Mapping Cell Types to Tissue Architecture: How to Use Single-Cell RNA-Seq References for Spatial Annotation of MERFISH Data

By Dr. Zubair Khalid, DVM, MS, PhD · 2026-08-20

Key Takeaways

• Reference selection is paramount: The MERFISH annotation workflow critically depends on selecting a single-cell (scRNA-seq) or single-nucleus (snRNA-seq) reference dataset that precisely matches the MERFISH sample's tissue, species, developmental stage, and disease state to avoid systematic annotation errors.
• Panel design dictates resolution: The MERFISH gene panel's composition directly limits the achievable cell type resolution; marker genes must be validated against the reference to ensure they can adequately distinguish the target cell populations, preventing collapse of distinct subtypes.
• Label transfer methods vary in complexity and output: Correlation-based methods offer speed and interpretability but lack uncertainty quantification, while probabilistic and integration methods provide confidence scores and better handle batch effects, respectively, at the cost of increased computational demands.
• Spatial validation is crucial for biological accuracy: Overlaying annotated cell types onto tissue coordinates and comparing their spatial organization to known anatomical patterns serves as a critical quality control step, identifying potential segmentation errors or misannotations that gene expression alone might miss.
• Panel bias and reference incompleteness are inherent limitations: MERFISH annotation cannot resolve cell types not represented by the targeted gene panel, nor can it identify cell populations absent from the reference dataset, necessitating careful consideration of these constraints during interpretation.

Researchers using multiplexed error-robust fluorescence in situ hybridization (MERFISH) face a distinct analytical problem: the platform produces spatially resolved single-cell transcriptomes, but those transcriptomes carry no cell type labels. Assigning cell types to each spatially barcoded cell requires a reference-based annotation workflow that transfers identities from single-cell RNA sequencing (scRNA-seq) or single-nucleus RNA sequencing (snRNA-seq) data. This article provides a step-by-step workflow for reference-based annotation of MERFISH data, covering reference selection, marker gene validation, label transfer, quality control, and interpretation limits. The workflow is designed for biology students, researchers, laboratory professionals, and life-science practitioners who need concrete decision criteria instead of abstract descriptions of available tools.

The Annotation Problem in Imaging-Based Spatial Transcriptomics

Imaging-based spatial transcriptomics platforms such as MERFISH measure gene expression directly in intact tissue sections. The output is a set of molecule coordinates, cell boundaries from segmentation, and a count matrix of transcripts per cell. Unlike sequencing-based spatial methods that capture whole transcriptomes, MERFISH uses a targeted gene panel. The panel typically includes several hundred genes selected for their ability to distinguish known cell types. This targeted design means the data contain enough information to classify cells, but only if the analyst brings in external knowledge about which genes define which cell types.

The core challenge is that MERFISH data do not exist in isolation. A cell's identity is defined by its transcriptional state, and that state must be interpreted against a reference. Single-cell RNA sequencing provides that reference because it captures full transcriptomes from dissociated cells. The reference contains cell type labels derived from unsupervised clustering and marker gene analysis. The annotation task is to project those labels onto MERFISH cells by matching gene expression patterns across the two platforms.

This problem is not unique to MERFISH. Spot deconvolution tools for sequencing-based spatial transcriptomics face similar challenges, and some of those tools have been benchmarked on MERFISH data. For example, SpatialPrompt integrates gene expression, spatial location, and scRNA-seq references to infer cell type proportions, and its benchmarking included MERFISH datasets alongside Visium and Slide-seq data [<a href="#ref-1">1</a>]. The existence of such tools confirms that reference-based annotation is a recognized analytical strategy across spatial transcriptomics platforms.

The scale of the problem matters. Large collaborative efforts such as the BRAIN Initiative Cell Census Network have aggregated reference databases spanning over 40 multimodal profiling techniques from more than 30 research groups [<a href="#ref-2">2</a>]. These references include MERFISH data mapped into common coordinate frameworks, demonstrating that spatial annotation is beyond a single-sample exercise but an atlas-scale integration problem. For individual researchers, the practical question is how to select an appropriate reference, validate that it applies to their tissue and species, and execute the label transfer without introducing systematic errors.

