acceptodds
Under review as a conference paper at ICLR 2027

FuzzyCell: Differentiable Cell Ontology Reasoning for Graded Cell Type Annotation

Abstract

*How reliable are single-cell foundation models?* We audit scGPT, Geneformer, CellFM, and scCello on Tabula Sapiens and find that of predictions violate a hard biological constraint encoded in the Cell Ontology (): neurons placed in liver, cells assigned simultaneously to mutually exclusive lineages. These errors are invisible to accuracy and F1, yet they follow from elementary reasoning over the formal axioms of . The root cause is that annotation is treated as flat classification over a label set. Even methods that do use use only its graph: random walks, hierarchical classifiers, PageRank, post-hoc probability propagation, hierarchy-weighted losses. The axioms themselves, disjointness and role restrictions that state which cell types cannot coexist and where each type must reside, are discarded. We recast annotation as ontology-constrained entity linking. FuzzyCell retrieves candidate types with a pretrained encoder and re-ranks them with a differentiable reasoner that evaluates extracted axioms, including asserted disjointness axioms that propagate to effective pairs, inside the training loop. Scoring uses a sigmoidal implication in the family of van Krieken et al., which remains trainable where Goguen and Lukasiewicz implications produce zero gradient on positive-measure regions, and satisfies the fuzzy implication boundary conditions at any finite steepness . On Tabula Sapiens, HLCA, and AIDA v2, FuzzyCell reduces axiom violations by over and improves Macro-F1 by to points over the respective backbone, as a backbone-agnostic module adding under ms per cell. Gains concentrate in the long tail: median per-type F1 rises on common types and on rare types, and the improvement is recall-led, with macro-recall gaining points against for macro-precision. Ablations isolate the source: switching off the reasoner at fixed checkpoint changes of predictions, of which corrections outnumber degradations to ; removing tissue context costs Macro-F1 while removing costs ; removing of axioms at random costs . Beyond hard labels, FuzzyCell outputs graded memberships that track differentiation: along two bone marrow trajectories, terminal-type membership correlates with pseudotime at Spearman and .

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
cell type annotation, Cell Ontology, description logic, differentiable fuzzy logic, single-cell foundation models

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_VxFAOpenReview ↗
Rating
3(+1)
Confidence
3(+1)
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes FuzzyCell, an ontology-aware framework for cell type annotation that incorporates Cell Ontology constraints into training and inference. It combines neural candidate retrieval with differentiable fuzzy logic re-ranking and introduces a Sigmoidal Reichenbach implication operator. Experiments on three single-cell datasets report improved Macro-F1, particularly for rare and zero-shot types, and substantially reduced Axiom Violation Rate (AVR).

Critical strengths

1. Meaningful research direction. The paper integrates structured biological knowledge from the Cell Ontology into cell type annotation, potentially improving biological consistency beyond conventional expression-based predictions.

2. Promising empirical results. The method reports consistent improvements across multiple datasets and foundation model backbones, with particularly notable gains for rare and zero-shot cell types.

Critical weaknesses

1. Insufficient justification of the claimed full ontology reasoning.

The consistency score in Section 3.2 primarily relies on binary subsumption indicators. How disjointness, existential restrictions, and tissue context are incorporated remains unclear. Moreover, treating non-entailed relationships as zero may conflict with open-world semantics. Notably, Iσ(a,1)=σ(s/2)I_\sigma(a,1)=\sigma(s/2) is independent of aa, raising questions about how positive semantic evidence contributes to consistency scoring.

2. Theoretical and evaluation inconsistencies.

Proposition 3.1 does not establish a clear advantage over the original Reichenbach implication, which also has a nonzero gradient norm on (0,1)2(0,1)^2. The proposed operator still exhibits vanishing antecedent gradients as b→1b\to 1.

More seriously, the subsumption violation definition in Appendix C.1 appears incompatible with the argmax prediction rule, making this violation term identically zero under the stated definitions, contrary to Figure 2. The reported Wilcoxon significance (p<0.01p<0.01) also appears inconsistent with an exact test based on five paired runs.

Questions for authors

Q1. How are disjointness axioms, existential restrictions, and tissue context explicitly incorporated into the reasoning module? How does the scoring function distinguish non-entailed from incompatible relationships and reward positive semantic evidence, given that Iσ(a,1)=σ(s/2)I_\sigma(a,1)=\sigma(s/2) is independent of aa?

