Normalization and selecting non-differentially expressed genes improve machine learning modelling of cross-platform...

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• Normalization is crucial for quantitative biological analyses.
• Cross-platform integration of RNA microarray and RNA-sequencing (RNA-seq) data for machine learning (ML) is challenging due to data heterogeneity.
• Previous studies lacked validation on independent datasets, leaving the improvement of ML performance on unseen data unclear.

Purpose of the Study:

• To test the hypothesis that non-differentially expressed genes (NDEG) can improve transcriptomic data normalization and subsequent cross-platform ML model performance.
• To evaluate the effectiveness of NDEG-based normalization for classifying breast cancer molecular subtypes using independent datasets.
• To compare normalization methods based on parametric versus nonparametric statistics for cross-platform ML.

Main Methods:

• Utilized The Cancer Genome Atlas (TCGA) breast cancer microarray and RNA-seq datasets as independent training and testing sets, respectively.
• Selected NDEG (p>0.85) and differentially expressed genes (DEG, p<0.05) using ANOVA for normalization and classification, respectively.
• Trained ML models on one platform's data and tested on the other, employing LOG_QN and LOG_QNZ normalization methods.

Main Results:

• NDEG and DEG selection significantly improved ML model classification performance for breast cancer subtypes.
• Nonparametric normalization methods outperformed parametric methods.
• The LOG_QN and LOG_QNZ normalization methods, combined with a neural network model, demonstrated superior performance.

Conclusions:

• NDEG-based normalization is a promising strategy for improving cross-platform ML model performance on independent transcriptomic datasets.
• This method enhances the classification of molecular subtypes.
• Further research is needed to validate NDEG normalization across diverse datasets and omics types.