AI summary by TwinMind
StatQuest: DESeq2, part 1, Library Normalization
StatQuest with Josh Starmer
12m 41s
Introduction to DESeq2
- StatQuest presentation from UNC Chapel Hill Genetics Department
- Focus: DESeq2 differential gene expression analysis program
- Complex program broken into parts - this is Part 1: Library normalization
- DESeq2 doesn't use RPKM, FPKM, TPM methods (neither does EdgeR)
Two Main Problems in Library Normalization
Problem 1: Library Size Differences
- Example: 6-gene genome simulation
- Sample 1: 635 total reads mapped
- Sample 2: 1,270 reads mapped (exactly double)
- Read counts per gene in Sample 2 = 2x Sample 1 counts
- Difference due to sequencing depth, not biology
- RPKM/FPKM/TPM/CPM handle this adequately
Problem 2: Library Composition Differences
- RNA-seq compares different tissue types (liver vs spleen)
- Liver-specific genes transcribed in liver but not spleen
- Same tissue comparisons: knockout transcription factors
- Specific example: both libraries 635 reads
- Sample 1: gene A2M highly expressed (563/635 reads)
- Sample 2: A2M not transcribed (0 reads)
- A2M's 563 reads redistributed to other genes in Sample 2
- Makes other genes appear highly expressed in Sample 2
- Only true differentially expressed gene: A2M
DESeq2 Normalization Method (7 Steps)
Step 1: Take Log of All Values
- Uses natural log (log base e) - R programming default
- Log(0) = negative infinity in R
- Could use log base 2 or 10, but chose base e
Step 2: Calculate Row Averages
- Average each gene's log values across samples
- Adding numbers to infinity = infinity
- Genes with zero counts have negative infinity averages
- Log averages = geometric averages
- Geometric averages less swayed by outliers
- Example: Gene 3 raw average = 96, geometric average = 73.7
Step 3: Filter Genes with Infinity
- Remove genes with infinity averages
- Eliminates genes with zero counts in ≥1 sample
- Liver vs spleen: removes tissue-specific genes
- Focuses scaling on housekeeping genes
Step 4: Subtract Average from Log Counts
- For each gene/sample: subtract average log from individual log
- Mathematically equivalent to log(ratio)
- Shows ratio of sample reads to average across samples
- Identifies genes expressed above/below/at average levels
Step 5: Calculate Median Ratios per Sample
- Calculate median of log ratios for each sample
- Median avoids extreme gene influence
- Huge expression differences don't sway median
- Emphasizes moderate differences and housekeeping genes
Step 6: Convert Medians to Scaling Factors
- Medians are log values (exponents for e)
- Raise e to median power = scaling factor
- Final scaling factors calculated for each sample
Step 7: Divide Original Counts by Scaling Factors
- Apply scaling factors to original read counts
- Include all genes (even those filtered during calculation)
- Gene 1 example: still scaled despite zero count in Sample 1
- Result: Sample 1 counts scaled up, Sample 3 scaled down
DESeq2 Method Summary
- Logs eliminate tissue-specific genes, smooth outliers via geometric mean
- Median downplays high-read genes, emphasizes moderate expression
- Combined approach focuses scaling on housekeeping genes
- Targets genes transcribed at similar levels across all samples
