acceptodds
Under review as a conference paper at ICLR 2027

Cross-Modality Representation Interpretation for Single-Cell Language Models

Abstract

Single-cell foundation models continue to improve on standard benchmarks, but it remains unclear whether their internal representations encode biological knowledge or merely capture subtle expression patterns correlated with benchmark labels. We investigate this question in Cell-to-Sentence-Scale models, which represent cells as gene-name sentences and thus provide a natural setting for examining whether language-style training helps models capture biological knowledge. Using two human peripheral blood mononuclear cell datasets, we train sparse autoencoders (SAEs) on gene-level residual-stream activations from several intermediate layers, and evaluate the learned feature dictionaries along three axes: representation quality, downstream utility, and biological interpretability. These learned features are broadly utilized, predominantly selective, and nonredundant, with these properties persisting under distributional shifts in the input data. Replacing gene embeddings with their SAE reconstructions largely preserves downstream task performance, suggesting that the information discarded by the SAEs has limited relevance to the tasks evaluated. Gene Ontology enrichment analysis further reveals that feature semantics sharpen with layer depth, progressing from generic immune to lineage-specific programs. Steering with interpretable biological concepts shifts cell-type predictions in the intended directions, providing further evidence that the foundation model learns biologically meaningful representations. This control varies across layers, while SAE-reconstructed representations preserve the model’s responsiveness to feature-level perturbations. Together, these findings provide a practical framework for interpreting cross-modal single-cell foundation models and diagnosing their learned representations.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
single-cell, interpretability

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_msqrOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the biological interpretability of Cell-to-Sentence (C2S) single-cell language models using sparse autoencoders (SAEs). The authors train SAEs on intermediate residual-stream activations from C2S Gemma-2-2B and examine their reconstruction fidelity, feature organization, and biological relevance. The learned features are associated with cell-type markers and Gene Ontology terms, while intervention experiments investigate whether manipulating these features can steer cell-type predictions. Experiments on PBMC datasets show that SAE reconstruction largely preserves cell-type annotation accuracy, deeper-layer features exhibit greater cell-type specificity, and selected feature interventions can redirect model predictions. Additional analyses examine feature utilization, redundancy, and reconstruction robustness across datasets.

Critical strengths
  • Interesting application of mechanistic interpretability to single-cell language models.
  • Complementary reconstruction and intervention analyses.
Critical weaknesses
  • Limited generalizability of the experimental evaluation. The study focuses exclusively on C2S Gemma-2-2B and three relatively well-separated immune cell types. It remains unclear whether the observed feature organization and steering behavior extend to other single-cell foundation models, more heterogeneous cell populations, or finer-grained biological distinctions.

  • Insufficient validation of biological feature interpretation. The authors associate SAE features with genes presented in isolation and use marker associations and GO enrichment to establish biological relevance. However, this procedure may primarily recover known gene semantics and cell-type markers rather than context-dependent biological knowledge. The relationship between isolated-gene feature associations and feature activation in complete cellular contexts is not sufficiently established.

  • Lack of appropriate controls for feature steering. The steering experiments demonstrate changes in cell-type predictions but do not compare against matched random-feature or alternative-direction interventions. Moreover, some interventions produce substantial off-target prediction changes, as illustrated in Appendix J.2. Without stronger controls and a systematic assessment of target-specific versus unintended changes, the results provide limited evidence for precise biological control.

  • Limited downstream validation of SAE representation fidelity. The main reconstruction experiments primarily evaluate annotation of three easily distinguishable cell types, where the original model already achieves approximately 96–97% accuracy. Preserving performance on this relatively simple task does not establish that SAE reconstruction retains more complex biological information, such as cellular states, regulatory programs, or perturbation responses. This limits the evidence supporting broader claims about the utility of the reconstructed representations.

Questions for authors
  • Could the authors provide a matched random-feature or random-direction steering baseline? Comparing target-directed and off-target prediction changes would help establish whether the observed interventions are specific to the identified biological features.

  • How consistent are the feature-to-gene associations when genes are evaluated within complete cell contexts rather than in isolation? Clarifying this distinction would strengthen the interpretation of the GO enrichment results.

  • Could the authors quantify the proportion of steering-induced changes that reach the intended target label versus unrelated labels across different layers? The supplementary results suggest substantial variation in steering specificity.

6 comments
AuthorsOpenReview ↗

Reply to Reviewer msqr

Reply to Q1: random-direction steering baseline

We thank the reviewer for this suggestion and have added a matched random-direction control (Section 5.3, Figure 7). For each set of 15 marker features, we sample 15 Gaussian random directions, remove their projection onto the span of the selected features, and rescale them to the mean decoder norm. The controls thus match marker steering in the number of directions, norm, intervention layer (layer 10), and steering magnitude (40). The 36 marker conditions cover three target cell types, marker sets from three layers, two datasets, and two splits; each is paired with three random seeds, giving 108 control runs.

