Dhruv Iyer
Dhruv Iyer
Dhruv Iyer
Helios Scholar
School: Hamilton High School
Hometown: Chandler, Arizona
Mentor: Suengchan Kim, Ph.D.
Email This Article Print This Page

Identifying pathways enriched with differential dependencies using LASSO

Recent advances in high-throughput transcriptomic profiling have given rise to several statistical methods for modeling and analyzing Gene Regulatory Networks (GRNs).  More recently, context-specificity of such GRNs, mainly their topology, has garnered significant interest.  Evaluation of Differential Dependency (EDDY) is one approach to analyzing these Differential Dependency Networks (DDNs), using Bayesian GRN inference and network likelihood distributions to identify pathways enriched with differential dependency relationships. Bayesian inference, however, requires that expression data be quantized prior to analysis, leading to a possible loss of information in the continuous-to-discrete transformation of the data. We propose an alternate approach for identifying differential dependency relationships based on the L1-penalized LASSO regression for GRN reconstruction. The LASSO is a shrinkage and selection operator that minimizes the sum of squared residuals with the added benefit of feature selection. This feature selection can be used for neighborhood selection and local dependency network construction for each gene in a gene set. By using LASSO, we eliminate the dependency on quantized expression data seen in Bayesian EDDY.  In addition, we also propose a new method to test the differential dependencies of GRNs between conditions, more customized for LASSO GRN models. Specifically, we use two different but complimentary approaches. For a statistical metric, we calculate Coefficients of Determination and compare the fit of each GRN to both its own data and that of the other class. For a topological metric, we compute the Hamming Distance and compare the network topologies of the two GRNs. Statistical significance of uncovered differential dependencies is assessed asymptotically by approximating a null distribution via a beta distribution for which model parameters are estimated from an initial permutation test. Using expression data from the Cancer Cell Line Encyclopedia (CCLE) and drug sensitivity data from the Cancer Therapeutics Response Portal (CTRP), we found that the LASSO implementation of EDDY identified 7 of 217 pathways from the Biocarta Database enriched with differential dependencies. In further inspection of the DDNs, we noted that HLA-DRB1 seemed to play a significant role in mediating different responses of cell lines to the CIL55 compound.  While further testing with more systematic analysis is required to assess the accuracy of LASSO EDDY, we have successfully removed the quantization limitation of Bayesian EDDY.