A probabilistic method to calculate, with improved precision, multiple cutoffs for a continuous biomarker and to classify individual patients into cohorts based on natural variability of a biomarker.
Biomarkers are a powerful tool for identification of a variety of health conditions and diseases as well as monitoring responses to therapies. Development of biomarkers is limited by the difficulty of determining cutoffs values to classify patients into subpopulations.
Candidate biomarkers, measured as continuous variables, are often dichotomized using single cutoff value that defines subpopulations with low and high biomarker levels. These single cutoff points are often a value already published, a certain sample value like the median, or an “optimized” cutoff, which correlates with clinical test and survival data. However, there is concern about the natural existence of a single cutoff and its accuracy for a given variable in disease diagnosis.
The presented technology is an original analytical method. It generates probability distribution for candidate biomarker in an overall patient population and then provides inference about the natural existence of subpopulations with different levels of expression of candidate biomarker. The inference about the characteristics of subpopulations is based on the concurrent analysis of datasets obtained from healthy and diseased samples. This method is expected to be more precise than traditional methods of dichotomization with a single cutoff.
Figure 1. Comparison of dichotomization using a single cutoff point – the currently used method – (left) with a multiclassification method using a set of cutoffs provided by the titled biomarker level analysis method (right). Image source: PMID: 31113820
- Analyzes raw sample data sets representing an overall patient population
- Predicts patient subpopulations with different levels of a biomarker
- Estimates proportions of patients in each subpopulation
- Classifies individual patients into respective subpopulations
- Robust: Validated on standard clinical markers in breast cancer
- Practical: Tested on a candidate tumor marker in breast cancer
- Performance: Detection/monitoring/targeted prevention of cancer
- Prospective: Other conditions and diseases
- Provides reliable estimates from a relatively small sample data sets
- Handles missing data and provides a confidence interval for a cutoff
- No generation of training and validation subsets
- More precise disease diagnosis (not limited to a single cutoff point)
- More precise molecularly targeted therapy (not limited to two cohorts)
Stage of Development:
- The classifier algorithm has been developed and tested on mRNA expression datasets from the human breast tissue for diagnosis of breast cancer.
- It is ready to be deployed in a clinical research setting with GMM capabilities.
Prabakaran et al. Cancer Res., 2019, 79, 3492. (PMID: 31113820).
Docket # 19-8759