Robust evaluation of deep learning-based representation methods for survival and gene essentiality prediction on bulk...

Area of Science:

• Computational biology
• Genomics
• Cancer research

Background:

• Deep learning (DL) shows promise for analyzing bulk RNA-sequencing (RNA-seq) data in cancer.
• Lack of consensus exists on how DL design choices (architecture, training, hyperparameters) affect learned representations.

Purpose of the Study:

• To evaluate DL representation learning methods for cancer RNA-seq data.
• To assess the impact of various DL design choices on performance.
• To provide guidelines for robust DL evaluation in cancer research.

Main Methods:

• Utilized TCGA and DepMap pan-cancer datasets.
• Evaluated diverse DL representation learning methods.
• Assessed predictive power for survival and gene essentiality.
• Investigated auto-encoder (AE) improvements with masking and multi-head training.

Main Results:

• Baseline DL methods performed comparably or better than complex models for survival prediction.
• DL representation methods excelled in predicting gene essentiality in cell lines.
• Auto-encoders benefited significantly from masking and multi-head training techniques.
• DL representation and pretraining impact is task- and architecture-dependent.

Conclusions:

• Rigorous evaluation guidelines are essential for DL in cancer research.
• The choice of DL model design significantly influences predictive performance.
• A standardized pipeline for robust DL evaluation is provided to the research community.