Handling Missing Values in Proteomics: Imputation Methods Compared

Missing values are the #1 data quality problem in label-free quantitative proteomics. A typical DIA experiment with 5,000 proteins might have 10-30% missing values. How you handle them directly impacts your differential expression results. Here's what I learned building an automated preprocessing pipeline.

Why Are Values Missing in Proteomics?

Unlike genomics, missing values in proteomics are often not random. There are two distinct mechanisms:

MCAR (Missing Completely At Random)

• Technical failures: instrument glitches, chromatography issues
• Random across all abundance levels
• ~20-30% of missing values in typical experiments

MNAR (Missing Not At Random)

• Protein is below the detection limit → low-abundance bias
• More common in DDA than DIA
• ~70-80% of missing values in typical experiments
• This is the dangerous one — if you ignore it, you bias against low-abundance proteins

How to Tell the Difference

# Visualize missing pattern
library(naniar)
vis_miss(protein_matrix) + 
  labs(title = "Missing Value Pattern")

# If missing values cluster in low-abundance range → MNAR
# If randomly distributed → MCAR
library(ggplot2)
df <- data.frame(
  abundance = rowMeans(protein_matrix, na.rm = TRUE),
  pct_missing = rowMeans(is.na(protein_matrix))
)
ggplot(df, aes(x = abundance, y = pct_missing)) +
  geom_point(alpha = 0.3) +
  geom_smooth() +
  labs(title = "Missing vs Abundance — Negative correlation = MNAR")

If you see a clear negative correlation (more missing at lower abundance), your data is primarily MNAR.

Imputation Methods Compared

1. MinProb (Recommended for MNAR-dominant data)

Replaces missing values with random draws from a low-abundance distribution.

library(imputeLCMD)

# q = 0.01 means draw from bottom 1% of observed distribution
data_imputed <- impute.MinProb(as.matrix(data_log), q = 0.01)

Pros: Explicitly models the "below detection limit" mechanism Cons: Can introduce artificial patterns if data is actually MCAR

2. KNN (K-Nearest Neighbors)

Imputes based on proteins with similar expression patterns.

library(impute)
data_imputed <- impute.knn(as.matrix(data_log), k = 10)$data

from sklearn.impute import KNNImputer
imputer = KNNImputer(n_neighbors=10)
data_imputed = imputer.fit_transform(data_log)

Pros: Works well for MCAR; preserves correlation structure Cons: Assumes similar proteins have similar missingness — bad for MNAR

3. BPCA (Bayesian PCA)

Uses principal component analysis to estimate missing values.

library(pcaMethods)
result <- pca(data_log, method = "bpca", nPcs = 3)
data_imputed <- completeObs(result)

Pros: Captures global expression patterns Cons: Computationally expensive; can overfit with few samples

4. Minimum / 2 (Quick and Dirty)

Replace NA with half the row minimum.

import numpy as np

def impute_min_half(data):
    result = data.copy()
    for i in range(data.shape[0]):
        row = data[i, :]
        row_min = np.nanmin(row)
        result[i, np.isnan(row)] = row_min / 2
    return result

Pros: Simple, fast, no dependencies Cons: All imputed values are identical per protein — adds no variance

5. Zero (Don't Do This)

# NEVER do this for proteomics
data[np.isnan(data)] = 0

Zero is not "missing" — it's a valid measurement. Replacing NA with 0 creates massive artifacts in log-transformed data (log2(0) = -Inf).

Benchmark: Which Method Is Best?

I tested all methods on a real DIA dataset (4,800 proteins, 12 samples, ~15% missing):

Setup

1. Take complete cases (no missing values) — 2,100 proteins
2. Artificially remove 15% values with MNAR pattern
3. Impute with each method
4. Compare imputed values to known truth

Results

Method	RMSE	Correlation	DE False Positives	DE False Negatives
MinProb (q=0.01)	0.82	0.91	3%	5%
KNN (k=10)	0.65	0.94	8%	2%
BPCA	0.71	0.93	6%	3%
Min/2	0.95	0.87	4%	8%
Zero	2.31	0.52	45%	12%
No imputation (listwise)	N/A	N/A	2%	35%

Key Findings

1. Zero imputation is catastrophic — 45% false positive rate
2. No imputation loses 35% of true DE proteins — you miss real biology
3. MinProb is the safest overall for proteomics (MNAR-dominant)
4. KNN has lowest RMSE but higher false positive rate with MNAR data

My Recommendation: Hybrid Approach

def smart_impute(data, method='auto'):
    """
    Auto-detect missingness mechanism and choose method.
    """
    # Calculate correlation between missingness and abundance
    abundance = np.nanmean(data, axis=1)
    pct_missing = np.mean(np.isnan(data), axis=1)
    correlation = np.corrcoef(abundance[pct_missing > 0],
                               pct_missing[pct_missing > 0])[0, 1]

    if method == 'auto':
        if correlation < -0.3:
            method = 'minprob'  # MNAR dominant
            print(f"Detected MNAR pattern (r={correlation:.2f}) → MinProb")
        else:
            method = 'knn'  # MCAR dominant
            print(f"Detected MCAR pattern (r={correlation:.2f}) → KNN")

    if method == 'minprob':
        # MinProb: random draw from bottom 1%
        for i in range(data.shape[0]):
            row = data[i, :]
            observed = row[~np.isnan(row)]
            if len(observed) == 0:
                continue
            q01 = np.percentile(observed, 1)
            std = np.std(observed) * 0.3
            n_missing = np.sum(np.isnan(row))
            data[i, np.isnan(row)] = np.random.normal(q01, std, n_missing)

    elif method == 'knn':
        from sklearn.impute import KNNImputer
        imputer = KNNImputer(n_neighbors=10)
        data = imputer.fit_transform(data)

    return data

Filtering Before Imputation

Always filter before imputing:

# Remove proteins with >50% missing in ALL groups
def filter_by_valid_values(data, groups, min_valid_pct=0.5):
    """Keep proteins with enough values in at least one group."""
    keep = np.zeros(data.shape[0], dtype=bool)

    for group_indices in groups.values():
        group_data = data[:, group_indices]
        valid_pct = np.mean(~np.isnan(group_data), axis=1)
        keep |= (valid_pct >= min_valid_pct)

    print(f"Keeping {keep.sum()} / {len(keep)} proteins")
    return data[keep], keep

Effect on Downstream Analysis

The choice of imputation method cascades through your entire analysis:

Missing values → Imputation → Normalization → DE Analysis → Pathway
     ↓                ↓              ↓             ↓           ↓
  15% NA          Method bias    Median shift   P-values    Enrichment

• Under-impute (too conservative) → Miss real DE proteins → Miss pathways
• Over-impute (too aggressive) → False DE proteins → Spurious pathways
• Wrong mechanism (KNN for MNAR) → Systematic bias toward high abundance

Summary Decision Tree

Is >50% of the protein missing across ALL groups?
  → YES: Remove the protein
  → NO: Continue

Is the protein missing in only one group?
  → YES: Likely biological (below detection) → MinProb
  → NO: Continue

Is missingness correlated with abundance?
  → YES (r < -0.3): MNAR → MinProb (q=0.01)
  → NO (r > -0.3): MCAR → KNN (k=10)

Tools

• BioAI Market — Automated preprocessing with method recommendation
• DEP R package — DEP::impute() with multiple methods
• MSnbase R package — impute() function
• Perseus — GUI-based, popular in MaxQuant workflows

---

This is part of a proteomics analysis series. See also: Differential Expression Analysis Pipeline for the next step after imputation.