Matched and Unmatched Omics Sampling

Why Subject Overlap Matters

A transcriptome dataset and a proteome dataset may look like two tables of molecular measurements, but their relationship depends on who was measured. If both assays include measurements from the same people, the data can be aligned at the subject level. If they include different people, there is no subject-level pair between the tables, even if the samples came from the same hospital or were collected under the same study protocol.

The distinction matters because a design determines which comparisons the data can support. Matched sampling preserves information about how two assay layers vary within the same subject. Unmatched sampling can describe the two groups of subjects separately, but it does not provide a within-subject transcriptome–proteome pair. Neither design is automatically better: the appropriate design depends on the scientific question and practical constraints. This tutorial focuses on identifying the relationship between the sampled subjects, not on choosing a study design or a method of integration.

The earlier material on subjects, assays, and cohorts established that these are distinct design units. Here, the key question is whether the subjects represented in the transcriptome and proteome datasets overlap in a way that supports pairing.

For a simple study with one relevant transcriptome and one relevant proteome measurement per subject, define \(n_T=|S_T|\), \(n_P=|S_P|\), and \(m=|S_T\cap S_P|\), where \(m\) is the number of subjects represented in both datasets. Then \(m\) counts the available subject-level pairs. The overlap fraction \(m/\min(n_T,n_P)\) describes what proportion of the smaller assay set can be paired, but it does not by itself say whether all observations in both datasets are matched.

This count follows by adding the sizes of the two sets: every subject in only one dataset is counted once, but each subject in both is counted twice. Subtracting \(m\) removes that second count. Thus, a fully matched design with the same subjects in both assays has \(m=n_T=n_P\); a fully unmatched design has \(m=0\). An overlap between these cases indicates partial overlap.

What Counts as a Match?

A match is a relationship between measurements from the same biological subject, not simply measurements with similar values, the same diagnosis, or nearby row numbers. If a subject contributes RNA-sequencing data and a proteomic measurement, those observations may form a subject-level pair. This does not require that both assays use the same physical aliquot: different specimens or aliquots can still come from the same subject.

However, subject identity alone may not be enough to define a scientifically meaningful pair. A blood transcriptome collected before treatment and a tissue proteome collected years later are both associated with the same person, but they do not necessarily represent measurements of the same biological state. Before calling observations matched for an intended comparison, the relevant condition and collection time should be compatible with that comparison. The definition of “relevant” depends on the study question; the data structure should not silently turn measurements from different states into equivalent pairs.

A match also does not mean that every transcript is paired one-to-one with a protein feature. The pairing described here is between subjects across assay datasets. Which molecular features correspond across layers is a separate issue. Likewise, multiple technical replicates from one subject do not create multiple independent subject pairs.

Unmatched and Partially Overlapping Samples

In an unmatched design, the two assay datasets represent distinct subjects. For instance, a transcriptome assay might be available for one set of participants and a proteome assay for another. Having equal sample sizes does not make those datasets matched: the relevant question is identity overlap, not whether the row counts agree. Similarly, collecting both datasets at one institution or from people with the same diagnosis does not establish a subject-level match.

Partial overlap deserves its own label. Some subjects may contribute to both assays, while others contribute to only one. Those shared subjects provide pairs; the remaining measurements do not. Calling the entire dataset “matched” can hide the unpaired observations, while calling it “unmatched” discards the fact that some subject pairs exist. Reporting the three counts \(n_T\), \(n_P\), and \(m\) makes the structure clearer.

Assumptions and Checks Before Classifying

The set-counting description is simple, but it depends on correctly identifying the sampling units. At minimum, classification assumes that subject labels refer consistently to biological people or other intended units, and that the entries being counted are the relevant assay observations rather than repeated technical measurements counted as new subjects. When a study includes multiple visits, tissues, or conditions, the comparison must specify which observations are being paired. Otherwise, a subject can appear to overlap across assays while the particular observations of interest do not correspond.

A practical audit can proceed in a fixed order:

This audit is a check of design structure, not a guarantee that every feature is observed in every sample. Missing values or detection limits can make some measurements unavailable within an otherwise matched set of subjects. Conversely, a measurement that is absent for a particular feature does not change the identity of the subject. Keep sample-level subject overlap distinct from feature-level missingness.

Why the Distinction Changes What Can Be Compared

With matched sampling, each shared subject supplies observations from both assays. This makes a within-subject comparison possible in principle: the analysis can use the fact that two measurements came from the same person. The subject may have stable characteristics that influence both molecular measurements, so retaining the pairing can matter. Matched sampling does not guarantee that a particular analysis is valid, nor does it remove the need to account for other design features.

With unmatched sampling, there is no observed pair for an individual across assays. One can describe each assay dataset, but a direct claim about how transcriptome and proteome values co-vary within subjects cannot be supported by nonexistent pairs. A statistical model may address other questions under additional assumptions, but it cannot recover the actual unobserved subject-level pairings from the two tables alone.

These categories describe sampling relationships, not quality rankings. A well-conducted unmatched study can be appropriate when separate cohorts are the available data, while a matched study can still be compromised by poor measurement or incompatible collection conditions. The classification tells readers what subject-level links are present; it does not by itself settle whether a study answers a particular scientific question.

Diagnostics and Common Pitfalls

• Equating equal sample sizes with matching. Two tables each containing 100 rows may represent the same 100 subjects or two disjoint groups of 100. Compare subject identity, not table dimensions.
• Treating a shared cohort label as a shared subject. “Cohort 1” names a group, not an individual. Cohort membership does not establish pairs.
• Using row order as identity. The first row of each file is not automatically the same participant. Ordering can change during filtering or export.
• Counting technical replicates as subjects. Repeated instrument measurements can increase the number of records without increasing the number of independent biological subjects.
• Ignoring time or condition. Measurements from the same person but different study states may not form the intended pair. Check the context of collection before interpreting overlap.
• Calling partial overlap simply matched. State how many subjects have both assays and how many are assay-specific; do not imply that every observation is paired.

A useful diagnostic is to compare the reported number of assay records with the number of unique subjects, then verify that the shared-subject count is plausible given recruitment and collection. Unexpectedly high overlap may reflect reused or nonunique labels; unexpectedly low overlap may indicate inconsistent identifiers. Resolve discrepancies before describing the design. Similarity of molecular values is not evidence that two samples came from the same person.