Q2. In what precise mathematical sense does the proposed implication eliminate gradient degeneracy compared with the original Reichenbach implication? The antecedent gradient still contains the factor (b−1)(b-1) and approaches zero as b→1b\to1.

Q3. How is the subsumption violation in AVR calculated? According to Section 3.2, the predicted type is selected by maximizing the final score. Its score therefore cannot be lower than that of the ground-truth type when both are candidates, contradicting the violation condition in Appendix C.1. When the ground-truth type is absent, the violation is explicitly set to zero. Could the authors reconcile this with Figure 2 and explain why some AVR values exceed classification error rates?

Q4. How were the reported p-values below 0.01 obtained from five runs using Wilcoxon signed-rank tests with Bonferroni correction? For five paired observations, the minimum two-sided exact p-value is 0.0625. Please clarify the statistical unit and test procedure.

4 comments
AuthorsOpenReview ↗

Response to Reviewer VxFA

We thank the reviewer! Q2 is answered in the general comment; Q1, Q3 and Q4 follow.


Q1. The full scoring function

Revision (Sec. 3.2, paragraph "Consistency scoring and score fusion", pp. 5-6): the full score is now written out: sonto(x,Cj)=∏τ∈ΓIσ(τ(x),τ(Cj))⋅∏D∈R(Cj)Iσ(1,κ(c,D)).s_{\text{onto}}(\mathbf{x},C_j) = \prod_{\tau\in\Gamma} I_\sigma\bigl(\tau(\mathbf{x}), \tau(C_j)\bigr)\cdot\prod_{D\in\mathcal{R}(C_j)} I_\sigma\bigl(1, \kappa(c,D)\bigr).

  • R(Cj)\mathcal{R}(C_j) collects the fillers of CjC_j's entailed restrictions C_j \sqsubseteq \exists\textit{part\\_of}.D.
  • κ(c,D)=1\kappa(c,D) = 1 iff DD contains the cell's tissue cc in the Uberon part_of closure.
  • The second factor is Definition 2.1 applied to these restrictions, with the tissue sample as the witness of the existential.

(a) How disjointness and tissue act (s=10s = 10). Take a liver cell with epithelial membership 0.9.

CandidateType factor (epithelial cell)Location factor (liver)
NeuronIσ(0.9,0)=0.018I_\sigma(0.9,0) = 0.018Iσ(1,0)=0.0067I_\sigma(1,0) = 0.0067
HepatocyteIσ(0.9,1)=0.993I_\sigma(0.9,1) = 0.993Iσ(1,1)=0.993I_\sigma(1,1) = 0.993
  • "Neuron" is disjoint from "epithelial cell", so it does not entail it, and its restriction to nervous tissue fails in liver. Every other factor is below 1, so sonto<0.018×0.0067≈1.2×10−4s_{\text{onto}} < 0.018 \times 0.0067 \approx 1.2\times10^{-4}.
  • "Hepatocyte" satisfies both.

Revision (Sec. 3.3, text after the definition of Lonto\mathcal{L}_{\text{onto}}, p. 6): the paper now states that Lonto\mathcal{L}_{\text{onto}} uses the same semantics in training: Iσ(C(x)D(x),0)I_\sigma(C(\mathbf{x})D(\mathbf{x}),0) for C⊓D⊑⊥C \sqcap D \sqsubseteq \bot, and Iσ(C(x),κ(c,D))I_\sigma(C(\mathbf{x}),\kappa(c,D)) for location restrictions.

(b) Open world. τ(Cj)=0\tau(C_j) = 0 means "not entailed", not "false". For example, "T cell" does not entail "regulatory T cell", because a T cell need not be regulatory. Formally, a satisfiable EL⊥\mathcal{EL}_\bot concept has a canonical model that is a model of T\mathcal{T}, and T⊨C⊑τ\mathcal{T} \models C \sqsubseteq \tau iff its root lies in τ\tau (Baader et al., 2005; Lutz & Wolter, 2010, Lemmas 12 and 13). So (τ(Cj))τ(\tau(C_j))_\tau is the profile of a genuine model, which commits CjC_j to what is entailed and nothing more.

Revision (Sec. 3.2, paragraph "Type inference", p. 5): this argument is added.

