Normalization

Normalization is a crucial step in proteomics data analysis to correct for systematic biases and ensure comparability across samples. msmu provides several normalization methods to address different experimental designs and data characteristics.

Log-transform intensities

The log2_transform() function applies a log2 transformation to the quantification data in the specified modality. This transformation helps stabilize variance and make the data more normally distributed, which is beneficial for downstream statistical analyses. msmu assumes that log2_transform() is applied on basal level of data before applying normalization methods.

mdata = mm.pp.log2_transform(
    mdata,
    modality="psm",  # or "peptide", "protein"
    layer=None,      # optional; default None transforms .X
)

Normalize sample intensities

The normalize() function (also available as normalise()) offers multiple normalization methods: median (median), median centering (median_center), quantile (quantile), total-intensity / constant-sum (total_sum, which rescales each sample so its summed intensity equals the median of the per-sample totals), and pairwise median (pairwise_median). Users can select the method that best suits their data and experimental design. All methods assume log2-transformed input. Normalization can also be performed independently within groups: pass group_obs (an adata.obs column, e.g., sample batch or type) and/or group_var (an adata.var column, e.g., "filename" for fractionated runs) to normalize within each group.

mdata = mm.pp.normalize(
    mdata,
    modality="psm",           # or "peptide", "protein"
    method="median",          # required; one of "median", "median_center", "quantile", "total_sum", "pairwise_median"
    group_obs=None,           # optional adata.obs column: normalize within each sample group
    group_var=None,           # optional adata.var column (e.g., "filename"): normalize within each feature group
    layer=None,               # optional; default None normalizes .X
)

Adjust PTM sites by protein abundance

The adjust_ptm_by_protein() function reads each PTM site relative to its parent protein's abundance in a matched global proteome dataset, so that a site's change is not simply its protein's change.

For the ratio method (the default), the protein's log intensity is subtracted from the site's — the slope-one relationship mass action predicts, and the estimator used by msqrob2PTM. It needs no fitting, which is what keeps it usable at the sample counts proteomics actually has.

For the ridge method, a slope is instead fitted per site and the residual is kept. Note that a single-predictor ridge retains only Sxx / (Sxx + alpha) of the least-squares slope, where Sxx = (n - 1) * var(protein); a large ridge_alpha therefore turns the residual into a plain mean-centering and removes no protein signal at all.

mdata = mm.pp.adjust_ptm_by_protein(
    mdata,
    global_mdata=global_mdata,   # the global proteome: a MuData, or the path of its .h5mu
    modality="phospho_site",     # ptm modality
    method="ratio",              # options: "ratio", "ridge". default "ratio"
    rescale=True,                # whether to rescale adjusted values. default True
    layer=None,                  # optional; default None adjusts .X
    ridge_alpha=None,            # ridge penalty; only used when method="ridge"
)

The global dataset must hold a protein modality and the .uns["protein_map"] that infer_protein() writes, quantified on the same sample names as the PTM data, in log2 space.