Phenotype driven data augmentation methods for transcriptomic data.

Area of Science:

• Biomedical data science
• Computational biology
• Machine learning in genomics

Background:

• High-dimensional transcriptomic data presents challenges for supervised learning, including overfitting and poor generalization.
• Existing data augmentation methods for transcriptomic data are often computationally intensive or produce limited sample diversity.
• Class imbalance and low sample sizes are common issues in transcriptomic datasets, hindering model performance.

Purpose of the Study:

• To introduce novel phenotype-driven data augmentation methods for transcriptomic data.
• To address the challenges of high dimensionality, overfitting, and limited generalization in supervised learning tasks.
• To improve patient stratification accuracy in cancer transcriptomic studies.

Main Methods:

• Developed two classes of phenotype-driven data augmentation: signature-dependent and signature-independent methods.
• Signature-dependent methods utilize gene signatures for non-parametric data augmentation.
• Signature-independent methods adapt established Gamma-Poisson and Poisson sampling techniques for gene expression data.

Main Results:

• Applied augmentation methods to colorectal and breast cancer transcriptomic data.
• Demonstrated improved patient stratification by 5-15% compared to existing augmentation methods.
• Showcased enhanced model generalization and reduced overfitting through discriminative and generative experiments with external validation.

Conclusions:

• Phenotype-driven data augmentation effectively enhances supervised learning on transcriptomic data.
• The proposed methods offer a computationally efficient and diverse approach to data augmentation.
• Over-augmentation can yield limited benefits, highlighting the importance of strategic application.