Reference Selection Criteria for scRNA-Seq and snRNA-Seq Data

The choice of reference dataset is the most consequential decision in the annotation workflow. A poorly matched reference will produce confident but incorrect labels. The reference must match the tissue, species, developmental stage, and disease state of the MERFISH sample. It must also have sufficient cell type resolution to distinguish the populations present in the tissue.

Tissue and Species Matching

The reference should come from the same tissue and species as the MERFISH sample. Cross-species annotation is possible but introduces additional uncertainty because orthologous gene expression may differ. Cross-tissue annotation within the same species is also problematic because cell types are often tissue-specific. A reference built from mouse brain should not be used to annotate mouse lung MERFISH data, even though both are mouse tissues.

Public repositories provide the primary access point for reference data. The National Center for Biotechnology Information maintains databases for sequence data, gene expression, and associated metadata [<a href="#ref-3">3</a>]. Researchers can search these databases for scRNA-seq and snRNA-seq datasets from their tissue of interest. The European Bioinformatics Institute offers training pathways that cover data retrieval and analysis practices for these resources [<a href="#ref-4">4</a>]. These official sources provide the infrastructure for finding and accessing reference datasets, but the researcher must evaluate whether a given dataset is appropriate for the annotation task.

Cell Type Resolution

The reference must contain the cell types present in the MERFISH sample. If the reference lacks a rare population that exists in the tissue, those cells will be mislabeled as their nearest transcriptional neighbor. Conversely, if the reference contains subtypes that are not distinguishable by the MERFISH gene panel, the annotation will collapse those subtypes into a single label.

Recent atlas projects demonstrate the level of resolution that is achievable. A multimodal human lung atlas profiled 746,047 nuclei from 49 mapped lung blocks across 11 healthy adults and resolved 70 molecularly distinct populations [<a href="#ref-5">5</a>]. That same study used MERFISH to map 25 cell populations across 7 structural neighborhoods, showing that the spatial platform can resolve a meaningful subset of the transcriptional diversity captured by single-nucleus profiling [<a href="#ref-5">5</a>]. The gap between 70 molecular populations and 25 spatially mapped populations reflects the limited gene panel of MERFISH, not a failure of the technology.

For the mouse brain, a quantitative three-dimensional atlas integrated transcriptomic, morphological, and electrophysiological cell type information, resulting in 5,274 transcriptomic clusters and 458 functional morphological-electrophysiological types [<a href="#ref-6">6</a>]. This level of resolution is far beyond what any MERFISH gene panel can distinguish. The annotation workflow must therefore decide which level of cell type granularity is biologically meaningful for the spatial question being asked.

Data Quality and Metadata Completeness

The reference dataset must have been processed with adequate quality control. Single-cell quality control typically involves filtering cells by total transcript count, number of detected genes, and mitochondrial read fraction. A reference that includes low-quality cells will introduce noise into the label transfer. The reference should also have complete metadata, including tissue source, dissociation protocol, sequencing platform, and any disease or treatment conditions.

The Carpentries offer foundational training in data handling and reproducible analysis practices that apply to evaluating and processing reference datasets [<a href="#ref-7">7</a>]. Bioconductor provides documented workflows for single-cell analysis, including quality control and normalization steps that are prerequisites for using a dataset as a reference [<a href="#ref-8">8</a>]. These resources help researchers assess whether a candidate reference meets basic quality standards.

Marker Gene Selection and Validation for MERFISH Panels

MERFISH measures a targeted set of genes. The gene panel determines which cell types can be distinguished. Marker gene selection is therefore a critical step that occurs before the annotation workflow begins. The selected markers must be expressed at levels detectable by MERFISH, must be specific to the cell types of interest, and must be robust across the tissue region being studied.

Panel Design Constraints

The number of genes in a MERFISH panel is limited by the imaging and coding scheme. Each gene is assigned a barcode, and the barcodes are read out through sequential hybridization rounds. The panel size affects the tradeoff between cell type resolution and technical feasibility. A panel with too few genes will not distinguish closely related cell types. A panel with too many genes increases imaging time and the risk of optical crowding.

