From Cohort to Molecular Question
A multi-omics case study begins before any integration algorithm is selected. The first task is to define what biological quantity is to be compared, in which tumors, and across which assay blocks. A result is only interpretable if the genomic, transcriptomic, and proteomic measurements refer to the same biological specimens and the question is fixed clearly enough to guide preprocessing and summaries.
Consider a hypothetical study of primary tumors with DNA copy-number, RNA abundance, and protein abundance measurements. A focused question might be: among tumors with an assay-defined copy-number gain at a specified locus, are the corresponding RNA and protein abundances also higher? This asks whether three measurements show a coordinated pattern in the sampled tumors. It does not, by itself, ask whether the copy-number change causes the RNA or protein change, whether the pattern defines a tumor subtype, or whether it predicts a clinical outcome.
The unit of analysis must also be explicit. A patient may contribute multiple biopsies, regions, or time points, while an assay record usually refers to a particular specimen or aliquot. If the question concerns primary tumors at diagnosis, for example, the analysis should identify the primary-tumor specimen for each patient rather than silently treating multiple samples from one patient as independent tumors. If repeated specimens are deliberately included, their dependence needs to be addressed by the analysis plan.
Establishing and Auditing the Match
A reliable match starts with a specimen manifest: one record per biological specimen, with a stable specimen identifier and the patient identifier kept in a separate field. Each assay block should have its own data dictionary, units, quality-control status, and identifier. Match on the stable specimen identifier, not on row number, patient name, or an identifier that changes between assay pipelines. Confirm that identifiers are unique at the level intended for analysis and investigate duplicates, unexpected suffixes, or inconsistent sample types before joining the blocks.
A complete-case analysis is often the clearest starting point for a question that requires all three blocks: retain specimens with usable genomic, RNA, and protein data. Report the number of specimens at each stage and why any were excluded. This makes the analysis population visible. It also avoids an invalid comparison in which, for example, genomic measurements come from one set of tumors and protein measurements from another, while the results are described as if they were paired.
Worked Example: Counting the Matched Cohort
Suppose a hypothetical study enrolled 12 primary-tumor specimens, labelled T01 through T12. All 12 pass genomic quality control. The RNA block lacks T03 after quality control, and the protein block lacks T08 and T11. Assume these identifiers refer to distinct specimens and that no other samples fail.
The complete matched set excludes the union of the missing specimens: T03, T08, and T11. Thus,
The three assay blocks therefore contribute 9 matched specimens to this particular cross-layer analysis. The 12 genomic measurements may still support a genomic-only description, and the 11 RNA measurements may support an RNA-only description, but those larger counts do not change the matched sample size. The report should state both the starting cohort and the complete matched count, with the assay-specific exclusions.
This count is valid because the missing identifiers are distinct. If the same specimen failed both RNA and protein quality control, it would be subtracted only once when calculating the complete matched count. In general, count the intersection of valid identifiers across all required blocks; do not subtract failure counts without checking their overlap.
Preparing Blocks for a Shared Interpretation
Matched identifiers do not make measurements directly comparable. A genomic copy-number estimate, an RNA abundance, and a protein abundance have different scales, measurement error, and quality criteria. Before comparing them, state how each was obtained and transformed. For example, the analysis might use a copy-number log-ratio, normalized log2 RNA abundance, and normalized log2 protein abundance. These are still distinct measurements; equal numeric values do not mean equal biological abundance.
The relevant assumptions include accurate specimen matching, sufficiently comparable tissue source and collection context, valid assay measurements, and preprocessing appropriate to each block. Batch, tumor purity, stromal or immune content, and sampling region can affect observed molecular profiles. If these features differ systematically between the groups being compared, the apparent cross-omics pattern may reflect those differences rather than the biological feature of interest. Record available covariates and inspect their distributions by group; adjustment can help only when the covariates are measured and the design contains enough overlap to separate their effects.
