Deep neural network based tissue deconvolution of circulating tumor cell RNA.

Area of Science:

• Bioinformatics
• Genomics
• Computational Biology

Background:

• Cell-free RNA deconvolution identifies tissue origin but conventional methods struggle with data variability.
• Existing gene panel approaches lack adaptability to real-world biological data.

Purpose of the Study:

• To develop a novel neural network-based method for cell-free RNA deconvolution that overcomes limitations of conventional approaches.
• To enhance the accuracy of tissue origin identification by effectively capturing inherent data variability.
• To explore the clinical utility of this deep-learning method in tracing metastatic cancer cell migration.

Main Methods:

• Developed and trained a neural network model incorporating 15 distinct tissue types.
• Validated the model using semi in silico datasets, custom normal tissue mixture RNA-seq data, and longitudinal circulating tumor cell RNA-seq (ctcRNA) data.
• Performed sensitivity analyses to assess model robustness against missing data.

Main Results:

• The deep-learning approach demonstrated enhanced accuracy in tissue origin deconvolution by capturing dataset variability.
• Neural network models showed increased resilience to missing data compared to conventional methods.
• Successfully traced circulating tumor cell-derived RNA (ctcRNA) migration in a metastatic cancer patient, demonstrating organotropism.

Conclusions:

• The novel deep-learning framework offers a more accurate and robust method for cell-free RNA deconvolution.
• This approach has significant potential for the early detection of cancer metastasis through RNA migration tracing.
• The method's ability to handle data variability and missing data makes it suitable for clinical applications.