acceptodds
Under review as a conference paper at ICLR 2027

Knowledge-Guided RNA Velocity Inference with Large Language Models

Abstract

Understanding how cells transition between states is central to reconstructing dynamic biological processes from single-cell data. RNA velocity provides local, expression-derived estimates of transcriptional dynamics, but these signals do not explicitly encode the population-level biological structure governing plausible transitions between cell states. We therefore systematically investigate whether large language models (LLMs) can provide complementary biological reasoning for cell-state transition inference. Our benchmark reveals a clear capability boundary: LLMs are effective at identifying cell states and reasoning about state-level transitions, but are substantially less reliable at predicting fine-grained gene-level dynamics. Motivated by this complementarity, we introduce veloLLM, a knowledge-guided RNA velocity framework that preserves local expression-based dynamics while incorporating population-level transition priors to guide the organization of the velocity field. Across multiple single-cell datasets and base velocity estimators, veloLLM produces velocity fields that better align with known cell-state transitions while maintaining local dynamical coherence. Together, these results establish a principled division of labor between quantitative RNA velocity estimation and high-level biological reasoning, providing a general strategy for integrating biological knowledge into cell-state dynamics inference.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
single-cell RNA velocity, trajectory inference, large language models, knowledge-guided learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_L33qOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper explores how to leverage biological prior knowledge from large language models (LLMs) to refine single-cell RNA velocity inference. The authors first construct an evaluation benchmark, TransBench, across three developmental datasets, confirming that LLMs excel at population-level cell-state transition discrimination but underperform traditional methods in fine-grained quantitative prediction and ranking of gene-level dynamics. Motivated by this capability boundary, the authors propose veloLLM, a post-processing refinement framework that modifies the output of a base velocity model by jointly optimizing three constraints: base velocity field reconstruction, intra-cluster local smoothness, and population-level transition prior edge guidance. Improvements in metrics such as cross-boundary directional accuracy (CBdir) and terminal state identification are demonstrated across four single-cell datasets and five base velocity estimators.

Critical strengths
  1. This paper constructs the TransBench evaluation benchmark, quantitatively characterizing the capability boundaries of large language models in population-state topology inference versus gene-expression dynamics prediction, and providing empirical evidence for a division of labor between LLMs and quantitative biophysical models.

  2. The proposed veloLLM framework adopts a model-agnostic post-processing refinement paradigm that jointly optimizes velocity reconstruction, manifold smoothness, and prior edge-guidance losses, making it compatible with diverse base estimators while preserving local splicing kinetics and improving the biological plausibility of global developmental trajectories.

Critical weaknesses
  1. RNA velocity inference itself is fundamentally constrained by underlying biophysical limitations, including constant kinetic rate assumptions, transcriptional bursting, and local k-NN smoothing and projection artifacts (Bergen et al., 2021; Gorin et al., 2022; Marot-Lassauzaie et al., 2022; Zheng et al., 2023). Rather than addressing velocity estimation from the perspective of transcriptional kinetics or measurement noise modeling, this work forcibly aligns boundary velocity vectors toward known population transition edges; because this training loss shares the exact same reference edges with the core evaluation metric (CBdir), the reported directional gains are largely self-fulfilling and suffer from circular evaluation.

  2. The methodological originality of the proposed framework is relatively limited. The overall architecture loosely couples standard retrieval-augmented generation (RAG) with a conventional multilayer perceptron (MLP), where the mean-squared error distillation, intra-cluster cosine smoothness, and cross-boundary directional alignment are straightforward heuristic regularizers rather than principled algorithmic innovations for high-dimensional stochastic single-cell manifolds.

  3. There is a notable disconnect between the main empirical results and the central methodological claims of the paper. All performance gains reported in Table 3 of the main text rely on manually curated literature transition graphs rather than priors generated end-to-end by LLMs; as shown in Appendix Table 7, replacing the manual graph with LLM-inferred directed edges omits critical developmental transitions and causes the CBdir score to drop sharply (from 0.7285 to 0.4029 or even 0.0252), undermining the core claim that LLMs can reliably serve as autonomous global prior providers.

