Chimera Separates Expression Deployment From Protein Conservation Across Species
A gene’s sequence can hint at what it is capable of in another species, but the cell type that turns it on determines how that capability is actually deployed. Chimera, a new cross-species single-cell generative model, is built around that split. The framework aligns single-cell atlases across species and evaluates each one-to-one ortholog on two axes at once: the evidence in the counts for a shared expression pattern, and Δɑ, the position of that evidence above or below what protein identity would predict.
The key result is conceptual and practical. Protein conservation and expression deployment do not collapse into the same biological signal. Across eight tissues, protein identity tracked dN/dS, but Δɑ tracked neither. In other words, sequence conservation described one dimension of gene biology, while the cell-state deployment of that gene occupied a separate coordinate that could diverge sharply from the protein sequence pattern.
A Second Coordinate for Orthologs
This distinction becomes most apparent when a gene’s protein conservation and its expression context point in different directions. Spermatocyte genes such as ZPBP2 preserved their stage-specific expression pattern even when their proteins were less conserved, and that pattern held across eight held-out species. Conserved hepatocyte enzymes such as ALAS1 looked different: they shared their pattern, but not their level. The model therefore does not simply ask whether a gene exists in another species; it asks whether another species uses that gene in the same cellular compartment, at the same developmental stage, and in the same expression regime.
One sentence captures the biological consequence: sequence conservation alone missed deployment differences that the single-cell counts could still recover.
That matters because comparative genomics has traditionally leaned on protein similarity as a proxy for functional transfer. Chimera separates the two, giving each gene a pair of coordinates instead of forcing function into a single score. The first coordinate is protein conservation. The second is deployment, measured through expression evidence in the relevant cell type. When those coordinates agree, cross-species inference is straightforward. When they diverge, the gene may be conserved in sequence but repurposed in a different cellular context, or it may be less conserved yet still preserved in the exact cell-state program that matters most.
Specialized Cells Sit Above the Protein-Based Expectation
Across 23 of 24 cohorts, genes from specialized cells sat above what their proteins predicted. The pattern was not subtle. It suggests that specialized cellular programs may preserve expression placement more tightly than protein identity alone would suggest, perhaps because these genes participate in cell-type-defining regulatory programs that remain essential even as coding sequence drifts. In practical terms, a cell type can keep using a gene in the same way even as the protein changes more than the expression profile would lead one to expect.
Approved drug targets also sat above matched genes more often than their controls. That observation gives the work immediate translational relevance. If disease-relevant targets tend to preserve deployment in a way that exceeds what protein sequence predicts, then cross-species model selection should not rest on sequence homology alone. A candidate target may look conserved enough on paper and still fail to reproduce the human expression program in the animal tissue where a disease model depends on it.
The model’s value grows even more when proteins are nearly identical. In that setting, sequence can no longer separate genes cleanly, but the count evidence still can. Chimera therefore retains discriminatory power precisely where conventional conservation metrics start to flatten out. It is in those edge cases—closely related proteins, similar coding sequence, different cellular deployment—where a second coordinate is most useful.
Independent Data Support the Count Evidence
The generative model did not stand alone. Independent data supported the count evidence in liver and testis, and a pre-specified test supported Δɑ in liver. Those confirmations matter because they tie the framework back to external observations rather than leaving it as a purely model-derived construct. In biological terms, the signal survives contact with data outside the training and inference scaffold.
This is especially relevant for tissues with strong cell-state structure. Liver and testis each contain expression programs that are tightly organized by cell identity and developmental stage, making them useful proving grounds for any method that claims to distinguish conserved protein sequence from conserved deployment. The fact that external evidence aligned with the model in these tissues suggests that the approach is capturing a real layer of cross-species biology rather than a statistical artifact of atlas alignment.
Why the Distinction Matters for Animal Models
Animal models stand or fall on whether another species uses a gene the way humans do. Chimera frames that problem more precisely than sequence conservation does. A model organism can carry the ortholog, express the ortholog, and still deploy it differently in the relevant cell population. For disease genes and drug targets, that gap can distort interpretation long before a phenotype is measured.
The framework also speaks to a broader principle in comparative biology: gene function is not exhausted by protein identity. Cellular context determines when, where, and at what level a gene participates in a biological program. By bringing single-cell data into the comparison, Chimera makes that context measurable. The result is a method that can tell researchers not only whether the ortholog is present, but whether the ortholog is used in a comparable way in the species being studied.
For preclinical translation, that is a sharper filter than conventional homology tables. A gene may be conserved enough to appear interchangeable, yet differ in deployment enough to alter pathophysiologic interpretation. Another gene may be less conserved in sequence but remain locked to the same specialized cell program across species. Chimera is built to detect both situations.
What This Preprint Adds, and What It Cannot Yet Prove
This work is a preprint and has not undergone peer review. Its claims therefore remain provisional, even if the biological framing is compelling. The report does not provide clinical outcome data, therapeutic intervention results, or patient-level validation. Its strength lies in method development and in the cross-species atlas evidence used to support it.
Still, the central advance is clear: Chimera treats protein conservation and expression deployment as separate biological coordinates. Across multiple tissues and held-out species, that separation revealed discrepancies invisible to sequence-based comparison alone. For researchers choosing animal models of disease genes or drug targets, the relevant question is no longer just whether a gene is conserved. It is whether the gene is deployed in the same cellular setting, with the same expression logic, where human biology depends on it.