Optical crowding is a known challenge in imaging-based spatial transcriptomics. When transcripts are densely packed, individual molecules become difficult to resolve, and this problem worsens as panel size increases [<a href="#ref-9">9</a>]. Tissue thickness and panel bias also contribute to computational difficulty [<a href="#ref-9">9</a>]. These technical constraints mean that marker selection must balance biological informativeness against the physical limits of the imaging system.

Validating Markers Against the Reference

Before running the annotation, the selected markers should be validated against the reference dataset. The validation checks that each marker is expressed in the expected cell type and that the marker combination can separate the cell types of interest. This validation uses the scRNA-seq or snRNA-seq reference as ground truth.

The validation procedure involves several concrete steps. First, extract the expression values for the panel genes from the reference dataset. Second, check that each marker is detected in the reference at a reasonable level. A marker that is not detected in the reference will not be detected by MERFISH either. Third, test whether the marker combination separates the reference cell types using a simple classifier or dimensionality reduction. If the markers cannot separate the cell types in the reference, they will not separate them in the MERFISH data.

The Cell Type and Marker Gene Dictionary from the human lung atlas provides an example of how marker information can be organized for reuse. That resource includes anatomically aligned nomenclature and marker gene information that supports cell type identification across platforms [<a href="#ref-5">5</a>]. Similar resources exist for other tissues and species, and they can serve as starting points for panel design.

Handling Panel Bias

Panel bias refers to the systematic overrepresentation or underrepresentation of certain genes in the panel relative to the full transcriptome. This bias affects the annotation because the label transfer algorithm only sees the panel genes. If the panel is biased toward genes that are highly expressed but not cell type specific, the annotation will be driven by expression level instead of identity.

The review of imaging-based spatial transcriptomics methods notes that panel bias increases computational difficulty [<a href="#ref-9">9</a>]. The practical implication is that marker selection should prioritize genes with high cell type specificity, even if their absolute expression levels are moderate. Genes that are expressed in many cell types contribute noise to the annotation and should be excluded from the panel.

Label Transfer Methods for MERFISH Annotation

Label transfer is the computational step that assigns cell type labels to MERFISH cells based on their expression of the panel genes. Several classes of methods exist, and the choice of method affects both accuracy and computational cost.

Correlation-Based Methods

The simplest approach is to compute a correlation between each MERFISH cell's expression profile and the average expression profile of each reference cell type. The cell is assigned the label of the reference type with the highest correlation. This method is fast and interpretable, but it assumes that the reference type averages are representative and that the panel genes capture the relevant variation.

Correlation-based methods work best when the cell types are well separated and the panel genes are highly informative. They fail when cell types are transcriptionally similar or when batch effects between the reference and MERFISH data distort the expression values.

Probabilistic Classification

Probabilistic methods model the expression of each cell type as a distribution and assign each MERFISH cell the label with the highest posterior probability. These methods can account for expression variability within cell types and can provide uncertainty estimates for each assignment. The review of imaging-based spatial transcriptomics identifies probabilistic cell typing as one of the key analytical steps in the interpretation framework [<a href="#ref-9">9</a>].

The advantage of probabilistic methods is that they produce confidence scores. A cell with a high posterior probability for one label is a confident assignment. A cell with similar probabilities for multiple labels is ambiguous and should be flagged for manual review. This uncertainty information is valuable for downstream analysis because it identifies regions of the tissue where the annotation is unreliable.

Deconvolution-Based Approaches

Deconvolution methods estimate the proportion of each cell type within a spatial unit. For MERFISH, where the spatial unit is a segmented cell, deconvolution is less relevant than for sequencing-based platforms where each spot contains multiple cells. However, deconvolution tools have been benchmarked on MERFISH data, and they can be useful when cell segmentation is imperfect.

