@bgicli/bgicli 2.1.1 → 2.2.0

package/data/workflows/experimental-design-statistics/references/batch_effect_mitigation.md

@@ -0,0 +1,756 @@
+# Batch Effect Mitigation
+
+This document provides comprehensive guidance on preventing, detecting, and
+controlling batch effects in genomics experiments through proper experimental
+design.
+
+---
+
+## What Are Batch Effects?
+
+### Definition
+
+**Batch effect:** Systematic non-biological variation introduced when samples
+are processed in separate groups (batches) at different times or under different
+conditions.
+
+**Key principle:** Batch effects are technical artifacts, not biological signal.
+They can severely confound or obscure true biological differences.
+
+### Common Sources of Batch Effects
+
+**Sample processing:**
+
+- Different days/weeks of sample processing
+- Different technicians or operators
+- Different reagent lots or batches
+- Different equipment or instruments
+- Different lab locations or facilities
+
+**Library preparation:**
+
+- Different library prep kits or lots
+- Different barcoding plates
+- Different adapters or indexes
+- Time of day effects (temperature, humidity)
+
+**Sequencing:**
+
+- Different sequencing runs
+- Different flow cells or lanes
+- Different sequencing instruments
+- Different base-calling software versions
+
+**Computational:**
+
+- Different alignment software versions
+- Different reference genome versions
+- Different analysis pipelines
+
+### Impact of Batch Effects
+
+**Magnitude:**
+
+- Can be LARGER than biological effects of interest
+- Often affect 20-50% of features in RNA-seq data
+- Can create entirely spurious patterns in data
+
+**Consequences:**
+
+- False positives (detecting non-existent differences)
+- False negatives (missing true biological signals)
+- Inflated or deflated effect size estimates
+- Poor reproducibility across studies
+- Invalid biological conclusions
+
+---
+
+## The Cardinal Rule: Prevent Confounding
+
+### What Is Confounding?
+
+**Confounded design:** Batch is completely or partially correlated with
+experimental condition
+
+**CRITICAL:** If batch is confounded with condition, you CANNOT distinguish
+biological effects from batch effects, even with statistical correction.
+
+### Examples of Confounding (❌ BAD)
+
+**Complete confounding (worst case):**
+
+```
+Batch 1: Control_1, Control_2, Control_3
+Batch 2: Treatment_1, Treatment_2, Treatment_3
+```
+
+❌ All controls in one batch, all treatments in another ❌ Impossible to
+separate batch effect from treatment effect ❌ Any difference could be 100%
+batch artifact
+
+**Partial confounding (still bad):**
+
+```
+Batch 1: Control_1, Control_2, Treatment_1
+Batch 2: Control_3, Treatment_2, Treatment_3
+```
+
+❌ Unbalanced conditions across batches ❌ Statistical adjustment has limited
+power ❌ Results remain questionable
+
+### Non-Confounded Design (✅ GOOD)
+
+**Balanced design:**
+
+```
+Batch 1: Control_1, Control_2, Treatment_1, Treatment_2
+Batch 2: Control_3, Control_4, Treatment_3, Treatment_4
+```
+
+✅ Each batch contains all conditions ✅ Conditions balanced across batches ✅
+Can estimate and remove batch effects ✅ Valid biological inference possible
+
+**Key principle:** Every batch must contain samples from ALL experimental
+conditions in balanced proportions.
