A robust transfer learning approach for high-dimensional linear regression to support integration of multi-source...

Area of Science:

• Genomics
• Statistical Learning
• Bioinformatics

Background:

• Transfer learning integrates multi-source data to enhance target data learning.
• Genomics data, particularly gene associations in specific tissues, benefits from integrating data from other tissues.
• Genomics data frequently exhibits heavy-tail distributions and outliers, challenging existing transfer learning methods.

Purpose of the Study:

• To develop a robust transfer learning method for high-dimensional linear models with t-distributed errors (Trans-PtLR).
• To improve estimation and prediction accuracy in genomics by leveraging useful source data while accommodating heavy tails and outliers.
• To introduce a method for selecting informative source datasets to avoid non-informative data integration.

Main Methods:

• Developed a transfer learning algorithm using penalized maximum likelihood and an expectation-maximization algorithm for the oracle case (known transferable sources).
• Proposed a cross-validation strategy for selecting transferable source datasets, enhancing robustness.
• Utilized t-distributed error models to address heavy-tail distributions and outliers in genomics data.

Main Results:

• Trans-PtLR demonstrated superior robustness and performance in estimation and prediction compared to standard linear regression transfer learning.
• The method effectively handles genomics data with heavy-tail distributions and outliers.
• Cross-validation successfully identified informative source datasets, improving overall model performance.

Conclusions:

• Trans-PtLR offers a robust and effective approach for transfer learning in genomics, particularly when dealing with complex data characteristics.
• The proposed method enhances data integration and variable selection capabilities in high-dimensional genomic studies.
• The findings highlight the importance of robust statistical modeling for accurate genomic data analysis and transfer learning.