A simple paired association, such as a correlation between RNA and protein abundance across tumors, also has statistical requirements if an inferential test or interval is reported. The pairs must represent independent biological units for the usual sample-level inference, the measurements must be paired correctly, and the chosen association measure must match the relationship being summarized. Pearson correlation describes linear association and can be strongly affected by outliers. A small cohort can yield a large sample correlation with substantial uncertainty. A descriptive correlation alone does not require a distributional model, but conventional tests and confidence intervals for Pearson correlation rely on additional sampling assumptions, commonly independent observations and an approximately bivariate-normal population for exact small-sample inference.
Worked Example: Calculating RNA–Protein Concordance
For six hypothetical matched tumor specimens, suppose a gene’s normalized RNA abundance and the corresponding normalized protein abundance are measured on log2 scales. The values below are illustrative, not from a real study.
| Specimen | RNA, \(x\) (log2 units) | Protein, \(y\) (log2 units) |
|---|---|---|
| T01 | 5 | 4 |
| T02 | 6 | 6 |
| T03 | 7 | 5 |
| T04 | 8 | 8 |
| T05 | 9 | 7 |
| T06 | 10 | 9 |
For these six matched pairs, the sample means are \(\bar{x}=7.5\) log2 units and \(\bar{y}=6.5\) log2 units. The Pearson sample correlation is the centered cross-product divided by the square root of the two centered sums of squares:
The centered RNA values are \((-2.5,-1.5,-0.5,0.5,1.5,2.5)\), and the centered protein values are \((-2.5,-0.5,-1.5,1.5,0.5,2.5)\). Their cross-product sum is \(6.25+0.75+0.75+0.75+0.75+6.25=15.50\). Each centered sum of squares is \(17.50\). Therefore,
The six tumors show a strong positive linear association in this sample: tumors with higher RNA abundance tend also to have higher protein abundance. The statistic is unitless. It does not show that RNA abundance causes protein abundance, and six pairs are not enough to make the sample correlation a precise population estimate. A scatterplot should accompany this summary to check whether one specimen dominates the pattern or whether the relationship is markedly nonlinear.
Interpreting Evidence Across Three Layers
A useful integrated interpretation keeps each block’s evidence identifiable. State the genomic feature and its operational definition, summarize RNA and protein on their own scales, and then describe whether their directions agree across matched specimens. If a threshold defines a genomic group, choose it using assay knowledge or a pre-specified analysis plan rather than selecting the threshold that produces the most striking RNA or protein contrast. The resulting group comparison is descriptive unless an appropriate inferential analysis and its assumptions are also supplied.
Worked Example: Reading a Copy-Number, RNA, and Protein Pattern
Suppose a hypothetical six-tumor analysis uses a pre-specified copy-number log2-ratio threshold of 0.30 to label a locus as gained. The RNA and protein values are normalized log2 abundances. All measurements pass their respective quality checks, and each row refers to the same tumor specimen.
| Specimen | Copy-number log2 ratio | RNA (log2 units) | Protein (log2 units) |
|---|---|---|---|
| T01 | 0.42 | 10 | 8.0 |
| T02 | 0.38 | 9 | 7.5 |
| T03 | 0.35 | 8 | 7.0 |
| T04 | 0.05 | 7 | 6.0 |
| T05 | 0.02 | 6 | 5.5 |
| T06 | -0.08 | 5 | 5.0 |
T01–T03 exceed the specified threshold, so the gained group has \(n=3\), and the other group also has \(n=3\). The gained-group mean RNA abundance is \((10+9+8)/3=9\) log2 units; the other-group mean is \((7+6+5)/3=6\) log2 units. Their difference is \(9-6=3\) log2 units. If these are log2 abundances on a common normalized scale, a 3-unit difference corresponds to a ratio of \(2^3=8\) on the original scale.
For protein, the gained-group mean is \((8.0+7.5+7.0)/3=7.5\) log2 units, while the other-group mean is \((6.0+5.5+5.0)/3=5.5\) log2 units. The difference is \(2.0\) log2 units, corresponding to \(2^2=4\) on the original scale under the stated log2 interpretation. In this small example, the genomic gain, RNA abundance, and protein abundance show a directionally consistent pattern. The conclusion is evidence of co-occurrence in these six tumors, not proof that the gain caused the RNA and protein differences. The group sizes are small, and differences in tumor purity, batch, or other biology could contribute.
