MannLabs Imputing Proteomics Data
Impute missing values in protein intensity matrices for downstream analysis requiring complete data.
When to Impute
Impute when:
- Downstream analysis requires complete data (e.g. PCA, COMBAT batch correction)
- Missingness rate is moderate (<30-50% per feature)
Do NOT impute when:
- Downstream methods handle missing data natively (mixed-effects models, limma)
- Feature missingness is high - instead remove highly missing features with low data support.
Key missingness patterns
- MNAR (Missing Not At Random): Low-abundance proteins below detection limit. Shows correlation between intensity and missingness.
- MAR (Missing At Random): Ion suppression, peptide competition. More common in DDA.
- MCAR (Missing Completely At Random): Stochastic dropouts. Random pattern.
Imputation Workflow
Copy this checklist and track progress:
Analysis step progress:
- [ ] **Prepare data**
- [ ] **Assess** missingness patterns
- [ ] **Select** method
- [ ] **Apply imputation**
- [ ] **Validate** imputation quality
If quality criteria not met: Restart at step `assess`
Workflow
Prepare data
Remove features with high missingness before imputation due to little data support. For typical datasets, features with <50% completeness should be removed. Assess missingness within biological groups, not globally. For example, a cell-type marker "missing" in 90% of cells but present in 100% of that cell type is informative.
Ensure that data is log-transformed data if not already as most methods assume log-scale.
Assess missingness patterns
Before imputation, assess missingness patterns.
Calculate missingness rates per feature.
Visualize intensity vs missingness. If missingness is higher for features with low intensity, this corresponds to a MNAR patterns. If the missingness is roughly independent, this corresponds to an MCAR or MAR pattern.
Method Selection
Rationale: Prefer methods that consider the global data structure over methods that only consider local structure over methods that provide a single point estimate for all samples.
| Missingness Pattern | Recommended Methods (in recommended priority) |
|---|---|
| MCAR dominant (high completeness) | BPCA > Random Forest > KNN > median imputation |
| MAR | BPCA > Random Forest > KNN > median imputation |
| MNAR dominant (many low-abundance) | Density Probability Estimation (DPC/LIMPA) > MinProb > MinDet |
| Datasets with many (ca. >500) samples | PIMMS (autoencoder) |
Quality Assessment
Evaluate imputation success:
- Distribution comparison: Imputed values should match the overall intensity distribution (not create artificial modes)
- PCA stability: Compare PCA before/after imputation using Procrustes analysis
- Covariance preservation: Frobenius norm between original and imputed covariance matrices
Red flags:
- Imputed values clustered at single point (MinDet/MinProb artifacts)
- Sample clustering changes dramatically after imputation
- Variance inflation in highly missing features