Why Do Drugs That Work in Mice Fail in Humans? A Cross-Species Striatal Atlas Offers Clues

Inquiry

Summary

Mouse models remain essential for testing mechanisms in intact biological systems, yet a drug target can occupy a different cellular and anatomical context in humans. A 2026 cross-species atlas of the striatum analyzed single-nucleus RNA profiles from 109 human and 22 mouse samples and paired them with spatial and disease-related measurements. The study found broad conservation alongside species-enriched receptor expression, subregional programs, and cell-type vulnerabilities that can change how a target should be interpreted [1].

The lesson is not that mouse results are unreliable or that transcript abundance predicts clinical success. It is that target conservation should be tested at several levels: gene, cell type, cell state, anatomical subregion, disease context, and, where possible, protein or functional activity. Cross-species single-cell and spatial analysis can identify where translation is well supported, where a model captures only part of the human biology, and which experiments should be added before a program advances.

Research-use note: This guide discusses preclinical and translational research design. The services described are for research use only and are not clinical tests, treatment recommendations, or patient-specific interpretation.

Figure 1. Human and mouse striata share major cellular architecture but can differ in the abundance, expression state, and subregional distribution of specific populations and drug-relevant genes.

Key Takeaways

• A shared gene is not a fully conserved target context. Expression may differ by cell type, state, or subregion even when an ortholog exists in both species.
• Anatomy can alter pharmacological interpretation. Dorsolateral–ventromedial gradients and discrete subregional programs can place receptors and responding populations in different functional contexts.
• Disease vulnerability is cell-specific. The same pathological process can affect selected populations or subregions more strongly than others.
• Cross-species concordance strengthens a model but does not prove efficacy. RNA data do not establish protein abundance, target engagement, dose-response, or clinical outcome.
• Single-cell and spatial layers answer different questions. One resolves cellular programs; the other tests where those programs occur in tissue.

The Same Striatum Does Not Mean the Same Cellular Map

The striatum is often discussed as a conserved brain region involved in movement, action selection, motivation, and reward. That description is useful at the level of anatomy, but it can conceal molecular differences among species and within the region itself. Cell types show continuous dorsolateral–ventromedial gradients and discrete subregional programs, and closely related neuronal populations can express different receptor, signaling, and disease-associated programs. A drug acting in one subregion or cell state may therefore have a different biological context in mouse and human tissue.

Linville and colleagues constructed a cross-species atlas using single-nucleus RNA sequencing of dorsal and ventral striatal samples from 109 humans and 22 mice. They combined the transcriptomic data with multiplexed fluorescence in situ hybridization, spatial transcriptomics, and a single-cell multi-omic approach. The study described conserved cell classes as well as rare populations and molecular gradients, creating a framework for comparing like with like rather than assuming that the same anatomical label denotes the same target environment [1].

The value of an atlas lies in its coordinate system. A target can be evaluated within a defined cell population and subregion, and a disease-associated program can be assigned to the populations that carry it. This reduces two common errors: comparing whole-region averages that are influenced by composition and mapping a human observation to the nearest mouse cell label without checking whether the broader program is conserved.

Figure 2. A cell type with the same label in mouse and human can differ in receptor expression, regulatory state, neighboring cells, or anatomical distribution. Translation requires checking these layers rather than relying on the label alone.

Computational resources are beginning to formalize this mapping problem. BrainAlign aligns spatial transcriptomic structure between human and mouse brains rather than relying only on gene-by-gene similarity [3], while ZEBRA integrates murine and human brain expression at single-cell resolution [4]. These tools expand the comparison space, but their alignments remain models that should be checked against anatomy, experimental metadata, and independent measurements.

Sampling can create apparent species differences even before computation begins. Human tissue may come from postmortem donors with variable intervals, agonal states, medications, ancestry, age, and coexisting pathology, while mouse tissue can be collected at a controlled time from genetically similar animals. A cross-species atlas gains credibility by representing donor variability and by separating effects that recur across humans from those driven by one specimen. The comparison should not ask whether mouse and human profiles are identical; it should ask which features are conserved beyond the variability present within each species.

Anatomical matching is equally demanding. Boundaries defined by gross dissection may include different proportions of adjacent subregions, and a spatial gradient can turn a small sampling displacement into an expression difference. Histological landmarks, spatial measurements, and transparent region definitions help determine whether a divergence is molecular or anatomical. This is one reason a combined single-cell and spatial atlas is more informative than either layer alone for regional brain translation.

If the Drug Target Is Different, Can Mouse Results Be Directly Translated?

Direct translation requires more than detecting a target transcript in both species. The target must be present in the relevant cell population, reachable in the relevant compartment, coupled to a sufficiently similar signaling network, and engaged at an exposure that can be compared across species. Single-cell RNA data address the first parts of that chain: which cells express the target and which molecular programs accompany it. They do not measure binding, protein abundance, downstream activity, pharmacokinetics, or toxicity.

