From a Biological Question to a Sampling Plan
An integration study is not defined by the number of omics assays it collects. It is defined by the biological question those measurements are intended to answer and by whether the sampling plan makes that answer possible. For example, asking whether RNA and protein measurements are related within individuals requires a subject-level link between the two assays. Asking whether a molecular pattern is reproducible across two patient cohorts requires comparable features and a defensible connection between the cohorts, but not necessarily the same subjects.
A useful plan makes these requirements explicit before recruitment or assay allocation. It identifies the target subjects and cohorts, the biological specimens and time points, the assays, the features that must be shared, and the subject links that will be retained. The earlier material on matched and unmatched omics sampling, horizontal integration, and vertical integration supplies the design vocabulary; here the task is to use that vocabulary to construct a plan from a research question.
Two integration directions help organize the plan. Vertical integration combines different omics layers, such as RNA and protein, typically to study complementary measurements on the same subjects or on explicitly defined subject sets. Horizontal integration combines observations across cohorts or studies that measure the same or sufficiently comparable features. A study may have elements of both, but its intended unit of integration should be stated rather than inferred from the fact that it contains several data tables.
A Planning Sequence
Start by expressing the biological question in terms of what will be compared, related, or generalized. Then ask what the answer needs to be about: a subject, a subject at a particular time, a cohort, or a population. This choice determines whether subject-level links are essential, whether multiple cohorts are needed, and which features must overlap. A question about within-person correspondence between layers cannot be answered by merely collecting both layers somewhere in the same study.
Name the eligible population, biological condition or exposure, and the comparison or relationship of interest. State whether the aim concerns individual-level relationships, cohort-level replication, or another target.
Specify who contributes samples, when they are collected, and whether cohorts are independent, overlapping, or repeated observations on the same subjects.
For every assay, identify the specimen, subject, time point, and planned feature set. Distinguish an assay not performed from a measurement that is missing or unusable.
For vertical integration, specify the subject links needed across layers. For horizontal integration, specify the features and measurement definitions that must be comparable across cohorts.
Count the planned subjects and assay allocations, mark expected missingness or non-overlap, and verify that the diagram supports the question rather than only the availability of samples.
The plan depends on several assumptions that should be made visible. The eligibility criteria must define a population relevant to the question. A proposed subject link must identify the same person and relevant sampling occasion, not merely a similar label. If cohorts will be compared horizontally, their feature names, biological meanings, and measurement definitions must be sufficiently aligned for the intended comparison. If repeated samples are involved, the time points and their interpretation must be specified. Finally, the planned assay allocation must leave enough usable observations for the intended analysis; a design diagram cannot make sparse overlap adequate.
Worked Plans
Worked Example: Matched Vertical Study of Two Layers
Question. Among adults with a newly diagnosed inflammatory disease, do baseline RNA and protein measurements show a subject-level relationship for a prespecified set of 40 biological targets? The following numbers are hypothetical and illustrate planning, not findings from a real study.
The question concerns a relationship between two layers within individuals, so the design should be vertical and matched at baseline. Suppose the plan enrolls 60 eligible adults and collects one blood specimen from each before treatment. RNA and protein assays are both planned for all 60 specimens. The assay panels are designed to cover the same 40 named targets, with the target identifiers and biological definitions documented in advance.
| Planned unit | Subjects | RNA assay | Protein assay | Shared targets |
|---|---|---|---|---|
| Baseline blood specimen | 60 adults | 60 planned | 60 planned | 40 |
Design diagram. Each arrow represents an assay on a specimen linked to the same subject at baseline.
Adult 001 at baseline → RNA assay (40 targets) and protein assay (40 targets)
Adult 002 at baseline → RNA assay (40 targets) and protein assay (40 targets)
⋮
Adult 060 at baseline → RNA assay (40 targets) and protein assay (40 targets)
The plan specifies 60 potential matched pairs of assay profiles, rather than 120 unrelated assay records. That distinction follows directly from the question: the relationship is defined across layers for the same individual. The study should retain an unambiguous subject identifier and a baseline time-point label in both assay tables. It should also record which assay was attempted and which profiles are usable. If, for example, only 54 subjects ultimately have both usable profiles, the number of complete matched pairs is 54, not 60. The planned sample count does not guarantee that every pair will be complete.