(c) Positive evidence, and why Iσ(a,1)I_\sigma(a,1) is independent of aa. Evidence rewards a candidate relative to its competitors. If two candidates differ only in whether they entail τ\tau, their score ratio is Iσ(a,1)/Iσ(a,0)I_\sigma(a,1)/I_\sigma(a,0):

Evidence aa for τ\tau0.10.50.9
Score ratio1.012.055

Stronger evidence for τ\tau therefore strictly favors the candidates that entail it. The neural term of sfinals_{\text{final}} and Ltype\mathcal{L}_{\text{type}} add positive evidence directly.

Revision (Sec. 3.2, p. 6): the score-ratio argument is added after the definition of sontos_{\text{onto}}.


Q2. What the sigmoid changes

See the general comment (parts 2 and 3).


Q3. The AVR definition

The subsumption term. Thank you for catching the typo. As the sentence after the definition says, the term counts over-specification, that is, predicting a proper subtype of the true type.

Revision (App. C.1, p. 33; summary in Sec. 4.1, p. 7): vsub=1[T⊨y^i⊑yi,y^i≠yi]v_{\text{sub}} = \mathbf{1}[\mathcal{T} \models \hat y_i \sqsubseteq y_i,\\ \hat y_i \neq y_i]. It depends only on labels, so it applies identically to every method.

AVR above the error rate. vrolev_{\text{role}} never uses the label, so it can flag a correct prediction whose type's location restriction fails in the recorded tissue. An example is conjunctival epithelial cells sampled from the cornea in Tabula Sapiens. AVR is therefore not bounded by the error rate.

Revision (App. C.1, p. 33): the definition is followed by this explanation and example.


Q4. Statistical significance

  • Why not a p-value. For the exact signed-rank test with five paired seeds, all 25=322^5 = 32 sign patterns are equally likely under the null. The smallest attainable two-sided p-value is therefore 2/32=0.06252/32 = 0.0625, so a p-value is not the informative summary here.
  • What the tables show. In all 60 backbone-metric comparisons, FuzzyCell's gain is at least twice the sum of the two standard deviations. For example, scGPT Macro-F1 on Tabula Sapiens goes from 52.8±0.652.8 \pm 0.6 to 60.4±0.460.4 \pm 0.4: a gain of 7.6 against a combined spread of 1.0.

Revision (caption of Table 1, p. 8; App. B.6, p. 28): the p-value statement is replaced by this one.

Dear Reviewer VxFA,

Thank you again for such careful questions. Two of them caught real inconsistencies in the submission, and we wanted to say plainly how each one was fixed, in case that helps when you look at the paper again.

  • Q3 (subsumption term in AVR). You were right: the definition as written could not produce the values in Figure 2. App. C.1 (p. 33) now matches the sentence that follows it, so the term counts predictions that are a proper subtype of the true type. The same passage explains why AVR can exceed the error rate.
  • Q4 (significance). You were right here too: five paired runs cannot give p < 0.01. We removed the claim. The caption of Table 1 (p. 8) now reports that every gain is at least twice the sum of the two standard deviations.
  • Q1 and Q2. The full scoring function is written out in Sec. 3.2 (pp. 5-6), and Proposition 3.1 (p. 6) is restated as an exact gradient formula.

One more change since our response. Following a comment from Reviewer 4ZrZ, the paper now says the graded scores are "scored according to their consistency with CL", not that they are consistent with it. That is a more accurate description of what the scoring function in Q1 does.

If any of our answers missed what you were really asking, or if you still have doubts about soundness, please tell us. We would be glad to address them in the time that is left.

Thank you for your time and dedication!

Best regards,
The Authors

Thank you for your detailed response and for your efforts in revising the manuscript. I greatly appreciate the clarifications provided by the authors, as well as the revisions to the theoretical analysis, evaluation metrics, and implementation details. These revisions have addressed my main concerns.

Therefore, I am raising my rating to 3.

Dear Reviewer VxFA,

Thank you for your support and for raising your rating! Your questions made the statement of Proposition 3.1 more precise, and the paper is better for them. We will keep strengthening it.

One small note, in case it is helpful: on our side, the review form still displays the earlier rating (2), while the confidence shows as updated. It may simply be a display delay. If it is not, the rating can be updated through the "Edit" button on the official review whenever it is convenient for you.

Thank you again for your time and for such a constructive discussion. It has meant a lot to us.