The 2026 atlas reported human-enriched opioid receptor sites in the striatum and mapped the cellular context of drug-relevant expression. It also examined chronic antipsychotic action and found a ventral-biased transcriptional pattern in the analyzed mouse model [1]. These findings illustrate why a whole-striatum average can be misleading. A receptor or response concentrated in a subregion can be diluted by neighboring tissue, while a mouse response localized to one circuit may not generalize to human cells with a different receptor distribution.

The correct interpretation is conditional. Species divergence can lower confidence that a mouse experiment alone represents the human target context, but it does not automatically invalidate the target or the model. It may indicate that the model is suitable for one mechanism, dose-limiting effect, or circuit while human-derived systems are needed for another. The next experiment should be chosen to resolve the specific mismatch.

Target assessment should also distinguish baseline expression from induced expression. A gene that is weak at baseline may become prominent in a disease or treatment-associated state, and that inducible program may differ across species. Conversely, constitutive expression does not show that the target participates in the modeled disease. Comparing baseline, disease, and perturbation within homologous populations provides a more useful translation test than a single cross-sectional abundance value.

Direction is not enough when response magnitude and context matter. A pathway may increase in both species but in different cell types, at different time points, or alongside opposing compensatory programs. Reporting effect sizes by specimen and population makes these differences visible. Where data permit, the model should be evaluated against a range of human responses rather than a single averaged reference.

Questions worth asking before treating a mouse result as directly portable include:

• Is the target enriched in the same homologous cell population in human and mouse?
• Does expression occur in the same striatal subregion and neighboring cellular environment?
• Are receptor subunits, pathway partners, and downstream response genes conserved?
• Does the disease or perturbation create the target-positive state in both species?
• Is transcript expression supported by protein localization or functional perturbation?
• Could a human-specific or human-enriched population change efficacy or safety interpretation?

For research programs that need to connect target biology with tissue location, spatial omics solutions for drug discovery can be considered alongside single-cell sequencing services. The related resource on spatial omics in drug discovery provides a broader view of target identification, response mapping, and translational study design.

Disease Biology Also Has Cell-Type and Spatial Vulnerability

Disease-associated genes are rarely uniform across an organ. Vulnerability can depend on cell identity, anatomical position, developmental history, connectivity, metabolic demand, and the molecular state created by disease. A whole-tissue signature can show that a pathway is altered without revealing which population initiates, carries, or responds to that alteration. Cross-species atlases add another question: does the model reproduce the human vulnerable population and its local context?

In Huntington disease analyses, the 2026 study paired transcriptomic measurements with somatic trinucleotide repeat expansion and mapped cell-type and subregional vulnerability [1]. Pairing these layers matters because gene expression alone does not describe the genomic change, and the genomic measurement alone does not identify the cellular program in which it occurs. Their association can prioritize hypotheses about selective vulnerability, while causal direction still requires additional experiments.

A larger human brain cell atlas has likewise shown how integration across brain regions can define shared and region-specific cell populations [5]. Such references help place a disease-focused striatal observation within broader human cellular diversity. They also reveal a limitation of model translation: a mouse experiment may reproduce a molecular pathway yet lack a human-enriched state or regional gradient that modifies its effect.

Disease translation can be reviewed across four layers:

• Baseline identity. Determine whether the homologous population exists and shares core markers.
• Disease state. Test whether the same population enters a comparable molecular state under pathology.
• Spatial vulnerability. Map whether the affected cells occupy analogous subregions or microenvironments.
• Perturbation response. Assess whether treatment changes the disease-associated program in the same direction and cell context.

The brain snRNA-seq resource for neuroscience research discusses why nuclei-based profiling is useful for archived and structurally complex brain tissue. When location is decisive, spatial transcriptomics services can complement nuclei data by mapping selected programs back into intact sections.

What This Means for Preclinical Study Design

A translational program should treat model relevance as a set of testable claims. Instead of asking whether the mouse is a good model in general, ask whether it represents the target-bearing cell, anatomical compartment, molecular pathway, and perturbation response needed for the specific decision. A model can be informative for one endpoint and weak for another. Stating that scope before efficacy experiments reduces the risk of interpreting a positive result more broadly than the biology supports.

Figure 3. A cross-species decision pathway moves from target conservation to cellular and spatial context, perturbation response, orthogonal validation, and a documented conclusion about which part of human biology the model represents.

Sample balance is crucial. Species should not be confounded with tissue region, sex, age band, processing method, sequencing run, or disease status. Human postmortem interval and tissue quality introduce variables that do not have exact mouse equivalents. The analysis should preserve donor identity and compare effects at the specimen level, with cell-level data nested within specimens.

