Why classify an integration study before analyzing it?
An integration study is shaped by two questions that should be answered separately: what overlaps across assays or cohorts, and which measurements come from the same biological subjects? These are properties of the sampling and measurement design, not conclusions drawn from a statistical method. Getting either question wrong can lead to an inappropriate data layout or to treating unrelated measurements as if they belonged together.
The first axis concerns assay feature spaces. Two data blocks may measure the same kind of features, such as the same gene-expression panel in two cohorts, or they may measure distinct feature blocks, such as gene expression and metabolite abundance. The second axis concerns subject correspondence. Measurements are matched when there is a reliable mapping showing that two or more assays were collected from the same biological subject. They are unmatched when no such correspondence is available or intended.
These labels describe the basic structure, not every detail. In practice, some designs have partial feature overlap, partial subject overlap, or both. Those cases should be reported explicitly rather than treated as perfectly matched or fully disjoint. The earlier material on the two-axis taxonomy of integration designs provides the organizing idea; here the task is to apply it carefully to concrete studies.
A reliable classification procedure
For each pair or group of data blocks, identify the measured features and the biological subjects before assigning a category. A feature is “shared” only when the measurements refer to corresponding entities in a compatible feature space. Similar-looking column labels do not establish correspondence: for example, two proteins with the same short label may refer to different isoforms, and a gene-level measurement is not automatically identical to a transcript-level measurement. Feature IDs and annotation are therefore part of the classification evidence.
Likewise, subject correspondence requires a defensible identity mapping. A sample identifier labels a specimen or assay record; it does not necessarily identify a person or other biological subject. Two samples with different labels can come from the same person, while similar labels can belong to different people. Duplicate specimens, repeated visits, or aliquots must be interpreted according to the study’s sampling plan.
Name each cohort or assay block, the assay type, and the features recorded in it.
Establish whether feature identities correspond across blocks. Record shared, distinct, and unmeasured features instead of relying on assay labels alone.
Use verified subject identifiers or documented links. Count shared subjects and distinguish them from specimens, visits, or technical replicates.
Use the two axes to state the basic category. Add “partial” or “mixed” when the observed overlap does not fit a fully shared or fully disjoint description.
The four basic combinations
| Feature relationship | Subject relationship | Basic category | Typical structure |
|---|---|---|---|
| Shared feature space | Same subjects linked across blocks | Matched-horizontal | Repeated or complementary measurements on the same people using corresponding features |
| Shared feature space | Different subjects, with no links across blocks | Unmatched-horizontal | Comparable measurements from separate cohorts |
| Distinct feature blocks | Same subjects linked across blocks | Matched-vertical | Different omics layers measured on the same people |
| Distinct feature blocks | Different subjects, with no links across blocks | Unmatched-vertical | Different omics layers collected on separate subject sets |
The labels should not be read as statements about data quality. A matched design can have missing assays for some subjects, and an unmatched design can still be useful. “Matched” refers to known subject correspondence, not to equal sample counts or a complete rectangular data table. “Horizontal” and “vertical” refer to the feature relationship, not to the physical orientation of a spreadsheet.
Worked Example: Two Cohorts With the Same Expression Panel
A hypothetical study measures a 40-gene expression panel in 60 subjects recruited at Clinic A and 45 subjects recruited at Clinic B. The cohort rosters are disjoint: no subject appears in both clinics. Both assays use the same gene-level panel and compatible measurements. Classify the design.
Let \(F_A\) and \(F_B\) denote the feature sets, and \(S_A\) and \(S_B\) the subject sets. Here, \(|F_A|=|F_B|=40\), and the annotation audit confirms \(|F_A\cap F_B|=40\). The subject sets satisfy \(|S_A|=60\), \(|S_B|=45\), and \(|S_A\cap S_B|=0\).
The feature spaces are shared, so the design is horizontal. The subjects do not correspond across cohorts, so it is unmatched. Joining the cohort matrices by columns for a shared feature panel does not create subject matching; the cohort origin and subject identity must remain explicit. If Clinic B instead remeasured 20 subjects from Clinic A, then \(S_A\cap S_B\) would contain 20 subjects and the design would have partial subject overlap. It should be described as partially matched horizontal integration, not as wholly unmatched.
Worked Example: Expression and Metabolite Measurements on Linked Subjects
In a hypothetical study, gene expression is measured for 32 patients and a metabolite panel is measured for 28 patients. The assays have distinct feature blocks: 500 genes and 75 metabolites. A verified subject crosswalk shows that 24 patients contributed to both assays. Classify the design and quantify the overlap.
Let \(F_G\) be the gene features and \(F_M\) the metabolites. By assay definition, these are distinct feature blocks, so there are no corresponding feature columns to treat as one shared feature space. The subject sets have sizes \(|S_G|=32\), \(|S_M|=28\), and \(|S_G\cap S_M|=24\).
Two features must remain distinct in the conclusion: 24 subjects are linked across assays, and 12 subjects appear in only one assay (8 expression-only and 4 metabolite-only). This is a partially matched vertical design: distinct feature blocks establish the vertical axis, while shared subjects establish subject correspondence for part of the data. It is not a complete-pair design because not every subject has both assays. The overlap proportion is a descriptive summary of the sampling pattern, not a measure of biological association.
Worked Example: Changing the Sampling Scheme Changes the Category
A hypothetical team plans to collect a 12-protein panel from 50 patients and a 20-feature lipid panel from another 50 patients. The feature blocks are distinct, and initially the patient groups are separate. The team considers two changes: first, recruit the same 50 patients for both assays; second, instead use the same 12-protein panel in two separate groups of 50 patients.
