Investigating the effects of imputation methods for modelling gene networks using a dynamic bayesian network from...

Area of Science:

• Bioinformatics
• Computational Biology
• Systems Biology

Background:

• Missing values in gene expression data are common.
• Imputation methods like k-nearest neighbour (kNN), local least squares (LLS), and Bayesian principal component analysis (BPCA) are used to address missing data.
• The impact of these imputation methods on gene regulatory network (GRN) modeling using dynamic Bayesian networks (DBNs) is not well understood.

Purpose of the Study:

• To investigate and compare the effects of kNN, LLS, and BPCA imputation methods on GRN modeling.
• To analyze how dataset size influences the performance of different imputation methods in GRN construction.
• To evaluate the efficacy of dynamic Bayesian networks (DBNs) in modeling GRNs after data imputation.

Main Methods:

• Gene expression datasets from Escherichia coli (S.O.S. DNA repair) and Saccharomyces cerevisiae (cell cycle) were used.
• Datasets were imputed using kNN, LLS, and BPCA methods separately.
• Discrete dynamic Bayesian networks (DBNs) were employed to generate GRNs from imputed datasets, with comparisons based on minimizing error.

Main Results:

• BPCA and LLS imputation methods showed superior performance for larger datasets (S. cerevisiae).
• kNN imputation method performed better on smaller datasets (E. coli).
• The effectiveness of imputation methods varied depending on the dataset size.

Conclusions:

• Imputation method performance in GRN modeling is dataset-size dependent.
• Dynamic Bayesian Networks (DBNs) are capable of identifying potential gene interactions and displaying gene regulatory relationships.