Random directions do perturb the model, but not toward the target. They change the predicted annotation of 17.5% of cells on average (range 1.0-51.7% across runs), compared with 50.7% (1.7-88.9%) for marker steering. Of the 6985 annotations they change, only 55 (0.8%) move to the target cell type, never more than 3 in a single run, and the average count of net change in target is -21.1 cells. Instead, 97.6% of the changed cells receive labels outside the three studied cell types, mainly dendritic cell, leukocyte, myeloid cell, and CD8+ T cell.

In contrast, 72.0% of the annotations changed by marker steering (4764 of 6616) move to the target cell type. Marker steering raises the target count in 35 of 36 conditions, by +127.4 cells on average; the only exception is layer-15 B-cell features on the purified-PBMC test split (-10 cells).

Target cell typeCells changed: randomCells changed: markerChanged cells reaching target: randomChanged cells reaching target: marker
B cell17.2% (1.0–44.0)51.3% (25.7–84.0)14 / 2,301 (0.6%)1202 / 2,221 (54.1%)
CD14 monocyte17.4% (1.0–47.3)57.8% (42.7–74.3)9 / 2,296 (0.4%)1,993 / 2,489 (80.1%)
NK cell18.0% (1.0–51.7)43.0% (1.7–88.9)32 / 2,388 (1.3%)1,569 / 1,906 (82.3%)
Overall17.5% (1.0–51.7)50.7% (1.7–88.9)55 / 6,985 (0.8%)4764 / 6,616 (72.0%)

The labels that random directions produce depend on the random seed, not on the marker set the directions were orthogonalized against. Seeds 42 and 1234 send most changed cells to dendritic cell (61% and 84% of changes), which is the most common new label in 69 of their 72 runs regardless of the target cell type. Seed 0 instead yields mostly myeloid cell, leukocyte, and CD8⁺ T cell. The share of cells changed also follows the seed: 1.0-8.1% for seed 42, 1.7-17.6% for seed 1234, and 29.7-51.7% for seed 0.

The shift toward the intended cell type therefore depends on the identified feature directions and is not a generic response to perturbing the residual stream.

Reply to Q2: feature-to-gene associations in context

We thank the reviewer for raising this distinction. The associations hold in context at layers 10 and 15 but not at layer 20, and we have added an analysis to demonstrate this distinction (Section 4.3, Figure 4).

We define feature-to-gene associations from isolated genes because an in-context gene token attends to the genes preceding it, so only a context-free input attributes a feature's response to gene identity alone. To further test how these associations behave in context, we track 65 marker genes across up to 50 occurrences each in held-out cells.

At layers 10 and 15, a marker-annotated feature separates its gene's tokens from all other gene tokens almost perfectly (median AUROC 1.00 and 0.98), and 59% and 46% of its top-activating tokens fall on its annotated genes, compared with 38% and 32% for random features. At layer 20, this token-level selectivity disappears (median AUROC 0.54; 19% of top tokens, the same as for random features). By this depth, gene representations are dominated by cellular context: occurrences of the same gene in different cells share about as few top features as different genes within the same cell.

The GO enrichments should therefore be read differently by depth. At layers 10 and 15, they describe the genes a feature responds to in context. At layer 20, they describe the cellular program a feature encodes: the marker gene presented alone evokes the feature, whereas in context the feature tracks cell type, consistent with the high cell-type selectivity of layer-20 features (Figure 2). We have revised Sections 2.3 and 4.3 to make this distinction explicit.

AuthorsOpenReview ↗

Reply to Reviewer msqr Cont'

Reply to Q3: on-target rate of steering across layers

We thank the reviewer for this question. We quantified it over all cross-layer steering runs in Table 19: features from layers 10, 15, and 20 added at layer 10, at magnitudes 30, 40, and 50, on both datasets and both splits (108 runs). We fix the intervention at layer 10 because same-layer steering at layers 15 and 20 produces little or no shift toward the target (Figure 5). For every cell whose annotation changes, we record whether the new label is the target, one of the other two studied cell types, or any other label. We report the on-target rate, i.e., the share of changed annotations that reach the target label; this is what the reviewer refers to as steering specificity.

Feature source layerCells changedReach targetReach another studied cell typeReach any other label
103,289 / 13,149 (25.0%)2,458 (74.7%)78 (2.4%)753 (22.9%)
159,362 / 13,149 (71.2%)7,853 (83.9%)0 (0.0%)1,509 (16.1%)
206,827 / 13,149 (51.9%)3,541 (51.9%)21 (0.3%)3,265 (47.8%)
Overall19,478 / 39,447 (49.4%)13,852 (71.1%)99 (0.5%)5,527 (28.4%)

The on-target rate does vary with the source layer, as the reviewer observes. Pooled over cell types, it is highest for layer-15 features (83.9%) and lowest for layer-20 features (51.9%). The ranking depends on the target, however: it is driven by CD14 monocytes and NK cells, whereas for B cells about half of the changes reach the target at every layer. For comparison, random directions reach the target in 0.7–0.8% of changed cells at every source layer (Q1). The table below gives the on-target rate for each target cell type:

Target cell typeLayer 10Layer 15Layer 20
B cell692 / 1,248 (55.4%)1,677 / 3,126 (53.6%)1,239 / 2,208 (56.1%)
CD14 monocyte1,727 / 1,940 (89.0%)2,636 / 2,696 (97.8%)1,336 / 2,625 (50.9%)
NK cell39 / 101 (38.6%)3,540 / 3,540 (100%)966 / 1,994 (48.4%)

With layer-15 features, nearly all changes reach the target for CD14 monocytes and NK cells: for NK, every cell not already annotated as NK is relabeled as NK at all three magnitudes on both datasets, so the annotation saturates. The on-target rate also rises with magnitude at layers 10 and 15 (from 63.3% to 86.7% and from 76.0% to 90.1% between magnitudes 30 and 50) but plateaus near 55% at layer 20.