Best regards,

The Authors

Reviewer_4ZrZOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The authors present a method, namely FuzzyCell, for cell type identification in single cell transcriptome (scRNA-seq) studies. Their architecture combines several components: an off-the-shelf engine for generating cell embeddings (tipically a published single cell transcriptome foundation model), a Cell Ontology that specifies constraints between cell types (e.g., neurons cannot reside in liver), and a differentiable fuzzy logic reasoner for inference. The work includes theoretical guarantess on the behaviour of the system, as well as a large evaluation on real data.

Critical strengths

The main strength of the work is adapting differentiable fuzzy description logic reasoning to the cell classification problem. Critically, this technique nicely accommodates for the intrinsic uncertainty of cell classification, especially in the contex of cell differentiation where the identity of a cell cannot always be unequivocally assigned to a single type. Furthermore, I would like to commend the authors for the extensive evaluation, which includes a number of baseline methods as well as direct competitors of the proposed technique.

Critical weaknesses

My main criticism is that the authors seem to overstate some of their contributions.

The proposed method is overall an innovative application of previous techniques. However, the authors state that their "key innovation is the Sigmoidal Reichenbach implication, a novel fuzzy operator". I would argue that sigmoid functions in this context have already been proposed at least since the work of van Krieken et al. (2020), a work that the authors cite several times within their manuscript.

Furthermore, in Proposition 3.1 the authors state that the gradient of the sigmoidal Reichenbach implication is strictly nonzero for all (a,b)∈(0,1)2(a, b) \in (0, 1)^2, while the the gradient with respect to a of the standard Reichenbach implication "vanishes on the entire line b = 1". This is confusing: they evaluate the sigmoidal Reichenbach implication in the open domain (0,1)2(0, 1)^2, and criticize the standard Reichenbach implication at b = 1. More importantly, the gradient of the sigmoidal Reichenbach implication would also vanish on the entire line b = 1. In summary, while introducing a sigmoid variant of the Reichenbach implication might have computational advantanges, it does not seem that it provably solve the issue of the vanishing gradient for b = 1.

Emile van Krieken, Erman Acar, and Frank van Harmelen. Analyzing Differentiable Fuzzy Implications. In Proc. KR’20, pp. 893–903, 2020.

Questions for authors

I would ask the authors to answer the concerns rised in the critical weaknesses section, particularly:

  • Clarify that sigmoid variants of the Reichenbach implication have aready been proposed in the past, and that innovative point is their application in the scRNA-seq context for cell classification
  • Better elaborate on the claims regading vanishing gradients for b approaching 1 or b = 1 in the standard and sigmoidal Reichenbach implication

Furthermore, there are a few minor issues that should be addressed as well:

  1. the authors uses in the introduction notations that will only be explained later in the text. For instance, the term C(x) in "FUZZYCELL produces a graded membership vector μ(x) ∈ [0, 1]|C(x)| that captures...", or the term s in "FUZZYCELL’s graded annotations converge to valid crisp EL⊥ models as s → ∞". The authors should avoid using symbols that have not being properly introduced.
  2. Some sentences should be tone down. For example: "Standard metrics (accuracy, F1) are completely blind to these failures". Several "standard" metrics support different costs depending on the the type of error. "Current methods output a single hard label": I am not aware of the full literature regarding cell classification methods for scRNA-seq data, however I would be surprised if all other current methods strictly provide only a single label with no indication of confidence whatsoever.
  3. Some typos are present in the text, for example the first citation should be (Regev et al., 2017), not (A et al., 2017).
3 comments
AuthorsOpenReview ↗

Response to Reviewer 4ZrZ

We thank the reviewer for the careful reading and the encouraging assessment! The summary of the contribution, that differentiable description logic reasoning handles the intrinsic uncertainty of differentiating cells, is exactly how we see the paper.


W1. Attribution of the sigmoidal implication

We agree.

Revision (abstract, p. 1; introduction and contribution list, p. 2; Sec. 3.2, p. 5): the operator is credited to van Krieken et al. wherever it is introduced, and "novel operator" and "key innovation" are removed. The contribution list (p. 2) now names what is ours:

  1. A reasoner that scores each candidate type against the ELK closure of the Cell Ontology, including disjointness axioms and part_of restrictions witnessed by the tissue, and that trains the encoder through these axioms.
  2. Proposition 3.1, the exact gradient of this ontology loss, and Theorem 3.2, its classical limit.