+
+---
+
+## Batch Design Strategies
+
+### Strategy 1: Completely Randomized Batch Assignment
+
+**When to use:**
+
+- Single experimental factor (e.g., case vs. control)
+- No important covariates to balance
+
+**Method:**
+
+1. List all samples with their condition labels
+2. Randomly assign samples to batches
+3. Ensure each batch has equal numbers of each condition
+4. Verify no confounding with `check_confounding()` function
+
+**R code:**
+
+```r
+source("scripts/batch_assignment.R")
+batch_design <- assign_samples_to_batches(
+  metadata = sample_metadata,
+  batch_size = 8,
+  balance_vars = "condition"
+)
+```
+
+### Strategy 2: Stratified Randomization
+
+**When to use:**
+
+- Important covariates that should be balanced (sex, age, site)
+- Want to ensure balance across batches
+
+**Method:**
+
+1. Identify variables to balance (condition + covariates)
+2. Create strata for each covariate combination
+3. Randomly assign samples within strata to batches
+4. Maximize balance across all variables
+
+**R code:**
+
+```r
+batch_design <- assign_samples_to_batches(
+  metadata = sample_metadata,
+  batch_size = 8,
+  balance_vars = c("condition", "sex", "age_group")
+)
+```
+
+**Important covariates to consider:**
+
+- Sex/gender (major effect in many tissues)
+- Age or developmental stage
+- Genetic background or ancestry
+- Collection site (multi-site studies)
+- Disease severity or subtype
+- Time of sample collection
+
+### Strategy 3: Blocked Design
+
+**When to use:**
+
+- Paired samples (before/after treatment)
+- Matched samples (case-control pairs)
+- Want to minimize within-pair batch differences
+
+**Method:**
+
+1. Identify paired or matched samples
+2. Process paired samples in same batch when possible
+3. If pairs must span batches, balance pair distribution
+4. Account for pairing in statistical model
+
+**Example:**
+
+```
+Batch 1: Patient1_Before, Patient1_After, Patient2_Before, Patient2_After
+Batch 2: Patient3_Before, Patient3_After, Patient4_Before, Patient4_After
+```
+
+✅ Paired samples in same batch (preferred) ✅ Or balanced across batches if
+necessary
+
+### Strategy 4: Reference Sample Design
+
+**When to use:**
+
+- Multiple batches that cannot be processed simultaneously
+- Want to empirically calibrate batch effects
+- Large studies spanning many processing days
+
+**Method:**
+
+1. Create pool of reference RNA/DNA from all samples
+2. Include 1-2 reference aliquots in EVERY batch
+3. Use reference samples to estimate and correct batch effects
+4. Helps detect technical drift over time
+
+**Benefits:**
+
+- Enables empirical batch correction
+- Detects processing quality issues early
+- Allows comparison across batches
+- Standard in large-scale studies (GTEX, TCGA)
+
+---
+
+## Batch Size Considerations
+
+### Optimal Batch Size
+
+**Key trade-offs:**
+
+- **Smaller batches:** More batches, higher risk of batch effects, harder to
+  balance
+- **Larger batches:** Fewer batches, easier to balance, but logistically
+  challenging
+
+**Recommendations by assay:**
+
+| Assay             | Typical Batch Size | Considerations                          |
+| ----------------- | ------------------ | --------------------------------------- |
+| Bulk RNA-seq      | 8-24 samples       | Library prep plate size (8, 12, 24, 96) |
+| scRNA-seq         | 4-8 samples        | 10X lane capacity, complexity           |
+| ATAC-seq          | 8-12 samples       | Tagmentation variability, library prep  |
+| ChIP-seq          | 4-8 samples        | ChIP success rate, antibody lot         |
+| Methylation array | 96 samples         | Array plate format (96-well)            |
+| Proteomics (TMT)  | 6-16 samples       | TMT plex size (6, 10, 16, 18-plex)      |
+
+### Determining Batch Size
+
+**Approach 1: Divide by number of conditions**
+
+- For k conditions, use batch size = k × m where m ≥ 2
+- Example: 3 conditions, batch size = 6, 9, or 12
+- Ensures multiple replicates per condition per batch
+
+**Approach 2: Match experimental constraints**
+
+- Library prep kit size (e.g., 12-plex barcoding)
+- Sequencing lane capacity
+- Processing time constraints (all samples same day)
+- Technician availability
+
+**Approach 3: Minimize number of batches**
+
+- Fewer batches = less batch effect variation
+- Process all samples together if possible (small experiments)
