Correcting for Cell Type RNA Fractions in MDD RNA-seq Data

Progress finding cell type specific marker genes and estimating cell type promotions in MDD bulk RNAseq data, utilizing the deconvolution algorithm MuSiC (Wang et al, Nat. Comms., 2019), with complementary snRNA seq data (Tran et al, bioRxiv, 2020) on MDDseq data.

Presentation from EuroBioC2020.

Transcript

1. Correcting for Cell Type RNA Fractions
   in MDD RNA-seq Data
   Louise Huuki
   Research Associate
   @lahuuki
   

   View Slide

2. Major Depressive Disorder (MDD)
   ● Wide range of symptoms, can include depressed mood, reduced energy
   and concentration, or suicidal thoughts
   ● Among largest cause of world wide disability
   ● Lifetime prevalence of 17%
   ● Heritability estimated to be 30-40%
   

   View Slide

3. Data
   Bulk RNA seq Data
   ● 1091 samples from 595 individuals and 2
   brain regions
   ● We want to explore differences in
   proportions of cell types between samples
   ○ Are there differences between Dx?
   ○ Control for differences in other analysis
   ● Data set be available on psychENCODE
   knowledge portal
   Single Nucleus RNA seq data
   ● Tran et al., bioRxiv, 2020 10.1101/2020.01.19.910976
   ● Identified 10 specific/6 broad cell types in
   sACC, 12 specific/6 broad in Amygdala
   Amygdala sACC
   MDD 231 228
   Control 187 200
   Bipolar 122 123
   Deconvolution Methodology
   ● Sosina et al., bioRxiv, 2020 10.1101/2020.10.07.329839
   ● Highest accuracy with MuSiC + snRNA seq
   reference data from same region + filtered
   markers
   

   View Slide

4. ● Marker genes evaluated by a 1vAll
   t-test with findMarkers
   ● Top 5 genes for each broad cell
   type selected, ranked by standard
   fold change
   ● Lots of noise between some cell
   types
   scran::findMarkers
   pheatmap::pheatmap
   Finding Marker Genes
   

   View Slide

5. ● Checked expression of
   markers by cell type
   ● Observed outliers in one or
   more non-target cell type
   causing noise
   scater::plotExpression
   Exploring Marker Expression
   

   View Slide

6. ● Checked expression of
   markers by cell type
   ● Observed outliers in one or
   more non-target cell type
   causing noise
   ● Developed metric “mean
   ratio” : ratio of mean target
   expression and highest
   mean non-target
   expression
   scater::plotExpression
   Exploring Marker Expression
   

   View Slide

7. Mean Ratio vs. Fold Change
   ● Top mean ratio genes
   also have high fold
   changes
   ● This metric helps
   remove “noisy” genes
   

   View Slide

8. Original Markers vs. New Markers: Much Cleaner
   

   View Slide

9. Apply to Specific Cell Types
   

   View Slide

10. Broad Cell Types vs. Specific: Noisier but Usable
    

    View Slide

11. MuSiC results MuSiC::music_prop
    

    View Slide

12. Whats next?
    ● Apply same methodology to Amygdala data
    ○ Proving to be more complex
    ● Check for biological reasoning on marker
    genes
    ● Assess marker finding method with
    external single nucleus data
    ● Control for differences in sample cell type
    proportions in MDDseq project
    

    View Slide

13. Acknowledgements
    LIBD
    ● Leonardo Collado Torres
    ● Keri Martinowich
    ● Kristen Maynard
    ● Andrew Jaffe
    ● Matt Tran
    Get in touch
    ● Any advice for deconvolution?
    ● Early bioinformatics career advice?
    Johns Hopkins University
    ● Fernando Goes
    ● Stephanie Hicks
    ● Patricia Braun
    Funding
    ● LIBD
    ● NIH/NIMH R01MH111721-04
    @lahuuki
    

    View Slide

14. View Slide

15. Expression Plots for sACC - specific cell types
    

    View Slide

16. Expression Plots for sACC - specific cell types
    

    View Slide

17. Expression Plots for sACC - specific cell types
    

    View Slide

18. Expression Plots for sACC - specific cell types
    

    View Slide

19. Expression Plots for sACC - specific cell types
    

    View Slide