Unsupervised and supervised learning with neural network for human transcriptome analysis and cancer diagnosis

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Transcriptomic data analysis is crucial for understanding biological and pathological changes.
• The structural complexity of transcriptomic data presents unique challenges compared to image and text data.
• Current methods may struggle with the nuanced analysis of large-scale transcriptomic datasets.

Purpose of the Study:

• To develop a deep learning approach for analyzing complex human transcriptomic data.
• To create informative feature representations from transcriptomic data using an autoencoder.
• To evaluate the efficacy of these features in distinguishing normal and tumor tissues.

Main Methods:

• Unsupervised training of a Deep-Autoencoder (DeepT2Vec) on over 20,000 human normal and tumor transcriptomic datasets.
• Extraction of 30-dimensional Transcriptomic Feature Vectors (TFVs) from the transcriptomic data.
• Development and training of a supervised classifier (DeepC) using the extracted TFVs.

Main Results:

• DeepT2Vec successfully extracted informative features and embedded transcriptomes into 30-dimensional TFVs.
• TFVs accurately recapitulated expression patterns and could track tissue origins.
• DeepC achieved 90% accuracy for Pan-Cancer and 94% average accuracy for specific cancers in distinguishing normal from tumor samples.
• Further accuracy improvements to 96% for Pan-Cancer were observed when training on a connected network.

Conclusions:

• Deep learning, particularly using autoencoders, is a suitable method for transcriptomic data analysis.
• DeepT2Vec provides effective transcriptomic feature extraction for biological analysis.
• The developed approach demonstrates significant potential for classifying cancers and normal tissues, even with limited samples.