W2. Gradients at b=1b = 1 and as b→1b \to 1

The general comment gives the full argument. In brief:

  • A zero antecedent gradient at b=1b = 1 holds for every fuzzy implication, since I(a,1)=1I(a,1) = 1, and it is semantically correct.
  • Subsumption axioms never reach b=1b = 1 in training.
  • What separates IσI_\sigma from IRI_R is the exponential damping of satisfied cells (property 2), which also describes the behavior as b→1b \to 1.

Revision (Sec. 3.4, pp. 6-7; App. A.1, pp. 16-17): Proposition 3.1 is restated on the closed square with these properties, and part (iv) of the proof treats b=1b = 1.


Minor points

Minor 1 (notation used before it is defined).

Revision (introduction, pp. 1-2; Sec. 2.2-2.3, pp. 3-4): the introduction now states the graded output and the theory in words, without notation. Each symbol is defined at first use in Sec. 2 and 3.

Minor 2 (overstatements).

Revision:

  • (Introduction, p. 1) "Completely blind" → accuracy and F1 score a neuron predicted for a hepatocyte like a confusion between two T cell subsets. Hierarchical measures (Kosmopoulos et al., 2015) grade errors by taxonomic distance, but no standard metric checks disjointness or location axioms.
  • (Sec. 2.1, p. 2) "Single hard label" → classifiers such as CellTypist (Domínguez Conde et al., 2022) report confidences over a flat label set, whereas FuzzyCell's scores are also constrained by the ontology.

Minor 3 (reference).

Revision (References, p. 14): fixed. The Regev et al. entry had a malformed author field.

I would like to thank the authors for their efforts in answering my requests. Overall, I think my main concerns have been satisfactorily addressed, and I have updated my recommendation accordingly.

One additional minor issue the authors should still resolve, in my opinion: FuzzyCell graded scores are not "consistent with CL by construction". The ontology penalty encourages consistency, but does not guarantee that the final graded scores satisfy all CL constraints. I would suggest using alternative expressions like "encouraged to respect CL constraints” or “scored according to consistency with CL".

Dear Reviewer 4ZrZ

Thank you for coming back to this, and for taking the time to read our response so closely. We are glad the main concerns are resolved, and we appreciate the updated recommendation.

You are right about the remaining point, and we have adopted your wording.

Revision (abstract, p. 1; Sec. 2.1, p. 3; Conclusion and Limitations, p. 10): the graded scores are now described as "scored according to their consistency with CL", and the Limitations paragraph says they are "encouraged but not guaranteed to respect CL". The phrase "by construction" no longer appears.

Thank you again for your time and for helping us improve the paper!

Best regards,
The Authors

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

This paper studies cell-type annotation with explicit Cell Ontology constraints. FuzzyCell retrieves candidate types using a pretrained encoder, then re-ranks them with a differentiable fuzzy-logic module. Across three datasets and four backbones, the authors report higher Macro-F1 and substantially fewer ontology violations.

Critical strengths
  • Bringing ontology constraints into annotation is a sensible way to use biological knowledge that flat classifiers often leave out.
  • The evaluation covers several datasets and backbones.
  • The crisp-versus-fuzzy comparison is helpful.
Critical weaknesses
  • The novelty is limited. Combining pretrained cell embeddings with ontology-based scoring is useful, but builds largely on established ideas. The proposed sigmoid modification is fairly simple, and its theoretical analysis does not establish the stronger claims made in the paper.
  • Some figures (such as figure 1) give the impression of raw AI-generated images rather than carefully revised scientific figures. If generative tools were used, the outputs should have been checked and edited for technical accuracy.
  • The main theoretical claim is not supported by the proof. Some propositions establish a nonzero gradient inside the domain, but standard Reichenbach already has this property.
  • The biological meaning of the graded memberships remains untested.
Questions for authors

Please refer to the Weaknesses section.

4 comments
AuthorsOpenReview ↗

Response to Reviewer tzdb

We thank the reviewer! Our novelty is not the sigmoid. It is that the Cell Ontology enters training as a logical theory rather than as a graph, and the crisp-versus-fuzzy comparison the reviewer highlighted is the evidence for this.


1. Novelty

Concern: limited novelty; the approach builds on established ideas.

Step 1. Prior methods use the ontology as a graph.