  4. The TransBench benchmark suffers from severe information asymmetry and construction bias across tasks, making it difficult to objectively establish the claimed capability boundary. In macroscopic state-transition tasks, prompts directly expose standardized cell-cluster names, allowing LLMs to achieve high scores purely through semantic memorization (as evidenced by the sharp performance drop in the hidden-state HTE track); conversely, in microscopic gene-ranking tasks (IGR/DGR), prompts not only strip all quantitative expression values but also mask out almost the entire eligible gene universe from the visible context (e.g., masking 100 candidate genes out of 109 total eligible genes in Hematopoiesis), forcing LLMs to blindly guess noisy single-cell expression differences from nearly empty contexts.

References:

Questions for authors

See weaknesses

3 comments
AuthorsOpenReview ↗

Response to Reviewer L33q (1/3)

We thank the reviewer for the detailed and constructive feedback. We apologize that the original presentation was misleading regarding the role of the transition graph, and we are grateful for the opportunity to clarify a key point: in practice, veloLLM always uses an LLM-generated graph. The human-curated graph serves as the evaluation reference, and veloLLM (human-curated graph) is listed in the performance tables only for comparison. We have now completed the experiments with LLM-generated graphs on all datasets and base estimators. Supplementary analyses use a fixed random seed, so their values differ slightly from the multi-run means in Table 3. Below, we address the reviewer's concerns one by one.

W1: Biophysical Limitations of RNA Velocity and Circular Evaluation

Thank you for this important comment. We respond to its three points in turn.

On kinetic limitations. We agree that veloLLM does not model transcriptional kinetics or measurement noise, and we do not claim to. The limitations listed by the reviewer make local velocity estimates unreliable, and different kinetic models can yield conflicting directions on the same data (Ancheta et al., 2026). veloLLM addresses this problem from a different angle, by adding population-level biological knowledge that kinetic modeling alone cannot provide. Advances in kinetic modeling are inherited through the base estimator, and veloLLM improves CBdir for all five base estimators, including dynamical (scVelo), probabilistic (veloVI), and GRN-informed (RegVelo) models.

On forcible alignment. The alignment is soft rather than forced. The MSE term anchors the refined field to the base estimates, and the edge loss aligns each boundary cell with expression displacements toward its own target-state neighbors rather than with a fixed population direction. A neighborhood-level ODE fit score shows that the refined field remains locally compatible with the spliced-RNA kinetic relation (Figure 4, Appendix A.5).

On circular evaluation. We agree that when the training and evaluation edges coincide, CBdir gains are partly self-fulfilling. This is why veloLLM (human-curated graph) is now listed only for comparison. Two experiments show that the gains of veloLLM are not self-fulfilling.

First, we remove the overlap entirely. We refine the same MouseBrain scVelo field five times, each time withholding one transition from the training graph, and evaluate CBdir only on the withheld transition. A control without edge loss retains velocity alignment and smoothness (Appendix A.7, Table 10):

Withheld transitionscVeloNo edge lossEdge withheld
RG, Astro, OPC → IPC0.2930.3680.495
IPC → Subplate0.3320.3440.583
IPC → Deeper Layer0.3130.3260.627
IPC → Upper Layer0.6230.6380.843
RG, Astro, OPC → Ependymal cells0.1360.2950.872
Mean0.3390.3940.684

veloLLM improves every withheld transition over both scVelo and the no-edge-loss control, even though the evaluated edge never appears in training.

