How do you translate a MuData H5MU multi-omics dataset guide without confusing RNA, protein, chromatin accessibility or the cells that connect them? MuData stores several annotated assay objects in one HDF5-based structure. The central challenge is keeping each modality’s features and measurements distinct while preserving their relationship to shared observations and variables.
Searches for MuData H5MU format explained, mudata mod obs var, MuData obsmap versus varmap, RNA ATAC protein modalities, MuData missing modalities, H5MU versus H5AD, multiome metadata localization and single-cell multimodal data conversion all raise the same question: when does a visible label mean a data key, an assay-specific measurement, a shared annotation or a later inference? Clear translation needs to answer that before replacing any term.
This educational guide is for multi-omics learners, bioinformatics teams, dataset stewards and scientific translators worldwide. It focuses on linked modalities and identity maps rather than repeating how AnnData stores one matrix or how HDF5 groups work in general. Examples use synthetic observations, and real donor-linked data should follow applicable access, consent and privacy rules.
The short explanation
Imagine one study that measures RNA for some cells and protein for some of the same cells. Each measurement has its own features, numbers and metadata. MuData helps store these together and express how observations and variables relate. A translated guide is trustworthy when it preserves the actual links, including missing pairs, instead of pretending the datasets are one undifferentiated table.
A useful reading sequence
First list the modalities in the .mod collection. Next identify what each modality calls an observation and a variable. Then inspect shared metadata and the relationships used to coordinate the local indices. Finally examine processing stages, inferred annotations and any joint embeddings. This sequence prevents the common error of assigning a measurement from one assay to another.
The question MuData was built to answer
MuData collects multiple annotated datasets into one coordinated structure, allowing different modalities to be interpreted together. To interpret this correctly, locate the boundary between a modality’s own data and the shared coordination layer. The file’s usefulness depends on keeping both levels explicit rather than flattening them into one generic table.
Calling a multi-omics object a single gene-expression matrix obscures its separate assays. RNA and protein observations can occupy the same container without being the same measurement. The mistake is not simply a poor translation choice. It changes which observations or features the reader believes were measured and can therefore alter scientific comparisons.
List modalities and the biological units they measure before writing a summary. Make a tiny RNA-plus-protein example before reviewing a large atlas. Trace a single observation and feature through both levels and note what happens when one assay is absent.
H5MU is an HDF5-based container
The .h5mu format uses an HDF5 hierarchy designed to store MuData objects and nested AnnData-like modalities. The underlying concept is alignment: measurements from different assays become comparable only through justified links. A label may be understandable yet still be unusable as a join key.
Treating the extension as a flat CSV or interchange text document leads to incorrect handling instructions. A binary H5MU container can hold multiple matrices and linked metadata without being one large table. Beware of explaining missingness as though it were a measured value. In multimodal analysis, absence from a modality can have a completely different cause from a zero recorded within that modality.
Inspect the hierarchy with an HDF5-aware tool and document reader versions. Ask someone else to follow the mapped identifiers with no access to your assumptions. If they cannot reproduce the same membership, explain the data model more precisely.
The /mod group distinguishes modalities
Named groups under /mod hold separate modality-specific data objects. An effective article separates the source assay, the container hierarchy and the analyst’s interpretation. Each provides different evidence and each can be mistranslated independently.
Renaming /mod/rna to translated prose may break software expecting the canonical modality key. The string rna may be a practical program key even when a narrative says transcript measurements. Smooth prose can create false certainty when it gives every feature the same biological unit or treats algorithmic groups as observed biological categories.
Translate a display description separately and preserve operational names when they are used as keys. Inspect mapping arrays and source dimensions rather than relying on chart appearance. A familiar-looking plot cannot substitute for a verified row-to-identity relationship.
Each modality has its own measured variables
An RNA modality can have genes while an ATAC modality can have genomic peaks and a protein modality can have antibody features. The reader should understand both what stays fixed and what may be reworded. Stable keys make the object interoperable; explanatory phrases make it accessible across languages.