+- For large studies, batch size 10-20 samples optimal
+
+### Multi-Site Studies
+
+**Special considerations:**
+
+- Site often becomes a major batch effect source
+- Each site may process samples independently
+
+**Design strategy:**
+
+1. Ensure all conditions represented at each site
+2. Balance conditions within each site
+3. Include cross-site reference samples
+4. Account for site as batch in analysis
+5. Consider centralized processing if feasible
+
+---
+
+## Temporal Batching Strategies
+
+### Time-Based Batching
+
+**When samples arrive over time:**
+
+- Clinical studies with rolling enrollment
+- Longitudinal studies with multiple timepoints
+- Sequential experiments
+
+**Strategy 1: Fixed batch intervals**
+
+- Process samples every N weeks on fixed schedule
+- Accumulate samples, process batch when full
+- Ensures balanced batch sizes
+- Randomize order within each batch
+
+**Strategy 2: Continuous processing with reference**
+
+- Process samples as they arrive (small batches)
+- Include reference sample in each batch
+- Use references to normalize across batches
+- Appropriate for clinical diagnostics
+
+### Dealing with Sample Delays
+
+**Problem:** Some samples delayed, arrive after main batches processed
+
+**Solution 1: Reserve slots**
+
+- Plan for 10-20% additional samples
+- Reserve slots in batches for late arrivals
+- Process reserves with pilot samples or controls
+
+**Solution 2: Follow-up batch**
+
+- Process delayed samples in separate batch
+- Ensure batch includes both conditions
+- Include reference samples from original batches
+- Account for batch in analysis
+
+**Solution 3: Drop and re-collect**
+
+- If critical to have in main batch, drop from study
+- Re-collect samples for future study
+- Document exclusion criteria
+
+---
+
+## Covariate Balancing
+
+### What to Balance
+
+**Priority 1 (must balance):**
+
+- Experimental condition (primary variable)
+- Major biological covariates with known large effects:
+  - Sex/gender
+  - Age group (if wide age range)
+  - Disease subtype or severity
+
+**Priority 2 (should balance if possible):**
+
+- Collection site or center
+- Genetic background or ancestry
+- Sample type or tissue
+- Time of collection (AM vs PM, season)
+
+**Priority 3 (nice to balance):**
+
+- Technician
+- Storage time
+- Other measured covariates
+
+### How to Balance
+
+**Perfect balance (ideal but often not achievable):**
+
+- Equal numbers of each covariate level in each batch
+- Example: Each batch has 4 males and 4 females
+
+**Approximate balance (more realistic):**
+
+- Similar proportions of each level across batches
+- Example: Batch 1 has 55% female, Batch 2 has 50% female
+- Acceptable if differences small
+
+**Statistical balance:**
+
+- No significant association between batch and covariate
+- Test with chi-square or Fisher's exact test
+- p > 0.05 indicates adequate balance
+
+### Balancing Algorithm
+
+Use provided script for automated balancing:
+
+```r
+source("scripts/batch_assignment.R")
+batch_design <- assign_samples_to_batches(
+  metadata = sample_metadata,
+  batch_size = 8,
+  balance_vars = c("condition", "sex", "age_group", "site"),
+  n_iterations = 10000  # Try many random assignments
+)
+
+# Validate balance
+source("scripts/batch_validation.R")
+validation <- check_balance(batch_design, vars = c("condition", "sex", "age_group"))
+print(validation)
+```
+
+**Algorithm approach:**
+
+1. Generate many random batch assignments
+2. Calculate balance score for each
+3. Select assignment with best balance
+4. Verify no confounding
+
+---
+
+## Validation of Batch Design
+
+### Pre-Processing Validation (CRITICAL)
+
+**Before starting experiment, validate batch design:**
+
+**Check 1: Confounding test**
+
+```r
+source("scripts/batch_validation.R")
+confound_result <- check_confounding(
+  batch_design,
+  condition_var = "condition"
+)
+# Must return "No confounding detected"
+```
+
+❌ If confounding detected, regenerate design - DO NOT PROCEED
+
+**Check 2: Balance test**
+
+```r
+balance_result <- check_balance(
+  batch_design,
+  vars = c("condition", "sex", "age_group")
+)
+# Should show similar proportions across batches
+```
+
+**Check 3: Visual inspection**
+
+```r
+visualize_batch_design(
+  batch_design,
+  output_file = "batch_layout.svg"
+)
+# Manually inspect: does each batch look similar?
