Evolution as a Composition-Free Anchor for Guessing What Uncharacterized Genes Do in Each Cell Type

Thousands of human genes have almost no functional annotation. Two cheap sources of evidence could fill the gap. One is phylogenetic profiling: record, for each gene, which species carry a copy of it, and treat genes that are gained and lost together across evolution as likely partners. The other is co-expression: genes whose activity rises and falls together are guessed to work together. Each source has a known blind spot. Phylogenetic profiles are one fixed vector per gene, so they cannot say in which cell type a conserved module is actually used. Co-expression measured in tissue samples is dominated by cell mixture. Two genes can co-vary across samples simply because both mark an abundant cell type whose share changes from sample to sample. Averaging over mixed cells hides the real signal and invents fake links.

This note proposes a test and a model. The test measures how much expression-based guessing about gene function reflects cell-type composition rather than function. The model uses evolutionary co-occurrence as a signal that composition cannot touch. It is a single-cell topic model whose gene embeddings are tied to phylogenetic profiles, and whose topics are fit on variation within each cell type. For an uncharacterized gene, the output is which conserved module it joins and in which cell types that module is active. That output is a cell-type-resolved gene set, which curated enrichment databases do not provide.

Core Claim

We make three claims. First, guilt-by-association from expression, whether computed on bulk samples or across single-cell clusters, largely recovers "same cell type" rather than "same pathway," and the confound can be measured. Second, removing cell-type means before learning modules, and anchoring gene embeddings to phylogenetic profiles, recovers held-out functional annotations better than either signal alone and better than simply averaging the two signals. Third, only the joint within-cell-type model gives context-specific module memberships. We will check those memberships against an independent physical map of protein assemblies.

Relation to Prior Work

Several single-cell methods already build pathway knowledge into the model. UNIFAN routes an autoencoder's decoder through a binary GO/KEGG gene-set matrix. The pathway-informed variant of scETM fixes its gene-embedding matrix to a curated pathway incidence matrix. DeepGSEA learns prototype heads, one per predefined gene set. All three can only reason about genes that curators have already placed in sets, and a gene outside every set is invisible to them. Our model supplies the method they lack, a way to embed unannotated genes, by using scETM's linear tri-factorized decoder with evolutionary rather than curated gene embeddings.

The motivation for targeting unannotated genes comes from the understudied-proteins review, which documents annotation bias toward a small set of well-studied proteins. It also comes from DeepSyn. DeepSyn works around database incompleteness by mining literature, but literature is itself biased toward studied genes, and DeepSyn operates on bulk gene lists.

The evolutionary signal comes from classical profiling work. Binary presence/absence profiles with alternative scores such as inverse homology and cross-species cluster co-conservation supply the method. The bacterial genomic-context comparison and the plasmid profiling study constrain it. They show that accuracy depends on which reference genomes are chosen and is degraded by paralogy and horizontal transfer, so reference-panel choice must be an explicit experimental factor.

On composition masking, deconvolution models such as POPPY and scMoE recover cell-type proportions and marker programs from bulk data. Neither asks whether gene–gene links are artifacts of those proportions, and neither infers function for uncharacterized genes. We borrow their pseudo-bulk logic to build a controlled confound test.

Finally, the multimodal U2OS cell map resolved 275 assemblies and assigned functions to 975 uncharacterized proteins, but in one cell line under static conditions. We use it as an external check that owes nothing to expression or phylogeny. We do not try to extend it.

Research Program

The model follows scETM. The expected expression of gene $g$ in cell $c$ is proportional to $\exp(\theta_c^\top \alpha \rho_g + \lambda_{k(c),g})$. Here $\theta_c$ is the cell's mixture over topics, $\alpha$ maps topics into a gene-embedding space, $\rho_g$ is gene $g$'s embedding, and $\lambda_{k(c),g}$ is an intercept for gene $g$ in cell type $k(c)$. Because this intercept is indexed by cell type rather than by batch alone, the topics are forced to explain covariation within cell types. That removes the composition confound from the model's structure. The gene embedding is tied to evolution:

$$ \rho_g = W p_g + r_g, \qquad \text{penalty } \gamma \lVert r_g \rVert^2 , $$

where $p_g$ is gene $g$'s phylogenetic profile across a reference panel of eukaryotic genomes, $W$ is a learned projection, and $r_g$ is a penalized gene-specific residual whose weight is set by $\gamma$.

Unannotated genes enter on equal footing with annotated ones. Their placement is pulled toward evolutionary partners and adjusted by within-type co-expression. Module membership comes from each gene's loading on each topic. Cell-type context comes from how much each topic's activity $\theta$ varies within each cell type.