The conversion from a log2 difference to a ratio follows directly from the definition of a base-2 logarithm. If \(z=\log_2(a)\), then \(a=2^z\). For two log2 values \(z_1\) and \(z_0\), their difference is \(z_1-z_0=\log_2(a_1)-\log_2(a_0)=\log_2(a_1/a_0)\). Thus the original-scale ratio is \(a_1/a_0=2^{z_1-z_0}\). This interpretation is appropriate only when the reported values really are on comparable log2 abundance scales; it should not be applied automatically to arbitrary standardized scores or assay-specific residuals.
Checks, Alternatives, and Failure Modes
A compact analysis workflow helps prevent matching and interpretation errors:
Specify the specimen type, molecular features, comparison, and whether the unit is a patient, tumor specimen, or repeated sample.
Check uniqueness, assay-specific validity, tissue source, and the actual intersection of usable specimen identifiers.
Document transformations, units, thresholds, missingness, and relevant batch or composition concerns.
Use matched comparisons or paired plots, report the complete matched count, and describe the result at the strength supported by the design.
Complete-case analysis gives a clear paired population, but it may lose information and may be biased if assay failure is related to tumor characteristics. An analysis using every available specimen within each block can improve the precision of block-specific summaries, but those summaries no longer necessarily describe the same tumors. Report which population each result represents. Do not imply that separately estimated block patterns are paired evidence.
A single concatenated feature table is another possible strategy, but raw blocks can differ greatly in scale, number of features, and measurement reliability. A block with many variables or larger numerical variance can dominate a combined analysis unless preprocessing and weighting are justified. For a focused case study, transparent block-specific summaries and matched visualizations may be easier to audit than a complex integration result. More elaborate methods can be appropriate, but their assumptions and tuning choices do not eliminate the need to verify specimen matching and assay quality.
Diagnostics should include a specimen-flow count, identifier checks, per-block missingness and quality-control summaries, and plots of the relevant paired measurements. Inspect scatterplots for outliers, nonlinear patterns, and batch clustering; compare group sizes and available covariates; and verify that the direction of a reported contrast agrees with the underlying values. When multiple loci or proteins are explored, distinguish planned analyses from exploratory findings and account for multiplicity in formal claims. A small, selected case study is often best presented as hypothesis-generating evidence.
- Mismatched rows: “The first RNA row and first protein row belong together because the tables have the same length.” Row order is not evidence of identity; join using audited specimen identifiers.
- Unmatched denominators: “Each assay has many observations, so the cross-omics result has the largest assay’s sample size.” A paired result can use only specimens valid in every required block.
- Overstated mechanism: “A copy-number gain accompanies higher RNA and protein, so it caused both.” The observed alignment is an association; confounding, measurement effects, and other molecular mechanisms remain possible.
- Scale confusion: “A two-unit protein difference means twice as much protein.” On a log2 scale, a two-unit difference corresponds to a fourfold ratio; on another scale the conversion may differ.
Check Your Understanding
Use the definitions and worked calculations above to answer these questions. Short solutions follow the questions.
- A cohort has 20 specimens. Two fail RNA quality control, and three fail protein quality control; one specimen is in both failure groups. How many complete RNA–protein pairs remain?
- Why is joining an RNA table and protein table by row number unsafe, even if both tables contain the same number of rows?
- If the mean RNA difference between two groups is 1.5 log2 units, what original-scale ratio does that represent, assuming comparable log2 abundance measurements?
- A matched cohort shows positive RNA–protein correlation for a gene. Name one conclusion supported by this result and one conclusion it does not establish.
- Why can a complete-case analysis and an analysis using all available measurements describe different populations?
Solutions. (1) Four specimens fail at least one block: \(2+3-1=4\), so \(20-4=16\) complete pairs remain. (2) Row order may differ across assay pipelines; only verified specimen identifiers establish which records refer to the same biological sample. (3) The ratio is \(2^{1.5}\approx2.828\), or about 2.83-fold. (4) It supports a positive association across the observed matched tumors; it does not establish that RNA abundance causes protein abundance. (5) Complete cases require valid measurements in every required block, whereas available-case summaries may include different specimens in each block, especially when assay failures are related to specimen characteristics.