  • OnClass, CellO, SOCAM and HierCE use the is_a hierarchy, and scCello a PageRank similarity on the graph.
  • None uses disjointness ("no cell is both a neuron and an epithelial cell") or location ("every hepatocyte is part of some liver").
  • Ontology embeddings such as EL Embeddings and Box$^2$EL embed the ontology itself to predict missing axioms; they do not score cells.
  • Generic differentiable logics (Logic Tensor Networks; van Krieken et al.) supply the operators, but evaluate hand-written formulas; they do not say how a 2,743-concept ontology constrains a cell.

Step 2. FuzzyCell compiles the axioms into per-cell constraints.

  • Each candidate gets its type profile from the ELK closure of 6,165 axioms, in which 214 disjointness axioms propagate to 8,246 pairs.
  • This profile is read off the root of the canonical model, so it commits the candidate to what the ontology entails and nothing more.
  • Location restrictions are checked against the sampled tissue.
  • The encoder is trained through the same axioms.

Step 3. Training through them is what pays off (scGPT backbone, Tabula Sapiens).

Use of the same axiomsMacro-F1
None (backbone)52.8
Crisp filtering53.5
Crisp Boolean training mask (zero gradient to the encoder, Table 22)54.8
Fuzzy objective (FuzzyCell)60.4

The fuzzy objective also lifts zero-shot F1 from 22.8 to 38.5 and rare-type F1 from 5.1 to 22.1. The axioms are shaping the representation, not filtering the output.

Revision (Sec. 4.3, "Fuzzy versus crisp use of the same axioms", p. 9; Table 3, p. 10): the comparison moves from App. B.6 into the main text; the new Table 3 lists what each method takes from the ontology.


2. Theory

Concern: the proof shows a nonzero gradient inside the domain, which standard Reichenbach already has.

We agree with this reading of the original statement, and Proposition 3.1 has been restated (general comment). In short, IσI_\sigma damps already-satisfied cells exponentially while keeping the gradient direction of IRI_R, and recovers the RILL hinge as s→∞s \to \infty. The ablation shows the effect (Table 2, p. 9).


3. Figure 1 and presentation

Concern: Figure 1 looks like a raw AI-generated image; generated outputs should be checked and edited for technical accuracy.

We agree, and the figure is replaced.

  • What we did, and why. The original Figure 1 used an image-generation tool for its illustrative artwork. We wanted lifelike cells and tissues in the overview, and could not draw them well by hand. It looked attractive, but its details were not tied to the method.
  • What replaces it. A vector figure drawn in TikZ, in which every displayed number is computed from the paper's equations. The decorative artwork is gone. The figure now shows four things:
    1. retrieval of candidate types in the shared logic space;
    2. the matrix of IσI_\sigma factors, with the types each candidate entails marked;
    3. fusion into the final scores and the graded output;
    4. the gradient field of Proposition 3.1, with the weight wsw_s falling from 9.9 to 0.07 as IRI_R rises from 0 to 1.
  • How it was checked. We verified every element against the equations; the caption says which parts are computed and which are schematic.

Revision (Figure 1, p. 4; AI use statement, p. 11): new Figure 1, drawn in TikZ; the image-generation tool used for the earlier version is recorded in the AI use statement.

Revision (Sec. 2-3, pp. 3-6): symbols are defined before use, less familiar terms are explained, and Section 3 gives the full score and loss.


4. Graded memberships

Concern: the biological meaning of the graded memberships remains untested.

  • What they are. μ(x)\boldsymbol\mu(\mathbf{x}) normalizes the fused scores over the retrieved candidates.
  • What holds by construction. Candidates that the expression data cannot separate keep comparable weight, while a candidate that contradicts an axiom is demoted. In Case 1 of the paper (App. B.4), the neuron candidate of a liver cell gets ontology score 0.003, against 0.91 for hepatocyte, which becomes the prediction.
  • What is a separate question. Whether μ\boldsymbol\mu also orders cells along a trajectory. Testing it on lineage-traced data is a natural next step.

Revision (introduction, p. 2; Limitations, p. 10; App. F, p. 41): the text claims only what holds by construction, and names trajectory validation as open.

Dear Reviewer tzdb,

A quick update that touches your fourth point, on what the graded memberships mean.

Reviewer 4ZrZ pointed out that we should not call the graded scores consistent with CL "by construction", and we agree. The ontology term rewards agreement with CL but does not guarantee it. The paper now says the scores are "scored according to their consistency with CL" and are "encouraged but not guaranteed to respect CL" (abstract, p. 1; Limitations, p. 10). The same goes for the phrase "holds by construction" in our response above, which we would now word more carefully. We think this is closer to the caution you were asking for.

