A comparison of RNA-Seq data preprocessing pipelines for transcriptomic predictions across independent studies

Area of Science:

• Oncology
• Bioinformatics
• Computational Biology

Background:

• RNA sequencing (RNA-Seq) and machine learning (ML) enable cancer molecular classification.
• Integrating multi-lab datasets for cancer predictors is challenging due to data variability and noise.
• Data preprocessing aims to harmonize datasets for improved ML model performance.

Purpose of the Study:

• Investigate the impact of data preprocessing (normalization, batch effect correction, scaling) on cancer tissue of origin predictions.
• Improve cross-study predictions using large-scale RNA-Seq datasets.
• Evaluate preprocessing effects on ML model performance for cancer classification.

Main Methods:

• Comparative analysis of data preprocessing techniques.
• Application of normalization, batch effect correction, and data scaling.
• Training ML models on RNA-Seq data from large patient cohorts.

Main Results:

• Data preprocessing choices significantly affected classifier performance for tissue of origin prediction.
• Batch effect correction enhanced prediction accuracy on the GTEx dataset.
• Preprocessing degraded performance on aggregated ICGC and GEO datasets.

Conclusions:

• Batch effect correction can improve RNA-Seq-based cancer classification in specific scenarios.
• Data preprocessing is not universally beneficial and can hinder performance.
• Careful consideration of preprocessing steps is crucial for robust ML pipelines in cancer research.