Many human diseases have no good single-cell reference: no dissociated atlas of the diseased tissue, profiled on the right platform, from enough donors. Mouse models of those diseases often have one. This proposal asks whether human bulk RNA-seq, meaning one averaged expression profile per tissue sample, can be split into its cell-type proportions using only a mouse single-cell reference. That split is called deconvolution. The obstacle is that the mouse reference is wrong for human tissue in two tangled ways. One is technical: platform, chemistry, and capture efficiency differ. The other is evolutionary: a gene may keep its sequence but change how much each cell type expresses it. Current practice restricts both datasets to genes with a single clear counterpart in the other species (one-to-one orthologs) and runs a standard deconvolution method. That treats both kinds of error as noise.
The central claim is that the two errors have different shapes, and a cohort of human bulk samples contains enough information to separate them. Technical and ortholog-mapping errors mostly act as a per-gene scale factor shared by every cell type. True regulatory divergence is sparse and specific to a cell type. A gene whose mouse profile is wrong for one cell type, say beta cells, leaves a residual across the cohort that rises and falls with that cell type's estimated proportion. So the proposal is a cohort-level deconvolution model that estimates proportions jointly with a dense per-gene scale and a sparse cell-type-by-gene correction. The correction also serves as an output in its own right: a map of which mouse cell-type profiles fail to transfer to human tissue.
Core Claim
Let $y_{ij}$ be the expression of gene $j$ in human bulk sample $i$. Let $x_{kj}$ be the mouse reference profile of cell type $k$ after ortholog mapping, and $c_{ik}$ the unknown proportion of cell type $k$ in sample $i$. The model is
$$y_{ij} \approx s_j \sum_{k} c_{ik}\,(x_{kj} + \delta_{kj}),$$
where $s_j$ is a gene-wise factor absorbing platform, gene-length, and mapping effects, and $\delta_{kj}$ is a correction to cell type $k$'s profile at gene $j$ that is penalized to be sparse. The $\delta$ terms are identifiable only because proportions vary across the $N$ samples. A single bulk sample cannot tell them apart from a change in $c$, but a cohort can, provided composition varies enough. This gives the method a testable failure condition: with low compositional variance, it should revert to gene-scaling only.
Two further ingredients address known limits of ortholog filtering. First, genes in one-to-many and many-to-many ortholog families are not discarded. Their counts are summed into family-level features, recovering markers lost when CACIMAR found only 139 conserved markers against nearly 2,500 species-specific ones in mouse–zebrafish retina. Second, the likelihood is a count model (Poisson or negative binomial) on raw bulk counts rather than log-transformed values, following the argument of Count Bridges that transcript data should be modeled as integers. We do not reuse its generative sampler, because its exact bulk-summation constraint does not hold across species.
Relation to Prior Work
Three lines of work motivate the gap. SCDC shows that the lab and protocol of a single-cell reference change deconvolution results, and it weights several references by how well they reconstruct the bulk data. It works only within one species, and its weights apply per whole reference, so it cannot mark individual genes in individual cell types as untrustworthy. scAdapt transfers cell-type labels from human to mouse pancreas using adversarial domain alignment, but it classifies single cells and needs target-species cells, which is exactly what our setting lacks. CACIMAR constrains our design. It argues that forcing species into a shared embedding creates false alignments. We therefore avoid a shared latent space and keep the mouse profiles explicit, correcting them only where the human data demand it.
On the deconvolution side, POPPY aligns bulk and single-cell data through ontology-aware contrastive learning, entirely within human data. The breast cancer benchmark in Performance of tumour microenvironment deconvolution methods shows that accuracy degrades as tumor purity rises and that cells retain expression programs from their cell of origin. That benchmark supplies our secondary stress test. Herschkowitz et al. matched mouse and human tumors through bulk ortholog profiles and never unmixed cell fractions.
Research Program
We will build the cohort model above and fit it by alternating two steps: a non-negative, sum-to-one proportion update, and a penalized update of $s$ and $\delta$. The $\delta$ update uses a group-lasso penalty on each cell type's column so that divergence stays sparse. Cell-type-specific total mRNA content is confounded with $s_j$ and with $c$, so we fix $s$ by a median-one normalization and report proportions in units of mRNA fraction.
The main analysis is a factorial design that disentangles the two error sources. We deconvolve the same human bulk data under four references: a human reference from the matched platform, a human reference from a different platform, a mouse reference, and the mouse reference passed through our correction. Comparing the second and third conditions with the first splits the total error into a platform part and a species part. That split is the measurement Section 1 of the research file identifies as missing. The fitted $\delta$ values will be checked against independent divergence evidence: genes that CACIMAR-style within-species marker analysis calls species-specific should be enriched among large $\delta$ values.
