Feature Selection and Machine Learning Strategies to Develop a Minimal Bioassay for Cholangiocarcinoma Subtype

To demonstrate that the combination of feature selection and machine learning methods can curate affordable, accurate RNA sequencing-based assays for cholangiocarcinoma subtype that achieve comparable results to expensive, time-intensive multiomic analysis.

From 43 cholangiocarcinomas treated at Mayo Clinic in Minnesota (May 1, 1996, to July 30, 2019) and 206 treated at Fudan University in China (2009 to 2015), 35,449 RNA, protein, and phosphoprotein features were extracted. These multiomic features were used to assign patients to 3 molecular classes. Important RNA sequencing features were selected by competing feature selection methods. Machine learning classification models were tested on each feature set. Accuracy, F1 score, area under the curve, precision, and recall were validated for each feature selection and predictive model combination, bootstrapping using up-sampled, class-balanced data. A custom function of physical assay cost and classification accuracy was optimized to find the best assay. The best feature selection and model combination was tested with reserved data.

A set of 20 RNA transcripts was identified as sufficient for classification with gradient boosted trees when optimizing number of features and overall accuracy simultaneously. Reserved data testing demonstrated an accuracy of 83.8% (75.1% to 92.3%); precision, 82.9% (73.3% to 92.5%); recall, 87.2% (74.8% to 99.7%); and area under the curve, 0.961 (0.925 to 0.997). Metabolic pathways were overrepresented among the features selected.

Feature selection methods and predictive machine learning models work synchronously to accurately reproduce the results of a complex multiomic subtype assay based on limited RNA sequencing. This methodologic framework is a potential pathway to integrate research multiomic subtyping for molecularly complex cancers into a realistic clinical setting.