Most off-target changes are not to unrelated labels. They rarely go to one of the other two studied cell types (0.5% of changed cells) and instead concentrate on a few labels that depend on the source layer and target:

  • Layer 15: 1,426 of the 1,509 off-target changes are malignant cell, all under B-cell steering. With beam-search decoding, B cell is the second-ranked label for most of these cells (Appendix K.2.2).
  • Layer 20, B-cell features: 865 of 969 off-target changes are dendritic cell. These features combine B-cell receptor genes with MHC class II genes, a program shared with dendritic cells.
  • Layer 20, CD14 features: 1,287 of 1,289 off-target changes are myeloid cell, leukocyte, or dendritic cell, i.e. coarser or neighboring labels within the same lineage.
  • Layer 20, NK features: these have the lowest on-target rate. Off-target cells receive CD4 T cell or generic labels (native cell, progenitor cell, leukocyte), and 859 of the 966 target transitions are mature NK T cell, which our matcher accepts as an NK label, so 48.4% is an upper bound (Appendix K.2.4).

The lower on-target rate of layer-20 features therefore reflects programs that the target shares with other cell types (antigen presentation, cytotoxicity) rather than a nonspecific perturbation.

We have also expanded the case studies in Appendix K.2 to document this in more detail. The appendix now reports, for every source layer, magnitude, dataset, and split, the number of changed annotations and the number that reach the target (Table 19). It adds a separate case study for each target cell type that traces which labels the off-target cells receive, and uses beam-search decoding (width 5) to test whether the target label is still among the model's top candidates when it is not the first choice. Finally, it explains why later-layer features redirect annotations (Appendix K.2.5): the most cell-type-specific layer-20 features encode programs shared with other cell types, and the injected vector is five to eight times the class-centroid offset at layer 15 but only about twice that offset at layer 20, so layer-15 features saturate the annotation whereas layer-20 features act as a partial edit.

AuthorsOpenReview ↗

Reply to Reviewer msqr Cont' 2

Additional weaknesses addressed

W1: Limited generalizability of the experimental evaluation

We agree that the evaluation covers one model and three cell types, and we now explicitly acknowledge this limitation in the Limitations section. The restriction to well-separated populations was deliberate, for the reason given in Appendix B.1. Testing whether SAE features align with biological concepts requires reliable ground truth. B cells, CD14 monocytes, and CD56 NK cells have high-confidence labels and well-established marker genes (Zheng et al., 2017), and together they span the lymphoid and myeloid lineages and the adaptive and innate arms of immunity. Finer distinctions, such as naive CD4 versus CD8 T cells, overlap transcriptomically to the point that protein or chromatin measurements are needed to resolve them (Hao et al., 2021). There, a mismatch between a feature and a label could reflect label noise or shared programs rather than the model, and we could not tell these apart. The controlled setting is what allows us to attribute our findings to the model. Extending to the 27B model exceeded our computational budget.

Within this scope, two results bear on generalization. First, SAEs trained only on purified-PBMC transfer to the PBMC-4k and PBMC-8k datasets, which were unseen during SAE training: reconstruction preserves annotation accuracy (Table 1), and steering and the random-direction controls behave consistently on both (Figures 6 and 7). Second, our analyses already expose structure finer than the three labels. Monocyte-selective features carry antigen-presentation programs shared with dendritic cells (Section 4.2), and steered cells that miss the target mostly receive neighboring labels such as dendritic cell and myeloid cell (Appendix K.2; see Q3). This suggests the framework can resolve finer distinctions where the labels support them.

Beyond these specific findings, we do not claim that these specific layer-wise findings generalize to other models or tissues. Rather, we encourage future work to apply this pipeline to investigate what biological information different models learn, across datasets of varying annotation granularity. The pipeline requires only gene-token activations and cell-type labels, and we release our code so that it can be readily adapted to new settings.

W2: Annotation is a simple task for validating SAE reconstruction

We agree on what the annotation result does not show, but we respectfully disagree that this weakens the role reconstruction plays in our study.

Cell-type annotation is not an incidental benchmark for this model. It is one of the core tasks on which Cell2Sentence models are trained and evaluated, from the original Cell2Sentence paper (Levine et al., 2024) to C2S-Scale, and it is the task for which a per-cell output can be scored objectively against ground truth. It is also harder than a three-way classification. The model generates a free-text label from an open vocabulary (its outputs include dendritic cell, naive B cell, and leukocyte), and the SAEs were trained on activations collected without the annotation prompt, so the test also measures transfer across prompt contexts.

