Why Alignment Comes Before Joint Analysis
Expression, methylation, chromatin, protein, and metabolite assays describe different molecular features, often on different sets of specimens. A combined analysis is meaningful only if its rows and columns refer to well-defined observational units. Two matrices can have the same number of rows and still be misaligned: row 12 in one file might represent a donor’s blood sample, while row 12 in another represents a different donor’s liver sample.
The goal here is to represent each assay layer explicitly and record the mappings that connect it to the others. Alignment does not mean that every layer must contain identical measurements or identical features. It means that each proposed link has a documented meaning, and that the unit being linked is suitable for the intended analysis.
Examples of layers include gene-expression measurements, methylation probes or regions, chromatin accessibility peaks, protein abundances, and metabolite abundances. A feature is the measured entity represented by a column; it is not necessarily a gene. A methylation probe, an accessibility peak, and a protein assay each have their own native feature identifiers, even if later annotations link them to the same gene.
Four Mappings to Keep Distinct
A robust data structure distinguishes the assay sample, the donor, the tissue, and the feature. These concepts are related but not interchangeable. A donor can contribute several specimens; a specimen can be assayed in several layers; a tissue can be sampled at multiple times; and multiple features can map to one annotated gene.
A variant is a separate kind of entity: it is a genetic locus and allele definition, not a molecular assay feature. When genotype data are included, the genotype matrix typically has donors as rows and variants as columns. The link from a molecular assay sample to genotype is therefore often two-step: assay sample to donor, then donor to genotype row. A variant identifier should specify the genome build and alleles, not only a gene name or an unqualified position.
Feature crosswalks may be one-to-one, one-to-many, many-to-one, or uncertain. For example, several methylation probes may be annotated to one gene, while a chromatin peak may overlap more than one gene. Keep the native feature columns and store such annotations in a separate mapping table. Collapsing them prematurely can silently change the data and the question being asked.
What Valid Sample Alignment Requires
Before placing rows from different layers side by side, define the target unit. If the target is a donor, each layer must contribute a measurement that has been selected or summarized at the donor level in a defensible way. If the target is a tissue specimen, the rows must correspond to the same specimen or to a clearly specified matched set of specimens. Matching donor identifiers alone does not establish that two measurements came from the same tissue, time point, or biological condition.
Here \(\mathcal{I}_{\ell}\) is the set of eligible identifiers for the chosen target unit in layer \(\ell\). The intersection \(\mathcal{I}_{\mathrm{shared}}\) identifies units represented in every required layer, but it does not by itself prove that the measurements are comparable. Before using it, check that identifiers truly refer to the same unit, that duplicate rows have been resolved according to a stated rule, and that tissue and collection conditions meet the analysis definition.
When all layers use the same target units but list them in different orders, row alignment is an ordering operation, not a statistical model. If \(P_\ell\) is the permutation matrix that puts layer \(\ell\)’s rows into a common order, the aligned matrix is \(P_\ell Y^{(\ell)}\). A permutation changes row order only; it does not change any measurement. This is why joining by verified identifiers is essential: matching by row position can attach the right value to the wrong person.
Worked Example: Reordering Two Layers by Verified Donor IDs
Suppose an expression matrix has the following values for gene \(G_1\), in the listed donor order, and a protein matrix has the following values for protein \(P_1\):
| Layer | Row order | Values in that order |
|---|---|---|
| Expression | D01, D02, D03 | 8, 5, 11 |
| Protein | D03, D01, D02 | 2.4, 1.8, 2.1 |
Assume the records confirm that these are the same donor-level units and that tissue and collection time are compatible. Choose the expression order, D01, D02, D03, as the common order. The required protein permutation is D01 first, D02 second, D03 third. Thus its aligned values are \(1.8, 2.1, 2.4\), not the original positional sequence \(2.4, 1.8, 2.1\).
For the third donor, the aligned pair is expression \(11\) for D03 and protein \(2.4\) for D03. The paired rows are therefore \((8,1.8)\), \((5,2.1)\), and \((11,2.4)\). The arithmetic is only a reordering: each value remains attached to its original donor. A positional join would instead pair D01 expression \(8\) with D03 protein \(2.4\), which is an incorrect link.
Variants and Tissues Need Explicit Keys
A genotype-to-molecular link requires more than a shared column name. Variant records should use a normalized key containing at least genome build, chromosome, position, reference allele, and alternate allele. Dosage must be defined consistently, typically as the number of copies of the specified alternate allele, from 0 to 2 for a diploid autosomal genotype. A strand or allele swap can reverse dosage interpretation; a position without a genome build can refer to different bases in different assemblies.
Variant harmonization across datasets is valid only after checking that the build and allele definitions agree, including strand conventions where relevant. Palindromic allele pairs can be ambiguous without reliable strand or frequency information. Do not infer that two variant columns are the same merely because their rsID or position text matches; preserve the original keys and record the harmonized key and any action taken.
Tissue is also part of the linkage definition. “Lung,” for example, may refer to bulk tissue, a particular cell type, or a particular anatomical region. Two records labelled with the same broad tissue term need not represent equivalent material. Use a controlled vocabulary or documented tissue crosswalk where possible, retain the original label, and specify whether matching is exact or based on an explicit broader grouping. For repeated collections, include time or condition in the sample mapping rather than assuming one measurement per donor.
