A pairwise strategy for imputing predictive features when combining multiple datasets.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Combining multiple genomic studies enhances predictive model generalizability.
• Variations in measurement platforms lead to different feature sets across studies.
• Using only common features discards potentially valuable data.

Purpose of the Study:

• Quantify performance loss from using only intersected features.
• Evaluate imputation methods for missing genomic data.
• Propose and validate a pairwise imputation strategy for cross-study analysis.

Main Methods:

• Characterized performance loss using linear and polynomial regression for imputation.
• Simulated data and used breast cancer gene expression datasets.
• Developed and tested a pairwise imputation strategy, averaging imputed features across pairs.

Main Results:

• Pairwise imputation significantly improves external predictive model performance compared to using only intersected features.
• The pairwise strategy outperforms merging all datasets before imputation.
• Identified optimal feature subsets for imputation to enhance cross-study replicability.

Conclusions:

• Discarding study-specific genomic features leads to substantial predictive performance loss.
• Pairwise imputation is a superior strategy for integrating heterogeneous genomic datasets.
• This approach maximizes information utilization for robust cross-study genomic prediction.