One-Class Detection of Cell States in Tumor Subtypes

Artem Sokolov, Evan O. Paull, Joshua M. StuartPSB 2016

Introduction

Tumors are rarely composed of a single cell type; they are heterogeneous "smoothies" containing various subclones, immune cells, and structural tissues. For precision medicine to be effective, clinicians must accurately identify the specific cell states present within a biopsy. Traditional machine learning approaches typically rely on binary classification, which requires clearly defined "positive" and "negative" training sets. However, in the context of cancer genomics, defining a "negative" set is notoriously difficult. Background samples often contain unknown proportions of the target subtype or other uncharacterized states, leading to "contamination" that causes binary models to fail.

Methodology: One-Class Logistic Regression (OCLR)

To address these challenges, the authors propose One-Class Logistic Regression (OCLR). Unlike standard binary classifiers, one-class models are trained exclusively on examples of the "positive" class to learn a characteristic signature.

OCLR improves upon existing one-class methods like $\nu$-SVM and Support Vector Data Description (SVDD) by utilizing a differentiable logistic loss function. This mathematical property allows for the seamless integration of standard regularization techniques, such as LASSO or Elastic Nets, enabling the creation of "sparse" models that focus on the most biologically relevant genes. The model effectively learns the "direction" of the positive data in feature space, creating a signature for a "pure" cell state that can score mixed or noisy samples.

Experimental setup for testing model robustness against background contamination.

The researchers validated their approach using high-throughput RNA-sequencing data from the Progenitor Cell Biology Consortium (PCBC) for training on Embryonic Stem Cells (ESCs) and The Cancer Genome Atlas (TCGA) for validation on real-world Breast Invasive Carcinoma (BRCA) and Bladder Urothelial Carcinoma (BLCA) subtypes.

Robustness to Contamination

A primary advantage of the one-class approach is its stability in the face of contaminated training data. In simulations where the target signal (e.g., ESC signature) was mixed into the background samples with an increasing coefficient $\alpha$, the performance of binary models plummeted as the background became more similar to the positive class.

Performance comparison (AUC) of binary vs. one-class models as mixing proportion increases.

As shown in the figure above, while binary Logistic Regression and SVMs lose predictive power at higher mixing proportions, OCLR and SVDD show improved or stable performance. This robustness is critical for analyzing real-world tumor samples where "pure" negative controls are non-existent.

Cross-Tissue Commonalities and Stemness

The study revealed that cancer subtypes often transcend tissue boundaries. A one-class model trained solely on BRCA Basal samples was able to predict Type III Bladder cancer with an impressive AUC of 0.89. This suggests a shared "cell-of-origin" or similar evolutionary trajectories between aggressive cancers in different organs.

Furthermore, the models highlighted a strong link between Basal-like breast cancers and embryonic stem cell signatures. This enrichment suggests that aggressive tumors undergo "de-differentiation," reverting to a stem-like state that promotes rapid growth and therapeutic resistance. Through OCLR's sparse modeling, the authors identified specific genes shared between these states, such as PHGDH and STMN1, which represent potential novel therapeutic targets.

Conclusion

One-Class Logistic Regression provides a flexible and robust framework for identifying cell states in complex biological mixtures. By requiring only positive training examples and offering superior interpretability through regularization, OCLR facilitates the discovery of cross-tissue molecular links and aggressive stemness signatures. Future work will focus on building comprehensive dictionaries of one-class signatures for all known cell types, providing a powerful toolkit for dissecting tumor heterogeneity.

Summary updated

BibTeX
@inproceedings{sokolov2016one,
  author = {Artem Sokolov and Evan O. Paull and Joshua M. Stuart},
  title = {{One-Class Detection of Cell States in Tumor Subtypes}},
  booktitle = {PSB},
  year = {2016},
}
Figure