+```
+
+### Post-Processing Validation
+
+**After sequencing, check for batch effects:**
+
+**PCA plot colored by batch:**
+
+- Batch clusters indicate batch effect
+- Ideally: samples cluster by biology, not batch
+
+**Heatmap of batch vs. sample:**
+
+- Should see no clear batch-related patterns
+
+**Statistical tests:**
+
+- `sva::ComBat` before/after comparison
+- Percentage of variance explained by batch (should be <10%)
+
+---
+
+## Common Batch Design Mistakes
+
+### Mistake 1: Sequential Processing by Condition
+
+**Problem:**
+
+```
+Week 1: Process all controls
+Week 2: Process all treatments
+```
+
+❌ Complete confounding of time with condition ❌ Any time-dependent effects
+look like treatment effects
+
+**Solution:**
+
+```
+Week 1: Process Controls 1-4, Treatments 1-4
+Week 2: Process Controls 5-8, Treatments 5-8
+```
+
+✅ Each week includes both conditions
+
+### Mistake 2: Ignoring Sex Imbalance
+
+**Problem:**
+
+```
+Batch 1: 6 females, 2 males
+Batch 2: 2 females, 6 males
+```
+
+❌ Sex partially confounded with batch ❌ Sex effects can be huge (thousands of
+genes)
+
+**Solution:**
+
+```
+Batch 1: 4 females, 4 males
+Batch 2: 4 females, 4 males
+```
+
+✅ Balanced sex across batches
+
+### Mistake 3: Filling Batches Sequentially
+
+**Problem:**
+
+- Samples arrive over time
+- Fill Batch 1 completely, then move to Batch 2
+- Early-arriving samples all in Batch 1, late arrivals in Batch 2
+- If sample timing correlates with condition, creates confounding
+
+**Solution:**
+
+- Reserve slots in each batch for each condition
+- Randomize arrival order within condition
+- Don't start processing until enough samples for balanced batch
+
+### Mistake 4: Unequal Batch Sizes
+
+**Problem:**
+
+```
+Batch 1: 12 samples (6 controls, 6 treatments)
+Batch 2: 4 samples (2 controls, 2 treatments)
+```
+
+⚠️ Very different batch sizes reduce power ⚠️ Harder to detect and correct batch
+effects
+
+**Solution:**
+
+- Keep batch sizes as equal as possible
+- Combine small batches if needed
+- Or split large batches to equalize
+
+### Mistake 5: Not Documenting Batch Structure
+
+**Problem:**
+
+- Process samples without recording which batch
+- Batch information lost
+- Cannot correct for batch effects in analysis
+
+**Solution:**
+
+- Document batch assignment BEFORE processing
+- Record batch ID for every sample in metadata
+- Include processing date, technician, reagent lots
+- Export batch design using `export_design.R`
+
+---
+
+## Statistical Adjustment for Batch Effects
+
+### When Design Prevention Is Not Enough
+
+Even with good design, some batch effects may remain. Statistical correction can
+help but is not a substitute for good design.
+
+**Methods:**
+
+**1. Linear model with batch covariate**
+
+```r
+# DESeq2 with batch
+dds <- DESeqDataSetFromMatrix(counts, colData, design = ~ batch + condition)
+```
+
+✅ Simple and transparent ✅ Works well for moderate batch effects ❌ May not
+remove all batch variation
+
+**2. ComBat (sva package)**
+
+```r
+library(sva)
+normalized_counts <- ComBat(
+  dat = log_normalized_counts,
+  batch = metadata$batch,
+  mod = model.matrix(~ condition, data = metadata)
+)
+```
+
+✅ Strong batch removal ❌ Can over-correct and remove biology ❌ Use with
+caution for DE analysis
+
+**3. RUVSeq (Remove Unwanted Variation)**
+
+```r
+library(RUVSeq)
+set <- RUVs(set, cIdx = negative_control_genes, k = 1)
+```
+
+✅ Data-driven approach ✅ Uses negative control genes ❌ Requires specification
+of controls
+
+**4. SVA (Surrogate Variable Analysis)**
+
+```r
+library(sva)
+svobj <- sva(dat, mod, mod0)
+# Include surrogate variables in DE model
+```
+
+✅ Discovers hidden batch variables ✅ Flexible approach ❌ Can be aggressive
+
+### When Statistical Correction Fails
+
+Statistical correction CANNOT fix:
+
+- ❌ Complete confounding (batch = condition)
+- ❌ Very strong batch effects (>50% variance)
+- ❌ Non-linear batch effects
+- ❌ Batch-specific biology (e.g., different protocols)
+
+**In these cases:**
+
+- May need to exclude problematic batches
+- Repeat experiment with better design
+- Acknowledge as major limitation
+
+---
+
+## Batch Design Checklist
+
+### Before Starting Experiment
+
+- [ ] Identified all batches (processing groups, sequencing runs, etc.)