Second, if the gains came simply from aligning velocities with the graph edges, directly interpolating the base velocity toward the same edges would achieve a similar CBdir. It does not. The table below reports CBdir, with the interpolation strength swept over five values (Appendix A.8, Table 12):

Dataset (estimator)Base fieldBest direct adjustmentveloLLM
MouseBrain (scVelo)0.3390.4160.729
Hematopoiesis (DeepVelo)0.3140.3540.546

Imposing the edges directly recovers only a small part of the gain, which shows that the improvement comes from how veloLLM integrates the graph with local expression dynamics.

AuthorsOpenReview ↗

Response to Reviewer L33q (2/3)

W2: Methodological Originality

Thank you for this comment. We respectfully clarify that the main contribution of this work is not a new network architecture, but a division of labor between LLMs and quantitative velocity estimators, established empirically by TransBench and realized by veloLLM. We intentionally keep the refiner simple so that it can be plugged into any base estimator without modification, as demonstrated across five estimators with different modeling assumptions.

The loss terms are not interchangeable heuristics. Each plays a distinct role (MouseBrain, DeepVelo, Table 11):

ConfigurationVeloCohICCohCBdir
DeepVelo base0.7660.3490.380
Full refinement0.8750.6620.823
Without edge loss0.8900.6360.461
Without smoothness0.7570.3800.805

The edge loss determines cross-boundary direction, while the smoothness term determines within-state coherence. In addition, learned refinement substantially outperforms direct interpolation toward the same graph (see W1), which shows that the design of the objective, rather than the graph alone, drives the improvement.

W3: Disconnect Between the Main Results and LLM-Generated Priors

Thank you for pointing out this disconnect. In the revision, Table 3 reports veloLLM (LLM-generated graph) with DeepSeek-generated graphs as the main result on all four datasets and five base estimators, while veloLLM (human-curated graph) is listed for comparison. Averaged over the five base estimators, CBdir is:

DatasetBaseveloLLM (LLM-generated graph)veloLLM (human-curated graph)
MouseBrain0.1770.7100.785
Hematopoiesis0.1040.4440.464
Dentate Gyrus0.5200.7860.786
Neural Crest0.1340.6390.696

veloLLM (LLM-generated graph) improves CBdir over the matched base field in all 20 dataset–estimator comparisons and, on average, recovers 88–100% of the gain achieved with the human-curated graph.

The CBdir drop cited by the reviewer comes from the generated graphs in the originally submitted Table 7. In the revision, we manually checked the reference edges for all four datasets against the original studies and prior CBdir evaluations, including those in GFG and TIvelo. DeepSeek and Kimi now recover all reference edges on all four datasets (Table 8). With the scVelo base field, their graphs match the human-curated graph on Hematopoiesis, Dentate Gyrus, and Neural Crest, and reach 0.729 and 0.725 on MouseBrain, compared with 0.771 for the human-curated graph (Table 7).

AuthorsOpenReview ↗

Response to Reviewer L33q (3/3)

W4: Information Asymmetry and Construction Bias in TransBench

Thank you for raising this concern. We address the two tasks separately.

TE and HTE are designed to measure different conditional capabilities. TE provides cell-state names and evaluates whether a model can apply biological knowledge given annotated identities. This is exactly the setting in which veloLLM uses the LLM, since veloLLM relies on annotated cell states. HTE withholds these names and requires the model to infer identities from gene signatures. The TE–HTE gap therefore does not undermine the capability boundary. It is part of it, showing that LLMs are reliable when reasoning over annotated states but less reliable when inferring states from expression, which is why veloLLM does not ask the LLM for expression-level inference. We will make this conditional interpretation explicit in the revision.

IGR and DGR do not present near-empty contexts. In Hematopoiesis, the 109 genes define the eligible ranking-answer universe, whereas the visible context is drawn from a separate 925-gene expression-context universe. After the 100 candidates are removed, context genes are selected from the remaining 825 genes. Across the Hematopoiesis ranking examples, the median context contains 100 genes on each side, and no list is empty. The expression values of candidate genes are withheld by design, since they would directly reveal the answer.