A prospective preclinical design can use an early translation checkpoint. Before a long intervention study, a small number of well-characterized samples can test whether the model expresses the target in the intended population and whether the proposed pharmacodynamic markers are measurable. The checkpoint does not need to establish efficacy. Its purpose is to confirm that the model can generate the evidence the later study expects to interpret.

Negative translation findings can still guide development. If the target is absent from the homologous mouse population, the team can select another species, engineer a targeted model, add a human-derived system, or restrict the claim to a conserved downstream pathway. If the target is present but spatially displaced, dosing and endpoint selection may need to reflect the affected circuit. Documenting that decision prevents the same uncertainty from resurfacing after a costly efficacy experiment.

Cross-species gene mapping also needs explicit rules. One-to-one orthologs are easier to compare than one-to-many relationships, and a conserved name does not ensure conserved regulatory logic. Analyses should report the orthology resource, excluded genes, aggregation strategy, and sensitivity of the result to the mapping. Cell-type alignment should be supported by multiple genes and, where possible, anatomical evidence rather than a single marker.

When Cross-Species Single-Cell Analysis Adds the Most Value

The approach is most valuable when the decision depends on cellular heterogeneity that bulk tissue would obscure. A target expressed in a rare neuronal population, a treatment response confined to ventral striatum, or a disease state shared by only a subset of cells can all be lost in an average. It is also useful when human tissue is limited and each specimen must answer several linked questions about identity, state, and model concordance.

High-value use cases include:

• prioritizing targets whose human cellular context is uncertain;
• selecting among animal models that reproduce different components of the biology;
• identifying species-specific populations or response programs before a long efficacy study;
• comparing disease-associated states across human tissue, models, and human-derived systems;
• separating conserved mechanism from model-specific adaptation;
• designing spatial validation for subregion- or niche-dependent effects.

The method adds less value when the primary endpoint is already measured directly and cell identity is unlikely to modify interpretation, or when the human comparison lacks enough independent specimens to distinguish species from donor variation. It is also premature when tissue sampling is anatomically inconsistent. More cells cannot repair a species comparison in which the sampled regions are not homologous.

Reference atlases are most useful as priors, not replacements for project-specific controls. Differences in age, disease stage, tissue processing, ancestry, strain, and platform can alter apparent cell states. A public atlas can identify likely target populations and help size a study, but a decisive program should verify the target in material that matches the intended model and human context as closely as possible. When only unmatched references are available, the resulting uncertainty belongs in the decision record.

Rare populations create another high-value case. Bulk assays may miss a human-enriched population that expresses a target or liability, while a standard mouse study may never sample enough homologous cells to reveal the gap. Enrichment, deeper cellular sampling, or targeted spatial imaging can test whether the population is reproducible. The analysis should report the number of independent specimens contributing those cells so that rarity is not confused with a donor-specific observation.

New computational frameworks may help scale the comparison. TransBrain, for example, was developed to translate brain-wide phenotypes between humans and mice by linking spatial and molecular information [2]. Such methods can generate testable mappings, but they do not remove the need for experimental confirmation in the target tissue and biological context.

What Cross-Species Concordance Cannot Prove

Concordant RNA expression does not prove that the target protein is present at the same abundance, occupies the same cellular compartment, or has the same activity. It does not establish that a compound reaches the site at a comparable exposure or binds with comparable affinity. Nor does it show that downstream physiology, behavior, adverse effects, or clinical response will match. These are separate links in the translational chain.

Cross-species agreement is also sensitive to analysis choices. Cell labels can be too broad, integration can suppress real divergence, and gene-set scores can appear conserved even when different genes carry the signal. Conversely, technical differences can exaggerate divergence. A credible analysis shows conserved and species-enriched features, reports uncertainty, and tests key findings in unintegrated data.

Concordance can support statements such as:

• the target transcript is detected in homologous populations;
• a cell-state program is shared across sampled specimens;
• a regional gradient has a comparable direction;
• a perturbation-associated expression response overlaps between species.

It cannot by itself support statements that a target is validated for therapy, a dose will translate, a response will occur in patients, or a model predicts clinical efficacy. Keeping these boundaries visible protects the value of the atlas: it becomes evidence for better experiments, not a substitute for them.

A Multi-Layer Strategy for Human-Relevant Translation

The strongest translation strategy combines layers selected for the uncertainty at hand. Single-cell or single-nucleus RNA sequencing defines cellular expression programs. Spatial transcriptomics or multiplex imaging locates those programs in tissue. Chromatin or multi-omic data can test regulatory context. Protein assays, pharmacology, and perturbation assess whether expression corresponds to function. No layer is automatically required; each should close a named evidence gap.