The plan is appropriate only if the two assays measure the intended targets in comparable biological material and baseline is the relevant occasion. If the aim instead were to compare average target profiles between disease groups, the question would need to name those groups and the design would need to ensure suitable representation of each; “relationship between layers” alone does not specify that contrast.
Worked Example: Horizontal Integration Across Two Cohorts
Question. Is a 120-feature metabolite profile associated with a clinical severity score in two independent cohorts, and is the association direction consistent across cohorts? All numbers here are hypothetical.
Because the question asks about consistency across cohorts for the same features, the plan is horizontal. Suppose Cohort A contributes 80 adults and Cohort B contributes 65 different adults. Both use the same eligibility definition, collect the same specimen type at the same clinical stage, measure the same 120 metabolites using a documented common feature definition, and record the severity score using the same scale. No person belongs to both cohorts.
| Cohort | Subjects | Metabolite profile | Clinical score | Shared features across cohorts |
|---|---|---|---|---|
| A | 80 | 120 features | Recorded for all planned subjects | 120 |
| B | 65 | 120 features | Recorded for all planned subjects | 120 |
Design diagram.
Cohort A: 80 subjects → 120-feature profile + severity score
Cohort B: 65 different subjects → 120-feature profile + severity score
Cohort A ↔ Cohort B: compare results for the same 120 defined features
The arrow between cohorts denotes comparison of feature-level results, not matching of individual subjects. Since the cohorts contain different people, a subject ID in Cohort A must not be treated as identifying a person in Cohort B. The shared elements needed for the planned comparison are the feature definitions and clinical-score meaning. If only 108 features can be aligned reliably between assay specifications, the plan should state that 108-feature overlap rather than imply that all 120 are comparable.
This design addresses cross-cohort consistency. It does not by itself establish that the cohorts represent the same population or that any observed association is causal. If the scientific aim were to estimate a within-person relationship between metabolite and protein measurements, this horizontal design would be insufficient: it lacks the second layer and the subject-level links required for that question.
Worked Example: Unmatched Vertical Sampling for a Population-Level Aim
Question. Do the average RNA profiles and average protein profiles differ between two treatment groups in a tissue bank when the RNA and protein assays must use separate specimens? The hypothetical plan has 50 treated subjects and 50 comparison subjects.
Here the goal is a group-level comparison within each layer, not a subject-level RNA–protein relationship. If tissue availability makes the assay sets disjoint, the plan can be unmatched vertical sampling, provided each assay has subjects in both treatment groups and the intended comparisons are defined separately by layer.
| Group | RNA-assayed subjects | Protein-assayed subjects | Subject overlap |
|---|---|---|---|
| Treatment | 25 | 25 different subjects | 0 |
| Comparison | 25 | 25 different subjects | 0 |
Design diagram.
Treatment group → 25 subjects for RNA; 25 different subjects for protein
Comparison group → 25 subjects for RNA; 25 different subjects for protein
RNA subjects ∩ protein subjects = 0 within each group
The arithmetic is explicit: each group contributes \(25+25=50\) distinct subjects, so the full plan uses \(50+50=100\) subjects, not 200. There are 50 RNA profiles and 50 protein profiles, but no matched RNA–protein pairs. This design can support group comparisons within the RNA data and within the protein data if the group definitions and sampling conditions are appropriate. It cannot support a claim about whether an individual’s RNA profile corresponds to that same individual’s protein profile, because no such pairs exist.
If the biological question requires cross-layer correspondence, the design must change: obtain both assays from each subject or explicitly revise the question to a group-level aim. Calling the samples “integrated” does not create a subject link that the sampling plan did not collect.
Justifying the Required Links and Overlap
The distinction between matched and unmatched designs can be justified by writing down the observational unit required by the question. Suppose the target is a relationship between two measurements \(A_i\) and \(B_i\) for subject \(i\). The data needed are pairs \((A_i,B_i)\) with a common subject index. If the first assay is observed for subjects in set \(S_A\) and the second for subjects in set \(S_B\), then the directly available paired records are indexed by the intersection \(S_A \cap S_B\).
