Canada Spends Logo

Explore our Budget 2026 analysis

Grant: Discovery Grants Program - Individual

University of Guelph

$12,311.00

Model-based clustering in bioinformatics

FY 2016-17

Summary
Introduction: With advances in computing, finite mixture models have become an increasingly popular basis for clustering microarray data. Model-based clustering involves the imposition of a parametric mixture distribution. In the past few years, non-Gaussian model-based clustering approaches are gaining popularity due to their non-elliptical shape and robustness. However, when dealing with high-dimensional data, these complex models tend to be highly parameterized and computationally inefficient using the traditional expectation-maximization (EM) approach for parameter estimation. Typically, analysis of microarray data using mixture models only uses gene expression values. Clinical data can sometimes contain critical information; therefore, simultaneous use of clinical and microarray data might lead to better clustering performance. Objectives: The long term goal of the applicant's research program is to develop cutting edge statistical algorithms for biological datasets to gain a comprehensive understanding of biological processes. The research over the lifetime of this proposal will focus mainly on developing efficient parameter estimation techniques for non-Gaussian model-based clustering for high-dimensional data using multivariate normal inverse Gaussian (MNIG) mixture models and extending the MNIG mixture models for longitudinal data. Secondly, cluster-weighted models (CWMs) will be developed to model simultaneously microarray and clinical data. In addition, regularization-based variable selection selection techniques will be implemented in the CWM framework. Methods: The MNIG mixture models provide flexibility of modelling skewed or symmetric components or a combination thereof, which makes them highly applicable in most scenarios. A factor analyzers structure will be incorporated in mixtures of MNIG distributions. This adds even more computational cost when the number of latent factors and the number of components are unknown, and the EM algorithm needs to be run for every possible combination of the number of components and factors. A computationally efficient parameter estimation approach using variational Bayes algorithm will be implemented to allow simultaneous estimation of the optimal number of components and latent factors for MNIG mixtures. Some work will focus on extending MNIG mixture models to incorporate a modified Cholesky-decomposed covariance parameter. This will aid in better understanding of the trend of the gene expression patterns over a period of time during some cellular process of interest. Cluster-weighted modeling utilizes the relationship between the predictor and response variables and their marginal and conditional distributions, respectively. CWMs will be extended to model a high dimensional multivariate response variable with a latent factor structure, and their dependence on the clinical variables can be introduced by using the continuous clinical variables as the predictor variables and a concomitant variable multinomial logit model for the component mixing weights to incorporate categorical clinical covariates. A regularization-based variable selection technique will be utilized to select a subset of relevant predictor variables. This will allow for simultaneous dimension reduction in the response variables and variable selection for predictor variables. Impact: This work will represent a major step forward in non-Gaussian mixture model-based clustering via a variational Bayes approach for high dimensional and complex models. The methodologies proposed here are suitable for high dimensional and longitudinal data. These algorithms will be made available publicly as R packages. The impact of this work will be felt in the computational statistics and bioinformatics communities.
Program
Discovery Grants Program - Individual
Location
Ontario, Canada
Recipient Type
Academia
Type
Grant
Source
NSERC's Awards Data

As published

Application id
RGPIN-2014-05470
Name nom
Dang, Sanjeena
Department département
Mathematics and Statistics
Organization id
24
Institution établissement
University of Guelph
Province en
Ontario
Province fr
Ontario
Country en
CANADA
Country fr
CANADA
Fiscal year exercice financier
2016
Competition year année de concours
2014
Award amount
12311
Program id
RGPIN
Program name en
Discovery Grants Program - Individual
Program name fr
Programme de subventions à la découverte - individuelles
Group en
Discovery grants programs
Group fr
Programmes de subventions à la découverte
Committee code
1508
Committee name en
Mathematics and Statistics
Committee name fr
Mathématiques et statistique
Area of application code
1101
Area of application group en
Health, education and social services
Area of application group fr
Santé publique, éducation et services sociaux
Area of application en
Human health (including medically-related psychological research)
Area of application fr
Santé publique (y compris recherche psychologique liée à la médecine)
Research subject code
3011
Research subject group en
Statistics and probability
Research subject group fr
Statistique et probabilité
Research subject en
Multivariate analysis
Research subject fr
Analyse multivariée
Partie
3
Nb partie
5
Application title
Model-based clustering in bioinformatics
Keyword
NSERC
Application summary
Introduction: With advances in computing, finite mixture models have become an increasingly popular basis for clustering microarray data. Model-based clustering involves the imposition of a parametric mixture distribution. In the past few years, non-Gaussian model-based clustering approaches are gaining popularity due to their non-elliptical shape and robustness. However, when dealing with high-dimensional data, these complex models tend to be highly parameterized and computationally inefficient using the traditional expectation-maximization (EM) approach for parameter estimation. Typically, analysis of microarray data using mixture models only uses gene expression values. Clinical data can sometimes contain critical information; therefore, simultaneous use of clinical and microarray data might lead to better clustering performance. Objectives: The long term goal of the applicant's research program is to develop cutting edge statistical algorithms for biological datasets to gain a comprehensive understanding of biological processes. The research over the lifetime of this proposal will focus mainly on developing efficient parameter estimation techniques for non-Gaussian model-based clustering for high-dimensional data using multivariate normal inverse Gaussian (MNIG) mixture models and extending the MNIG mixture models for longitudinal data. Secondly, cluster-weighted models (CWMs) will be developed to model simultaneously microarray and clinical data. In addition, regularization-based variable selection selection techniques will be implemented in the CWM framework. Methods: The MNIG mixture models provide flexibility of modelling skewed or symmetric components or a combination thereof, which makes them highly applicable in most scenarios. A factor analyzers structure will be incorporated in mixtures of MNIG distributions. This adds even more computational cost when the number of latent factors and the number of components are unknown, and the EM algorithm needs to be run for every possible combination of the number of components and factors. A computationally efficient parameter estimation approach using variational Bayes algorithm will be implemented to allow simultaneous estimation of the optimal number of components and latent factors for MNIG mixtures. Some work will focus on extending MNIG mixture models to incorporate a modified Cholesky-decomposed covariance parameter. This will aid in better understanding of the trend of the gene expression patterns over a period of time during some cellular process of interest. Cluster-weighted modeling utilizes the relationship between the predictor and response variables and their marginal and conditional distributions, respectively. CWMs will be extended to model a high dimensional multivariate response variable with a latent factor structure, and their dependence on the clinical variables can be introduced by using the continuous clinical variables as the predictor variables and a concomitant variable multinomial logit model for the component mixing weights to incorporate categorical clinical covariates. A regularization-based variable selection technique will be utilized to select a subset of relevant predictor variables. This will allow for simultaneous dimension reduction in the response variables and variable selection for predictor variables. Impact: This work will represent a major step forward in non-Gaussian mixture model-based clustering via a variational Bayes approach for high dimensional and complex models. The methodologies proposed here are suitable for high dimensional and longitudinal data. These algorithms will be made available publicly as R packages. The impact of this work will be felt in the computational statistics and bioinformatics communities.