🧬 biology

Cross-species comparison delineates transferability limits of myeloid transcriptional measurements in human osteosarcoma

This study demonstrates that while a fixed 24-gene myeloid transcriptional module yields reproducible, correlated scores across human and canine osteosarcoma datasets, it fails to establish invariant cell identities or independent transcriptional programs, thereby highlighting significant limits to its cross-species interpretability and clinical utility without further prospective validation.

Original authors: Riyou Wen, Hanwen Luo, Zhifa Huang, Heng Huang, Muyun Tan, Guoping Le

Published 2026-09-28

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Riyou Wen, Hanwen Luo, Zhifa Huang, Heng Huang, Muyun Tan, Guoping Le

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Bone cancer in humans, specifically a type called osteosarcoma, is a fierce and unpredictable disease. It does not just attack the bone; it creates a complex neighborhood of cells around the tumor, including immune cells that are supposed to fight the cancer but often end up helping it grow. Scientists have long tried to read the genetic instructions inside these immune cells to understand what they are doing. They look for specific patterns in the DNA messages, or transcripts, that act like a scorecard, telling them if the cells are in a state of "activation" (fighting) or "antigen presentation" (showing the enemy to other immune cells). The hope has been that if we can measure these scores accurately, we might predict how a patient will respond to treatment. However, these measurements are tricky. The immune system is messy, and the same set of genes might mean different things in different people or even in different species.

To test whether these genetic scorecards actually work across species, a team of researchers from Liuzhou Workers' Hospital in China turned to dogs. Dogs naturally develop osteosarcoma in a way that is strikingly similar to humans, making them a powerful tool for comparison. The researchers wanted to see if a specific set of twenty-four genes, which had been used to create a scorecard for human immune cells, would tell the same story in dogs. They were not trying to prove that dogs and humans are identical, but rather to find the limits of how far we can stretch a human measurement tool before it stops making sense. The study is a careful examination of whether a tool built for one species can reliably measure the same biological state in another, without making false promises about what the numbers actually mean.

The researchers started with a pre-existing list of twenty-four genes known to be involved in myeloid cells, a broad category of immune cells that includes macrophages and monocytes. They split this list into three groups: four genes linked to activation, three linked to antigen presentation, and the remaining seventeen as a general background group. They then applied this same grouping to two different sets of data. The first set came from eleven human osteosarcoma samples, where they combined the genetic data from individual cells to create a single, representative profile for each patient. The second set came from one hundred and eighty-six primary bone tumor samples from dogs. By calculating a score for each group of genes in every sample, they could see how the different parts of the scorecard moved together.

In the human samples, the results showed a very strong connection. The score for the activation genes and the score for the antigen-presentation genes both rose and fell in perfect step with the general background score. This means that in these human tumors, when the general immune activity went up, the specific "fighting" and "showing" signals went up with it. The researchers compared this pattern against thousands of random sets of genes to ensure the connection wasn't just a fluke of the data, and the result held up. However, when they looked at the relationship between the activation genes and the antigen-presentation genes themselves, they found no special link; they moved together only as much as any two random groups of genes would.

When the team turned to the dogs, the pattern was similar but weaker. The activation and antigen-presentation scores still moved in the same direction as the general background score, but the connection was not as tight as it was in humans. One of the connections in the dogs was just barely strong enough to be considered statistically significant, sitting right on the edge of what scientists consider a reliable signal. Crucially, just like in the human data, there was no strong, special link between the activation and antigen-presentation scores in the dogs. This suggests that while the tool can measure a general trend in both species, it does not capture a single, unchanging biological identity that is identical in humans and dogs.

The study also looked at other data sources to see where the tool might break down. They examined detailed maps of individual cells from dogs and data from a different type of genetic test called NanoString, which only looks at a small selection of genes. These checks revealed that the tool is sensitive to how the data is collected. If the genetic test does not cover all the necessary genes, or if the cells are not counted correctly, the score changes meaning. The researchers found that the tool could measure a consistent pattern of movement across different samples and species, but it could not prove that the cells were doing the exact same thing, nor could it confirm that the scores predicted how a patient would survive or respond to therapy.

The most important takeaway from this work is a lesson in caution. The researchers demonstrated that a genetic scorecard can be measured reliably across different species and different types of tests, but that does not mean the score represents a fixed, universal truth about the immune system. The tool measures a reproducible pattern of gene activity, but it does not prove that the cells are in a specific state of "fighting" or that the tool can be used to make clinical decisions for patients. The study explicitly rules out the idea that this scorecard is a definitive biomarker for a specific type of immune cell identity that works the same way in humans and dogs. Instead, it shows that the relationship between these genes is complex and depends heavily on the context of the sample.

Ultimately, the paper concludes that before such a tool can be used to guide treatment for cancer patients, it must be tested in a new, forward-looking study where the rules are set in advance. The current analysis, which looked at existing data, is not enough to claim that the tool works for predicting outcomes. The researchers emphasize that while the cross-species comparison was useful for understanding the limits of the measurement, it cannot replace the need for direct, prospective testing in patients. The scorecard works as a measurement of gene activity, but it is not yet a map of the biological reality it is trying to describe.