Worked Example: Linking a Tissue Assay to a Variant
Consider a hypothetical donor D17 with two assay samples: S17-B, from blood, and S17-L, from liver. A genotype table has one row for D17. It contains variant key GRCh38:chr4:123456:A:G with alternate-allele dosage 1. Expression is measured in both tissue samples.
The valid mapping is S17-B \(\rightarrow\) D17 and S17-L \(\rightarrow\) D17 for assay sample to donor, with separate tissue labels for blood and liver. The genotype row maps to D17, and the variant column is identified by the full build-and-allele key. Thus both assay samples can be linked to D17’s genotype, but they remain two distinct tissue observations; they must not be treated as duplicate donor rows without accounting for tissue.
Suppose the variant is instead recorded in a second file as GRCh37:chr4:123456:A:G. The matching text position does not establish a valid match, because the genome builds differ. The variant must be lifted or otherwise harmonized using a documented procedure, and its alleles checked, before the genotype column can be linked. If a verified conversion maps it to GRCh38:chr4:123456:A:G with the same allele definition, the keys then agree; if it maps elsewhere or the alleles are uncertain, the link remains unresolved rather than being guessed.
Worked Alignment With Missing and Unmatched Samples
The sample intersection is often smaller than any individual layer. Keep track of this attrition and distinguish a missing assay from a measured value of zero. Missingness is not itself an observed molecular measurement, and replacing it with zero changes the data’s meaning.
Worked Example: Finding the Complete-Case Set
A hypothetical study has expression samples for donors A, B, C, and D; methylation samples for A, C, D, and E; and protein samples for A, B, C, and E. Assume all rows refer to the same tissue and collection window, and that each donor appears at most once in each layer. Find the donor set represented in every layer and the number retained.
The sets are \(\mathcal{I}_{E}=\{A,B,C,D\}\), \(\mathcal{I}_{M}=\{A,C,D,E\}\), and \(\mathcal{I}_{P}=\{A,B,C,E\}\). Their intersection is:
Two donors have all three layers. Relative to the four expression donors, the retained fraction is \(2/4=0.50\), or 50%; relative to the five donors appearing in at least one layer, it is \(2/5=0.40\), or 40%. The denominators answer different questions, so a report should identify which was used. Donor B lacks methylation, D lacks protein, and E lacks expression. Their absent values should remain missing, not be encoded as zero.
This complete-case intersection is a bookkeeping result, not a guarantee that the retained donors represent the original study population. If availability differs by tissue, phenotype, batch, or donor characteristics, alignment can select a nonrepresentative subset. Record the number excluded at each step and the reason, and preserve layer-specific data if the planned analysis can legitimately use incomplete sets. The statistical treatment of missing layers is a separate modeling question; the data organization should make the pattern visible rather than conceal it.
A Practical Alignment Audit
A reliable workflow makes the joins reproducible. Before constructing any combined table, verify that each matrix row key is unique at its stated level, each feature key is unique within its layer, and mapping tables have documented sources and targets. If several assay samples map to one donor, decide whether the analysis requires one sample per donor, a specified tissue-specific record, or repeated observations; do not silently keep the first row.
State whether a row represents a donor, specimen, aliquot, or donor-by-tissue-by-time observation.
Check uniqueness, missing IDs, duplicate mappings, and agreement between sample sheets and assay files.
Compare tissue, cell type, time, condition, and any other attributes required for the intended linkage.
Use explicit donor-to-genotype, tissue, variant, and feature crosswalks; keep native identifiers and document ambiguous links.
Join by validated keys, inspect unmatched records and missingness, and report the final unit and retained sample count.
Common Mistakes and Why the Mapping Matters
- Matching rows by their position. Equal row counts do not imply equal sample order. Join on validated identifiers and verify paired records.
- Treating donor IDs as specimen IDs. One donor can contribute multiple tissues or times. Preserve the distinct specimen rows and the donor link.
- Equating annotation with identity. A probe or peak annotated to a gene is not the same measured feature as that gene’s expression column. Retain native features and make annotations explicit.
- Using an incomplete variant key. A chromosome-position label without genome build and alleles may identify the wrong variant or allele orientation.
- Replacing missing measurements with zero. Zero is a measured value on an assay’s scale; missing means no usable recorded value. They are not interchangeable.
- Assuming shared donor means shared tissue. A genotype can be donor-level while molecular measurements are tissue-specific. The link supports donor matching, not tissue equivalence.
Good alignment is substantive, not cosmetic. It determines which biological observations are being compared and whether an apparent cross-layer pattern could be an artifact of mismatched donors, tissues, or feature definitions. A reproducible analysis should be able to trace every matrix entry to its assay sample, every sample to its biological unit, and every mapped feature or variant to a documented key.
Check Your Understanding
Answer each question using the definitions and alignment principles above.
- A donor has expression measured in blood and liver. Why are these not automatically interchangeable rows, even though both map to the same donor?
- Three assay matrices have different row orders but share the same validated specimen IDs. What should determine the row alignment?
- Why should a methylation probe annotated to a gene remain distinct from the gene’s expression feature?
- What information is needed to make a genotype variant key more reliable than chromosome and position alone?
- In a three-layer study, one donor is absent from the protein assay. Why is coding that missing protein value as zero inappropriate?