To test whether the lack of quantitative values explains the low ranking scores, we add the observed expression values of all visible context genes to the prompts, keeping the cell pairs, candidate pools, and reference rankings unchanged (GPT-5.6-Sol, P1):

DatasetTaskExamplesOriginal NDCG@20With expression values
MouseBrainIGR3000.2040.197
MouseBrainDGR3000.2480.241
Mouse GastrulationIGR2,0000.2190.198

Adding quantitative expression values does not improve ranking performance in any setting. The limited gene-level performance is therefore not an artifact of missing context, which supports the granularity gap reported in TransBench.

We greatly appreciate your insightful comments, which will help improve our paper. If our response has successfully addressed your concerns and clarified any ambiguities, we respectfully hope that you consider raising the score. Should you have any further questions or require additional clarification, we would be delighted to engage in further discussion.

Reviewer_c6kYOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper construct a benchmark Transbench and a deep learning refiner veloLLM.

Critical strengths
  • Transbench systematically evaluates LLMs on cell-state transition reasoning across 6 tasks and 3 datasets.
  • VeloLLM is a flexible plug-in design. It can be applied on top of any existing velocity estimator.
Critical weaknesses
  • The paper needs more information on how the human-curated priors were constructed, since they are used in the main results (Table 3).
  • The storyline is not coherent. TransBench tests LLMs' ability in cell-state transition reasoning, but veloLLM uses human-curated priors rather than LLM-generated ones, so the claim that LLMs can provide useful transition priors is not supported by the main results.
Questions for authors
  • How does the quality of the prior affect performance? What would the results in Table 3 look like if LLM-generated priors were used instead beyond the Dentate Gyrus/scVelo setting in Table 7? A controlled robustness analysis (e.g., randomly removing, reversing, or adding edges in the prior) would also help clarify how sensitive veloLLM is to prior errors.
  • Why does the LLM prior perform worse? Is the performance drop mainly due to missing edges, or other factors?
  • Please provide the priors for MouseBrain, Hematopoiesis, and Neural Crest, as Table 7 does for Dentate Gyrus.
  • I would be willing to increase my score if the authors could make the paper more coherent, e.g., by providing a stronger LLM prior, or by adding a TransBench task in which the LLM outputs the transition prior and its accuracy is evaluated against ground-truth/human-curated labels.
2 comments
AuthorsOpenReview ↗

Response to Reviewer c6kY (1/2)

We thank the reviewer for the constructive feedback and for the concrete suggestions on how to make the paper more coherent. We sincerely apologize that the original presentation was misleading, and we are grateful for the opportunity to clarify a key point: in practice, veloLLM always uses an LLM-generated graph. The human-curated graph serves as the evaluation reference, and veloLLM (human-curated graph) is listed in the performance tables only for comparison. We have now completed the experiments with LLM-generated graphs on all datasets and base estimators. Below, we address the reviewer's concerns one by one.

W1: Construction of the Curated Priors

Thank you for raising this point. The human-curated graph is constructed by mapping the directed transitions reported in the original study of each dataset onto its annotated cell states, and it serves as the reference against which CBdir is computed. Since training with this graph exposes the evaluation reference to the refiner, veloLLM (human-curated graph) is listed in our performance tables only as a comparison setting, rather than as a practical setting of veloLLM. The originally submitted Table 3 placed this setting in the main comparison, which obscured the intended workflow. In the revision, the main results use LLM-generated graphs (see W2), and the human-curated graphs for all datasets are listed in Q3.

W2: Coherence Between TransBench and the Main Results

We appreciate this comment. In the revision, Table 3 reports veloLLM (LLM-generated graph) with DeepSeek-generated graphs as the main result on all four datasets and five base estimators, while veloLLM (human-curated graph) is listed for comparison. The CBdir results are summarized below in the format Base / veloLLM (LLM-generated graph) / veloLLM (human-curated graph). The full Table 3, including VeloCoh and ICCoh, is provided in the revised paper.