+- [ ] Created batch assignment that includes all conditions in each batch
+- [ ] Balanced important covariates (sex, age, site) across batches
+- [ ] Ran `check_confounding()` - confirmed no confounding
+- [ ] Ran `check_balance()` - confirmed adequate balance
+- [ ] Visualized batch design - manual inspection looks good
+- [ ] Documented batch assignment in metadata spreadsheet
+- [ ] Exported batch layout for lab use (`export_batch_layout()`)
+- [ ] Lab has clear batch processing plan with dates
+
+### During Processing
+
+- [ ] Recording batch ID, date, technician, reagent lots
+- [ ] Following randomized processing order within batches
+- [ ] Including any reference samples in each batch
+- [ ] Noting any processing deviations or issues
+
+### After Processing, Before Analysis
+
+- [ ] All samples have batch information in metadata
+- [ ] Checked QC metrics - similar across batches
+- [ ] Generated PCA plot colored by batch - no obvious clustering
+- [ ] Calculated variance explained by batch - documented
+
+### During Analysis
+
+- [ ] Including batch as covariate in statistical model
+- [ ] Considering batch correction methods if needed
+- [ ] Comparing results with/without batch correction
+- [ ] Documenting batch effect handling in methods
+
+---
+
+## Special Cases
+
+### Single Sample Arrives Late
+
+**Problem:** One treatment sample arrives after main batches processed
+
+**Options:**
+
+**Option A: Drop the sample**
+
+- If already well-powered (n ≥ 6 per group)
+- One sample won't substantially change results
+- Simplest solution
+
+**Option B: Process in follow-up batch**
+
+- Include controls from freezer in the batch
+- Must have both conditions in follow-up batch
+- Account for batch in analysis
+- Works if n is critical
+
+**Option C: Save for future experiment**
+
+- Set aside for replication study
+- Avoid confounding issues entirely
+
+### Batch Size Doesn't Divide Sample Size
+
+**Problem:** 22 samples, batch size = 8
+
+**Options:**
+
+**Option A: Variable batch sizes (8, 8, 6)**
+
+- Minimize size variation
+- Balance conditions in each batch
+- Acceptable solution
+
+**Option B: Include controls/references to fill (8, 8, 8)**
+
+- Include replicates of controls or references
+- Makes batches equal size
+- More samples can increase power
+
+**Option C: Reduce to equal batches (8, 8)**
+
+- Only if can exclude 6 samples without losing power
+- Not recommended if samples are precious
+
+### Batches Span Multiple Sequencing Runs
+
+**Problem:** Library prep batch different from sequencing run batch
+
+**Solution:**
+
+- Record BOTH batch types in metadata
+- Library prep batch usually more important
+- Can include both in model: `~ seq_run + prep_batch + condition`
+- Or use `prep_batch` only if `seq_run` effect small
+
+---
+
+## Additional Resources
+
+**Key Papers:**
+
+- Leek JT et al. (2010) "Tackling the widespread and critical impact of batch
+  effects." _Nat Rev Genet_ 11(10):733-739
+- Goh WW et al. (2017) "Why Batch Effects Matter in Omics Data." _Trends
+  Biotechnol_ 35(6):498-507
+- Hicks SC et al. (2015) "Smooth quantile normalization." _Biostatistics_
+  19(2):185-198
+
+**Workflow Scripts:**
+
+- [batch_assignment.R](../scripts/batch_assignment.R) - Generate balanced batch
+  designs
+- [batch_validation.R](../scripts/batch_validation.R) - Validate and visualize
+  designs
+
+**Related Guides:**
+
+- [experimental_design_best_practices.md](experimental_design_best_practices.md) -
+  General design principles
+- [qc_guidelines.md](qc_guidelines.md) - Batch-specific QC checks
+
+---
+
+**Last Updated:** 2026-01-28 **Version:** 1.0