Missing value imputation for epistatic MAPs.

Area of Science:

• Genetics and Genomics
• Systems Biology
• Bioinformatics

Background:

• Epistatic mini-array profiling (E-MAPs) generates symmetric pairwise matrices of genetic interaction scores.
• E-MAP datasets often contain a substantial percentage (up to 35%) of missing values, hindering downstream analyses.
• Existing imputation methods for microarray data may not be directly applicable to E-MAPs due to their pairwise nature and high missingness.

Purpose of the Study:

• To evaluate the effectiveness of imputation strategies for missing values in E-MAP datasets.
• To compare local (nearest neighbor-based) and global (PCA-based) imputation methods.
• To determine if imputation can increase the identification of novel functional gene pair interactions.

Main Methods:

• Categorization of missing data based on underlying causes.
• Evaluation of four imputation strategies: three local (nearest neighbor) and one global (PCA-based).
• Modification of imputation methods to accommodate symmetric pairwise data structures.

Main Results:

• Missing values from the largest category were effectively imputed.
• Both local and global imputation approaches were competitive and superior to zero-filling.
• Imputation methods demonstrated effectiveness across different species' E-MAP datasets.
• Nearest neighbor imputation successfully enriched for gene annotations, similar to measured interactions.
• Alleviating interactions were more challenging to predict than aggravating interactions.

Conclusions:

• Symmetric nearest neighbor-based imputation offers accurate and tractable solutions for missing values in E-MAPs.
• Imputation techniques enhance the potential for discovering novel gene pair interactions.
• Developed algorithms are made available to the research community.