Treating every modality’s variables as genes falsifies what the experiment recorded. One assay’s variable may be a genomic interval rather than a gene symbol. When global and modality-local metadata share similar labels, the scope can disappear in translation. A statement that is true for one modality should not silently become a claim about the whole experiment.
Inspect each modality’s var and measurement descriptions before labeling its feature axis. State the expected modality, observation scope and feature type alongside your check. If a relationship changes, repair the mapping or prose rather than editing measured values.
Top-level observations form a joint view
The MuData object may expose a global obs view for observations linked across constituent modalities. To interpret this correctly, locate the boundary between a modality’s own data and the shared coordination layer. The file’s usefulness depends on keeping both levels explicit rather than flattening them into one generic table.
Assuming all top-level observations occur in every modality can distort comparisons. A cell present in RNA results may be absent from a protein-specific matrix. The mistake is not simply a poor translation choice. It changes which observations or features the reader believes were measured and can therefore alter scientific comparisons.
Study how global observation identifiers map to each modality. Make a tiny RNA-plus-protein example before reviewing a large atlas. Trace a single observation and feature through both levels and note what happens when one assay is absent.
Top-level variables form a joint view
Global var information collates feature annotations across modalities in the multi-omics object. The underlying concept is alignment: measurements from different assays become comparable only through justified links. A label may be understandable yet still be unusable as a join key.
Presenting the global variable list as one universal count matrix may misrepresent the data model. An ATAC peak and an RNA gene can both appear in global metadata but remain different feature types. Beware of explaining missingness as though it were a measured value. In multimodal analysis, absence from a modality can have a completely different cause from a zero recorded within that modality.
Relate variable entries to the modality from which they arise. Ask someone else to follow the mapped identifiers with no access to your assumptions. If they cannot reproduce the same membership, explain the data model more precisely.
The .mod collection stores AnnData objects
Each modality is represented by an annotated object with its own data matrix and associated metadata. An effective article separates the source assay, the container hierarchy and the analyst’s interpretation. Each provides different evidence and each can be mistranslated independently.
The phrase one matrix for all modalities may suggest a false shared unit or orientation. A protein count matrix and an RNA count matrix need not have matching feature counts. Smooth prose can create false certainty when it gives every feature the same biological unit or treats algorithmic groups as observed biological categories.
Explain modality-local X, obs and var independently from top-level attributes. Inspect mapping arrays and source dimensions rather than relying on chart appearance. A familiar-looking plot cannot substitute for a verified row-to-identity relationship.
Observation identifiers are joining keys
Observation names and modality mappings provide alignment across data collected from cells or samples. The reader should understand both what stays fixed and what may be reworded. Stable keys make the object interoperable; explanatory phrases make it accessible across languages.
Translating a cell barcode or trimming its prefix can break links between measurements. A cell barcode is not a sentence needing linguistic replacement. When global and modality-local metadata share similar labels, the scope can disappear in translation. A statement that is true for one modality should not silently become a claim about the whole experiment.
Keep original observation IDs intact and use explicit reversible aliases for display. State the expected modality, observation scope and feature type alongside your check. If a relationship changes, repair the mapping or prose rather than editing measured values.
Variable identifiers need modality context
Feature names may repeat across modalities or refer to distinct molecular entities. To interpret this correctly, locate the boundary between a modality’s own data and the shared coordination layer. The file’s usefulness depends on keeping both levels explicit rather than flattening them into one generic table.
Deduplicating variables solely because their printed names match can merge different assays. An antibody target label matching a gene symbol does not make the measured values equivalent. The mistake is not simply a poor translation choice. It changes which observations or features the reader believes were measured and can therefore alter scientific comparisons.
Use modality-qualified feature identity in joins and documentation. Make a tiny RNA-plus-protein example before reviewing a large atlas. Trace a single observation and feature through both levels and note what happens when one assay is absent.
The obsmap relationship
MuData includes observation mappings that relate modality-specific observations to global observations. The underlying concept is alignment: measurements from different assays become comparable only through justified links. A label may be understandable yet still be unusable as a join key.