Figure 4. Human-relevant translation can combine gene orthology, cell identity, spatial context, regulatory state, protein evidence, perturbation, and pharmacology. Confidence grows when independent layers address the same specific claim.

A staged design can control complexity:

1. Screen conservation. Use existing atlases to assess orthology, cell-type expression, and regional distribution.
2. Generate matched profiles. Sample homologous regions and retain specimen metadata across species.
3. Test the mismatch. Add spatial, protein, or functional assays where the cross-species comparison diverges.
4. Perturb the mechanism. Measure cell-resolved responses under a controlled genetic or pharmacological intervention.
5. Validate in human-derived material. Use independent tissue, organoids, or cellular systems appropriate to the claim.
6. Document residual uncertainty. State what the model represents and which human-specific features remain untested.

Integrating dissociated and spatial data requires attention to reference quality, platform resolution, and the distinction between measured and computationally mapped cells. The resource on integrating scRNA-seq with spatial transcriptomics describes these considerations. Researchers examining treatment response and pathway localization can also consult spatial omics for mechanism-of-action studies.

Practical Questions to Ask Before Advancing a Mouse-Derived Drug Target

A review meeting benefits from questions that lead to an experimental action. "Is the target conserved?" is too broad. A useful review identifies the precise cell type, tissue subregion, disease state, and evidence layer behind the target claim. It should also separate a missing measurement from evidence of genuine species divergence.

Use the following checklist before advancing the target:

• Which human cell populations express the target, and across how many independent donors?
• Is the expression pattern observed in the relevant disease or perturbation state?
• Does the mouse model contain a homologous population with comparable pathway partners?
• Are mouse and human tissues sampled from anatomically matched subregions?
• Is the signal supported by spatial or imaging evidence rather than dissociated profiles alone?
• Does protein localization agree with RNA expression?
• Has target engagement been demonstrated in the relevant cells?
• Are treatment-response signatures conserved at matched dose context and time point?
• Which result would cause the team to change models, add a human-derived system, or stop?

Answers can be recorded in an evidence matrix with one row per claim and columns for human, mouse, spatial, protein, perturbation, and uncertainty. An empty cell then becomes a visible study-design need. This is more informative than compressing all evidence into a single model-relevance score.

Cross-Species and Spatial Omics Support at CD Genomics

Cross-species projects require coordinated decisions about homologous tissue sampling, cell or nucleus preparation, genome annotation, ortholog mapping, batch design, and downstream integration. CD Genomics supports research workflows that combine single-nucleus RNA sequencing, spatial transcriptomics, and single-cell RNA-seq data analysis. The service scope can be aligned to target conservation, disease-state comparison, mechanism-of-action mapping, or model-selection questions.

Planning should begin with a claim-evidence map. For each intended conclusion, specify the species, anatomical region, independent specimens, target cell type, required molecular layer, and validation assay. This makes it possible to distinguish measurements that are essential from those that are merely available. It also guides whether a paired species design, a human-focused validation stage, or a spatial follow-up is the most efficient next step.

Information useful for project planning includes:

• species, strain or donor source, age, sex, condition, and tissue-region definitions;
• preservation, post-collection interval, and proposed cell or nuclei workflow;
• target genes, expected populations, and known orthology complications;
• independent specimen numbers and batch allocation;
• available human reference datasets or matched tissue;
• the decision that the cross-species evidence is intended to support.

CD Genomics services described here are for research use only. They are not clinical tests and do not provide diagnosis, treatment recommendations, or predictions of patient response.

FAQs

References

1. Linville RM, James BT, Galani K, et al. Cross-species single-cell atlas of the striatum defines cell-type and subregion disease vulnerabilities. Cell. 2026. Advance online publication, September 1, 2026.
2. Huang S, Zhang T, Dong C, et al. TransBrain: a computational framework for translating brain-wide phenotypes between humans and mice. Nature Methods. 2026;23(2):426–437.
3. Zhang B, Zhang S, Zhang S. Whole brain alignment of spatial transcriptomics between humans and mice with BrainAlign. Nature Communications. 2024;15(1):6302.
4. Flotho M, Amand J, Hirsch P, et al. ZEBRA: a hierarchically integrated gene expression atlas of the murine and human brain at single-cell resolution. Nucleic Acids Research. 2024;52(D1):D1089–D1096.
5. Chen X, Huang Y, Huang L, et al. A brain cell atlas integrating single-cell transcriptomes across human brain regions. Nature Medicine. 2024;30(9):2679–2691.

Research Use and Trust Statement

This article is intended for research use only. It does not provide medical advice and is not intended for diagnostic, prognostic, preventive, or therapeutic use. Cross-species expression and spatial associations do not establish target engagement, causal mechanism, drug efficacy, safety, or patient response without appropriate functional and clinical evidence.