SpatialPrompt uses non-negative ridge regression and graph neural networks to infer cell type proportions while incorporating spatial location information [<a href="#ref-1">1</a>]. Its benchmarking on MERFISH datasets demonstrated superior performance over 15 existing tools, and it achieved deconvolution and domain identification for 50,000 spots in under 2 minutes on a mouse hippocampus dataset [<a href="#ref-1">1</a>]. The tool also includes a database of over 40 curated scRNA-seq datasets for integration [<a href="#ref-1">1</a>]. This example shows that deconvolution approaches can be adapted to MERFISH data, particularly when the analysis goal includes spatial domain identification instead of single-cell labeling alone.

Integration-Based Methods

Integration methods embed the reference and MERFISH data into a shared low-dimensional space and then transfer labels based on neighborhood relationships. These methods are more complex than correlation or probabilistic classification, but they can handle batch effects and nonlinear relationships between the platforms.

The challenge with integration methods is that they require the two datasets to share enough genes. MERFISH panels typically include several hundred genes, which is sufficient for integration if the panel was designed with the reference in mind. If the panel genes do not overlap well with the reference, integration will fail.

At a Glance: Reference Annotation Workflow for MERFISH

Practical Workflow for Reference-Based Annotation

The following workflow assumes the researcher has already generated MERFISH data with cell segmentation and transcript assignment. The steps focus on the annotation process from reference selection through spatial validation.

Step 1: Define the Cell Type Taxonomy

Before selecting a reference, define the cell type taxonomy that the annotation should produce. This taxonomy should be based on the biological question. A study of lung structure might require distinguishing alveolar type 1 cells, alveolar type 2 cells, endothelial cells, and immune cell subtypes. A study of brain organization might require distinguishing dozens of neuronal subtypes.

The taxonomy should be informed by the reference dataset. If the reference contains 70 molecular populations, the annotation can attempt to distinguish all 70, but the MERFISH panel may only support 25 [<a href="#ref-5">5</a>]. The taxonomy should therefore be set at a level that the panel can support. Trying to annotate more cell types than the panel can distinguish will produce unreliable labels.

Step 2: Acquire and Process the Reference

Download the reference dataset from a public repository such as NCBI [<a href="#ref-3">3</a>]. Process the reference with standard single-cell quality control procedures. Filter low-quality cells, normalize expression values, and identify highly variable genes. The reference should be processed consistently with the MERFISH data to minimize platform differences.

Bioconductor provides documented workflows for single-cell analysis that cover these processing steps [<a href="#ref-8">8</a>]. The Galaxy Training Network offers accessible tutorials for running these analyses without extensive programming experience [<a href="#ref-10">10</a>]. These resources support reproducible processing of reference datasets.

Step 3: Validate the Gene Panel

Extract the panel genes from the reference expression matrix. Check that each gene is detected in the reference and that the genes collectively separate the cell types in the taxonomy. This validation can be done with a simple classifier or by visualizing the reference cells in a dimensionality reduction colored by the taxonomy labels.

If the panel genes do not separate the reference cell types, the annotation will fail regardless of the label transfer method. The researcher must either adjust the taxonomy to a coarser level or redesign the panel. Redesigning the panel requires new MERFISH experiments, so this validation should occur before committing to the full imaging run.

Step 4: Run Label Transfer

Apply the chosen label transfer method to assign labels to MERFISH cells. Record the confidence score for each assignment. For probabilistic methods, this score is the posterior probability. For correlation methods, it may be the correlation coefficient or the difference between the top two correlations.

The label transfer should be run with the same gene set in both the reference and the MERFISH data. Any gene that is missing from either dataset should be excluded from the analysis. The gene set should be limited to the panel genes to avoid introducing genes that were not measured in the MERFISH data.

Step 5: Filter and Review Assignments

Filter out cells with low confidence scores. The threshold depends on the method and the data, but a reasonable starting point is to retain cells where the top label probability exceeds 0.5 or where the margin between the top two labels is substantial. Cells that fall below the threshold should be labeled as ambiguous instead of forced into a cell type.

Review the ambiguous cells to determine whether they represent a real cell type that the panel cannot distinguish or technical artifacts. This review may involve examining the spatial distribution of ambiguous cells. If ambiguous cells cluster in a specific tissue region, they may represent a distinct population that was not captured by the taxonomy.