The purpose of the test is to license the analyses that follow, all of which concern cell-type identity and the gene programs behind it: feature selectivity, marker and GO associations, and steering of annotations. For this purpose, annotation is the relevant behavior. We do not claim that the reconstructions retain cellular states, regulatory programs, or perturbation responses.

Three further results support the reconstructions beyond annotation accuracy:

  • Task-agnostic fidelity. The SAEs explain 95.1-97.6% of activation variance on held-out cells, more than PCA with the same number of directions (88.9-95.8%; Table 7).
  • Preserved response to interventions. Reconstructed activations respond to feature steering in the same way as the original activations (Figure 5), so reconstruction preserves the causal role of the features and not only the final label.
  • A more complex task. We also evaluated natural-language interpretation of groups of cells (Appendix H). The base model itself produced near-identical text that was not specific to the input cells, so this task cannot distinguish original from reconstructed representations for the 2B model. We excluded it from the main analysis for that reason.
Reviewer_Tv5QOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper uses sparse autoencoders (SAEs) to interpret the internal representations of the C2S single-cell language model. It evaluates reconstruction fidelity, links learned features to biological functions, and tests whether editing these features changes cell-type predictions. Experiments suggest that deeper layers contain more cell-type-selective features, while earlier layers allow stronger steering.

Critical strengths

1.The paper addresses a meaningful question: what biological information do single-cell language models encode beyond prediction accuracy?

2.The framework connects feature interpretation with direct interventions, providing stronger evidence than enrichment analysis alone.

  1. Reconstruction largely preserves annotation accuracy, including on a separate PBMC dataset. The supplementary analyses provide useful diagnostics and openly discuss limitations.
Critical weaknesses

1.Limited generalizability. Experiments cover one model and three well-separated immune populations. This provides limited evidence for broader claims about single-cell foundation models.

2.Insufficient controls for biological interpretation. Changing predicted labels demonstrates control over model outputs, but does not establish biological reasoning. Comparisons with norm-matched random directions, simple cell-type steering vectors, and the original Gemma backbone would strengthen this claim.

3.Uncertain feature semantics. Features are annotated using genes presented individually, whereas training uses genes within cell sentences. It remains unclear whether these annotations reflect the same features in context.

4.Limited novelty and stability evidence. The approach mainly combines existing SAE, enrichment, and steering methods. Single-seed training also leaves the reproducibility of individual features and layer-wise trends uncertain.

Questions for authors

Do the selected SAE features outperform norm-matched random directions and simple cell-type mean-difference vectors for steering?

Are biological feature annotations consistent between isolated-gene inputs and genes within real cell sentences?

Do the main findings hold across SAE training seeds and more closely related cell populations?

Can comparison with the original Gemma backbone clarify which biological associations arise from single-cell training?

5 comments
AuthorsOpenReview ↗

Reply to Reviewer Tv5Q

Reply to W1 and Q3: generalizability

We agree that the evaluation covers one model and three cell types, and we acknowledge this in the Limitations section. The restriction to well-separated populations was deliberate, for the reason given in Appendix B.1. Testing whether SAE features align with biological concepts requires reliable ground truth. B cells, CD14 monocytes, and CD56 NK cells have high-confidence labels and well-established marker genes (Zheng et al., 2017), and together they span the lymphoid and myeloid lineages and the adaptive and innate arms of immunity. Finer distinctions, such as naive CD4 versus CD8 T cells, overlap transcriptomically to the point that protein or chromatin measurements are needed to resolve them (Hao et al., 2021). There, a mismatch between a feature and a label could reflect label noise or shared programs rather than the model, and we could not tell these apart. Extending to the 27B model exceeded our computational budget.

Within this scope, SAEs trained only on purified-PBMC transfer to the PBMC-4k and PBMC-8k datasets, which were unseen during SAE training: reconstruction preserves annotation accuracy (Table 1), and steering and the random-direction controls behave consistently on both (Figures 6 and 7).

On more closely related populations, we did not add populations, but our results indicate what to expect. Selectivity is relative to the populations evaluated (Section 4.2). The layer-20 CD14 features are highly specific in purified-PBMC only because monocytes are the sole myeloid population there. Their top genes include dendritic-cell and macrophage markers, and when used for steering they send cells to myeloid cell and dendritic cell as well as to monocytes (Appendix K.2.3). Likewise, layer-20 NK features carry a cytotoxic program shared with T cells and mostly yield mature NK T cell (Appendix K.2.4). We therefore expect features that encode shared programs not to separate closely related populations, and steering with them to redirect cells among those populations. 

Beyond these specific findings, we do not claim that these specific layer-wise findings generalize to other models or tissues. Rather, we encourage future work to apply this pipeline to investigate what biological information different models learn, across datasets of varying annotation granularity. The pipeline requires only gene-token activations and cell-type labels, and we release our code so that it can be readily adapted to new settings.

AuthorsOpenReview ↗

Reply to Reviewer Tv5Q Cont'

Reply to W2, Q1, and Q4: controls for biological interpretation

We agree that changing a predicted label does not establish biological reasoning, and we do not make this claim. Our claim is narrower: features that carry a biological annotation causally influence the model's output in a direction-specific way, and the annotation predicts the outcome. We address the three requested comparisons in turn.

