A statistical framework for extraction of structured knowledge from biological/biotechnological systems

Abstract

Despite enormous efforts to understand complex biological/biotechnological systems, a significant amount of knowledge has still remained unraveled. However, recent advances in high throughput technologies have offered new opportunities to understand these complex systems by providing us with huge amounts of data about these systems. Unlike traditional tools, these high throughput detection tools: (1) permit large-scale screening of formulations to find the optimal condition, and (2) provide us with a global scale of measurement for a given system. Thus, there has been a strong need for computational tools that effectively extract useful knowledge about systems behavior from the vast amount of data. This thesis presents a comprehensive set of computational tools that enables us to extract important information (called structured knowledge) from this huge amount of data to improve our understanding of biological and biotechnological systems. Then, in several case studies, this extracted knowledge is used to optimize these systems. These tools include: (1) optimal design of experiments (DOE) for efficient investigation of systems, and (2) various statistical methods for effective analyses of the data to capture all structured knowledge in the data. These tools have been applied to various biological and biotechnological systems for identification of: (1) discriminatory characteristics for several diseases from gene expression data to construct disease classifiers; (2) rules to improve plasma absorptions of drugs from high-throughput screening data; (3) binding rules of epitopes to MHC molecules from binding assay data to artificially activate immune responses involving these MHC molecules; (4) rules for pre-conditioning and plasma supplementation from metabolic profiling data to improve the bio-artificial liver (BAL) device;

(cont.) (5) rules to facilitate protein crystallizations from high-throughput screening data to find the optimal condition for crystallization; (6) a new clinical index from metabolic profiling through serum data to improve the diagnostic resolution of liver failure. The results from these applications demonstrate that the developed tools successfully extracted important information to understand systems behavior from various high-throughput data and suggested rules to improve systems performance. In the first case study, the statistical methods helped us identify a drug target for Multiple Scleroses disease through analyses of gene expression data and, then, facilitated finding a peptide drug to inhibit the drug target. In the fifth case study, the methodology enabled us to find large protein crystals for several test proteins difficult to crystallize. The rules identified from the other case studies are being validated for improvement of the systems behavior.