Where Cross-Species Cell Annotation Breaks: Soft Orthology in a Distilled Hybrid Retriever

Labelled single-cell reference data, meaning per-cell measurements of which genes are switched on, exist mostly for humans and mice. A common way to annotate cells from another species is to translate its genes into human genes and then reuse a human pipeline. A recent example of such a pipeline is scRAG. It writes each cell as a "sentence" of its top 100 expressed genes, retrieves similar labelled cells and candidate cell types, and asks a large language model to choose a label. Each step makes a hard decision. A gene with no clean one-to-one human counterpart (ortholog) is dropped. Genes ranked below 100 are cut. Only about four neighbour cells and two candidate types go forward. The model can only choose among what survived. Work on biomedical question answering has shown how costly this kind of chain is: feeding a reader retrieved text instead of gold text cost 7 to 14 absolute points and in one case halved strict accuracy (Measuring Domain Portability and Error Propagation in Biomedical QA). No one has measured where the chain breaks when the query cell comes from another species.

This note proposes two things. The first is a diagnostic: substitute gold inputs one stage at a time to locate where cross-species annotation loses accuracy. The second is a model aimed at the stage we expect to dominate, the gene-translation step. It is a compact retrieval-and-annotation model with three features. It replaces hard one-to-one translation with soft, learned gene correspondences. It combines exact gene-name matching with learned embeddings in a single index. It is trained by distillation, meaning a large "teacher" model's judgments are used as training targets for a smaller "student." If the diagnostic shows that translation is not the bottleneck, the method half of the proposal is redirected. The diagnostic result stands on its own either way.

Core Claim

We hypothesise that accuracy in cross-species cell annotation is lost mainly at the gene-translation stage, not at retrieval or label selection, and that this loss grows with evolutionary distance from the reference species.

If that is true, two changes should recover most of the gap between hard-translated queries and a within-species upper bound. The first is to let each query gene spread its weight over several reference genes through a learned correspondence. The second is to replace the hard choice of four neighbours with a weighted vote over many neighbours.

The model draws on three existing ideas, adapted to genes rather than words:

  • Cross-lingual distillation as a template. Cross-lingual Knowledge Transfer via Distillation for Multilingual Information Retrieval aligns token embeddings between an English teacher and a multilingual student using optimal transport (a method for finding the cheapest way to move weight from one set of items to another) over parallel sentence pairs. We treat species as languages and orthology tables as noisy word alignments. Matched cell types across species play the role of parallel text.
  • One index for exact and learned matching. Hybrid First-stage Retrieval Models for Biomedical Literature shows that keyword scoring (BM25) can be written as a dot product over sparse vectors. Concatenating those sparse vectors with dense embeddings lets one similarity search do both kinds of matching.
  • Retrieval as a latent variable. Claim-Dissector treats evidence relevance as a hidden variable trained jointly with the final prediction, which removes the hard boundary between ranking and classifying.

None of these three works addresses biology or species. scRAG is human-only and keeps every hard stage.

Research Program

Stage 1: the cascade audit. We copy the gold-versus-predicted design of the biomedical QA study and apply it to a multi-species blood immune-cell atlas spanning 12 species (Molecular mechanisms governing circulating immune cell heterogeneity across different species revealed by single‐cell sequencing), with human blood cells as the reference. We first map the atlas's cell-type labels into one shared label set. Some types will lack a clean match across species, and we report those separately.

For each held-out species, we run a scRAG-style pipeline and then remove one hard boundary at a time:

  • Translation. Compare the cross-species query against a within-species reference, which serves as an upper bound when that species has enough labelled cells. Also compare one-to-one-only translation against translation that keeps one-to-many orthologs.
  • Truncation. Compare top-100 cell sentences against the full expression profile.
  • Retrieval. Insert the gold cell type into the candidate set, as gold snippets were inserted in the QA study.
  • Label choice. Score oracle selection from the retrieved candidates.

We plot the accuracy lost at each stage against evolutionary distance. This audit costs little, needs no new model, and answers a question none of the cited work asks.

Stage 2: the student model. Each reference-species gene gets a learned embedding. For a query species, we compute a soft correspondence matrix $T$ by entropy-regularised optimal transport. Its transport cost is low for annotated orthologs and is refined by co-expression within the matched cell types of the training species. A query gene $g$ is therefore spread over reference genes $h$ with weights $T_{gh}$ instead of being kept or dropped.

The sparse half of the index is a BM25-style weight vector over these spread-out reference genes. It keeps exact gene identity, which dense encoders tend to blur (Sparse, Dense, and Attentional Representations for Text Retrieval). The dense half is a small encoder applied to the expression profile. The two vectors are concatenated into one index, following Ma et al.

