/brain_eqtl_2022

Primary LanguageRMIT LicenseMIT

Single-cell eQTL calling on 29 Brain cell types

Step 1. Merge data across 14 different batches

Before we merge data sets, we performed gentle Q/C steps to remove "empty" features (genes) in the following criteria:

  • the standard deviation across cells $<$ 0.01

  • the mean expression $<$ 1e-4

  • the number of non-zero cells $<$ 10

We built a quite large data set consisting of gene expression vectors across 2.4M cells measured on 26k+ features

Step 2. Sort the full data into smaller data sets for each cell type

Step 3. Expression data preprocessing

Streamlined data processing by PCA

  1. Constructed pseudobulk expression matrix (gene x samples/individuals/projid)
  2. Perform sample by sample quantile normalization to values generated by standard normal distribution, namely $N(0,1)$
  3. For each chromosome, adjust non-genetic effects on the genes located in the chrososome by taking top 50 principal components estimated by genes in the other chromosomes (leaving one chrososome out at a time, LOCO)

${celltype}/${celltype}_PC50_all.bed.gz

a. AD vs. non-AD

We can partition the samples into two groups to carry out condition-specific eQTL analysis:

${celltype}/${celltype}_PC50_AD.bed.gz ${celltype}/${celltype}_PC50_noAD.bed.gz

b. APOE (e4) vs. APOE WT (without e4)

${celltype}/${celltype}_PC50_APOE.bed.gz ${celltype}/${celltype}_PC50_noAPOE.bed.gz

c. Female vs. male

${celltype}/${celltype}_PC50_female.bed.gz ${celltype}/${celltype}_PC50_male.bed.gz

Experimental data processing

Confounder adjustment by cell-level kNN matching

We can match AD vs. non-AD cells and selectively remove putative confounding factors present in both types of cells derived from AD and non-AD samples:

${celltype}/${celltype}_AD_all.bed.gz

Confounder adjustment by individual-level pairwise matching

We can also remove individual-level confounding factors by matching cells between individuals:

${celltype}/${celltype}_PINE_all.bed.gz

Data sharing

Naming convention:

${celltype}/${celltype}_${data_processing}_${condition}.bed.gz

How to extract gene expression vectors:

$ tabix Mic/Mic_AD_all.bed.gz 19:45409000-45410000 -h | cut -f 1-10

We can retrieve genes by their chromosome name and approximate locations (e.g., LD block):

#chromosome_name	tss	tes	ensembl_gene_id	hgnc_symbol	11409232_Mic	11336574_Mic	10260309_Mic	10248033_Mic	20207013_Mic
19	45409011	45412650	ENSG00000130203	APOE	-0.00219808405473058	0.0633145290915309	0.28734561628366	0.123355747962476	0.264699779472434

Columns:

  1. #chromosome_name : chromosome name (between 1 and 22)
  2. tss : transcription start site (the left most location in hg19)
  3. tes : transcription end site (the right most location in hg19)
  4. ensembl_gene_id : ENSEMBL gene ID
  5. hgnc_symbol : human-readable gene symbol
  6. ${projid}_${celltype} : many samples

2. Expression QTL calling within each cell type

3.