Transformer-based feature extraction approach for hematopoietic cancer subtype classification.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Hematopoietic cancer subtype classification is challenging due to cell multipotency and lack of definitive genetic markers.
• Gene expression data holds potential for improved classification but requires effective feature extraction.

Purpose of the Study:

• To develop and evaluate a Transformer-based Autoencoder for unsupervised feature extraction from gene expression data for hematopoietic cancer subtype classification.
• To compare the proposed method against established feature extraction techniques.

Main Methods:

• A Transformer-based Autoencoder utilizing multi-head self-attention was developed to learn gene interactions.
• Transcriptomic data from 2452 The Cancer Genome Atlas samples across five hematopoietic cancer subtypes were used.
• Feature embeddings were evaluated using eight multi-class classifiers, with performance metrics including F1-score, accuracy, precision, recall, and balanced accuracy.

Main Results:

• The Transformer-based Autoencoder embeddings achieved superior performance, reaching an F1-score of 0.969 and accuracy of 0.986 when combined with Light Gradient Boosting Machine.
• The method outperformed Principal Component Analysis, Non-negative Matrix Factorization, Autoencoder, and Variational Autoencoder.
• Shapley Additive exPlanations identified key genes related to endoplasmic reticulum function, antigen processing, and RNA regulation as important biomarkers.

Conclusions:

• Transformer-based unsupervised feature extraction significantly improves the accuracy of hematopoietic cancer subtype classification.
• The approach provides biologically informative embeddings and identifies potential diagnostic biomarkers.
• This study highlights the potential of attention-driven representation learning for tabular biomedical data, particularly gene expression.