Knowledge-based computational models

In our recent paper [1], we describe a biology-driven approach to translate cancer genomic data into clinically actionable information for personalized patient therapy selection. Despite the increasing knowledge on intracellular signal transduction pathways as drivers of tumor growth, and the corresponding development of a large number of " targeted drugs " to correct aberrant pathway behavior in cancer, it appears to be a tremendous challenge to select treatment protocols with a sustainable outcome for many patients. As the number of treatment options and their combinations increases, it becomes even more necessary to develop approaches for such " precision diagnostics ". Although extensive amounts of genomic microarray and sequencing data have been generated, which hold the key for optimal treatment selection for an individual patient, it appears to be difficult with presently available approaches to translate complex genomic data into clinically meaningful results. Many genomic studies focus on identifying DNA mutations associated with therapy response and prognosis, however many results have failed to be clinically actionable [2]. First, most mutations found in tumors are passenger mutations, while only a few are driving tumor growth. Second, of only a limited number of tumor-driving mutations the incidence is sufficiently high to allow clinical validation, while most mutations are highly patient specific. Furthermore, the cancer genotype is increasingly recognized as providing only part of the puzzle. Other aspects such as the tumor micro-environment are thought to be equally important in determining functional behavior of cancer cells. As result, combined interpretation of cancer genotype and molecular phenotype is required to fully characterize an individual tumor and enable reliable prediction of therapy response [2]. In addition to the above, most efforts to identify diagnostic or predictive biomarkers are data driven, and fail to exploit the rapidly expanding biological cancer knowledge. We argue that this may be a costly omission, since using genomics data within a cancer biology knowledge framework enables reduction of data noise and focused extraction of relevant information for tumor characterization. Furthermore, the huge number of available data features versus the number of patient samples, combined with the fact that such sample sets are generally very heterogeneous, makes a data-driven approach prone to finding spurious patterns. Limiting biomarker search by using biological knowledge leads to more robust findings, which can be translated more straightforwardly into clinical practice as diagnostic assays. In our paper [1], we address the primary question to be answered when predicting …