Initial plan. The protein and lipid panels are distinct, and their subject sets are disjoint: \(|F_P\cap F_L|=0\) as feature blocks and \(|S_P\cap S_L|=0\). This is unmatched-vertical integration.
Change 1: assay the same patients. If both assays are collected on all 50 people, then \(|S_P\cap S_L|=50\), while the feature blocks remain distinct. The category changes to matched-vertical integration. The change affects subject correspondence, not the feature axis.
Change 2: use one shared protein panel in two separate groups. If the lipid assay is removed and the same 12-protein panel is measured in two disjoint cohorts, then the shared feature space makes the design horizontal, while the disjoint cohorts make it unmatched-horizontal. This change affects the feature axis; it does not make the cohorts matched.
The exercise shows why classification should be revisited whenever the sampling plan changes. A change to which assays are measured can change the feature axis; a change to which subjects receive each assay can change the subject axis.
Justification: why two axes give four basic categories
For a simple design with two blocks, summarize each axis as a binary distinction: shared versus distinct feature spaces, and linked versus unlinked subjects. Each axis has two possible states, so the number of basic combinations is \(2\times2=4\). The classification is therefore a cross-classification of two independent design questions, not a single continuum from “less integrated” to “more integrated.”
This simple derivation also explains a useful diagnostic: changing only the subject allocation cannot by itself change a horizontal design into a vertical one, and changing only the assays cannot by itself establish subject matching. If a classification changes on both axes, the study plan changed on both dimensions.
Partial overlap, diagnostics, and common errors
Real studies often fall between the four clean cases. A panel may share most, but not all, features across cohorts. Some subjects may have multiple assays while others have only one. In these settings, report the relevant overlap directly: for example, “the assays share 36 of 40 annotated gene features, and 18 of 25 subjects in one assay have a linked measurement in the other.” The basic category can be named with a qualification, but the counts and mapping rules carry essential detail.
A useful diagnostic is to construct two explicit overlap summaries. For features, compare the number of corresponding feature IDs and inspect unmatched or ambiguous annotations. For subjects, compare the number of verified linked subjects and check whether a subject has multiple specimens, visits, or replicates. These checks expose common failures before integration:
- Using assay names as proof of shared features. Two “RNA-seq” blocks may differ in feature definition, reference annotation, or measured unit. Verify feature correspondence rather than assuming it.
- Using sample labels as proof of a match. A similar label is not a biological identity link. Confirm the crosswalk and document duplicate or repeated samples.
- Calling unequal sample sizes unmatched. Unequal counts do not rule out matching. A 30-subject block and a 25-subject block may share 20 verified subjects.
- Calling any shared subjects a complete match. Partial correspondence is not complete pairing. State how many records are linked and how many are not.
- Confusing shared features with shared subjects. The same genes in two cohorts establish feature-space overlap, not that the same people were measured.
When a design has multiple cohorts or assays, classify the relevant pairs or report the full incidence pattern rather than assigning one label to the entire study too quickly. For instance, one pair of blocks may be matched while another is not. The four-way table is a useful summary for simple pairs; it does not erase more complicated sampling structure.
How this classification compares with other descriptions
A study can also be described by assay count, cohort count, number of subjects, or whether data are stored as separate matrices or a concatenated table. Those descriptions are useful but do not replace the two axes. A single assay type measured in two independent cohorts is horizontal and unmatched even if the matrices are stored in one file. Two assay types collected from the same people are vertical and matched even if each assay is stored in a separate file.
Likewise, “complete” versus “incomplete” refers to whether measurements are available across the planned blocks, while matched versus unmatched concerns whether observations are linked to the same subjects. They are related design properties but are not interchangeable. A matched study can have missing assay blocks for some subjects; an unmatched study can have complete data within each separate cohort. Classification also does not choose an analysis method or guarantee that a particular method’s assumptions are satisfied. It identifies the structure that later method choices must respect.
The main limitation of the two-axis summary is compression: binary labels can hide the extent and quality of overlap. Feature correspondence may be uncertain or many-to-one, and subject links may be incomplete. Always retain the detailed mapping, counts, and design notes alongside the category. A label is a starting point for communicating design, not a substitute for auditing it.
Check Your Understanding
For each question, determine the feature relationship and subject relationship before naming the design. Solutions follow the questions.
- A hypothetical 25-gene panel is measured in two cohorts of 40 and 35 people. The cohorts are disjoint and use the same gene definitions. What is the category?
- Two distinct assays measure 100 transcripts and 30 metabolites. Both assays are performed on the same 18 subjects. What is the category?
- A 60-feature protein panel and a 20-feature RNA panel are measured in 30 and 24 subjects, respectively. A verified crosswalk links 15 subjects. Is this fully matched, unmatched, horizontal, or vertical? Give the most precise description.
- Two assay files each contain 50 rows with labels that look similar, but no documented subject crosswalk exists. Can the design be called matched? What evidence is needed?
Solutions. (1) The panel definitions are shared, so the feature relationship is horizontal; the disjoint cohorts have no linked subjects, so the design is unmatched-horizontal. (2) Transcripts and metabolites are distinct feature blocks, while the same 18 subjects are measured in both, so the design is matched-vertical. (3) The feature blocks are distinct, so it is vertical; 15 of the subjects are linked, so it is partially matched vertical, not fully matched. The remaining subjects should be reported as assay-specific. (4) No. Similar sample labels alone do not establish biological identity. A documented crosswalk or another validated link between assay records and the same subjects is needed before calling the design matched.