Norm-matched random directions. We have added this control (Section 5.3, Figure 7). For each set of 15 marker features, we sample 15 Gaussian random directions, remove their projection onto the span of the selected features, and rescale them to the mean decoder norm, so the controls match marker steering in the number of directions, norm, intervention layer (layer 10), and magnitude (40). The 36 marker conditions (three targets, marker sets from three layers, two datasets, two splits) are each paired with three random seeds, giving 108 control runs.

Random directions do perturb the model, but not toward the target. They change the annotation of 17.5% of cells on average, compared with 50.7% for marker steering. Of the 6,985 annotations they change, only 55 (0.8%) move to the target cell type, never more than 3 in a single run; 97.6% receive labels outside the three studied cell types. In contrast, 72.0% of the annotations changed by marker steering (4,764 of 6,616) move to the target, and marker steering raises the target count in 35 of 36 conditions.

Target cell typeCells changed: randomCells changed: markerChanged cells reaching target: randomChanged cells reaching target: marker
B cell17.2% (1.0-44.0)51.3% (25.7-84.0)14 / 2,301 (0.6%)1,202 / 2,221 (54.1%)
CD14 monocyte17.4% (1.0-47.3)57.8% (42.7-74.3)9 / 2,296 (0.4%)1,993 / 2,489 (80.1%)
NK cell18.0% (1.0-51.7)43.0% (1.7-88.9)32 / 2,388 (1.3%)1,569 / 1,906 (82.3%)
Overall17.5% (1.0-51.7)50.7% (1.7-88.9)55 / 6,985 (0.8%)4,764 / 6,616 (72.0%)

Cell-type mean-difference vectors. We have not run a head-to-head steering comparison with mean-difference vectors, but two existing results bear on it (Appendix K.2). First, a direction closely aligned with the class-mean direction can fail to steer: the summed layer-10 NK feature direction has cosine 0.76 with the NK class-mean direction at layer 10 (the offset of the NK centroid from the mean of the class centroids), yet it moves at most 6 cells to NK in any setting. Second, such alignment is not necessary: the layer-15 and layer-20 feature directions are only weakly aligned with the layer-10 class directions (cosine 0.10-0.24), yet layer-15 features added at layer 10 relabel every non-NK cell as NK. The directions that steer most effectively are therefore not the ones a mean-difference vector at the intervention layer would identify. More generally, our aim in steering is not to find the strongest steering vector but to test whether features that carry a biological annotation have a causal role. A mean-difference vector is a single dense direction with no gene-level annotation. SAE features decompose it into annotated programs, and this is what lets us explain where steered cells go: feature sets enriched for a program shared across cell types redirect cells toward the cell types that share it (MHC class II for B cells and dendritic cells, cytotoxicity for NK and T cells; Appendix K.2.5).

AuthorsOpenReview ↗

Reply to Reviewer Tv5Q Cont' 2

Original Gemma backbone. We agree that this comparison would separate associations inherited from text pretraining from those acquired through single-cell training. Our existing results indicate which associations depend on single-cell training:

  • Cell-level associations. The features we interpret are selected by their selectivity for cell types in cell sentences built from measured expression profiles, and validated by steering annotation, the task the model acquires through single-cell training. Steered outputs follow the ontology-style labels of single-cell atlases (classical monocyte, mature NK T cell, native cell), and B-lineage steering at intermediate strength yields malignant cell, which we tentatively attribute to malignant B-lineage cells in the training data (Appendix K.2.2).
  • Context dependence. The same gene name is represented differently depending on the cell it occurs in, and this dependence is structured by cell type at every layer (Section 4.3, Figure 4a). Because the gene name is identical across occurrences, this information comes from the expression-ranked context, the input format introduced by single-cell training, and not from knowledge about the gene name.
  • Gene-level associations at shallow layers. This is where backbone knowledge is most plausible. At layers 10 and 15, features respond to gene identity, and gene names and their functions also appear in the text on which Gemma was pretrained. We therefore do not claim that these gene-level associations arise from single-cell training.

Reply to W3 and Q2: feature annotations in context

The annotations hold in context at layers 10 and 15 but not at layer 20, and we have added an analysis that shows this (Section 4.3, Figure 4).

We define feature-to-gene associations from isolated genes because an in-context gene token attends to the genes preceding it, so only a context-free input attributes a feature's response to gene identity alone. To test how these associations behave in context, we track 65 marker genes across up to 50 occurrences each in held-out cells.

At layers 10 and 15, a marker-annotated feature separates its gene's tokens from all other gene tokens almost perfectly (median AUROC 1.00 and 0.98), and 59% and 46% of its top-activating tokens fall on its annotated genes, compared with 38% and 32% for random features. At layer 20, this token-level selectivity disappears (median AUROC 0.54; 19% of top tokens, the same as for random features). By this depth, gene representations are dominated by cellular context: occurrences of the same gene in different cells share about as few top features as different genes within the same cell.