An absent mapping should not be converted into a biological zero or a missing cell by assumption. A shared ID must be traced to the correct row in each modality. Beware of explaining missingness as though it were a measured value. In multimodal analysis, absence from a modality can have a completely different cause from a zero recorded within that modality.
Check membership and mapping rules before assembling joint tables. Ask someone else to follow the mapped identifiers with no access to your assumptions. If they cannot reproduce the same membership, explain the data model more precisely.
The varmap relationship
MuData variable mappings help relate modality features to the global variable index. An effective article separates the source assay, the container hierarchy and the analyst’s interpretation. Each provides different evidence and each can be mistranslated independently.
Reordering feature labels without updating the mapping detaches values from their annotations. The same displayed feature name can require distinct local positions. Smooth prose can create false certainty when it gives every feature the same biological unit or treats algorithmic groups as observed biological categories.
Validate local-to-global variable relationships after any transformation. Inspect mapping arrays and source dimensions rather than relying on chart appearance. A familiar-looking plot cannot substitute for a verified row-to-identity relationship.
Mappings are not translated labels
An integer or positional map expresses data alignment rather than human-facing vocabulary. The reader should understand both what stays fixed and what may be reworded. Stable keys make the object interoperable; explanatory phrases make it accessible across languages.
Replacing mapping values with translated descriptions destroys the purpose of the field. A bilingual report can describe membership without modifying mapping arrays. When global and modality-local metadata share similar labels, the scope can disappear in translation. A statement that is true for one modality should not silently become a claim about the whole experiment.
Preserve map types and indices; explain their semantics outside the stored mapping. State the expected modality, observation scope and feature type alongside your check. If a relationship changes, repair the mapping or prose rather than editing measured values.
Modalities can have unequal cell sets
Multi-omics datasets are not always perfectly paired across all assay types. To interpret this correctly, locate the boundary between a modality’s own data and the shared coordination layer. The file’s usefulness depends on keeping both levels explicit rather than flattening them into one generic table.
Assuming a complete RNA-protein-ATAC triplet for every observation invents measurements. A cell seen only in RNA should not receive fabricated protein counts. The mistake is not simply a poor translation choice. It changes which observations or features the reader believes were measured and can therefore alter scientific comparisons.
Count observations per modality and state their intersection or union when known. Make a tiny RNA-plus-protein example before reviewing a large atlas. Trace a single observation and feature through both levels and note what happens when one assay is absent.
Missing modality and zero signal differ
No assay measurement for an observation is not the same statement as a recorded zero signal. The underlying concept is alignment: measurements from different assays become comparable only through justified links. A label may be understandable yet still be unusable as a join key.
A missing protein modality wrongly rendered as zero protein abundance changes evidence. No protein library for a sample means no measurement, not proof of absent protein. Beware of explaining missingness as though it were a measured value. In multimodal analysis, absence from a modality can have a completely different cause from a zero recorded within that modality.
Check whether an observation is present in that modality before interpreting numerical values. Ask someone else to follow the mapped identifiers with no access to your assumptions. If they cannot reproduce the same membership, explain the data model more precisely.
An RNA layer is not an ATAC layer
A layer inside one modality is a parallel representation of that modality’s matrix. An effective article separates the source assay, the container hierarchy and the analyst’s interpretation. Each provides different evidence and each can be mistranslated independently.
Treating modality-specific layers as separate global assays confuses two axes of organization. RNA raw counts and normalized RNA may be layers, while ATAC is another modality. Smooth prose can create false certainty when it gives every feature the same biological unit or treats algorithmic groups as observed biological categories.
Describe the role of modality and layer independently. Inspect mapping arrays and source dimensions rather than relying on chart appearance. A familiar-looking plot cannot substitute for a verified row-to-identity relationship.
Top-level obs metadata has a scope
Global observation annotations can describe cells or samples across the multi-omics object. The reader should understand both what stays fixed and what may be reworded. Stable keys make the object interoperable; explanatory phrases make it accessible across languages.
