Statistical modeling strategies for linking multiple molecular sources to time-to-event endpoints
Year: 2014
Type: dissertation
Abstract: Modern microarray and sequencing technologies allow to measure a large number of patients' characteristics at various molecular levels simultaneously, typically ten thousands to more than one million, e.g. single nucleotide polymorphism (SNP), methylation and gene expression. In this thesis, a sequential complementary strategy as a tool for integrative analysis of two different types of genome-wide data sets is developed. The work is motivated by a study on acute myeloid leukemia (AML) patients as current AML molecular-based classification built on single genomic data is not robust enough to predict the prognosis of AML patients. Therefore, multiple sources of molecular measurements and clinical information have to be integrated into a risk prediction model with respect to clinical endpoints such as time-to-event. The sequential complementary strategy is a stepwise procedure by investigating previous knowledge obtained from one of the molecular data sources to then stabilize the statistical model for the next molecular source. An integrative analysis will be complicated by having only a small overlap in the biological samples, i.e. measurements from all levels are available for a small proportion of patients. To deal with such a situation and to allow the use of all available data including the non-overlapping cases, an imputation approach is developed within the sequential complementary strategy. The approach is based on the imputation of the linear predictor of the model using data from one molecular source alone. To impute the linear predictor values, componentwise likelihood-based boosting and aggregated random forest are evaluated as prediction techniques. The sequential complementary strategy is illustrated in a real application to survival data from AML patients considering methylation and gene expression with almost complete overlap in the biological samples. Unfortunately, such an almost complete overlap in the biological samples does not always have to be present. To provide guidance on how to deal with different sizes of overlap in the biological samples, these public available data sets are used to perform a systematic analysis of the effect of the overlap size on the sequential complementary strategy. Stability, quantified by resampling techniques, and prediction performance are considered to judge model quality. One can see a considerable effect of the overlap size on molecular characteristics selected by the sequential complementary strategy for a risk prediction signature and as well on the prediction performance of the latter. Investigating the systematic analysis of the effect of the overlap size on the sequential complementary strategy, imputation by prediction of the linear predictor values is seen to be feasible. This is, the sequential complementary strategy is found to affect the selected molecular characteristics and to improve prediction performance across all varying overlap data. In addition, the sequential complementary strategy identifies molecular entities that could not be seen in the analysis of a single molecular data source alone while applying it to SNP microarray and microarray-based gene expression measurements with relative small overlap in the biological samples. In summary, the sequential complementary strategy provides reasonable results by including all available biological samples, not only the overlap cases, and is an effective tool for integrative data analysis of molecular measurements from two different types of data sources to identify additional molecular signatures beyond a single source analysis even for very small overlap sizes.
Author: Stefanie Hieke
Language: English
Subfield: Molecular Biology
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W2562074372