Supplementary materials and code for the BMC Bioinformatics journal article "Machine learning with the TCGA-HNSC dataset: Improving performance by addressing inconsistency, sparsity, and high-dimensionality" by authors Michael C. Rendleman, B.S.E.; John M. Buatti, MD; Terry A. Braun, Ph.D.; Brian J. Smith, Ph.D.; Bart Brown; Chibuzo Nwakama; Reinhard Beichel, Ph.D.; Thomas L. Casavant, Ph.D.
To install the necessary dependencies for the R scripts, we supply the install_prereqs.R script.
Any questions about this analysis or the manuscript can be sent to [email protected].
Preprocessed pre-imputation and post-imputation datasets are provided in .arff format (WEKA's attribute-relation file format) in clinical_NO_imp.arff and clinical_rf_imp.arff, respectively. Importance values for these datasets are provided in raw_importance_noimp.csv and raw_importance_rfimp.csv.
Tumor grading variables and corresponding patient outcomes are stored in clintum_tx_grade.Rda for convenience of use in R-based SPCA experiments. Only the 520 patients with tumor expression data are included in this data frame.
RNA expression data for the 520 patients (alongside tumor grading and treatment information) is supplied in rnatum_tx_grade_surv.Rda.
Transformations of RNA expression data via SPCA can be found in spcaXXcomponents.Rda, where XX is the number of components. These data were transformed from the rnatum_tx_grade_surv.Rda data using the SPCA_generation.r script.
Classifier training on pre- and post-imputation data can be done in WEKA as described in our manuscript: https://pubmed.ncbi.nlm.nih.gov/31208324/
Importance values for these variables can be calculated with CIRF_importance.r, though the raw (pre_averaged) results can be examined in the raw_importance_noimp.csv and raw_importance_rfimp.csv files.
Classifier training on the full set of solid-tumor RNA expression data can be replicated with the Full_RNA_training.R script. The models from this script are not supplied, as they can be quite large (on the order of hundreds of MB to GB). This script requires a high-performance computing environment, and we recommend no less than 100 GB of memory to ensure training will complete.
Training of classifiers, calculations of variable importance, and training timing for the SPCA-transformed data can be performed using the SPCA_training_and_importance.R script, though the resulting models are also available in the model_fits/ directory.
SPC gene weights can be obtained from the SPCA_generation.r file. For our analysis, genes with absolute weight greater than 0.1 are considered contributors.
After obtaining the genes comprising the SPC under consideration, GOEA can be performed here: http://geneontology.org/page/go-enrichment-analysis