Step 6: Validate Spatial Organization

Overlay the annotated cell types on the tissue coordinates to check that the spatial organization matches known anatomy. For example, in the mouse olfactory system, MERFISH revealed stereotypical gradients of sensory neuron distribution along central-to-peripheral and apical-to-basal axes in the main olfactory epithelium [<a href="#ref-11">11</a>]. If the annotation produces a spatial pattern that contradicts known anatomy, the annotation is likely wrong.

Spatial validation can also identify segmentation errors. If a cell is annotated as a neuronal subtype but is located in a region where that subtype is never found, the cell boundary may have been drawn incorrectly, mixing transcripts from two adjacent cells.

Options and Tradeoffs in Label Transfer Methods

The choice of label transfer method involves tradeoffs between accuracy, computational cost, interpretability, and robustness to batch effects.

Correlation Methods: Speed and Simplicity

Correlation-based methods are the fastest and most interpretable. They require minimal computational resources and produce results that can be explained to collaborators. The main limitation is that they do not account for expression variability within cell types. A cell that is at the edge of a cell type's expression distribution may be assigned to the wrong type.

Correlation methods are appropriate when the cell types are well separated and the panel genes are highly informative. They are less appropriate when the tissue contains transcriptionally similar cell types that differ by subtle expression patterns.

Probabilistic Methods: Uncertainty Quantification

Probabilistic methods provide confidence scores that support downstream filtering and review. The uncertainty information is valuable for identifying regions of the tissue where the annotation is unreliable. The main limitation is that probabilistic methods make distributional assumptions about gene expression that may not hold for all genes.

Probabilistic methods are appropriate when the analysis requires confidence scores or when the tissue contains cell types with overlapping expression profiles. They are less appropriate when the reference dataset is small or when the panel genes have unusual expression distributions.

Integration Methods: Batch Effect Handling

Integration methods can handle systematic differences between the reference and MERFISH data, such as differences in sequencing depth, dissociation artifacts, or platform-specific biases. The main limitation is complexity. Integration methods require careful parameter tuning and are more difficult to troubleshoot when they fail.

Integration methods are appropriate when the reference and MERFISH data come from different laboratories, different protocols, or different sequencing platforms. They are less appropriate when the reference and MERFISH data are highly consistent, because the added complexity does not improve accuracy.

Deconvolution Methods: Spatial Domain Focus

Deconvolution methods are designed for estimating cell type proportions within spatial units. For MERFISH, they are most useful when the analysis goal includes identifying spatial domains or when cell segmentation is imperfect. SpatialPrompt demonstrated that deconvolution can be fast and accurate on MERFISH data, completing analysis of 50,000 spots in under 2 minutes [<a href="#ref-1">1</a>].

Deconvolution methods are appropriate when the analysis focuses on tissue regions instead of individual cells. They are less appropriate when the analysis requires single-cell resolution labels, because deconvolution estimates proportions instead of assigning discrete identities.

Records and Measurements for Annotation Quality

Maintaining records of the annotation process supports reproducibility and troubleshooting. The following measurements should be recorded for each annotation run.

Reference Dataset Metadata

Record the accession number, tissue source, species, dissociation protocol, sequencing platform, and cell type taxonomy of the reference dataset. This information allows other researchers to reproduce the annotation and to assess whether the reference was appropriate.

Panel Gene Statistics

Record the number of panel genes, the detection rate of each gene in the reference and MERFISH data, and the results of the marker validation. Genes with low detection rates should be flagged because they may not contribute to the annotation.

Label Transfer Parameters

Record the label transfer method, the gene set used, any normalization or scaling steps, and the confidence threshold applied. These parameters determine the annotation output and must be documented for reproducibility.

Assignment Quality Metrics

Record the distribution of confidence scores, the number and fraction of cells assigned to each cell type, and the number of ambiguous cells. These metrics provide an overview of annotation quality and can be compared across samples or experiments.

