Cancer as a Tissue Anomaly: Classifying Tumor Transcriptomes Based Only on Healthy Data.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Cancer diagnosis often relies on gene expression signatures for classification.
• Traditional classification methods struggle with the vast diversity of cancer, as they require prior examples of cancerous samples.
• Existing methods may fail to identify novel or rare cancer types not represented in training data.

Purpose of the Study:

• To develop a novel method for cancer detection using anomaly detection on transcriptomic data.
• To create a
• tissue detector
• capable of identifying cancerous tissues without prior exposure to cancer examples.
• To evaluate the efficacy of anomaly detection for classifying cancer across different tissue types.

Main Methods:

• Utilized an established surveillance method for anomaly detection, adapted for transcriptomic data.
• Trained an anomaly detection model (a "tissue detector") on normal tissue samples from the Generative Transcriptomics Explorer (GTEx) dataset.
• Applied the trained model to classify samples from The Cancer Genome Atlas (TCGA) dataset.

Main Results:

• The
• tissue detector
• achieved an area under the curve (AUC) greater than 0.90 for 3 out of 6 tested tissues when classifying TCGA samples.
• Classification accuracy improved with the inclusion of a larger number of healthy samples in the training set.
• Demonstrated the capability of anomaly detection to identify cancer without needing pre-existing cancer data.

Conclusions:

• Anomaly detection offers a conceptually advantageous approach for cancer classification, particularly for identifying diverse or novel cancer types.
• The
• tissue detector
• shows promise as a sensitive tool for cancer identification in transcriptomic data.
• Future research should explore the expansion and refinement of anomaly detection techniques for broader cancer diagnostics.