Methods

Data Cleaning and Preprocessing

First, we obtained 2 feature tables from the original study examining cancer type classification. These feature tables consisted of the final, cleaned TCGA fungal counts and metadata used in the study with 12773 total samples. Next, we preprocessed the metadata using a combination of One Hot Encoding, Ordinal Encoding, Scaler, passthrough, and dropping features. For days to death regression specifically, we filtered outliers greater than 10,000 days to prevent them from skewing the data. We then imputed missing values, which were primarily from the passthrough features, with the column mean and combined the transformed metadata table with the fungal counts table. Lastly, we consolidated certain columns, such as pathologic_stage_label, down to a handful of categorgies effectively eliminating substages. This proved to be useful in both of our models.

Regression

For regression, we used scikit-learn to run lasso ridge regression with 10-folds cross validation on both the fungal data and preprocessed metadata to predict the patients days to death. We tried out other regression models as well including simple linear regression, bayesian regression, and decision tree regression as well. The parameter for our primary model, lasso was an alpha of 0.1

Classification

For classification, we made a gradient boosting classifier with stratified 10-folds cross validation using scikit-learn. For the gradient boost classifier, we used exponential loss, learning rate of 0.1, n-estimators 150, and max depth of 3.