Detecting novel lung cancer groups with Mapper

Published Paper DOI
10.1371/journal.pone.0284820

Genomics data analysis via spectral shape and topology

Erik Amézquita1,2

Farzana Nasrin3 Kathleen Storey4 Masato Yoshizawa5

1. Division of Plant Sciences & Technology, University of Missouri, Columbia, MO
2. Department of Mathematics, University of Missouri, Columbia, MO
3. Department of Mathematics, University of Hawaii, Manoa, HI
4. Department of Mathematics, Lafayette College, Easton, PA
5. School of Life Sciences, University of Hawaii, Manoa, HI


Materials and methods

  • FPKM counts of RNAseq data from human lung tissue:
    ↪ 19,648 genes per sample.
    ↪ 314 healthy samples (GTEx).
    ↪ 500 cancerous samples (TCGA).
  • Fit a Gaussian Mixture Model (GMM):
    ↪ Accurate transformation to a unimodal Gaussian.
    ↪ Consider Z-scores onward.
  • Compute pairwise correlation across all samples
    ↪ Filter data by mean correlation value.

Novel subgroups revealed with Mapper

  • Vary the number of bins 60 ≤ b ≤ 110.
  • Vary the overlap percentage 30 ≤ p ≤ 80.
  • Bright yellow: 100% cancerous samples.
  • Deep purple: 100% healthy samples.
  • Regardless of parameters:
    ↪ Healthy samples tend to stay at the center.
    ↪ Cancerous samples are split between both ends.

Split not captured by tSNE

  • tSNE separates healthy from cancerous samples but,
  • Fine structural details are lost with tSNE:
    (a) Using FPKM, (b) Using GMM Z-scores.

GO Enrichment Analysis

  • Two possible processes for forming lung cancer.
  • Strand -1: Mostly tumor cells
    ↪ Primarily upregulating inflammatory reactions.
  • Strand +1: Mixed bag but high risk
    ↪ Environmental factors and tumor gene interactions.
  • Similar conclusions when analyzing KEGG pathways.

Acknowledgements

This research was supported by National Institute of General Medical Sciences - Centers of Biomedical Research Excellence (COBRE) grant number P20GM125508.