Technical Summary: Cross-species comparison delineates transferability limits of myeloid transcriptional measurements in human osteosarcoma

Problem Statement
Myeloid transcriptional scores are increasingly utilized to characterize the tumor microenvironment in human osteosarcoma. However, the biological interpretability of these scores remains uncertain, particularly regarding whether they represent invariant cell states (e.g., a fixed M1 identity) or independent transcriptional programs across different species, cohorts, and assay platforms. While naturally occurring canine osteosarcoma serves as a valuable comparative model, the authors argue that cross-species relevance does not automatically validate that a transcriptional score measures the same biological state in both species. The core problem addressed is the lack of rigorous separation between transcriptomic measurement, cellular identity, and clinical prediction, necessitating a framework to evaluate the limits of transferability for myeloid signatures.

Methodology
The study employs a secondary analysis of public transcriptomic datasets, utilizing a "human-first" design to separate measurement questions from cellular identity claims.

Gene Module Construction: A pre-defined 24-gene myeloid module was partitioned into three mutually exclusive components based on one-to-one human-dog orthologue mapping:
- Activation-oriented: 4 genes (CYBA, TCIRG1, SPP1, CD68).
- Antigen-presentation-oriented: 3 genes (HLA-DRA/DLA-DRA orthologue, CD74, LYZ).
- Residual: The remaining 17 genes.
- Note: The partition labels are operational descriptors fixed prior to analysis, not claims of functional coherence for individual genes.
Datasets and Units:
- Human Discovery: GSE152048 ( $n=11$ specimens), analyzed as specimen-level pseudobulk profiles derived from single-cell data to avoid treating cells as independent biological replicates.
- Canine External Test: GSE238110 ( $n=186$ primary osteosarcoma samples), analyzed as bulk RNA-sequencing data.
- Descriptive/Contextual: Canine single-cell data (GSE252470/Zenodo) for descriptive localization and a NanoString cohort (GSE289059) for platform coverage analysis.
Statistical Approach:
- Scores were calculated as the arithmetic mean of normalized expression values (log1p CPM).
- Associations were assessed using two-sided Spearman correlation.
- Robustness Checks: Uncertainty was quantified via non-parametric bootstrap intervals. To determine if correlations exceeded random chance, the authors compared observed correlations against 5,000 expression-matched random gene-set pairs.
- Multiplicity Correction: Benjamini–Hochberg false-discovery-rate ( $q$ ) adjustments were applied across six contrasts.
- Exclusions: No cell-composition or tumor-purity covariates were available for adjustment; thus, correlations are interpreted as specimen-level co-variation rather than specific biological associations.

Key Results

Orthologue Mapping: All 24 genes mapped one-to-one between human and dog.
Human Pseudobulk Cohort: In the human discovery matrix, both the activation score and the antigen-presentation score showed strong positive correlations with the residual 17-gene score (Spearman $\rho = 0.927$ and $0.873$, respectively; $q < 0.0006$ ). These correlations significantly exceeded expression-matched random gene-set distributions. However, the correlation between the activation and antigen-presentation scores ( $\rho = 0.782$ ) did not exceed the matched random background ( $q = 0.564$ ).
Canine Bulk Cohort: In the independent canine dataset, the activation and antigen-presentation scores also correlated with the residual score ( $\rho = 0.696, q = 0.018$ ; and $\rho = 0.672, q = 0.050$ , respectively). The antigen-presentation correlation sits exactly at the multiplicity-adjusted significance threshold. The mutual activation–antigen correlation was weak ( $\rho = 0.134$ ) and did not exceed random background.
Descriptive Localization: Canine single-cell data confirmed that the module genes are enriched in annotated myeloid-associated states, but this was treated as descriptive only due to the small number of source samples (six dogs) and lack of clinical endpoints.
Platform Limitations: Analysis of the NanoString cohort (GSE289059) highlighted that restricted feature coverage (only 9 of 24 module genes covered) limits the ability to compute the full module score, preventing direct clinical validation in this specific dataset.

Significance and Claims
The paper makes a deliberately narrow and modest conclusion regarding the utility of the 24-gene myeloid module:

Reproducible Measurement Property: The module demonstrates reproducible specimen-level transcriptomic covariance across two species (human and canine) and two assay contexts (single-cell-derived pseudobulk and bulk RNA-seq). The scores consistently co-vary with the residual module genes in a manner exceeding random genomic background.
Limits of Transferability: The study explicitly does not establish that the module measures an invariant M1 cell identity, independent transcriptional programs, or a clinically validated biomarker. The weak correlation between activation and antigen-presentation scores in the canine cohort, and the lack of excess over random background in human data for this specific contrast, suggests these components are not biologically independent programs in the way the score decomposition might imply.
Interpretation of Cross-Species Data: The authors emphasize that cross-species similarity in score behavior supports the transferability of a measurable transcriptional structure but does not prove equivalence of cell states, regulatory mechanisms, or therapeutic associations.
Future Requirements: The paper concludes that prospective, prespecified testing with orthogonal cell-state validation (e.g., protein or spatial measurements) in independently collected, clinically annotated cohorts is required before any clinical application of these scores. The current analysis identifies a robust measurement property, not a clinically usable biomarker.

In summary, the work delineates the boundaries of current myeloid transcriptional measurements, confirming their statistical reproducibility across species while rigorously refuting claims of invariant biological identity or immediate clinical utility without further validation.

Technical Summary: Cross-species comparison delineates transferability limits of myeloid transcriptional measurements in human osteosarcoma

More like this