Can We Hack Cancer Systems by Supercomputer?
Satoru Miyano, Professor, Human Genome Center, Institute of Medical Science, University of Tokyo, Japan
We present computational methods which boost the challenge for hacking cancer systems by the power of supercomputer. The first is digitalizing personal cancer systems as gene networks. It is impossible to build an individual gene network from one patient sample. However, given many patient samples, it turns possible. We developed a computational method to exhibit how gene networks vary from patient to patient according to, for example, the gene expression levels of some specific genes. The method uses structural equation model for discovering associations between the differences in molecular mechanisms and the diversity of phenotype traits. We applied this method to microarray gene expression data from 762 cancer cell lines and unraveled global differences of networks with 20000 genes of different EMT (epithelial-mesenchymal transition) expression levels. The computation took 90 days using 1024 cores on the supercomputer system at Human Genome Center (6000 cores; 75 TFLOPS at peak). Out of 1732 possible regulators of E-cadherin, a cell adhesion molecule that modulates the EMT, we identified 25 candidate regulators, of which about half have been reported in the literature. Within these 25 candidates, we also found a novel gene which is validated by siRNA knockdown experiment. We also present computational methods and results using the supercomputer for modeling dynamics of cancer systems from time-course gene expression data. We developed a series of computational methods based on Bayesian network with nonparametric regression which can mine gene networks of size from 30 (optimal) to genome-wide 20000 (locally optimal). The dynamic Bayesian network method unveiled how the gene network focused on 2500 genes in melanoma cell changes in time-course after dosing the anti-cancer drug Paclitaxel, and identified Paclitaxel resistant genes. This method was also applied to lung cancer cell and the anti-cancer drug Gefitinib. Our thorough analyses took a week on the supercomputer. For dynamic system modeling, we devised a state space model (SSM) with dimension reduction method for reverse-engineering gene networks from time-course data, with which we can view their dynamic changes over time by simulation. We succeeded in computing a gene network with prediction ability focused on 1500 genes from data of about 20 time-points. We applied this SSM model to human normal lung cell treated with (case)/without (control) Gefitinib, and we identified genes under differential regulations between case and control. This signature of genes was used to predict prognosis for lung cancer patients and showed a good performance for survival prediction. We also developed a software NetComparator for inferring dynamic gene networks under varying conditions based on regularized weighted recursive elastic net. NetComparator uncovered subnetworks showing the difference between Gefitinib sensitive lung cancer and resistant lung cancer.