Related Experiment Videos
Complementary structure of statistical significance and predictive relevance in explainable machine learning-based
Dogyeong Lee1, Junyoung Lee2, Inchul Choi3
1Department of Bio-Big Data, Chungnam National University, Daejeon, Republic of Korea.
Frontiers in Genetics
|July 20, 2026
Summary
This study introduces an explainable machine learning framework to classify cattle tissues using RNA sequencing data. The model identifies key genes, revealing a hierarchical transcriptional structure crucial for livestock genomics and breeding.
Area of Science:
- Livestock genomics
- Transcriptomics
- Machine learning in biology
Background:
- Tissue-specific transcriptomic structures are key to understanding economically important traits in livestock.
- Conventional differential gene expression analysis has limitations in capturing complex gene interactions and tissue identity.
- Understanding these structures is vital for livestock breeding programs.
Purpose of the Study:
- To develop an explainable machine learning framework for classifying seven Hanwoo cattle tissues using RNA sequencing data.
- To compare the contributions of statistical and model-derived signals in tissue classification.
- To provide insights into the transcriptional basis of tissue identity in livestock.
Main Methods:
- Trained a Random Forest-based one-versus-rest classification model on 130 Hanwoo transcriptomes.
- Externally validated the model using 231 independent Bos taurus samples from public datasets with batch correction.
- Employed model-based feature attribution to interpret gene contributions and compared them with differential expression analysis.
Main Results:
- Achieved high classification performance (mean accuracy 0.907, macro-average AUC 0.963) with stable generalization.
- Found that differentially expressed genes form the primary discriminatory structure, while model-prioritized genes enhance classification, especially for related tissues.
- Interpreted gene contributions, revealing consistency with known tissue functions and non-linear, context-dependent patterns.
Conclusions:
- Tissue identity is characterized by a hierarchical transcriptional structure with dominant differential signals and refined multivariate interactions.
- The developed framework effectively distinguishes statistical significance from predictive relevance in transcriptomic data.
- The findings have practical implications for developing molecular markers in livestock genomics and breeding.