EstimatorMouseBrainHematopoiesisDentate GyrusNeural Crest
scVelo0.370 / 0.721 / 0.792-0.311 / 0.370 / 0.3660.401 / 0.693 / 0.693-0.091 / 0.610 / 0.696
veloVI0.344 / 0.797 / 0.7880.121 / 0.391 / 0.3850.637 / 0.828 / 0.8280.087 / 0.575 / 0.591
DeepVelo0.380 / 0.766 / 0.8230.329 / 0.532 / 0.5390.669 / 0.817 / 0.8170.362 / 0.745 / 0.797
RegVelo-0.791 / 0.568 / 0.7960.346 / 0.527 / 0.6320.327 / 0.713 / 0.713-0.126 / 0.550 / 0.681
GFG0.580 / 0.697 / 0.7270.033 / 0.398 / 0.3980.568 / 0.879 / 0.8790.436 / 0.717 / 0.716
Mean0.177 / 0.710 / 0.7850.104 / 0.444 / 0.4640.520 / 0.786 / 0.7860.134 / 0.639 / 0.696

veloLLM (LLM-generated graph) improves CBdir over the matched base field in all 20 dataset–estimator comparisons and, on average, recovers 88–100% of the gain achieved with the human-curated graph. The full pipeline is therefore LLM-driven, consistent with the capability boundary identified by TransBench. LLMs supply cell-state-level transition knowledge, while the base estimator supplies quantitative kinetics.

Q1: Effect of Prior Quality on Performance

Thank you for this important question. The results of veloLLM (LLM-generated graph) on all four datasets are reported in W2. To further assess how sensitive veloLLM is to prior errors, we perturb the MouseBrain human-curated graph by removing, reversing, or adding edges, while fixing the scVelo base field and evaluation reference (Appendix A.7, Table 9):

Training graphnVeloCohICCohCBdir
Human-curated graph10.8660.6410.771
One edge removed50.864 ± 0.0030.640 ± 0.0010.761 ± 0.023
Two edges removed50.864 ± 0.0030.641 ± 0.0020.718 ± 0.075
One edge reversed30.862 ± 0.0040.640 ± 0.0030.730 ± 0.004
Two edges reversed20.858 ± 0.0050.636 ± 0.0110.682 ± 0.007
One spurious edge added10.8670.6430.725

veloLLM degrades gracefully. Coherence is essentially unchanged, and CBdir remains far above the scVelo base field (0.339) under all perturbations. Moreover, in a leave-one-edge-out experiment, refinement still improves CBdir on every withheld transition, with a mean of 0.684 compared with 0.394 for a control without edge loss (Table 10). This shows that the prior helps beyond the edges it explicitly supervises.

AuthorsOpenReview ↗

Response to Reviewer c6kY (2/2)

Q2: Causes of the Performance Drop with LLM-Generated Priors

Thank you for this question. Across the four datasets, MouseBrain is the only one where veloLLM (LLM-generated graph) falls noticeably below veloLLM (human-curated graph). Using the scVelo base field (Table 7), the gap can be attributed to two factors:

Training graphEdge recallAdditional edgesCBdir
Human-curated5/500.771
DeepSeek5/510.729
Kimi5/510.725
GLM3/530.657
Qwen1/550.510
  • Missing edges cause the largest drop. Qwen recovers only one of the five reference edges and reaches a CBdir of 0.510.
  • Additional edges cause a smaller drop. DeepSeek and Kimi recover all reference edges but each adds one transition, reaching 0.729 and 0.725. This closely matches the controlled setting in Q1, where adding one spurious edge to the human-curated graph yields 0.725.

Q3: Priors for MouseBrain, Hematopoiesis, and Neural Crest

Thank you for this request. The human-curated graphs and the DeepSeek-generated graphs used in the main results are listed below. The graphs generated by all four LLMs are provided in Table 8.

