Verifying explainability of a deep learning tissue classifier trained on RNA-seq data.

Area of Science:

• Computational Biology
• Genomics
• Machine Learning

Background:

• Machine learning (ML) explainability is crucial for decision-making, but its reliability in high-dimensional data is uncertain.
• SHapley Additive exPlanations (SHAP) is a model to provide transparency in ML algorithms.

Purpose of the Study:

• To test the reliability of SHAP for explaining a deep learning model applied to transcriptome data.
• To assess the utility of SHAP in identifying biologically relevant features for tissue classification.

Main Methods:

• Developed a convolutional neural network classifier for tissue classification using Genotype-Tissue Expression (GTEx) RNA-seq data (16,651 samples, 47 tissues).
• Applied SHAP to identify discriminatory genes driving the classifier's predictions.
• Compared SHAP-identified genes with those from differential expression analysis and random gene sets.

Main Results:

• The classifier achieved a high average F1 score of 96.1% on held-out data.
• SHAP identified 2423 discriminatory genes, with 98.6% overlap with genes found via differential expression analysis.
• SHAP-identified genes reflected known biological processes and clustered tissues more effectively than other gene sets.

Conclusions:

• SHAP is a reliable and useful tool for explaining deep learning models in the context of high-dimensional transcriptome data.
• Machine learning, particularly with explainability methods like SHAP, offers powerful insights into biological data and gene function.