The GO enrichments should therefore be read differently by depth. At layers 10 and 15, they describe the genes a feature responds to in context. At layer 20, they describe the cellular program a feature encodes: the marker gene presented alone evokes the feature, whereas in context the feature tracks cell type, consistent with the high cell-type selectivity of layer-20 features (Figure 2). We have revised Sections 2.3 and 4.3 to state this distinction explicitly.

Reviewer_j5twOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper looks at whether the internal representations of a single-cell language model capture recognizable biological concepts. The authors train sparse autoencoders on three layers of C2S Gemma-2-2B. SAE reconstructions largely preserve annotation accuracy, while selected deeper-layer features show clearer cell-type specificity.

Critical strengths
  • The paper asks a useful question beyond benchmark accuracy. It looks at what the model represents and whether those representations actually affect its predictions.
  • The reconstruction checks make the interpretation more credible. The authors first show that replacing hidden states with SAE reconstructions largely preserves annotation performance.
  • The layer-wise results are interesting. More selective features appear deeper in the model, but earlier layers are easier to steer, showing that selectivity and steering effectiveness are different things.
Critical weaknesses
  • I noticed a significant formatting violation: the authors have used an incorrect font throughout the paper, which does not comply with the official conference submission guidelines.
  • The novelty is limited, and the analysis does not quite make up for it. SAEs are already widely used in interpretability research, so applying them to biological concepts is reasonable but is not, by itself, a strong contribution.
  • The study covers one model and three well-separated PBMC populations, so it is unclear whether the findings hold for finer cell states or other tissues.
  • Gene associations are obtained from isolated gene names rather than full cell contexts, and steering is not compared with matched random or non-biological directions.

My main concern is the formatting violation. It shows a lack of care in preparing the submission and is irresponsible toward the review process.

Questions for authors

Please refer to the Weaknesses section.

5 comments
AuthorsOpenReview ↗

Reply to Reviewer j5tw

Reply to W1: font

Thank you so much for bringing this to our attention. The deviation was caused by an inadvertently included \usepackage{lmodern} command, which overrode the Times font set by the ICLR style file. We have removed this command and uploaded a revised PDF. A diff against the original confirms that this correction did not introduce any changes to the manuscript’s content. We would also like to clarify that, because Latin Modern is wider than Times, the original formatting did not provide any additional space advantage. The corrected main text remains within the page limit. 

We sincerely apologize for the oversight and any concern it may have caused. We would be very grateful if you would consider evaluating the revised submission on its technical merits. We have also carefully considered your other comments and address them below.

Reply to W2: novelty

We agree that applying SAEs to biological concepts, in itself, does not constitute a sufficiently strong contribution. The novelty of our work lies in the analyses built on the SAEs, which yield findings specific to single-cell language models that, to our knowledge, have not been reported. We have revised the contributions in the Introduction to make this explicit.

  • Gene tokens shift from gene identity to cellular context (Section 4.3). Because C2S represents cells as sentences of gene names, a gene token can encode the identity of the gene or the cellular context in which it occurs. We show that this balance changes with model layer depth: a marker-annotated feature separates its gene's tokens from all others almost perfectly at layers 10 and 15 (median AUROC 1.00 and 0.98,) but not at layer 20 (0.54). This determines how feature-to-gene annotations, and hence GO enrichments, should be read at each layer (see W4a).
  • Selectivity does not imply steerability, and features transfer across layers (Section 5.2). Layer-10 features selected for NK cells fail to steer predictions at any tested magnitude. In contrast, injecting later-layer features at layer 10 succeeds: with layer-15 features, every cell not already annotated as NK is relabeled as NK.
  • Steering effects are direction-specific (Section 5.3). Norm-matched random directions orthogonal to the selected features change many predictions but almost never steer cells toward the target cell type (0.8% of changed cells, against 72.0% for marker features; see W4b).
  • Cell-type features converge across SAE training seeds, and GO enrichment predicts where steering sends cells (Appendices I and K.2). We describe these two analyses under “Additional revisions” below.

Reply to W3: one model and three well-separated populations

We agree that the evaluation covers one model and three cell types, and we acknowledge this limitation in the Limitations section. The restriction to well-separated populations was deliberate, for the reasons given in Appendix B.1. Testing whether SAE features align with biological concepts requires reliable ground truth. B cells, CD14 monocytes, and CD56 NK cells have high-confidence labels and well-established marker genes (Zheng et al., 2017), and together they span the lymphoid and myeloid lineages and the adaptive and innate arms of immunity. Finer distinctions, such as naive CD4 versus CD8 T cells, overlap transcriptomically to the point that protein or chromatin measurements are needed to resolve them (Hao et al., 2021). There, a mismatch between a feature and a label could reflect label noise or shared programs rather than the model, and we could not tell these apart. The controlled setting is what allows us to attribute our findings to the model. Extending to the 27B model exceeded our computational budget.