  • MouseBrain. Human-curated graph: IPC → Deeper Layer, Subplate, Upper Layer. RG, Astro, OPC → IPC, Ependymal cells. Additional edge in the DeepSeek-generated graph: V-SVZ → RG, Astro, OPC.
  • Hematopoiesis. Human-curated graph: HSC → MEP-like, GMP-like. MEP-like → Meg. GMP-like → Mon. Additional edges in the DeepSeek-generated graph: MEP-like → Ery. GMP-like → Bas, Neu.
  • Dentate Gyrus. Human-curated graph: nIPC → Neuroblast → Granule immature → Granule mature. OPC → OL. Additional edge in the DeepSeek-generated graph: Radial Glia-like → nIPC.
  • Neural Crest. Human-curated graph: Neural crest (PNS glia) → Myelinating Schwann cells. Neural crest (PNS neurons) → Dorsal root ganglion neurons, Otic sensory neurons. Myelinating Schwann cells → Myelinating Schwann cells (Tgfb2+). Additional edges in the DeepSeek-generated graph: Neural crest (PNS glia) → Melanocyte cells, Olfactory ensheathing cells. Neural crest (PNS neurons) → Enteric neurons.

Q4: Evaluating LLM-Generated Transition Priors Against Curated Labels in TransBench

Following the reviewer's suggestion, we add a graph-level track to TransBench. Each LLM proposes directed edges from annotated cell states and retrieved evidence, with the human-curated edges withheld during generation, and the proposals are filtered by undirected neighborhood support in expression space. We score the resulting graphs against the human-curated graphs:

LLMMouseBrainHematopoiesisDentate GyrusNeural CrestTotal recall
DeepSeek5/54/44/44/417/17
Kimi5/54/44/44/417/17
GLM3/54/44/4no output11/13
Qwen1/54/44/44/413/17

DeepSeek and Kimi recover every reference edge across all four datasets. The complete generated graphs, including additional proposed edges, are listed in Table 8, and their effect on refinement is analyzed in Q2. We will add this track to TransBench in the revision.

We greatly appreciate your insightful comments, which have substantially improved the coherence of our paper. If our response has successfully addressed your concerns and clarified any ambiguities, we respectfully hope that you consider raising the score. Should you have any further questions or require additional clarification, we would be delighted to engage in further discussion.

Reviewer_aX5wOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes veloLLM, an method to enhance classical single-cell velocity estimation methods with biological knowledge of cell-state transitions. veloLLM learns a neural network that maps each cell’s expression profile to a velocity vector which is trained to:

  1. Stay close to velocities estimated by an existing method based on RNA splicing dynamics, such as scVelo or veloVI
  2. Produce similar velocity directions for neighboring cells that share a cell-state label
  3. The velocity from a cell to a neighboring cell with successor states (for example, immature neuron → mature neuron) should match the difference in gene expression. These proposed transitions form a directed prior graph, which the framework obtains through LLM reasoning.

In addition, the paper introduces TransBench, a benchmark evaluating LLM predictions of cell-state transitions and associated gene-expression changes across three single-cell datasets. Its six tasks cover transition existence, transitions with hidden cell identities, next-state prediction, gene-change direction, and rankings of increasing or decreasing genes.

Critical strengths

While the use of LLM-based knowledge to enhance and improve single-cell analysis is not new, the way veloLLM improves splicing-dynamics based velocity estimation with direct, easy to verify, directed cell-state transition priors is interesting.

Despite the Hematopoiesis, Dentate Gyrus and Mouse Brain datasets themselves being well-known in the field, the TransBench tasks and evaluation framework could be of some interest to the field.