A global field should not be called an RNA-specific measurement unless its origin supports that claim. A donor field may attach to global observations even though no donor property was measured by RNA. When global and modality-local metadata share similar labels, the scope can disappear in translation. A statement that is true for one modality should not silently become a claim about the whole experiment.
Identify where the annotation came from and which observations it covers. State the expected modality, observation scope and feature type alongside your check. If a relationship changes, repair the mapping or prose rather than editing measured values.
Modality-local obs metadata has another scope
An individual modality can store its own observation annotations and quality-control measures. To interpret this correctly, locate the boundary between a modality’s own data and the shared coordination layer. The file’s usefulness depends on keeping both levels explicit rather than flattening them into one generic table.
Copying an assay-specific metric into a universal cell-quality claim is misleading. RNA mitochondrial fraction cannot be substituted for protein library quality. The mistake is not simply a poor translation choice. It changes which observations or features the reader believes were measured and can therefore alter scientific comparisons.
Keep quality-control fields linked to the relevant assay and pipeline. Make a tiny RNA-plus-protein example before reviewing a large atlas. Trace a single observation and feature through both levels and note what happens when one assay is absent.
Modality-local var metadata
A modality’s feature metadata records annotations relevant to that particular assay. The underlying concept is alignment: measurements from different assays become comparable only through justified links. A label may be understandable yet still be unusable as a join key.
Renaming a feature class across modalities can erase the distinction between genomic intervals and protein targets. A genomic peak’s chromosome position is not a protein’s antibody identifier. Beware of explaining missingness as though it were a measured value. In multimodal analysis, absence from a modality can have a completely different cause from a zero recorded within that modality.
Use an assay-aware glossary with stable identifiers and units. Ask someone else to follow the mapped identifiers with no access to your assumptions. If they cannot reproduce the same membership, explain the data model more precisely.
Shared embeddings need provenance
Global obsm data can hold observation-aligned multidimensional representations, sometimes produced by integration. An effective article separates the source assay, the container hierarchy and the analyst’s interpretation. Each provides different evidence and each can be mistranslated independently.
Calling integrated coordinates raw physical locations overstates their evidence. Two cells close in an embedding were positioned by an algorithm, not necessarily side by side in tissue. Smooth prose can create false certainty when it gives every feature the same biological unit or treats algorithmic groups as observed biological categories.
Record the integration method and how observations were included. Inspect mapping arrays and source dimensions rather than relying on chart appearance. A familiar-looking plot cannot substitute for a verified row-to-identity relationship.
Modality-specific embeddings remain distinct
Each modality can carry its own dimensionality-reduction outputs with different feature spaces. The reader should understand both what stays fixed and what may be reworded. Stable keys make the object interoperable; explanatory phrases make it accessible across languages.
Treating RNA UMAP and ATAC UMAP values as interchangeable loses method context. Two plots with axes named UMAP1 and UMAP2 need not share the same geometry. When global and modality-local metadata share similar labels, the scope can disappear in translation. A statement that is true for one modality should not silently become a claim about the whole experiment.
Label each embedding with its modality and processing origin. State the expected modality, observation scope and feature type alongside your check. If a relationship changes, repair the mapping or prose rather than editing measured values.
Global unstructured metadata
MuData includes unstructured information that can document processing, experimental choices or additional annotations. To interpret this correctly, locate the boundary between a modality’s own data and the shared coordination layer. The file’s usefulness depends on keeping both levels explicit rather than flattening them into one generic table.
Translating software keys within uns can break compatibility even if values appear readable. A tool configuration string may need preservation while a narrative description can be localized. The mistake is not simply a poor translation choice. It changes which observations or features the reader believes were measured and can therefore alter scientific comparisons.
Distinguish free-text notes from keys referenced by software. Make a tiny RNA-plus-protein example before reviewing a large atlas. Trace a single observation and feature through both levels and note what happens when one assay is absent.