Within this scope, two results bear on generalization. First, SAEs trained only on purified-PBMC transfer to the PBMC-4k and PBMC-8k datasets, which were unseen during SAE training: reconstruction preserves annotation accuracy (Table 1), and steering and the random-direction controls behave consistently on both (Figures 6 and 7). Second, our analyses already reveal structure beyond the three annotated cell types. Monocyte-selective features carry antigen-presentation programs shared with dendritic cells (Section 4.2), and steered cells that miss the target mostly receive neighboring labels such as dendritic cell and myeloid cell (Appendix K.2). This observation suggests that the framework can resolve finer-grained biological distinctions when supported by the available labels.

Beyond these specific findings, we don't claim that these specific layer-wise findings generalize to other models or tissues. We instead encourage future work to apply the pipeline to investigate what biological information different models learn, across datasets of varying annotation granularity. The pipeline requires only gene-token activations and cell-type labels, and we release our code so that it can be readily adapted to new settings.

AuthorsOpenReview ↗

Reply to Reviewer j5tw Cont'

Reply to W4a: gene associations from isolated genes rather than full cell contexts

The associations hold in context at layers 10 and 15 but not at layer 20, and we have added an analysis to demonstrate this distinction (Section 4.3, Figure 4).

We define feature-to-gene associations from isolated genes because an in-context gene token attends to the genes preceding it, so only a context-free input attributes a feature's response to gene identity alone. To test how these associations behave in context, we track 65 marker genes across up to 50 occurrences each in held-out cells.

At layers 10 and 15, a marker-annotated feature separates its gene's tokens from all other gene tokens almost perfectly (median AUROC 1.00 and 0.98), and 59% and 46% of its top-activating tokens fall on its annotated genes, compared with 38% and 32% for random features. At layer 20, this token-level selectivity disappears (median AUROC 0.54; 19% of top tokens, the same as for random features). By this depth, gene representations are dominated by cellular context: occurrences of the same gene in different cells share about as few top features as different genes within the same cell.

The GO enrichments should therefore be read differently by depth. At layers 10 and 15, they describe the genes a feature responds to in context. At layer 20, they describe the cellular program a feature encodes: the marker gene presented alone evokes the feature, whereas in context the feature tracks cell type, consistent with the high cell-type selectivity of layer-20 features (Figure 2). We have revised Sections 2.3 and 4.3 to state this distinction explicitly.

Reply to W4b: matched random-direction steering control

We have added a matched random-direction control (Section 5.3, Figure 7). For each set of 15 marker features, we sample 15 Gaussian random directions, remove their projection onto the span of the selected features, and rescale them to the mean decoder norm. The controls thus match marker steering in the number of directions, norm, intervention layer (layer 10), and steering magnitude (40). The 36 marker conditions cover three target cell types, marker sets from three layers, two datasets, and two splits; each is paired with three random seeds, giving 108 control runs.

Random directions do perturb the model, but not toward the target. They change the predicted annotation of 17.5% of cells on average (range 1.0–51.7% across runs), compared with 50.7% (1.7–88.9%) for marker steering. Of the 6,985 annotations they change, only 55 (0.8%) move to the target cell type, never more than 3 in a single run, and the net change in target count averages −21.1 cells. Instead, 97.6% of the changed cells receive labels outside the three studied cell types, mainly dendritic cell, leukocyte, myeloid cell, and CD8⁺ T cell.

In contrast, 72.0% of the annotations changed by marker steering (4,764 of 6,616) move to the target cell type. Marker steering raises the target count in 35 of 36 conditions, by +127.4 cells on average; the one exception is layer-15 B-cell features on the purified-PBMC test split (−10 cells).

Target cell typeCells changed: randomCells changed: markerChanged cells reaching target: randomChanged cells reaching target: marker
B cell17.2% (1.0–44.0)51.3% (25.7–84.0)14 / 2,301 (0.6%)1,202 / 2,221 (54.1%)
CD14 monocyte17.4% (1.0–47.3)57.8% (42.7–74.3)9 / 2,296 (0.4%)1,993 / 2,489 (80.1%)
NK cell18.0% (1.0–51.7)43.0% (1.7–88.9)32 / 2,388 (1.3%)1,569 / 1,906 (82.3%)
Overall17.5% (1.0–51.7)50.7% (1.7–88.9)55 / 6,985 (0.8%)4,764 / 6,616 (72.0%)

The labels that random directions produce depend on the random seed, not on the marker set the directions were orthogonalized against. Seeds 42 and 1234 send most changed cells to dendritic cell (61% and 84% of changes), which is the most common new label in 69 of their 72 runs regardless of the target cell type. Seed 0 instead yields mostly myeloid cell, leukocyte, and CD8⁺ T cell.

The shift toward the intended cell type therefore depends on the identified feature directions and is not a generic response to perturbing the residual stream.

AuthorsOpenReview ↗

Reply to Reviewer j5tw Cont' 2

Additional revisions

Beyond the points above, we add two additional analyses in the revision.

Reproducibility across SAE training seeds (Section 2.3, Appendix I). SAEs trained with different seeds are known to learn different dictionaries (Paulo & Belrose, 2026), so we trained SAEs with three seeds at every layer and repeated the feature selection. The cell-type features are largely reproducible. In every layer and cell type, 28–43 of the 45 top features (15 per seed) are linked to a top feature for the same cell type in another seed, and at least three recur in all three seeds. The closest counterpart of a top feature is itself a top feature for the same cell type in 47–67% of cases, compared with at most 4% for frequency-matched controls. The seeds also recover the same marker genes (overlap 0.84–1.00, against 0.25–0.39 for the remaining top genes). Our interpretations therefore do not depend on a single training run.

