Bio Multi Omics Data Harmonization Agent Skill

library(MultiAssayExperiment)

# Load individual assays
rna <- SummarizedExperiment(assays = list(counts = rna_matrix), colData = sample_info)
protein <- SummarizedExperiment(assays = list(intensity = protein_matrix), colData = sample_info)
methylation <- SummarizedExperiment(assays = list(beta = meth_matrix), colData = sample_info)

# Create experiment list
exp_list <- ExperimentList(RNA = rna, Protein = protein, Methylation = methylation)

# Sample map (links samples to assays)
smap <- data.frame(
    assay = rep(c('RNA', 'Protein', 'Methylation'), each = nrow(sample_info)),
    primary = rep(sample_info$SampleID, 3),
    colname = c(colnames(rna_matrix), colnames(protein_matrix), colnames(meth_matrix))
)

# Create MAE
mae <- MultiAssayExperiment(experiments = exp_list, colData = sample_info, sampleMap = smap)

# RNA-seq: VST normalization
library(DESeq2)
dds <- DESeqDataSetFromMatrix(countData = assay(mae, 'RNA'),
                               colData = colData(mae),
                               design = ~ 1)
vst_rna <- assay(vst(dds))

# Proteomics: Log2 + median centering
log2_protein <- log2(assay(mae, 'Protein'))
log2_protein[is.infinite(log2_protein)] <- NA
medians <- apply(log2_protein, 2, median, na.rm = TRUE)
norm_protein <- sweep(log2_protein, 2, medians - median(medians))

# Methylation: M-value transformation
beta <- assay(mae, 'Methylation')
m_values <- log2(beta / (1 - beta))

library(sva)

# Combine normalized matrices for joint batch correction
# Only use common samples
common_samples <- Reduce(intersect, colnames(mae))

combined <- rbind(
    vst_rna[, common_samples],
    norm_protein[, common_samples],
    m_values[, common_samples]
)

# Add omics type as covariate
omics_type <- c(rep('RNA', nrow(vst_rna)),
                rep('Protein', nrow(norm_protein)),
                rep('Methylation', nrow(m_values)))

# ComBat for batch correction
batch <- colData(mae)[common_samples, 'Batch']
mod <- model.matrix(~ Condition, data = colData(mae)[common_samples, ])

corrected <- ComBat(dat = combined, batch = batch, mod = mod)

# Split back into separate matrices
idx_rna <- 1:nrow(vst_rna)
idx_prot <- (nrow(vst_rna) + 1):(nrow(vst_rna) + nrow(norm_protein))
idx_meth <- (nrow(vst_rna) + nrow(norm_protein) + 1):nrow(combined)

corrected_rna <- corrected[idx_rna, ]
corrected_protein <- corrected[idx_prot, ]
corrected_meth <- corrected[idx_meth, ]

library(biomaRt)

# Map protein IDs to gene symbols
ensembl <- useEnsembl(biomart = 'genes', dataset = 'hsapiens_gene_ensembl')

# Protein to gene mapping
protein_ids <- rownames(norm_protein)
protein_mapping <- getBM(attributes = c('uniprotswissprot', 'hgnc_symbol'),
                          filters = 'uniprotswissprot',
                          values = protein_ids,
                          mart = ensembl)

# Aggregate proteins to gene level (mean)
protein_gene <- norm_protein
rownames(protein_gene) <- protein_mapping$hgnc_symbol[match(rownames(protein_gene), protein_mapping$uniprotswissprot)]
protein_gene <- protein_gene[!is.na(rownames(protein_gene)), ]
protein_gene <- aggregate(. ~ rownames(protein_gene), data = as.data.frame(protein_gene), FUN = mean)

# Map methylation probes to genes
# (requires annotation package, e.g., IlluminaHumanMethylation450kanno.ilmn12.hg19)
library(IlluminaHumanMethylation450kanno.ilmn12.hg19)
anno <- getAnnotation(IlluminaHumanMethylation450kanno.ilmn12.hg19)
probe_genes <- anno[rownames(m_values), 'UCSC_RefGene_Name']

# Per-assay missing value analysis
missing_summary <- function(mat) {
    data.frame(
        total_missing = sum(is.na(mat)),
        pct_missing = mean(is.na(mat)) * 100,
        samples_complete = sum(colSums(is.na(mat)) == 0),
        features_complete = sum(rowSums(is.na(mat)) == 0)
    )
}

lapply(list(RNA = vst_rna, Protein = norm_protein, Methylation = m_values), missing_summary)

# Filter features with too many missing values
filter_missing <- function(mat, max_missing_pct = 50) {
    keep <- rowMeans(is.na(mat)) * 100 < max_missing_pct
    mat[keep, ]
}

protein_filtered <- filter_missing(norm_protein, max_missing_pct = 30)

# Imputation (MinProb for proteomics)
impute_minprob <- function(mat) {
    for (i in 1:ncol(mat)) {
        nas <- is.na(mat[, i])
        if (any(nas)) {
            q01 <- quantile(mat[, i], 0.01, na.rm = TRUE)
            mat[nas, i] <- rnorm(sum(nas), mean = q01, sd = abs(q01) * 0.1)
        }
    }
    mat
}

protein_imputed <- impute_minprob(protein_filtered)

# Find complete samples across all assays
complete_samples <- intersectColumns(mae)
cat('Samples in all assays:', ncol(complete_samples), '\n')

# Subset to common samples
mae_matched <- mae[, complete_samples, ]

# Alternative: keep samples with N-1 assays
subsetByColData(mae, mae$has_at_least_2_assays)

# Z-score transformation (per feature)
scale_matrix <- function(mat) {
    t(scale(t(mat)))
}

scaled_rna <- scale_matrix(vst_rna)
scaled_protein <- scale_matrix(norm_protein)
scaled_meth <- scale_matrix(m_values)

# Verify scaling
cat('RNA mean:', mean(scaled_rna, na.rm = TRUE), 'sd:', sd(scaled_rna, na.rm = TRUE), '\n')

# Save as list for integration tools
harmonized <- list(
    RNA = scaled_rna,
    Protein = scaled_protein,
    Methylation = scaled_meth,
    sample_info = colData(mae)[common_samples, ]
)

saveRDS(harmonized, 'harmonized_multiomics.rds')

# Or as separate CSVs
write.csv(scaled_rna, 'harmonized_rna.csv')
write.csv(scaled_protein, 'harmonized_protein.csv')
write.csv(scaled_meth, 'harmonized_methylation.csv')