Vertical Integration Across Omics Layers

What Vertical Integration Means

A study may measure RNA expression, protein abundance, and metabolite signals to describe different molecular layers of the same biological system. The central design question is not whether these measurements have the same units or feature names. It is whether their rows can be linked to the same underlying subjects. When distinct omics feature types are measured on shared subjects and the data are brought together by that subject correspondence, the design is called vertical integration.

For example, one study might collect a blood sample from each participant and generate transcriptomic, proteomic, and metabolomic measurements. Each assay produces its own sample-by-feature matrix. Vertical integration means that the measurements from a given participant can be identified across those matrices. It does not mean that all assays measure the same molecules, use the same scale, or necessarily use the same physical aliquot.

Let \(i\) index subjects, \(l\) index omics layers, and \(j\) index features within a layer. Write \(X^{(l)}_{ij}\) for the value of feature \(j\) from layer \(l\) for subject \(i\). The number and meaning of features may differ by layer: a transcriptomic table may have genes or transcripts as columns, a proteomic table may have proteins or peptides, and a metabolomic table may have measured peaks or annotated compounds. The row identity \(i\), not the feature index \(j\), is what links these tables vertically.

This notation describes a design, not a requirement that data have already been transformed or stacked into one matrix. Keeping the layer-specific matrices distinct is often clearer because their measurement scales and feature definitions differ. The earlier material on sample-by-feature matrices and metadata establishes how to read each table; here the focus is on the cross-table subject link.

What Must Be Matched

A valid vertical match requires more than two rows with similar-looking labels. The row identifiers must refer to the same underlying subject, and the measurements must correspond to a meaningful biological unit and context. In a clinical study, the biological unit may be a participant; in another design it could be an animal or a specimen source defined at the subject level. A sample identifier can identify a particular aliquot or assay run rather than the person or organism from which it came, so identifiers need an interpretable mapping.

Measurements from the same subject need not come from the same physical tube. Different assays may require separate aliquots or collection procedures. Nor must measurements be taken at exactly the same clock time in every possible design. However, if the assays are separated in time, the time difference and any intervening treatment or change in condition matter: the measurements then describe the same subject at different occasions, not a simultaneous molecular profile. A claim about cross-layer relationships at one occasion would require suitably aligned occasions.

The basic design checks are therefore specific:

• Identity: confirm that cross-layer identifiers map to the same subject, not merely the same site, visit label, or group.
• Availability: establish which subjects actually have usable measurements in each layer. Planned collection is not the same as observed assay data.
• Context: check that specimen source, visit, condition, and timing support the intended comparison across layers.
• Layer meaning: retain the feature type and assay context for each table; a column in one layer need not correspond one-to-one to a column in another.

These conditions support a defensible subject-level link. They do not guarantee that assay measurements are unbiased, biologically comparable in every respect, or suitable for a particular inferential method. They establish the design structure that any later analysis must respect.

Partial Matching and the Analysis Set

Vertical integration does not require every subject to have every assay. A subject can contribute to one layer or to a subset of layers. It is important, though, to say which overlap is relevant. A comparison of transcriptomic and proteomic data needs subjects present in those two layers; an analysis that uses transcriptomic, proteomic, and metabolomic measurements together needs subjects present in all three. These sets can have different sizes.

For two layers \(A\) and \(B\), define their shared-subject set as the intersection of their subject sets. The shared count is the number of subjects in that intersection. For several layers, the complete matched set is the intersection across all required layers. This is a bookkeeping definition: it does not choose a statistical method or resolve why a measurement is unavailable.

Here \(S_l\) is the set of subjects with usable measurements in layer \(l\), \(L\) is the number of layers under consideration, and \(|S_{\mathrm{all}}|\) is the number of subjects in the complete intersection. A two-layer overlap is calculated using just those two sets. Reporting the set or count alongside the layer names prevents the phrase “matched sample size” from hiding which assays are included.

Why the Link Matters

The motivation for vertical integration is that distinct molecular layers can provide complementary measurements on the same biological units. A transcript measurement and a protein measurement are not duplicate readings of one feature: they describe different molecular entities and may reflect different stages or processes. Observing them on shared subjects makes it possible to ask whether between-subject patterns in one layer align with patterns in another, subject to the measurement properties and assumptions of the eventual analysis.

The cross-layer connection is meaningful only when the row alignment is meaningful. If transcript values for one person are accidentally paired with protein values for another, a computed relationship would describe artificial pairings rather than subject-level biology. Duplicate rows can create a similar problem if repeated assay records are mistaken for distinct subjects. Valid design descriptions therefore record the subject identifier and, where relevant, specimen and visit identifiers separately.

Distinguishing Vertical and Horizontal Integration

The earlier material on horizontal integration considered combining information across cohorts. The distinction here is the direction of the link. Vertical integration connects different feature types across omics layers for shared subjects. Horizontal integration connects observations across cohorts or studies, often with related feature types. A project may have both structures—for example, several cohorts each measured on several omics layers—but the two kinds of correspondence should be described separately.

This comparison is about study structure, not a ranking of approaches. A study with several layers measured on nonoverlapping groups does not have the same subject-level vertical link as one with shared participants, even if the groups have the same diagnosis. Conversely, measuring one layer in multiple cohorts does not by itself create a vertical integration design.

Assumptions, Diagnostics, and Limitations

Before interpreting a combined data set as vertically integrated, the design description should make explicit the matching assumptions: subject identifiers are correct and unique at the intended unit; each assay row maps to an identifiable subject; and the biological context of the measurements is appropriate for the question. If observations are intended to represent one visit or condition, that context must also align. These are design assumptions, not properties that can be inferred simply from the shape of the matrices.

Practical diagnostics include counting distinct subject identifiers in each layer, checking that identifiers map consistently to metadata, listing the shared sets for the layers needed, and inspecting repeated subject-by-layer records. A useful audit also compares collection visit and specimen information across assays. An unexpected zero overlap, a sudden loss of subjects after mapping, or a count larger than the number of distinct subjects can indicate a coding, join, or duplicate-record problem.

Even with correct matching, the complete set may be much smaller than the enrolled cohort. The subjects with all assays may differ from those with only one assay, so complete-case analysis can limit representativeness. Assay availability can also vary by layer, and a correct link does not remove the consequences of missing data, detection limits, or differences in measurement scales. Those issues require their own assessment; vertical integration names the cross-layer subject structure rather than solving every measurement problem.

• Wrong reasoning: “The row labels look similar, so they must represent the same people.”
• Correction: verify the subject-level mapping using identifiers and metadata; resemblance of sample, visit, or batch labels is not evidence of identity.
• Wrong reasoning: “All assays were planned for the same cohort, so the integrated sample size is the cohort size.”
• Correction: calculate the intersection of subjects with usable measurements in every required layer. Planned measurement and observed availability are different.