Modality order is not arbitrary decoration
The on-disk specification discusses a modality-order attribute used to retain the intended ordering. The underlying concept is alignment: measurements from different assays become comparable only through justified links. A label may be understandable yet still be unusable as a join key.
Alphabetically displaying modalities while implying that was the original order can mislead a reader. An RNA-protein heading order can be a design choice, not a change to measurement values. Beware of explaining missingness as though it were a measured value. In multimodal analysis, absence from a modality can have a completely different cause from a zero recorded within that modality.
Record the stored modality order and explain when views deliberately reorder for presentation. Ask someone else to follow the mapped identifiers with no access to your assumptions. If they cannot reproduce the same membership, explain the data model more precisely.
Cell IDs can collide across experiments
Barcodes and local observation IDs may be reused across libraries or studies. An effective article separates the source assay, the container hierarchy and the analyst’s interpretation. Each provides different evidence and each can be mistranslated independently.
Removing sample prefixes or library context can merge unrelated observations. The same barcode sequence in two libraries does not by itself identify one physical cell. Smooth prose can create false certainty when it gives every feature the same biological unit or treats algorithmic groups as observed biological categories.
Use globally unambiguous identifiers before cross-study joins. Inspect mapping arrays and source dimensions rather than relying on chart appearance. A familiar-looking plot cannot substitute for a verified row-to-identity relationship.
The number of variables differs by modality
RNA genes, ATAC regions and protein features often have very different feature counts. The reader should understand both what stays fixed and what may be reworded. Stable keys make the object interoperable; explanatory phrases make it accessible across languages.
A single feature count given without naming the modality can make the dataset seem inconsistent. Ten proteins and twenty thousand genes are not contradictory feature dimensions. When global and modality-local metadata share similar labels, the scope can disappear in translation. A statement that is true for one modality should not silently become a claim about the whole experiment.
Report dimensions separately for each modality and distinguish global metadata counts. State the expected modality, observation scope and feature type alongside your check. If a relationship changes, repair the mapping or prose rather than editing measured values.
Sample and cell are different observation units
Some multi-omics experiments connect modalities at single-cell level while others represent different sample structures. To interpret this correctly, locate the boundary between a modality’s own data and the shared coordination layer. The file’s usefulness depends on keeping both levels explicit rather than flattening them into one generic table.
Calling every sample-level measurement a matched single-cell value invents pairing. Two assays from the same tissue are not automatically measurements of the same individual cells. The mistake is not simply a poor translation choice. It changes which observations or features the reader believes were measured and can therefore alter scientific comparisons.
State experimental unit and pairing method in the article and figure captions. Make a tiny RNA-plus-protein example before reviewing a large atlas. Trace a single observation and feature through both levels and note what happens when one assay is absent.
Pipeline-derived labels need their source
Cell types, clusters and assay quality annotations may come from downstream algorithms. The underlying concept is alignment: measurements from different assays become comparable only through justified links. A label may be understandable yet still be unusable as a join key.
Treating such labels as primary experimental observations creates false certainty. An inferred label can be helpful while remaining subject to validation. Beware of explaining missingness as though it were a measured value. In multimodal analysis, absence from a modality can have a completely different cause from a zero recorded within that modality.
Differentiate captured measurements, computational inference and manual curation. Ask someone else to follow the mapped identifiers with no access to your assumptions. If they cannot reproduce the same membership, explain the data model more precisely.
Data types and scale cannot be unified by prose
RNA counts, chromatin accessibility scores and protein-derived signals can have different units and transformations. An effective article separates the source assay, the container hierarchy and the analyst’s interpretation. Each provides different evidence and each can be mistranslated independently.
Combining them under a single expression value heading can erase analytical assumptions. A zero in accessibility data and a zero in RNA counts should not be declared equivalent. Smooth prose can create false certainty when it gives every feature the same biological unit or treats algorithmic groups as observed biological categories.
