Advanced technique for optimizing sgRNA-to-cell assignment sensitivity in large-scale CRISPR screens. --- ## Overview The UMI threshold for assigning sgRNAs to cells affects the trade-off between sensitivity (capturing more cells per perturbation) and specificity (avoiding false assignments from ambient RNA). Testing multiple thresholds helps identify the optimal balance for your specific screen. **Based on:** Wang et al. (2025) "Integrative analysis of pooled CRISPR screens provide functional insights into AD GWAS risk genes" --- ## When to Use UMI Optimization ### Use this approach when: - ✅ Large screens (>500 perturbations) where optimization has high payoff - ✅ Low guide capture efficiency (<40% mapping rate) - ✅ Sufficient computational resources for multiple parallel analyses - ✅ Critical screens where maximizing statistical power is essential ### Skip this step when: - ❌ Small screens (<100 perturbations) - ❌ High guide capture (>60% mapping rate with standard thresholds) - ❌ Standard 10X Feature Barcoding with good quality metrics - ❌ Time-limited pilot studies --- ## Optimization Approach ### Step 1: Generate Multiple sgRNA Assignments Test a range of UMI cutoffs to identify the optimal threshold: ``` # Test multiple UMI thresholds umi_thresholds = [3, 5, 10, 15, 20] results_by_threshold = {} for threshold in umi_thresholds: # Assign sgRNAs with this threshold adata_temp = assign_guides_with_threshold( adata, mapping_file, min_umi=threshold ) # Count cells per perturbation cells_per_gene = adata_temp.obs.groupby('gene').size() # Store results results_by_threshold[threshold] = { 'n_cells': adata_temp.n_obs, 'mean_cells_per_gene': cells_per_gene.mean(), 'median_cells_per_gene': cells_per_gene.median(), 'genes_with_sufficient_cells': (cells_per_gene >= 50).sum(), 'genes_with_low_cells': (cells_per_gene < 20).sum() } ``` ### Step 2: Run QC and DE Analysis For each UMI threshold, run the full analysis pipeline: ``` for threshold, adata_temp in adata_by_threshold.items(): # Apply QC filters adata_filtered = apply_qc_filters(adata_temp) # Normalize and scale adata_norm = normalize_and_scale_data(adata_filtered) # Run initial DE screening de_results = screen_all_perturbations( adata_norm, control_group='non-targeting' ) # Count hits (perturbations with significant DE) n_hits = sum(1 for gene, df in de_results.items() if len(df[df['qval'] < 0.05]) >= 10) results_by_threshold[threshold]['n_hits'] = n_hits ``` ### Step 3: Select Optimal Threshold Choose the threshold that maximizes hits while maintaining sufficient cells per perturbation: ``` import pandas as pd # Create summary table summary = pd.DataFrame(results_by_threshold).T summary.index.name = 'UMI_threshold' print(summary) # Selection criteria: # 1. Maximize hits # 2. Maintain ≥50 cells per perturbation for most genes # 3. Balance between sensitivity and specificity # Typical result: 5 UMI is optimal for most screens optimal_threshold = 5 ``` --- ## Expected Results ### Typical Optimization Outcome | UMI Threshold | Total Cells | Mean Cells/Gene | Genes ≥50 Cells | Hits Detected | | --- | --- | --- | --- | --- | | 3 | 45,000 | 65 | 850 | 125 | | **5** | **42,000** | **58** | **820** | **148** | | 10 | 35,000 | 48 | 720 | 142 | | 15 | 28,000 | 38 | 580 | 118 | | 20 | 22,000 | 30 | 450 | 95 | **Optimal:** UMI threshold = 5 maximizes hits (148) while maintaining good coverage (820 genes with ≥50 cells) ### Interpretation Guidelines **Lower thresholds (3-5 UMI):** - ✅ Higher sensitivity (more cells captured) - ✅ Better statistical power (more cells per perturbation) - ⚠️ Risk of false assignments from ambient RNA - **Use when:** Capture efficiency is low, need maximum power **Higher thresholds (10-20 UMI):** - ✅ Higher specificity (fewer false assignments) - ✅ Reduced ambient RNA contamination - ⚠️ Lower sensitivity (fewer cells per perturbation) - **Use when:** High ambient RNA, need high