In case it is useful, here is where each of your points landed in the revision:

  • Novelty: Sec. 4.3 (p. 9) and the new Table 3 (p. 10).
  • Figure 1: redrawn in TikZ (p. 4).
  • Theory: Proposition 3.1, restated (p. 6), with the proof in App. A.1.
  • Graded memberships: the Limitations paragraph (p. 10).

Your review made us state our claims more carefully, and we think the paper is better for it. If something still does not sit right with you, we would really like to hear it, and we will do our best to address it in the time that is left.

Best regards,
The Authors

Dear Reviewer tzdb,

As a follow-up to our previous response, this note completes our response to your fourth point, which our first reply left open, and then summarizes where each of your four points now stands. Your comments are quoted in italics.


→ "The biological meaning of the graded memberships remains untested."

This was a fair criticism: the submission described the graded scores in biological terms without testing them. We have now run the test, and the claims in the paper are limited to what it shows.

Setup. Tabula Sapiens test set (121,000 cells), FuzzyCell with the scGPT backbone, five seeds, no retraining. FuzzyCell is trained on CL, so a test against CL alone would be circular. We therefore test against marker genes, which the ontology never sees, and always compare with the backbone over the same candidates.

1. Graded scores follow graded marker expression. For cells whose two leading candidates are a textbook pair, we correlate the balance of the scores, log⁡(μB/μA)\log(\mu_B/\mu_A), with the balance of the defining markers (Spearman ρ\rho).

Pair (A vs. B)Markers (A vs. B)CellsFuzzyCell ρ\rhoBackbone ρ\rho
CD4-positive vs. CD8-positive alpha-beta T cellCD4 vs. CD8A, CD8B3,1470.470.33
Classical vs. non-classical monocyteCD14 vs. FCGR3A (CD16)6380.580.44
Type II vs. type I pneumocyteSFTPC, LAMP3, ABCA3 vs. AGER, HOPX, PDPN1270.520.39

Monocyte subsets are defined by graded CD14 and CD16 levels (Ziegler-Heitbrock et al., 2010), and type II pneumocytes become type I through a transitional state (Strunz et al., 2020). In both cases the score balance moves with the markers, and more closely than the backbone's does.

2. Scores spread where annotators stopped. In blood, some cells are labelled classical or non-classical monocyte and others only monocyte: the experts did not resolve them. Comparing within the same tissue, such parent-level cells (18,236 in total) have more spread-out scores than finely labelled ones in 19 of 22 tissues, with effect size 0.62 (0.5 means no effect), and a median 0.84 of their weight stays on the expert's label and its subclasses.

3. Two supporting checks. When scores are spread out, the two leading candidates are ontological neighbours in 0.71 of cells, against 0.56 for the backbone and 0.07 at random. The score margin detects misclassified cells with AUROC 0.87, against 0.83 for the backbone. For these two checks we report effect sizes only, since with tens of thousands of cells any difference is statistically significant.

Scope. These are snapshots. They show that the graded output follows continuous biology along known axes. They do not show ordering in time, and the paper does not claim it.

Revision: the results above, with the full protocol and per-seed values, will be made more informative and accurate, and then added to App. F in the final version.


Your other three points. For convenience, here is where each one is addressed in the revised PDF:

→ "The novelty is limited. Combining pretrained cell embeddings with ontology-based scoring is useful, but builds largely on established ideas."

Table 3 (p. 10) and Sec. 4.3 (p. 9) now state what FuzzyCell takes from the ontology that prior methods do not.

→ "Some figures (such as figure 1) give the impression of raw AI-generated images rather than carefully revised scientific figures."

Figure 1 is redrawn in TikZ, and every number in it is computed from the equations (p. 4).

→ "The main theoretical claim is not supported by the proof. Some propositions establish a nonzero gradient inside the domain, but standard Reichenbach already has this property."

We agree. The earlier claim is withdrawn, and Proposition 3.1 now states and proves what the sigmoid changes relative to standard Reichenbach (p. 6; App. A.1).


We hope this addresses your concerns. If anything remains unclear or unconvincing, we would be glad to follow up.