Document unit, normalization and transformation per modality. Inspect mapping arrays and source dimensions rather than relying on chart appearance. A familiar-looking plot cannot substitute for a verified row-to-identity relationship.
Cross-modal integration is not a file guarantee
A valid H5MU container can hold multiple modalities without proving that they have been biologically integrated. The reader should understand both what stays fixed and what may be reworded. Stable keys make the object interoperable; explanatory phrases make it accessible across languages.
Calling stored-together datasets integrated without method evidence exaggerates the analysis. Sharing a file is not the same as sharing a learned latent representation. When global and modality-local metadata share similar labels, the scope can disappear in translation. A statement that is true for one modality should not silently become a claim about the whole experiment.
Look for an actual integration method and output before making that claim. State the expected modality, observation scope and feature type alongside your check. If a relationship changes, repair the mapping or prose rather than editing measured values.
Version and read/write compatibility
H5MU behavior depends on its specification and the library versions that produced and consume it. To interpret this correctly, locate the boundary between a modality’s own data and the shared coordination layer. The file’s usefulness depends on keeping both levels explicit rather than flattening them into one generic table.
A file that fails in one older reader is not automatically scientifically corrupt. A reader limitation may require a documented conversion rather than manual editing of groups. The mistake is not simply a poor translation choice. It changes which observations or features the reader believes were measured and can therefore alter scientific comparisons.
Record software versions and test readback with a compatible implementation. Make a tiny RNA-plus-protein example before reviewing a large atlas. Trace a single observation and feature through both levels and note what happens when one assay is absent.
Round-tripping an H5MU object
A reader should be able to save and reopen a multi-omics object without losing modality membership or key annotations. The underlying concept is alignment: measurements from different assays become comparable only through justified links. A label may be understandable yet still be unusable as a join key.
A conversion that retains matrices but drops maps can silently break downstream joins. The same total observation count is not enough if local-to-global alignment changed. Beware of explaining missingness as though it were a measured value. In multimodal analysis, absence from a modality can have a completely different cause from a zero recorded within that modality.
Compare modality names, dimensions, observation mappings and variable mappings before and after. Ask someone else to follow the mapped identifiers with no access to your assumptions. If they cannot reproduce the same membership, explain the data model more precisely.
Writing a bilingual data dictionary
A bilingual glossary can explain obs, var, mod, obsmap and varmap while retaining exact software keys. An effective article separates the source assay, the container hierarchy and the analyst’s interpretation. Each provides different evidence and each can be mistranslated independently.
Renaming these keys inside the file for language consistency may make it unreadable. A reader can learn a local term for observation without changing the obs key. Smooth prose can create false certainty when it gives every feature the same biological unit or treats algorithmic groups as observed biological categories.
Keep separate columns for canonical key, reader-facing translation and scientific meaning. Inspect mapping arrays and source dimensions rather than relying on chart appearance. A familiar-looking plot cannot substitute for a verified row-to-identity relationship.
Privacy and donor information
Joint multimodal data can link molecular profiles to donors and clinical annotations. The reader should understand both what stays fixed and what may be reworded. Stable keys make the object interoperable; explanatory phrases make it accessible across languages.
A localization exercise does not grant permission to publish patient-associated data. A unique donor label may remain identifying when linked across modalities. When global and modality-local metadata share similar labels, the scope can disappear in translation. A statement that is true for one modality should not silently become a claim about the whole experiment.
Use synthetic, appropriately licensed or de-identified examples within consent constraints. State the expected modality, observation scope and feature type alongside your check. If a relationship changes, repair the mapping or prose rather than editing measured values.
A small mismatched-cell teaching case
A tiny two-modality dataset with unequal cell lists can teach why maps matter. To interpret this correctly, locate the boundary between a modality’s own data and the shared coordination layer. The file’s usefulness depends on keeping both levels explicit rather than flattening them into one generic table.
Filling unmatched cells with zero to make rectangular tables look complete creates false measurements. Three RNA observations and two protein observations need an explicit alignment map. The mistake is not simply a poor translation choice. It changes which observations or features the reader believes were measured and can therefore alter scientific comparisons.