Spatial Validation Results

Record the results of the spatial validation, including any comparisons to known anatomy or to independent spatial measurements. This information supports the biological interpretation of the annotation.

Common Failure Patterns in MERFISH Annotation

Several failure patterns recur across MERFISH annotation projects. Recognizing these patterns early can save substantial time and prevent incorrect biological conclusions.

Reference Mismatch

The reference does not match the MERFISH sample in tissue, species, developmental stage, or disease state. This failure produces confident but incorrect labels. The annotation assigns cells to cell types that do not exist in the sample or misses cell types that are present.

Prevention requires careful reference selection and validation. The reference should be checked for the presence of expected cell types before running the annotation. If the reference lacks a cell type that is known to exist in the tissue, a different reference should be used.

Panel Gene Inadequacy

The MERFISH panel does not contain enough informative genes to distinguish the cell types in the taxonomy. This failure produces ambiguous assignments or collapses distinct cell types into a single label. The annotation may appear successful because the confidence scores are high, but the labels do not reflect the true cell type diversity.

Prevention requires marker validation against the reference before the imaging run. The panel should be tested to confirm that it can separate the reference cell types. If the panel cannot separate them, the taxonomy must be adjusted or the panel redesigned.

Batch Effects Between Platforms

Systematic differences between the reference and MERFISH data distort the expression values and produce incorrect labels. These differences can arise from dissociation artifacts, sequencing depth, gene capture efficiency, or platform-specific biases. The review of imaging-based spatial transcriptomics notes that multimodal complexity increases computational difficulty [<a href="#ref-9">9</a>].

Prevention requires careful normalization and, when necessary, integration methods that can handle batch effects. The annotation should be validated by checking that the spatial organization of annotated cell types matches known anatomy.

Segmentation Errors

Cell segmentation errors mix transcripts from adjacent cells, producing chimeric expression profiles that do not match any reference cell type. These cells may be assigned to the wrong type or flagged as ambiguous. Segmentation errors are more common in dense tissue regions where cell boundaries are difficult to identify.

Prevention requires careful review of segmentation results before annotation. Cells with unusual morphology or unexpected transcript counts should be examined individually.

Overfitting to the Reference

The annotation method overfits to the reference dataset, producing labels that match the reference but do not generalize to the MERFISH data. This failure is more common with complex integration methods that have many parameters. The annotation may perform well on a training subset but poorly on the full dataset.

Prevention requires cross-validation and careful evaluation of the annotation on held-out data. The annotation should be tested on a subset of MERFISH cells that were not used for parameter tuning.

Limitations of Reference-Based Annotation

Reference-based annotation has inherent limitations that cannot be fully overcome by methodological improvements. These limitations should be communicated clearly in any report or publication.

Panel Bias Limits Resolution

The MERFISH gene panel determines the maximum cell type resolution. Even with a perfect reference and a perfect label transfer method, the annotation cannot distinguish cell types that are not separable by the panel genes. The human lung atlas resolved 70 molecular populations by single-nucleus profiling but mapped only 25 populations by MERFISH [<a href="#ref-5">5</a>]. This gap reflects the information content of the panel, not the quality of the analysis.

Reference Completeness

The reference may not contain all cell types present in the tissue. Rare cell types are often underrepresented in scRNA-seq and snRNA-seq datasets because they are difficult to capture during dissociation. If the reference lacks a rare population, the annotation will assign those cells to their nearest transcriptional neighbor.

Platform-Specific Expression Differences

Gene expression measurements differ between scRNA-seq and MERFISH. Sequencing-based methods capture full transcripts, while MERFISH detects barcode probes that bind to specific regions of the transcript. These differences can produce systematic biases in expression values that affect the annotation.

Spatial Context Is Not Used in Label Transfer

Most label transfer methods use only gene expression information and ignore spatial coordinates. This design choice means that the annotation does not benefit from the spatial organization of the tissue. A cell that is surrounded by cells of a particular type is more likely to be that type, but standard label transfer methods do not use this information.

Spatially aware methods such as SpatialPrompt incorporate spatial location into the analysis [<a href="#ref-1">1</a>]. These methods can improve annotation accuracy by using local microenvironment information, but they are more complex and require additional parameter tuning.