Baselines
The naive reference point is one-to-one ortholog filtering followed by MuSiC or CIBERSORTx on the mouse reference, with default settings. This is the procedure a practitioner would run today. Beating it is necessary but not sufficient.
The strongest competitor is BayesPrism run on the same ortholog-filtered mouse reference. The breast benchmark in Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures found BayesPrism among the most stable methods across purity levels and among the lowest in error for granular immune cell types, with no reported number found on Lacuna for those claims. We will also run SCDC's ensemble over all available mouse pancreas references, since cross-reference weighting is its designed strength. For the islet cohort, the strongest human-reference result is SCDC's significant negative association between beta-cell fraction and HbA1c, $p = 0.0019$, in SCDC. On the controlled cell-line mixture, SCDC reports Pearson $r = 0.991$. These are oracle ceilings obtained with same-species references. No reported result was found on Lacuna for any cross-species bulk deconvolution. Every method gets the same gene universe option, the same bulk counts, and the same tuning budget: a grid search of equal size on a held-out pseudo-bulk split, with no access to test proportions.
Two ablations isolate the proposal's key ingredient. The first fits $s_j$ alone, without $\delta$. The second uses one-to-one genes only, without family features. The claim holds only if the full model beats the $s$-only ablation specifically on the cell types where species divergence, not platform error, dominates the factorial error split. Beating BayesPrism means lower root-mean-square error of proportions on cross-species pseudo-bulk, with a paired bootstrap confidence interval excluding zero, and closing at least half the gap to the human-reference oracle.
Why not just…?
Why not just use BayesPrism, which already adapts reference profiles to each bulk sample? This is the real threat. Per-sample profile updating may absorb much of the species divergence. The difference is that our correction is pooled across the cohort and constrained to be sparse per cell type, so it should be more stable when many genes are wrong. That is a hypothesis, and the factorial experiment settles it. If BayesPrism with ortholog mapping matches our model in the species-dominated condition, we will report that the cohort correction adds nothing and redirect the contribution to the error decomposition itself.
Why not just batch-correct the mouse reference toward human data with Harmony or scAdapt and then deconvolve? Both need human single cells to align against, and the target setting has none. Where some human cells exist, full alignment is the move CACIMAR warns erases real divergence, and BERMUDA shows that global alignment merges populations when compositions differ. We will still run the sequential pipeline in the pancreas setting, where human cells are available, as an upper-bound comparison.
Why not just restrict to conserved markers and rescale genes? Selecting CACIMAR-style conserved markers and fitting only $s_j$ may capture most of the gain. That combination is our first ablation paired with marker selection, and we will report it as a separate arm. If it matches the full model, the sparse correction is unnecessary and we say so. We expect it to fail for cell types whose distinguishing markers are mostly species-specific, since 139 conserved markers against roughly 2,500 species-specific ones suggests fine subtypes will lack conserved markers.
Proposed Experiments
The first experiment uses pancreas. We deconvolve human islet pseudo-bulk built from human pancreas scRNA-seq profiled across labs and protocols (the resources used by SCDC and BERMUDA), using the mouse pancreas atlas from the scAdapt cross-species benchmark as the only reference, and score proportion error against known composition. The decisive real-data check is the 77-sample human islet cohort. Using mouse references only, does the estimated beta-cell fraction still correlate negatively and significantly with HbA1c, as SCDC's human-reference ensemble achieved at $p = 0.0019$? The naive pipeline should lose this association; the corrected model should recover it.
Second, we vary compositional variance in pseudo-bulk cohorts and cohort size $N$ to test the identifiability prediction: the $\delta$ recovery error should fall as proportion variance and $N$ grow. We will also plant known synthetic divergence into a human reference to measure how precisely and completely the model recovers it.
Third, as a stress test, we deconvolve pseudo-bulk from the Wu et al. breast atlas, from 5% to 95% purity, using a mouse mammary reference for non-malignant cell types. This asks whether species divergence worsens the purity trap, especially in luminal lineages, where the benchmark already found subtype-driven misassignment.
References
- CACIMAR: cross-species analysis of cell identities, markers, regulations, and interactions using single-cell RNA sequencing data
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- SCDC: Bulk Gene Expression Deconvolution by Multiple Single-Cell RNA Sequencing References
- scAdapt: virtual adversarial domain adaptation network for single cell RNA-seq data classification across platforms and species
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures
- Identification of conserved gene expression features between murine mammary carcinoma models and human breast tumors
- Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures
- BERMUDA: a novel deep transfer learning method for single-cell RNA sequencing batch correction reveals hidden high-resolution cellular subtypes