Critical weaknesses
  1. LLM only gives cell state transitions. Nothing about gene-regulation or direction, or any other granular detail. This is 'simple' and sometimes textbook information, definitely in the dataset they used. It is therefore not obvious what is the merit of using an LLM over manually curated relationships, ontology-based rules, or a simpler method given the same transition graph.
  2. TransBench only includes small datasets. This provides limited evidence of generalization across tissues, organisms, disease states, and unfamiliar developmental processes. Its applicability to large, heterogeneous atlases with many cell states, inconsistent annotations, and ambiguous transitions therefore remains untested.
  3. VeloLLM is only useful where we a-priori know the system and it's cell-state transition graph. This creates a circular logic, where velocity and trajecotry inference is commonly used to discovery of ordered cell-state relationships in new data and systems, VeloLLM needs them to be already pre-determined to be useful. This significantly limits, if not making totally obsolete, any potentially usefulness of the proposed algorithm.
Questions for authors
  1. What values does the LLM add beyond the transition graph itself?
  2. How well does the approach generalize beyond the three TransBench systems? Specifically to larger and more heterogeneous datasets.
  3. Does veloLLM work in a semi-supervised setting, where we only know the direction of some cell-state transitions in a dataset? Can it generalize and fix incorrect transitions in classical velocity estimation algorithms.
2 comments
AuthorsOpenReview ↗

Response to Reviewer aX5w (1/2)

We thank the reviewer for the thoughtful and critical feedback. We apologize that the original presentation was misleading regarding the role of the transition graph, and we are grateful for the opportunity to clarify a key point: in practice, veloLLM always uses an LLM-generated graph. The human-curated graph serves as the evaluation reference, and veloLLM (human-curated graph) is listed in the performance tables only for comparison. We have now completed the experiments with LLM-generated graphs and added a large-scale atlas. Below, we address the reviewer's concerns one by one.

W1 & Q1: Added Value of the LLM Beyond the Transition Graph

We appreciate this question. The LLM adds value in two ways.

First, the LLM constructs the transition graph, which would otherwise have to be curated by hand for each dataset. Cell-state labels vary across datasets in nomenclature and granularity (e.g., "RG, Astro, OPC" or "MEP-like"), so even well-known lineages must be mapped onto each annotation scheme. The LLM aligns these labels with ontology records, regulatory priors, and literature, and proposes directed edges without dataset-specific mapping rules. With the human-curated edges withheld, DeepSeek recovers all 17 reference transitions across the four datasets (Table 8). On the Mouse Gastrulation atlas (see W2 & Q2), it constructs a 25-edge graph over 34 cell states, a setting where manual curation becomes costly.

Second, restricting the LLM to state-level transitions is a deliberate design supported by TransBench. LLMs are reliable at judging state-level transitions (Best TE Macro-F1 reaches 0.996 on Hematopoiesis and 1.000 on both MouseBrain and DentateGyrus.). In contrast, their gene-level direction predictions are close to chance (GVD Macro-F1 between 0.48 and 0.64), and their gene rankings remain weak (NDCG@20 at most 0.28). We therefore leave gene-level and cell-level dynamics to the expression-based estimator rather than relying on unreliable fine-grained LLM predictions.

Regarding a simpler method given the same graph, we compare veloLLM with direct interpolation of the base velocity toward the graph-specified directions, sweeping the interpolation strength and reporting CBdir (Appendix A.8, Table 12):

Dataset (estimator)Base fieldBest direct adjustmentveloLLM
MouseBrain (scVelo)0.3390.4160.729
Hematopoiesis (DeepVelo)0.3140.3540.546

Learned refinement substantially outperforms the best interpolation setting on both datasets, showing that the gain does not come from trivially imposing the graph on the base velocity.

W2 & Q2: Generalization to Larger and More Heterogeneous Datasets

Thank you for raising this point. To address scale and heterogeneity, we add Mouse Gastrulation, a multi-lineage embryonic atlas with 89,267 cells, 53,801 genes, and 34 annotated cell states, which is more than ten times larger than the original TransBench datasets combined.