Quality Controls and Professional Escalation Criteria

Quality controls should be applied throughout the annotation workflow. The following controls are recommended for each stage.

Reference Quality Control

The reference should pass standard single-cell quality control metrics before use. These metrics include the number of cells, the number of genes detected per cell, the total transcript count per cell, and the fraction of reads mapping to mitochondrial genes. A reference with poor quality metrics should not be used for annotation.

Marker Validation Control

The marker validation should demonstrate that the panel genes separate the reference cell types. This validation can be quantified by the accuracy of a simple classifier trained on the reference and tested on held-out reference cells. If the classifier accuracy is below an acceptable threshold, the panel should be revised.

Label Transfer Quality Control

The label transfer should produce confidence scores that are inspected before downstream analysis. The distribution of confidence scores should be examined to identify any systematic problems. A large fraction of cells with low confidence scores indicates that the panel or the reference is inadequate.

Spatial Validation Control

The annotated cell types should show spatial organization consistent with known anatomy. This validation can be qualitative, by visual inspection of the annotated tissue, or quantitative, by comparing the spatial distribution of annotated cell types to independent measurements.

Professional Escalation Criteria

Escalate to a bioinformatics specialist or collaborator when any of the following conditions are met:

• The marker validation fails to separate reference cell types, indicating a fundamental problem with the panel or taxonomy.
• The label transfer produces a large fraction of ambiguous cells, and the ambiguous cells cluster in specific tissue regions.
• The spatial validation reveals patterns that contradict known anatomy, suggesting systematic annotation errors.
• The reference dataset is from a different tissue, species, or condition than the MERFISH sample, and no alternative reference is available.
• The annotation results will be used for clinical or regulatory decisions, requiring additional validation and documentation.

Welfare and Safety Context for Biological Interpretation

The annotation of MERFISH data has implications for biological interpretation that extend beyond the computational workflow. The spatial organization of cell types provides information about tissue architecture, cell-cell interactions, and disease mechanisms. These interpretations should be made with appropriate caution.

Spatial Context in Disease Studies

The human lung atlas demonstrated that spatial transcriptomics can map cell populations across structural neighborhoods, providing a foundation for interrogating the origins of lung pathophysiology [<a href="#ref-5">5</a>]. The annotation of cell types in diseased tissue requires a reference from the same disease state or a careful assessment of how disease alters gene expression. Using a healthy reference to annotate diseased tissue may produce incorrect labels if the disease changes the expression of panel genes.

Cross-Sample Comparisons

Comparing annotated cell types across samples requires consistent annotation parameters. The CODA framework for cross-sample alignment and spatially differential gene analysis addresses the challenge of comparing multiple spatial transcriptomics samples with nonlinear distortions and limited spatial overlap [<a href="#ref-12">12</a>]. This framework supports the identification of spatially informative genes associated with normal and disease conditions [<a href="#ref-12">12</a>]. Researchers planning cross-sample comparisons should use consistent references and annotation parameters across all samples.

Atlas-Scale Integration

Atlas-scale projects face additional challenges because they integrate data from multiple platforms and laboratories. The BRAIN Initiative Cell Census Network aggregated data from over 40 multimodal profiling techniques and more than 30 research groups [<a href="#ref-2">2</a>]. The variation in acquisition, tissue processing, and imaging techniques across data types requires tailored mapping approaches [<a href="#ref-2">2</a>]. Individual researchers should be aware that atlas-scale annotations may not be directly applicable to their specific samples.

Frequently Asked Questions

What is the difference between scRNA-seq and snRNA-seq references for MERFISH annotation?

Single-cell RNA sequencing profiles whole cells, while single-nucleus RNA sequencing profiles isolated nuclei. The choice depends on the tissue and the cell types of interest. snRNA-seq is often preferred for tissues where dissociation is difficult or where large cells are fragile. The reference should match the biological question and the tissue type. For example, the human lung atlas used single-nucleus profiling to construct its reference [<a href="#ref-5">5</a>]. The key is that the reference must contain the cell types present in the MERFISH sample, regardless of whether it was generated by scRNA-seq or snRNA-seq.

