Novel multi-omics deconfounding variational autoencoders can obtain meaningful disease subtyping

TCGA pan-cancer mRNA and DNA data augmented with artificial confounders utilised in "Novel multi-omics deconfounding variational autoencoders can obtain meaningful disease subtyping" by Zuqi Li and Sonja Katz (manuscript in preparation).

The following data curation steps were carried out: 

• Step 1. Download data from TCGA
  • R package `TCGAbiolinks`
  • 2547 patients (after step 2) with 6 cancer types:
    • BRCA (731)
    • THCA (408)
    • BLCA (387)
    • LUSC (297)
    • HNSC (412)
    • KIRC (312)
  • mRNA expression profiles
  • DNAm expression profiles
  • Clinical data:
    • tumor stage: i, ia, ib, ii, iia, iib, iii, iiia, iiib, iiic, iv, iva, ivb, ivc, x
    • age at diagnosis
    • race: 'white', 'black or african amarican', 'asian', 'american indian or alaska native'
    • gender
      
      
• Step 2. Removal criteria
  • Patients with
    • NA or 'not reported' clinical data
    • race 'american indian or alaska native'
    • tumor stage x
  • mRNA and DNAm probes with
    • 0 variance across all included patients
    • not shared across all cancer types
    • with missing values
      
      
•  Step 3. Encode clinical vairables and save datasets
  • mRNA dataset: 2547 patients x 58,456 mRNAs
  • DNAm dataset: 2547 patients x 232,088 DNAm
  • clinic dataset: 2547 patients x 6 variables
        1. patient ID
        2. tumor stage: 1, 1, 1, 2, 2, 2, 3, 3, 3, 3, 4, 4, 4, 4
        3. age at diagnosis
        4. race: asian(1), black or african amarican(2), white(3)
        5. gender: female(0), male(1)
        6. cancer type: BRCA(1), THCA(2), BLCA(3), LUSC(4), HNSC(5), KIRC(6)
        
•  Step 4. Pre-process the datasets
  • mRNA dataset: 'TCGA_mRNAs_processed.csv'
    
    • Take the 2000 mRNAs with highest variance
    • Rescale every feature to [0,1]
    • --> 2547 patients x 2000 mRNAs
  • DNAm dataset: 'TCGA_DNAm_processed.csv'
    
    • Take the 2000 DNAm with highest variance
    • Rescale every feature to [0,1]
    • --> 2547 patients x 2000 DNAm
  • clinic dataset: 'TCGA_clinic.csv'
    
    
• Step 5. Simulate confounders (instructions can be found in Methods section of manuscript)
  • Linear confounder:
    • 'TCGA_confounder_linear.csv' - linear confounding classes
      
    • 'TCGA_DNAm_confounded_linear.csv' - linearly confounded DNAm data
      
    • 'TCGA_mRNA2_confounded_linear.csv'  - linearly confounded mRNA data
      
  • Squared confounder
    • 'TCGA_confounder.csv' - squared confounding classes
      
    • 'TCGA_DNAm_confounded.csv' - squared confounded DNAm data
      
    • 'TCGA_mRNA2_confounded.csv'  - squared confounded mRNA data
  • Categorical confounder 
    • 'TCGA_confounder_categ2.csv' - categorical confounding classes
      
    • 'TCGA_DNAm_confounded_categ2.csv' - categorically confounded DNAm data
      
    • 'TCGA_mRNA2_confounded_categ2.csv'  - categorically  confounded mRNA data
  • Multiple confounders - combined effect (linear + squared + categorical)
    
    • 'TCGA_confounder_multi.csv' - confounding classes for combined effect
      
    • 'TCGA_DNAm_confounded_multi.csv' - DNAm data with combined effect
      
    • 'TCGA_mRNA2_confounded_multi.csv'  - mRNA data with combined effect