From a Diagram to a Design
An integration diagram is useful only when its symbols correspond to the actual study units. A box may represent an assay, a cohort, or a data block; a line between boxes may represent shared subjects, a common feature set, or simply an analytical connection. Those meanings are not interchangeable. Reading a design therefore begins by identifying what each box and connection denotes, then checking the underlying subject and feature identifiers.
The key distinction is between the subject axis and the feature axis. The subject axis asks which biological units were measured in each block and whether the same units appear in multiple blocks. The feature axis asks whether blocks contain the same measured variables or different omics features. Horizontal versus vertical integration describes the relationship between these axes; matched versus unmatched describes the relationship between subjects.
A sample is not automatically identical to a subject: a subject can contribute more than one sample or assay specimen. When a diagram names only “samples,” the subject relationship cannot be inferred unless the design provides a sample-to-subject mapping. Likewise, two columns with similar labels are not necessarily the same feature; feature IDs and annotation determine whether their measurements refer to the same entity.
These are two different classification questions. Horizontal or vertical describes how the blocks relate through their features and sample groups. Matched, unmatched, or partially matched describes subject overlap. A diagram should not be classified as matched merely because boxes are drawn side by side, nor as vertical merely because it contains several blocks.
A Reliable Reading Procedure
For each block, determine whether each row is a subject, a sample, or another unit. Record the mapping from sample IDs to subject IDs where available.
Compare feature IDs and annotations across blocks. A shared feature space supports cohort-wise combination; distinct feature sets suggest integration across layers.
Use the subject IDs to find \(S_a\cap S_b\), the subjects represented in both blocks. No overlap means unmatched; complete correspondence means matched; a nonempty but incomplete overlap means partially matched.
Name the feature relationship and the subject relationship separately. If IDs or feature definitions are missing, state that the diagram does not determine the classification.
For more than two blocks, make the comparison pairwise or use a subject-by-block incidence table. An entry of 1 indicates that a subject appears in a block; 0 indicates absence. The table makes clear whether the design has complete matching across all blocks or only selected overlaps.
Worked Example: Two Cohorts Measured on One Panel
A hypothetical study combines two cohorts measured with the same four-feature assay panel. Cohort A has sample IDs A01, A02, and A03, mapped to subjects A01, A02, and A03. Cohort B has sample IDs B01 and B02, mapped to subjects B01 and B02. The feature IDs are \(f_1,f_2,f_3,f_4\) in both blocks. How should the design be classified?
Each block has four feature columns, and the feature IDs agree, so the blocks are in a common feature space. The row counts are \(n_A=3\) and \(n_B=2\). Their subject sets are \(S_A=\{A01,A02,A03\}\) and \(S_B=\{B01,B02\}\), with \(S_A\cap S_B=\varnothing\). Therefore the subject groups are unmatched.
The combined cohort table has \(3+2=5\) rows and four aligned feature columns, so its dimensions are \(5\times4\). This is horizontal integration across cohorts and is unmatched with respect to subjects. The arithmetic describes the table dimensions; it does not itself establish that the cohorts are comparable in every scientific respect.
Reading Block Shapes and Subject Correspondence
A block-shaped table helps reveal the feature relationship, but dimensions alone are not enough. Two blocks can both have four columns while measuring different four-feature panels. Conversely, two blocks may have different numbers of columns yet share a subset of features. Read the feature names, not just \(p_k\). Similarly, equal row counts do not establish matching: row \(i\) in one block corresponds to row \(i\) in another only if subject or sample identifiers justify that correspondence.
Worked Example: Distinct Omics Layers on Matched Subjects
In a hypothetical study, an RNA block has three subjects, S1, S2, and S3, and two features, \(r_1,r_2\). A protein block has those same three subjects and three different features, \(q_1,q_2,q_3\). Both blocks have one sample per subject, and their subject IDs agree.
The RNA block has dimensions \(3\times2\), and the protein block has dimensions \(3\times3\). Their feature sets are distinct, while their subject sets are equal: \(S_{\text{RNA}}=S_{\text{protein}}=\{S1,S2,S3\}\). The design is therefore vertical and matched.
If the feature columns are joined after aligning rows by subject ID, the resulting table has three subject rows and \(2+3=5\) feature columns, or dimensions \(3\times5\). The row alignment is justified here because the IDs establish the correspondence. The same conclusion would not follow from the fact that both blocks happen to have three rows.
The alignment requirement can be stated algebraically. Suppose two blocks have the same \(n\) subjects but list them in different row orders. Let \(P\) be a permutation matrix that reorders the second block to the first block's subject order. Then the valid column-wise combination is \([X^{(1)},PX^{(2)}]\), not \([X^{(1)},X^{(2)}]\) unless the row orders already agree. This follows because each row of a combined table must describe one subject; without the permutation, measurements from different subjects can be placed in the same row.