confidence assignments **Balanced thresholds (5-8 UMI):** - ✅ Good balance of sensitivity and specificity - ✅ Typical optimal range for most screens - **Recommended for:** Standard 10X Feature Barcoding screens --- ## Quality Metrics to Monitor ### Assignment Quality ``` # Calculate assignment quality metrics for threshold, adata in adata_by_threshold.items(): # Mapping rate mapping_rate = adata.n_obs / total_cells_before_filtering # sgRNA doublet rate doublet_rate = doublets_detected / cells_with_sgrna # Cells per perturbation distribution cells_per_gene = adata.obs.groupby('gene').size() cv = cells_per_gene.std() / cells_per_gene.mean() print(f"UMI threshold {threshold}:") print(f" Mapping rate: {mapping_rate:.1%}") print(f" Doublet rate: {doublet_rate:.1%}") print(f" Coverage CV: {cv:.2f}") ``` ### DE Quality ``` # Check control outlier rates (should be <5%) for threshold, de_results in results_by_threshold.items(): control_de = de_results.get('non-targeting', pd.DataFrame()) n_control_de = len(control_de[control_de['qval'] < 0.05]) print(f"UMI threshold {threshold}:") print(f" Control DE genes: {n_control_de} (should be <500)") # If control shows excessive DE, threshold may be too permissive ``` --- ## Implementation Notes ### Computational Considerations **Runtime:** Testing 5 thresholds increases analysis time by ~5x - Small screen (100 genes): +30 minutes - Large screen (1000 genes): +2-3 hours **Memory:** Each threshold creates independent AnnData object - Estimate: ~2-5 GB per threshold for 50k cells - Consider processing thresholds sequentially if memory-limited **Parallelization:** ``` from concurrent.futures import ProcessPoolExecutor # Run thresholds in parallel (if resources available) with ProcessPoolExecutor(max_workers=5) as executor: futures = { executor.submit(analyze_threshold, adata, t): t for t in umi_thresholds } for future in futures: threshold = futures[future] results_by_threshold[threshold] = future.result() ``` ### Validation Checks After selecting optimal threshold, validate the choice: 1. **Check target gene validation rate** (should be >60%) 2. **Compare hit overlap** with adjacent thresholds (should be >70% overlap) 3. **Inspect cell count distributions** (should be balanced across genes) 4. **Review control outlier rates** (should be <5%) --- ## Alternative Approaches ### Adaptive UMI Thresholding Instead of a single global threshold, use gene-specific thresholds based on expression: ``` # Higher threshold for highly expressed genes (more ambient RNA) # Lower threshold for lowly expressed genes (maximize capture) def adaptive_threshold(gene_expression_level): if gene_expression_level > 100: return 10 # High expression elif gene_expression_level > 10: return 5 # Medium expression else: return 3 # Low expression ``` **Pros:** Balances sensitivity and specificity per gene **Cons:** More complex, harder to validate ### Probabilistic Assignment Use UMI counts to assign probabilities rather than hard cutoffs: ``` # Assign probability based on UMI count distribution p_assignment = umi_count / (umi_count + background_umi) # Include cells with p > 0.7 (70% confidence) ``` **Pros:** Retains more information, accounts for uncertainty **Cons:** Requires careful downstream handling of probabilities --- ## References **Method Source:** - Wang et al. (2025) "Integrative analysis of pooled CRISPR screens provide functional insights into AD GWAS risk genes" *bioRxiv* - GitHub: https://github.com/Alector-BIO/CRISPR\_PerturbSeq\_pHM\_publication **Related Methods:** - Replogle et al. (2020) "Combinatorial single-cell CRISPR screens by direct guide RNA capture and targeted sequencing" *Nature Biotechnology* - Discusses guide capture efficiency optimization **Statistical Considerations:** - Storey & Tibshirani (2003) "Statistical significance for genomewide studies" *PNAS* - Multiple testing correction with variable sample sizes