Three analyses surround the model. The first is a composition-confound audit. We build pseudo-bulk samples from annotated single-cell data, as in the Kassandra-style simulations used by POPPY, with controlled cell-type proportions. We then measure how the share of top co-expression links that are "same-marker" pairs grows with proportion variance. The second is a reference-panel sweep over taxonomic breadth of the genomes used to build $p_g$, following the sensitivity reported for bacterial profiling. The third is a set of ablations over $\gamma$ and over the cell-type intercept, which isolate the two key ingredients.

Baselines

The naive reference point has two parts. One is guilt-by-association with no learning: assign each held-out gene to the GO terms of its nearest neighbors under phylogenetic-profile correlation alone. The other does the same under Pearson co-expression computed across all cells, which is the composition-confounded version. These show what each raw signal gives.

The strongest direct competitor is late fusion. It computes the phylogenetic neighbor score and a within-cell-type co-expression neighbor score separately, then averages their calibrated ranks. It uses the same data and the same deconfounding as our model, with no joint embedding, so it is the fairest threat.

The strongest published learned model is p-scETM from the scETM paper. To give it a chance on unannotated genes, we run it with pathway-fixed embeddings for annotated genes and embeddings learned freely from expression for the rest. We also include UNIFAN, which is known for clustering on pbmc28k, where UNIFAN reports an adjusted Rand index of 0.81 against 0.38 for Leiden. Clustering is not our target task. We report it only to confirm that the evolutionary tie does not damage cell representations. On our main metric, annotation recovery for held-out genes, no reported result was found on Lacuna for any of these methods, so every baseline will be run by us.

For fairness, all methods see identical count matrices, cell-type labels, and held-out splits. All receive the same hyperparameter search budget, tuned on a validation split of annotated genes that is disjoint from the test genes.

Two ablations isolate the proposal's ingredients. One sets $\gamma \to \infty$, leaving phylogeny only through $W$. The other replaces the cell-type intercept with a batch-only intercept, which restores the composition confound.

Beating the baselines means two things. The model must show a significant gain in area under the precision–recall curve for held-out GO membership over late fusion, with a paired bootstrap across genes. It must also show higher agreement with U2OS assemblies for uncharacterized genes than both p-scETM and late fusion.

Why not just…?

Why not just run phylogenetic profiling first, extend GO sets with the predicted members, and feed them to UNIFAN or ordinary enrichment? This two-step pipeline is cheap and could match us on global annotation recovery. It is exactly the late-fusion baseline, with the phylogenetic part used alone. Its limit is that it yields one context-free membership per gene, because profiles are species-level. It cannot say that a conserved module is recruited in, say, NK cells but not B cells. If late fusion matches our model on held-out annotation, we will report that honestly and narrow the claim to the cell-type-resolved output. That output would then need to prove itself on the U2OS agreement test and the within-type test described below.

Why not just use single-cell co-expression, since it already measures function in context? The composition audit is designed to answer this. Our hypothesis is that across-cell co-expression mostly recovers marker pairs. Within-type co-expression fixes that, but it is noisy for lowly expressed genes, and dropouts are severe for exactly the understudied genes we care about, as the scETM paper stresses. The within-type co-expression baseline and the $\gamma \to \infty$ ablation show directly whether phylogeny adds recovery on low-expression genes.

Why not just use gene embeddings from a single-cell foundation model? Tabula shows such models annotate cells well, reporting 0.9810 accuracy on hPancreas against 0.9680 for scGPT. Their gene embeddings, however, are learned from the same composition-confounded expression. We will add a Tabula-style gene embedding as a neighbor-based baseline if one can be trained on our atlases within budget. We expect it to behave like across-cell co-expression in the audit, and this is testable rather than assumed.

Proposed Experiments

The primary benchmark uses the Tabula Muris atlas, which spans many tissues. Mouse genes are profiled against a eukaryotic reference panel. We hold out GO annotations for 20% of annotated genes under two splits. The random split masks genes uniformly. The bias-mimicking split masks the least-cited, lowest-expressed genes, so it resembles real understudied genes. The metric is AUPRC. pbmc28k serves as a second, human dataset, and clustering ARI there acts as a sanity check against UNIFAN's 0.81.

The composition audit uses pseudo-bulk mixtures drawn from pbmc28k and from the Curated Cancer Cell Atlas used by POPPY. For external validation, the model's module assignments for human uncharacterized genes are compared with U2OS assembly co-membership using enrichment odds ratios. This comparison uses a signal neither input saw.

The decisive checks are the following. The same-marker link fraction must rise with proportion variance for across-cell co-expression but not for our model; if it does not, the composition-masking premise fails. The model must beat late fusion on the bias-mimicking split; if it does not, the joint-embedding claim fails and we fall back to the context-resolved output. Finally, a within-type test must show that topics active in only some cell types are supported by U2OS assemblies at rates above a permuted-context control. The permuted-context control shuffles which cell types each topic is assigned to, keeping topic sizes fixed. If support does not exceed that control, the "in which cell type" half of the claim fails.