A machine learning one-class logistic regression model to predict stemness for single cell transcriptomics and...

Area of Science:

• Oncology
• Bioinformatics
• Genomics

Background:

• Cell annotation is vital for interpreting single cell and spatial omics data but is often biased and unproven in spatial omics.
• Existing methods for cell annotation in omics data can be labor-intensive and may not capture all relevant cell states.

Purpose of the Study:

• To apply a stemness model using machine learning for assessing oncogenic states in single cell and spatial omics cancer datasets.
• To identify dedifferentiated cell states in tumors without the need for manual cell annotation.

Main Methods:

• Utilized a one-class logistic regression machine learning algorithm trained on non-transformed stem cells.
• Extracted transcriptomic features to identify dedifferentiated cell states in both single cell and spatial omics data.
• Applied the machine learning tool across five emerging spatial transcriptomic and proteomic technologies.

Main Results:

• The machine learning model successfully identified single cell states in metastatic tumor populations without prior cell annotation.
• Discovered stem-like cell populations that were not identified by existing single cell or spatial transcriptomic analysis methods.
• Demonstrated the model's capability to identify oncogenic stem-like cell types within the tumor microenvironment.

Conclusions:

• The developed machine learning approach offers a robust method for identifying oncogenic stem-like cells in cancer omics data.
• This technique overcomes limitations of traditional cell annotation, particularly in the context of spatial omics.
• The study highlights the potential of machine learning in uncovering novel cell states relevant to cancer progression and the tumor microenvironment.