At prediction time there is no hard cut-off and no language model. The predicted label is a soft vote over the top $K$ reference cells, with $K$ in the hundreds:

$$p(c \mid x) = \sum_{j=1}^{K} \mathrm{softmax}_j\big(\alpha\, s_{\text{sparse}}(x, r_j) + \beta\, s_{\text{dense}}(x, r_j)\big)\, \mathbf{1}[y_j = c]$$

Here $x$ is the query cell, $r_j$ is the $j$-th retrieved reference cell with label $y_j$, $s$ denotes the sparse and dense similarity scores, $\alpha$ and $\beta$ are learned mixing weights, and $c$ is a cell type. Because the vote is differentiable, the annotation loss trains the retriever directly. This is the Claim-Dissector move applied to cell annotation.

Distillation targets. The student learns from two sources. On human data, it matches the cell-type probability distributions of a large teacher. The teacher is either scRAG with GPT-4 run over full profiles or a human-trained cross-encoder, whichever is stronger on a human validation split. On the training species, it is trained against the atlas's labels. It sees no labels from the held-out species. The intended outcome is a student that runs as a single index lookup, in the latency range Contextualized Query Embeddings for Conversational Search reports for collapsed retrieval (about 11 ms per query, versus more than 500 ms for multi-stage pipelines). Those are text-retrieval figures; we expect the same order of magnitude here but will measure it.

Baselines

Naive reference point. Translate genes one-to-one, build top-100 cell sentences, and run BM25 over gene tokens with a majority vote of the top four neighbours. This is the cheapest version of the existing pipeline and the floor any method must clear.

Strongest existing method. scRAG with GPT-4-Turbo, run on one-to-one-translated query cells with the human reference and its published settings (about four neighbour cells and two candidate types). On human cross-tissue annotation it reports 80.56% Fully Matched accuracy, meaning the predicted label exactly matches the gold label (scRAG). On Lacuna we found no reported cross-species result for scRAG or any comparable system, so its cross-species number must be measured here. It is the strongest competitor because it already combines structured and similarity-based retrieval with a frontier language model. To keep the comparison fair, every system gets the same reference cells, the same held-out species and test cells, and the same label mapping. scRAG also gets a tuning sweep over neighbour count and candidate count matching the number of student hyperparameter configurations.

Dense-only reference. A GenePT-style dense embedding, which scRAG reports at 67.11% Fully Matched accuracy on its human benchmark, run through the same one-to-one translation.

Ablations. These isolate the claimed ingredients by changing one thing at a time:

  • the student with hard one-to-one translation in place of the soft correspondence;
  • the student with a hard top-4 vote in place of the soft vote;
  • sparse-only and dense-only students;
  • a student trained on labels alone, with no teacher distributions.

What counts as beating the baselines. The student must exceed scRAG's measured cross-species Fully Matched accuracy on the more distant held-out species. The soft-translation ablation must account for most of that margin. If the margin comes mainly from the soft vote or from distillation, the core claim is wrong even if the system wins.

Why not just…?

Why not just retrieve more candidates in the existing pipeline? scRAG itself found that more neighbours and candidates hurt the language model; its best setting was about four cells and two types. This is a real threat to the soft-vote half of the proposal, though. A cheap trained reranker over a large candidate pool might recover the same accuracy. We include that variant (one-to-one translation, large $K$, learned reranker) as an extra baseline. If it matches the full student, we conclude the gain comes from retrieval width and drop the claim about soft translation.

Why not just prompt GPT-4 with the species name and let it use what it knows about that species' marker genes? For mouse this may work well, and we expect it to. The open question is whether it degrades for distant species with sparse literature. scRAG reports that language models without structured context hallucinate unfamiliar gene associations. We run this prompt-only baseline on every held-out species and report accuracy against evolutionary distance. A flat curve would undercut the motivation for the student model.

Why not just use exact gene-name matching with ortholog tables? That is the naive baseline. The text evidence suggests keyword matching alone falls short: BM25 reached a MAP of 0.2718 against 0.4056 for the hybrid system on BioASQ (Hybrid First-stage Retrieval Models for Biomedical Literature). Genes do not behave exactly like words, so the audit measures this gap directly rather than assuming it.

Why not just use a pure dense cell embedding? Dense retrievers blur exact identifiers. On BioASQ, a pure dense encoder scored a MAP of 0.1173 against BM25's 0.2718 (Ma et al.). The dense-only and sparse-only ablations test whether the same holds for gene tokens.

Proposed Experiments

We use a leave-one-species-out protocol on the 12-species atlas, with species ordered by evolutionary distance from human. A secondary within-human check on scRAG's PBMC data confirms the student does not lose accuracy at distance zero. Metrics are Fully Matched accuracy, recall of the gold type among retrieved candidates, and latency per cell.

The decisive check is the stage-1 audit combined with the soft-translation ablation. The core claim survives only if two things hold. First, translation is the largest single source of loss for distant species. Second, soft correspondence closes at least half of the gap between hard-translated queries and the within-species upper bound. If retrieval turns out to dominate instead, the audit itself becomes the main contribution, and the method effort shifts to the reranker variant.