Best regards,
The Authors

Hi reviewer tzdb,

Thank you for your reviewing service. The authors have provided fairly extensive rebuttals. Please confirm that you have read them, state which of your concerns are or are not addressed, and whether that changes your recommendations.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the discussion

Dear Area Chair and Reviewers,

Thank you for the discussion. As it closes, we summarize what the paper contributes and how each point raised was resolved.


The idea. Single-cell foundation models annotate cells by flat classification. The Cell Ontology (CL) already states, as logical axioms, what each cell type is: what it is a kind of, what it cannot be, and where it must be located. Our audit shows that one in every five to seven predictions of current models contradicts such an axiom, for example a neuron placed in the liver, and that accuracy and F1 do not see it. FuzzyCell treats annotation as entity linking against CL and makes the axioms part of the training signal.


What is new.

  1. The ontology enters as a logical theory, not as a graph. Earlier ontology-aware annotators use the is_a hierarchy or a graph similarity. FuzzyCell scores every candidate against the ELK closure of CL, including disjointness axioms and part_of restrictions witnessed by the tissue (Table 3).
  2. Training through the axioms matters far more than checking them. With the same axioms and the same backbone (scGPT, Tabula Sapiens), crisp filtering reaches 53.5 Macro-F1 and a crisp training mask 54.8, against 52.8 for the backbone. The fuzzy objective reaches 60.4, and lifts zero-shot F1 from 22.8 to 38.5 and rare-type F1 from 5.1 to 22.1. The axioms shape the representation; they do not merely filter the output.
  3. An exact account of the loss. Proposition 3.1 gives the gradient of the ontology loss in closed form and shows that it damps already-satisfied groundings exponentially, which is the mechanism behind implication bias. Theorem 3.2 gives the classical limit. The operator itself is credited to van Krieken et al.

Evidence. Across three atlases and four backbones, FuzzyCell cuts axiom violations by over 80% and raises Macro-F1 by 4 to 8 points, as a plug-in that adds less than 4 ms per cell.


Where the reviews agree. All three reviews name the direction as a strength: "adapting differentiable fuzzy description logic reasoning to the cell classification problem" (4ZrZ), a "meaningful research direction" (VxFA), and "a sensible way to use biological knowledge that flat classifiers often leave out" (tzdb), who also notes that "the crisp-versus-fuzzy comparison is helpful". That comparison is point 2 above.


How each point was resolved. The last column says whether the resolution rests on a proof, on experiments, or on a correction to the text.

Reviewer 4ZrZ: confirmed as addressed; rating raised.

Point raisedResolution in the revised PDFRests on
Sigmoidal implication presented as novelCredited to van Krieken et al.; contributions restated (pp. 1, 2, 5)Attribution
Gradient claim at b = 1Proposition 3.1 restated on the closed square, with the case b = 1 proved (Sec. 3.4; App. A.1)Proof
Notation, overstatements, a reference, "by construction"Corrected; scores are now "scored according to their consistency" with CL (pp. 1 to 3, 10)Text

Reviewer VxFA: confirmed as addressed; rating raised.

Point raisedResolution in the revised PDFRests on
How disjointness, restrictions and tissue enter the scoreFull scoring function and canonical-model argument (Sec. 3.2)Definition and proof
Advantage over ReichenbachProposition 3.1 (Sec. 3.4); ablation: replacing the operator costs 4.6 Macro-F1 and raises AVR from 2.8 to 8.9 (Table 2)Proof and experiment
Subsumption term of AVRDefinition corrected to match its description (App. C.1)Definition
p < 0.01 from five runsRemoved; every gain is at least twice the summed standard deviations (Table 1)Experiment

Reviewer tzdb: a response is posted for each of the four points.

Point raisedResolution in the revised PDFRests on
NoveltyWhat FuzzyCell takes from CL that prior methods do not (Table 3); fuzzy versus crisp use of the same axioms (Sec. 4.3)Experiment
Proof does not support the claimEarlier claim withdrawn; Proposition 3.1 restated and proved (p. 6; App. A.1); ablation (Table 2)Proof and experiment
Figure 1Redrawn in TikZ, with numbers computed from the equations (p. 4)Presentation
Graded memberships untestedTested against marker genes (reply of 17 Nov), to be added to App. FExperiment

No experimental result reported in the submission has changed. Every change we committed to in this discussion will be in the final version.

Thank you all for the careful reading and the time you gave to this paper!

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.