AI summary by TwinMind
StatQuest: edgeR, part 1, Library Normalization
StatQuest with Josh Starmer
14m 16s
Introduction and Background
- StatQuest video on EdgeR library normalization part 1
- Presented by genetics department at University of North Carolina Chapel Hill
- EdgeR doesn't use standard normalization (RPKM, TPM)
- Must adjust for sequencing depth AND library composition
- Different samples contain different active genes affecting results
Step 1: Remove Untranscribed Genes
- Remove all genes with 0 read counts across ALL samples
- Example: 3 samples, 5 genes - gene 5 has 0 counts in all samples → removed
Step 2: Select Reference Sample
- Pick one sample as reference for normalizing all others
- Goal: identify most "average" sample to avoid extreme outliers
- Bad reference example: sample with single noisy measurement
Reference Sample Selection Process:
- Scale each sample by total read counts
- Sum read counts per sample
- Divide original counts by sample total
- Calculate 75th quantile for each sample
- Find value where 75% of scaled data ≤ that value
- Example: Sample 1 = 0.26, Sample 2 = 0.36, Sample 3 = 0.13
- Average the 75th quantiles
- Example average = 0.25
- Select sample with 75th quantile closest to average
- Sample 1 selected (0.26 closest to 0.25)
Step 3: Gene Selection for Scaling Factors
- Different gene sets used for each sample vs reference
- Separate scaling factors calculated for each sample
Gene Type Categories (XY plot visualization):
- Left side: genes primarily transcribed in reference
- Right side: genes primarily transcribed in comparison sample
- High: genes with many reads in both samples
- Low: genes with few reads in both samples
- Target: middle genes with moderate reads, minimal bias
Filtering Process:
- Start with scaled read counts (counts ÷ sample total)
- Calculate log2 fold differences
- High reference vs sample → positive value
- High sample vs reference → negative value
- Set threshold ±6, remove extreme biases
- Remove genes with infinite values (0 reads in either sample)
Bias Detection Tables:
- Log ratio table: identifies biased genes
- Example: gene 1 = log2(0/0.9) = -∞, gene 2 = log2(0.8/X) = -0.32
- Mean of logs table: identifies high/low transcribed genes
- Geometric mean calculation (average of log2 values)
- Less influenced by outliers
Final Gene Filtering:
- Sort both tables low to high
- Remove top 30% and bottom 30% of biased genes
- Remove top 5% and bottom 5% of highly/lowly transcribed genes
- Use genes remaining in BOTH lists for scaling factor
Step 4: Calculate Weighted Average
- Method: "weighted trimmed mean of log2 ratios"
- Trimming removes extreme genes to avoid outlier effects
- Weighting scheme: genes with more reads get higher weight
- Log ratios have more variance with low read counts
- Example: 4-read difference in high-count genes = small log difference
- Same 4-read difference in low-count genes = large log difference
Step 5: Convert to Scaling Factors
- Raise 2 to power of weighted average: 2^(weighted_average)
- Calculate scaling factor for each sample relative to reference
- Example actual scaling factors from RNA-seq experiment shown
- WT2 used as reference sample
Step 6: Center Scaling Factors Around 1
- Raw scaling factors typically centered around ~0.95
- Centering process: divide each factor by geometric mean of all factors
- Final factors centered on 1
- Centering provides "nice mathematical properties" per method author Mark Robinson
- Doesn't change results, described as "artistic signature on mathematical formula"