The corresponding logic for combining cohorts by rows is different: the feature columns must first be in the same order and represent the same harmonized features. If two blocks have different feature sets, stacking their rows without resolving the columns would assign incompatible measurements to the same column. Thus, a block operation is valid only when the axis being joined has been aligned by identifiers and the other axis has a defensible shared meaning.
Worked Example: Distinct Layers With Partial Subject Overlap
A hypothetical metabolite block contains subjects S1, S2, and S3 with features \(m_1,m_2\). A lipid block contains subjects S2, S3, and S4 with features \(l_1,l_2,l_3\). The subject IDs are verified, and the feature IDs identify distinct measurement sets.
The feature sets differ, indicating vertical integration. The subject intersection is \(S_{\text{metabolite}}\cap S_{\text{lipid}}=\{S2,S3\}\), while the union is \(\{S1,S2,S3,S4\}\). There are two shared subjects out of four represented overall, so the design is partially matched, not fully matched and not unmatched.
A complete-case table containing only subjects measured in both layers would have \(2\) rows and \(2+3=5\) feature columns, hence dimensions \(2\times5\), assuming those five features are retained. The original blocks each contain three rows; concatenating them as though all three rows corresponded would be incorrect because S1 and S4 have no counterpart in the other block. The diagram should therefore show the shared IDs or explicitly mark the unmatched subjects.
Diagnostics, Alternatives, and Limits
A diagram or table is sufficiently informative for classification when it specifies the row unit, subject mapping, feature identity, and cross-block correspondence. Several checks help expose designs that are unclear or inconsistent:
- Check identifier uniqueness and meaning. Confirm that the displayed IDs are subject IDs or that sample IDs map to subject IDs. Do not infer identity from similar labels or row positions.
- Check feature harmonization. Confirm that columns called the same feature refer to the same entity and measurement definition. A common column count is not evidence of a common feature space.
- Check the diagram against the tables. The number of rows and features shown should agree with the block dimensions and the listed IDs. An apparent full match may actually be partial if some subjects are absent.
- Mark uncertainty rather than guessing. If a diagram says only “same cohort” but gives no sample-to-subject mapping, it may not establish that the same individuals were measured in each assay.
A useful alternative to a schematic diagram is a block inventory: one row per assay block, with columns for sample IDs, subject IDs, feature IDs, and row and column counts. For complicated overlap, a subject-by-block incidence table is more explicit than a Venn diagram. A schematic is often better for communicating the overall design; the inventory and incidence table are better for verifying exact correspondence. These tools complement rather than replace one another.
| Evidence in the design | Feature relationship | Subject relationship | Likely classification |
|---|---|---|---|
| Same harmonized panel; separate cohort IDs | Common feature space | No shared subjects | Horizontal, unmatched |
| Different omics feature sets; same verified IDs | Distinct feature blocks | Same subjects | Vertical, matched |
| Different feature sets; some verified IDs overlap | Distinct feature blocks | Partial overlap | Vertical, partially matched |
| Feature labels or subject mappings absent | Not established | Not established | Insufficient information |
These classifications describe design structure, not data quality or scientific validity. A matched design can still have missing measurements, and an unmatched design can still be useful for appropriate questions. The introductory material on matched and unmatched omics sampling and on horizontal and vertical integration gives the broader design context; here, the practical task is to read the evidence in the diagram and block tables accurately. Detailed consequences of incorrectly duplicating or matching subjects require separate attention.
Check Your Understanding
For each question, identify the design features from the stated evidence. Solutions follow so that the classifications can be checked.
- Two cohorts each have six subjects measured on the same annotated 100-feature panel. Their subject IDs are disjoint. Classify the design by feature relationship and matching status.
- Two blocks have five rows each and different feature sets. No subject IDs or mapping are provided. Can you conclude that the blocks are matched?
- Block A contains subjects P1, P2, P3, and P4; block B contains P3, P4, P5. Both measure different omics layers. Classify the design.
- Two matched blocks contain the same four subjects and have two and seven distinct features, respectively. What are the dimensions after valid column-wise alignment?
- Two blocks share a feature panel, and one has subjects T1, T2 while the other has T2, T3. Is row-wise combination justified without further work? Explain.
Solutions. (1) The feature space is common and the cohorts have no shared subjects: horizontal, unmatched integration. (2) No. Equal row counts do not establish identity; without subject correspondence the matching status is undetermined. (3) The feature sets are distinct, so the design is vertical; the intersection is \(\{P3,P4\}\), so it is partially matched. (4) Align rows using the four verified subject IDs. The result has four rows and \(2+7=9\) columns, or \(4\times9\). (5) Not as a single set of matched rows without further work. The blocks have a common feature panel, but their subject sets overlap only at T2. The design is horizontal in feature relationship and partially matched in subject relationship; row-binding requires consistent feature columns and careful retention of the distinct subject rows.