If \(S_A \cap S_B\) is empty, no individual-level cross-assay relationship is observed. Separate measurements from the same broad cohort do not alter this result. In contrast, a horizontal comparison of a feature \(f\) across cohorts requires that the feature have a comparable definition in both data sets; subject overlap is not the relevant requirement when the cohorts are independent. These requirements follow from the target of the question, not from a preference for one design label.
A design diagram is therefore more than a visual summary. It is an audit of whether the required units and overlaps exist. A useful diagram labels each cohort, indicates subject counts, places assays under the subjects or specimens that receive them, marks shared features, and uses explicit links for repeated or matched observations. If a line in the diagram cannot be justified by a recorded identifier and sampling plan, it should not be interpreted as a subject link.
Diagnostics, Alternatives, and Limitations
Before data collection, check the plan against the question and against practical constraints. Ask whether each group in the scientific contrast will contribute subjects to every assay required for that contrast; whether the selected features are genuinely shared across cohorts; whether time points and specimens correspond to the stated biological stage; and whether the subject identifiers distinguish people from samples and assay records. Clarify anticipated assay failures and sample exclusions. The diagram should show planned measurements and known structural absence, not conceal them as generic missing data.
- Question–design mismatch. If the question asks about within-person cross-layer association but the assays use disjoint subjects, revise the allocation or the scientific question.
- False feature overlap. Similar labels do not guarantee the same biological feature. Define the required feature correspondence before treating measurements as shared.
- Unclear sampling unit. Multiple specimens or time points from one person are not automatically independent subjects. Label the person, specimen, and occasion separately.
- Unbalanced assay allocation. A group represented in only one assay cannot support the intended group comparison in the other assay. Check group-by-assay counts explicitly.
- Planned counts mistaken for complete data. Recruitment targets do not guarantee usable profiles or matched pairs. State the expected attrition and the minimum overlap the scientific aim needs.
The main alternative designs answer different questions. Matched vertical sampling is the direct choice for subject-level relationships across layers, but it can be constrained by specimen availability or assay cost. Unmatched vertical sampling may serve distinct group-level comparisons by layer, but it gives up direct within-subject correspondence. Horizontal integration is useful for checking whether feature-level patterns recur across cohorts, but it depends on sufficiently comparable features, measurements, and populations. A mixed design can combine these aims, but it adds coordination requirements and should be diagrammed by cohort, assay, and subject link rather than described only as “multi-omics.”
No sampling plan removes all uncertainty. Recruitment and assay failures may reduce the achieved overlap; cohorts may differ in ways that matter to interpretation; and a feature panel may not capture the biology needed for the question. These are reasons to specify assumptions and limitations up front, not reasons to blur the design categories. The central planning rule is simple: collect the units and links that the intended interpretation requires.
Check Your Understanding
For each question, identify the design requirement and compare your reasoning with the solutions below.
- A study asks whether RNA and protein abundance are associated within the same patients. What sampling link is essential?
- Two independent cohorts measure 75 of the same metabolites. The goal is to assess whether feature-level results are consistent across cohorts. Is subject matching required, and what must be aligned?
- A planned study has 30 RNA-assayed subjects and 28 protein-assayed subjects, with 20 people in both sets. How many directly linkable pairs are available?
- A team wants to compare treatment groups in RNA and protein separately, but the assays use distinct subjects. Name one requirement for this design and one claim it cannot directly address.
- Why is a shared assay label alone insufficient evidence that two cohorts have comparable features?
Solutions. (1) The two assay profiles must be linked to the same patients at the relevant sampling occasion; the subject identifiers and specimen/time-point mapping must be retained. (2) No subject matching is needed for independent cohorts when the target is cross-cohort feature-level consistency. The metabolite definitions and measurement meanings must be sufficiently comparable, and the cohorts must be described so the scope of comparison is clear. (3) There are 20 directly linkable pairs, because the relevant count is the intersection of the RNA and protein subject sets, not the smaller assay count or their sum. (4) Each treatment group must be represented in each assay for separate group comparisons to be planned. The design cannot directly estimate a within-person RNA–protein relationship because the assay subjects are distinct. (5) The same label can refer to different biological entities, specimen definitions, or measurement conventions; feature correspondence must be checked and specified rather than assumed from names alone.