TransBench. We construct 2,000 examples each for Transition Existence (TE) and Increasing-Gene Ranking (IGR), covering 11 forward reference transitions among 20 cell states. TE includes 1,333 forward transitions and 667 reversed negatives. Both tasks retain the original format and metrics. Using GPT-5.6-Sol under P1:

DatasetTE: Macro-F1 ↑IGR: NDCG@20 ↑
Hematopoiesis0.9560.152
MouseBrain0.8580.204
DentateGyrus1.0000.236
Mouse Gastrulation0.9680.219

Performance on the large atlas falls within the range observed on the original datasets, and the gap between state-level and gene-level tasks persists. This supports the capability boundary that motivates veloLLM.

veloLLM. We refine DeepVelo with a 25-edge LLM-generated graph and evaluate on 11 human-curated transitions:

SettingVeloCohICCohCBdir
DeepVelo0.7570.4980.192
veloLLM (LLM-generated graph)0.9230.9280.591

veloLLM improves all three metrics on the large atlas and corrects the direction of all three transitions on which the base field points backward. We will incorporate these experiments into the revision.

AuthorsOpenReview ↗

Response to Reviewer aX5w (2/2)

W3: Dependence on a Pre-Determined Transition Graph

Thank you for this important comment. We respectfully clarify that veloLLM does not require a pre-determined transition graph. Beyond the spliced and unspliced matrices used by standard velocity estimators, veloLLM only uses cell-state clusters and their annotations, which are routinely obtained in standard single-cell workflows before velocity analysis. The LLM then proposes the transition graph from existing biological knowledge (see W1 & Q1). This graph is coarse and can be incomplete. Everything at finer resolution is still inferred from the data, including cell-wise velocities, gene-level dynamics, and directions on transitions that the graph does not supervise.

Two results support this. First, refinement improves directions on transitions withheld from the graph (see Q3). Second, the perturbation analysis of the refined velocity identifies transition-associated genes, such as Igfbpl1 for the nIPC → Neuroblast transition, which are not encoded in the graph (Section 5.5).

veloLLM is therefore designed to complement, rather than replace, data-driven discovery. Classical estimators frequently produce directions that contradict well-established biology. Base CBdir is negative in four of the 20 dataset–estimator settings in Table 3. veloLLM uses existing knowledge to correct such failures, while leaving the data to determine the remaining dynamics.

Q3: Semi-Supervised Setting and Correction of Incorrect Transitions

Thank you for this question. Yes, veloLLM works with a partially known graph. In a leave-one-edge-out experiment on MouseBrain, we refine the same scVelo field five times, each time withholding one transition from the graph, and evaluate CBdir on the withheld transition. A control without edge loss retains velocity alignment and smoothness (Appendix A.7, Table 10):

Withheld transitionscVeloNo edge lossEdge withheld
RG, Astro, OPC → IPC0.2930.3680.495
IPC → Subplate0.3320.3440.583
IPC → Deeper Layer0.3130.3260.627
IPC → Upper Layer0.6230.6380.843
RG, Astro, OPC → Ependymal cells0.1360.2950.872
Mean0.3390.3940.684

veloLLM improves every withheld transition over both scVelo and the no-edge-loss control, showing that supervision on the known transitions generalizes to unsupervised ones.

veloLLM also corrects incorrect transitions in classical estimators. In every setting of Table 3 where the base field points predominantly backward (negative CBdir), veloLLM (LLM-generated graph) restores the correct direction:

Estimator (dataset)Base CBdirveloLLM (LLM-generated graph)
RegVelo (MouseBrain)-0.7910.568
scVelo (Hematopoiesis)-0.3110.370
scVelo (Neural Crest)-0.0910.610
RegVelo (Neural Crest)-0.1260.550

The same holds on Mouse Gastrulation, where all three backward transitions in the base field are corrected (see W2 & Q2).

We greatly appreciate your insightful comments, which will help improve our paper. If our response has successfully addressed your concerns and clarified any ambiguities, we respectfully hope that you consider raising the score. Should you have any further questions or require additional clarification, we would be delighted to engage in further discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.