Trace which observation exists in which assay before plotting or exporting. Make a tiny RNA-plus-protein example before reviewing a large atlas. Trace a single observation and feature through both levels and note what happens when one assay is absent.
Distinguishing neighboring formats
H5MU solves a multi-modality linking problem; H5AD represents an individual annotated object, while Loom has its own matrix conventions. The underlying concept is alignment: measurements from different assays become comparable only through justified links. A label may be understandable yet still be unusable as a join key.
Explaining all HDF5-based omics files as if they shared one schema leads to incompatible tutorials. The .mod group is an important H5MU distinction rather than a generic HDF5 feature. Beware of explaining missingness as though it were a measured value. In multimodal analysis, absence from a modality can have a completely different cause from a zero recorded within that modality.
Keep format names, top-level hierarchy and axis semantics separate. Ask someone else to follow the mapped identifiers with no access to your assumptions. If they cannot reproduce the same membership, explain the data model more precisely.
Checking labels after localization
A localized multi-omics guide should preserve every stable observation and variable identity and maintain modality membership. An effective article separates the source assay, the container hierarchy and the analyst’s interpretation. Each provides different evidence and each can be mistranslated independently.
An attractive translated figure can conceal a broken join or false assertion about paired measurements. A reader should recover the same matched-cell subset from both language versions. Smooth prose can create false certainty when it gives every feature the same biological unit or treats algorithmic groups as observed biological categories.
Verify key sets, assay dimensions, category legends and sample pairing using a synthetic comparison. Inspect mapping arrays and source dimensions rather than relying on chart appearance. A familiar-looking plot cannot substitute for a verified row-to-identity relationship.
A practical quality-control handoff
A research team needs to know which labels are translated, which identifiers are fixed and which claims depend on analysis choices. The reader should understand both what stays fixed and what may be reworded. Stable keys make the object interoperable; explanatory phrases make it accessible across languages.
A handoff that says translated all metadata without classification is not reproducible. An analyst can then evaluate scientific claims without guessing what the localization changed. When global and modality-local metadata share similar labels, the scope can disappear in translation. A statement that is true for one modality should not silently become a claim about the whole experiment.
Document each human-facing change and the corresponding unchanged machine field. State the expected modality, observation scope and feature type alongside your check. If a relationship changes, repair the mapping or prose rather than editing measured values.
Worked example: RNA and protein observations
Imagine a synthetic RNA dataset containing cells A, B and C and a protein dataset containing cells B and C. The joint list may contain A, B and C, but A is not measured by the protein assay. A translated diagram should make that missing relationship visible rather than fill protein columns for A with zeros. Trace the original cell identifiers through the observation maps; the two assays can then be described as partially paired without fabricating results.
Worked example: shared word, different feature
Suppose an RNA modality has a feature named CD3D while a protein modality has an antibody feature whose displayed name includes CD3. The two labels are related in a biological discussion, but the modalities measure different things and have different feature identifiers. A local-language explanation can introduce the relationship without merging their rows, equating their units or replacing either stable key.
Frequently asked questions
Is H5MU the same as H5AD? No. H5MU coordinates multiple modality-specific annotated objects, whereas H5AD commonly stores one annotated object. Does every cell occur in every modality? Not necessarily. Can I translate modality names? Explain them in readable prose but preserve actual keys used in the file and software. Does absent protein mean zero protein? No: first determine whether there was a measurement. What should I compare after a conversion? Modality names, observations, features, maps and measurements.
Official specifications and further reading
MuData H5MU on-disk specification; MuData H5MU output documentation; scverse MuData project and API overview; AnnData H5AD as a separate single-object guide. Consult the version-specific source documentation before converting files.
Conclusion
Good multi-omics translation preserves differences as carefully as similarities. Keep modality-local meaning attached to its assay, protect the identifiers that join shared observations and mark missing measurements honestly. The result should help a multilingual reader see the same evidence, including its uncertainty, that the original dataset contains.