this kind of data statistical models for gene interactions seem to be fruitless