How many genes should a MERFISH panel include for reliable cell type annotation?

The number of genes depends on the cell type diversity of the tissue and the desired resolution. A panel must include enough informative markers to separate the cell types of interest. The human lung atlas mapped 25 cell populations with MERFISH, while single-nucleus profiling resolved 70 molecular populations [<a href="#ref-5">5</a>]. The panel size should be matched to the taxonomy. A panel with too few genes will not distinguish closely related cell types, while a panel with too many genes increases imaging time and optical crowding [<a href="#ref-9">9</a>].

Can I use a reference from a different species to annotate my MERFISH data?

Cross-species annotation is possible but introduces additional uncertainty. Orthologous genes may have different expression patterns across species, and cell type definitions may not transfer directly. The annotation should be validated carefully if a cross-species reference is used. Whenever possible, use a reference from the same species as the MERFISH sample.

What should I do if my annotation produces many ambiguous cells?

Ambiguous cells are those where the confidence score is low or where multiple cell types have similar probabilities. First, check whether the ambiguous cells cluster in specific tissue regions. If they do, they may represent a distinct cell type that the panel cannot distinguish. Second, check whether the panel genes are informative for the ambiguous cells. If the panel lacks informative markers for a particular population, the annotation cannot resolve it. Third, consider whether the reference contains the cell type that the ambiguous cells represent. If the reference lacks the population, the annotation will fail to identify it.

How do I validate that my annotation is correct?

Validation involves multiple checks. First, confirm that the panel genes separate the reference cell types before running the annotation. Second, check that the annotated cell types show spatial organization consistent with known anatomy. Third, compare the annotation to independent measurements if available, such as immunohistochemistry or in situ hybridization for specific markers. The human lung atlas used multiplexed immunofluorescence to localize cell subtypes and validate the MERFISH annotation [<a href="#ref-5">5</a>].

What is the role of spatial coordinates in annotation?

Most label transfer methods use only gene expression information and ignore spatial coordinates. Spatially aware methods such as SpatialPrompt incorporate spatial location into the analysis and can improve accuracy by using local microenvironment information [<a href="#ref-1">1</a>]. These methods are particularly useful for identifying spatial domains and for handling imperfect cell segmentation. However, they are more complex and require additional parameter tuning.

How do batch effects between the reference and MERFISH data affect annotation?

Batch effects are systematic differences between datasets that are not related to biology. They can arise from differences in tissue processing, sequencing depth, gene capture efficiency, or platform-specific biases. Batch effects can distort expression values and produce incorrect labels. Integration methods can handle batch effects by embedding the datasets into a shared space. The CODA framework addresses cross-sample alignment and can extract spatially informative genes associated with normal and disease conditions [<a href="#ref-12">12</a>].

When should I escalate to a bioinformatics specialist?

Escalate when the marker validation fails, when the annotation produces a large fraction of ambiguous cells that cluster in specific regions, when the spatial validation contradicts known anatomy, when the reference is mismatched to the sample, or when the results will be used for clinical or regulatory decisions. These situations indicate that the annotation may be unreliable and that additional expertise is needed to diagnose and correct the problem.

Related Bioinformatics Guides

• Single-Cell Annotation: A Workflow for Cell Type Identification
• Single-Cell RNA-seq Clustering and Cell-Type Annotation Pipelines
• Single-Cell Sequencing Depth: How Much Is Enough?
• Single-Cell Sequencing Workflow: From Sample Preparation to Data Analysis
• RNA-Seq Data Analysis Workflow: From Raw Reads to Insights

Related Clinical & Scientific Guides

• A Practical Guide to Detecting Antimicrobial Resistance Genes in Shotgun Metagenomic Data
• Computational Immunology: Modeling the Immune System
• How to Set Hard Filters for Germline Variant Calling: A Practical Guide to GATK Best Practices

References and Further Reading

This article is educational and does not replace validated analysis plans, institutional policy, clinical interpretation, or specialist review.