A Pan-Cancer Signature Catalog to Classify Tumor Mixtures: Application to Recognition of Metastatic Disease in Prostate Cancer Kiley Graim UC Santa Cruz Motivation TCGA has many high quality primary tumor samples, but metastasis kills Which primaries will metastasize? Image courtesy of wikimedia commons 1 3 Possible Scenarios Primary Subtype Metastatic Subtype ? ? ? ? Do a restricted subset of primary subtypes share gene expression signatures with metastatic disease? If true, may use signature as an early sign of aggressive disease. 2 Multiple Datasets to Define Primary and Metastatic Gene Expression Signatures • n Dataset # Normal # Primary Cai (2011) Chandran (2007) Grasso (2012) GTEx (2014) Monzon (2007) Taylor (2010) TCGA Joint 0 0 28 42 52 29 21 172 22 10 59 0 65 131 246 533 # Metastatic 29 21 32 0 25 19 0 126 831 Samples (659 Tumor) # Genes 10,523 14,997 15,830 13,256 9,383 19,923 20,500 4,895 3 Removal of Batch and Dataset Effects • << newBefore before/after combat PCA plots After>> • << add indicat Batch effect removal via COMBAT (R package ‘sva’) 4 Removal of Batch and Dataset Effects • << newBefore before/after combat PCA plots After>> • << add indicat Batch effect removal via COMBAT (R package ‘sva’) 5 K=4 4 Primary Subtypes Identified from Multiple Datasets (Including TCGA) 6 Primary Subtype Predictors • Multinomial elastic net to predict primary subtypes • Trained using primary data • Leave-one-out cross-validation • Apply to metastatic samples Samples Subtype 1 vs. Not Subtype 2 vs. Not Subtype 3 vs. Not Subtype 4 vs. Not Not Subtype Subtype 7 How Robust Are the Predictors? True Predicted cluster 1 2 3 4 4 0 0 0 1 3 0.010 0 0.990 0 2 0.017 0.983 0 0 1 0.990 0.010 0 0 Balanced Success Rate = 0.991 8 K=3 3 Met Subtypes Identified from Multiple Datasets (None TCGA) 9 The Majority of Mets Are Predicted to Be Primary Subtype 2 Met-like primaries Predicted Primary Cluster 10 Met-Like Primaries Have Higher Gleason and Higher Tumor Grade Met-like primaries pval = 0.0e-3 FDR = 0.0e-3 pval = 0.0e-3 FDR = 0.0e-3 11 Predisposition for Movement and Metastasis MILI_PSEUDOPODIA_HAPTOTAXIS_UP BIDUS_METASTASIS_UP pval = 0.0e-3 pval = 0.0e-3 FDR = 0.0e-3 FDR = 0.0e-3 12 Are There Networks that Distinguish Differential pathway activity analysis Met-like Primaries from the Others? Activity up in Group A (relative to group B) Activity down in Group A (relative to Group B) Pathway Activities Pathway Activities Pathway Signature high activity low activity 13 Josh Stuart Group PathMark Overview of Distinguishing Networks of Met-Like Primaries 14 Proliferation-Related Subnetwork 15 MYB/MYC Subnetwork 16 Acknowledgements Yulia Newton Adrian Bivol Robert Baertsch Artem Sokolov Christina Yau (Buck Institute) Joshua M. Stuart 17 18 TCGA Taylor Joint primaries, mets, normals Exponential Normalization Combat Normalized joint Batch effect adjusted joint … Primaries Subtype Pipeline Mets Consensus Clustering 19 20 21