Where steered cells go (Appendix K.2). For all cross-layer steering runs (features from layers 10, 15, and 20 added at layer 10; magnitudes 30–50; both datasets and splits; 108 runs), we measured the on-target rate, i.e. the share of changed annotations that reach the target label.

Feature source layerCells changedReach targetReach another studied cell typeReach any other label
103,289 / 13,149 (25.0%)2,458 (74.7%)78 (2.4%)753 (22.9%)
159,362 / 13,149 (71.2%)7,853 (83.9%)0 (0.0%)1,509 (16.1%)
206,827 / 13,149 (51.9%)3,541 (51.9%)21 (0.3%)3,265 (47.8%)

Pooled over cell types, the on-target rate is highest for layer-15 features and lowest for layer-20 features, although for B cells about half of the changes reach the target at every layer. The off-target changes are mostly not to unrelated labels, and the expanded case studies in Appendix K.2 trace them for each target, using beam-search decoding (width 5) to inspect the five most probable labels:

  • B cells: layer-15 features yield malignant cell at intermediate magnitude, with B cell as the second-ranked label for most of these cells; a stronger push yields B cell. Layer-20 features, which combine B-cell receptor genes with MHC class II genes, send most off-target cells to dendritic cell.
  • CD14 monocytes: layer-20 features move cells along the lineage, from leukocyte to myeloid cell to monocyte as magnitude increases, with most stopping at myeloid cell.
  • NK cells: layer-10 features fail because 12 of the 15 are dense features active in every cell type. Layer-20 features carry a cytotoxic program shared with T cells and mostly yield mature NK T cell.

Across these cases, the GO enrichment of a feature set predicts where steering sends the cells: sets enriched for a program shared across cell types redirect cells toward the cell types that share it. Target-directed steering requires features whose programs are specific to the target, applied at sufficient strength.

References

  • Yuhan Hao, Stephanie Hao, Erica Andersen-Nissen, William M. Mauck, Shiwei Zheng, Andrew Butler, Maddie J. Lee, Aaron J. Wilk, Charlotte Darby, Michael Zager, et al. Integrated analysis of multimodal single-cell data. Cell, 184(13):3573–3587, 2021.
  • Daniel Levine, Syed Asad Rizvi, Sacha Lévy, Nazreen Pallikkavaliyaveetil, David Zhang, Xingyu Chen, Sina Ghadermarzi, Ruiming Wu, Zihe Zheng, Ivan Vrkic, et al. Cell2Sentence: Teaching large language models the language of biology. In Proceedings of the 41st International Conference on Machine Learning, PMLR 235, 2024.
  • Gonçalo Paulo and Nora Belrose. Sparse autoencoders trained on the same data learn different features. In The Fourteenth International Conference on Learning Representations, 2026.
  • Grace X. Y. Zheng, Jessica M. Terry, Phillip Belgrader, Paul Ryvkin, Zachary W. Bent, Ryan Wilson, Solongo B. Ziraldo, Tobias D. Wheeler, Geoff P. McDermott, Junjie Zhu, et al. Massively parallel digital transcriptional profiling of single cells. Nature Communications, 8(1):14049, 2017.

Comments on the paper

3 comments
~Nicolas_Yanovsky1OpenReview ↗

Related work on SAE interpretation and steering in scFMs

Hi, I came across this paper skimming through single cell submissions for ICLR. I'd like to respectfully point out a closely related contribution of mine presented at two bio workshops (MLGenX and Gen2) at last year's ICLR, titled “Dissecting and Steering Cell Identity in a Single-Cell Foundation Model Using Sparse Autoencoders”. (https://iclr.cc/virtual/2026/10021643)

Our study's main contribution was training TopK SAEs on the gene-level residual-stream activations of a scFM, where we interpreted features through Gene Ontology enrichment, and intervened on those features to assess their influence on predicted gene expression. Even though your study uses another model which concerns natural language + single-cell modeling, and has several different experiments, there's decent overlap in the interpretation and steering approach. I'd very much appreciate if you could cite our work which I think would help readers understand the shared methodology and your specific contributions.

Thank you for bringing your work to our attention. We agree that the overlap in how our studies use sparse autoencoders to interpret and steer single-cell foundation models makes it a relevant reference. We’ve added a citation to your paper in the Related Work section of the revised manuscript to help readers better understand the connection between the two studies.

Dear Area Chair and Reviewers,

As the rebuttal period draws to a close, we thank the AC for encouraging discussion and all reviewers for their thoughtful feedback. We are especially grateful for the reviewers’ recognition of the strengths of our work. Their suggestions have helped us improve the manuscript, and we have addressed each concern and described the corresponding revisions in our responses. We hope the reviewers can take these responses and revisions into account during their discussion with the AC in the next phase. We appreciate the time and care that everyone has devoted to our submission.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.