Associate Professor Lin Zhang

This group is working on several projects on Bayesian analysis of high-dimensional genomics and image data.

• Develop novel Bayesian methods for set-based genome-wide association studies (GWAS). Classical GWAS tests the association between a single genetic variant and a single phenotypic trait, which is usually moderate in effect size and thus hard to detect with statistical significancy. Joint tests of association between multiple variants and phenotype is attractive in that they are able to borrow strength and thus boost powers. The researchers are developing novel Bayesian methods for set-based GWAS, which, compared to frequentist methods, have the capability of incorporating complex hierarchical structure and correlations in the multiple variants in a test. Future work will extend for rare variant tests and multi-variants and multi-trait association tests.
• Develop novel static and dynamic graphical modeling methods for functional connectivity (FC) analysis of neuroimaging data. FC analysis is important for shedding information on the coordinated activities among multiple brain regions at rest or performing a task. Graphical modeling methods, which infer sparse precision matrices of multiple random variables and corresponding network structures, have been used as important tools for joint inference of FC networks among multiple brain regions based on high-throughput neuroimaging data from event-related potentials (ERP) and functional magnetic resonance imaging (fMRI). These researchers are developing several novel methods for bi-level static and dynamic graphical modeling for multi-subject neuroimaging data analysis, which simultaneously learns the static/dynamic graphical models at both the subject and group level for a group of subjects having common clinical characteristics. Such bi-level graphical modeling has the advantage of pooling information to infer FC features characteristic to the group and borrowing strength from other subjects in subject-level FC network construction.
• Develop novel Bayesian cancer classification methods for imaging data from MRI. The researchers have developed several Bayesian cancer classification methods for voxel-wise cancer classification based on multi-parametric MRI data of prostate cancer patients. These methods account for the regional heterogeneity and spatial correlations present in the imaging data. The researchers are currently developing novel Bayesian methods for detecting lesions using MRI imaging data by modeling nonstationary spatial process.

Since these projects involve analysis and high-dimensional genomics and imaging data and thus extensive computations, MSI computing resources are essential to ensure that the group can conduct the simulations and data analyses in an efficient way. They are developing software (MATLAB or R functions) that implements their methods and that can be parallelized for computing on multiprocessor systems. Since the proposed Bayesian methods involve Markov chain Monte Carlo sampling algorithm which collects posterior samples of parameters in the models, this group is also using MSI's high-capacity storage capabilities.