acceptodds
Under review as a conference paper at ICLR 2027

RhinoNet: Unsupervised learning for phenoscaping and generation of high dimensional point clouds

12 views

Abstract

We present RhinoNet, a simplicial complex-based point cloud neural network, that creates a geometry-preserving embedding of high-dimensional point clouds. RhinoNet takes collections of datasets, also called data cohorts, and processes each dataset (point cloud) within the cohort as a simplicial complex. These simplicial complexes have datapoints as vertices and edges learned by Vietoris-Rips filtration. Using simplicial wavelets and MLP layers, RhinoNet computes a vector representation of each dataset in a latent space produced by an autoencoder. The latent space is optimized such that an optimal transport distance between pairs of point clouds is preserved between the latent representations. We apply RhinoNet to single-cell datasets collected on patients—a common source of such high-dimensional data. RhinoNet creates a patient level representation, known in biological contexts as a phenoscape. We establish that RhinoNet generates organized phenoscapes, whereas other methods such as static maps do not. Next we show that we can generate entire point clouds by sampling and decoding this phenoscape. We show two modes of generation via 1) latent flow matching and 2) geodesic interpolation in latent space, followed by a decoding back into the point cloud space. RhinoNet achieves state-of-the-art performance on single-cell and spatial data cohorts relative to tested baselines. Overall, RhinoNet lifts the concepts of representation learning and generative modeling to the point cloud or population level and defines a neural model for phenoscaping.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Geometric Deep Learning, Point Clouds, High-Dimensional Data Analysis, Single-Cell Data, Optimal Transport.

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_z7G4OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper presents RhinoNet, a framework for representation learning and generation at the level of entire high-dimensional point clouds, motivated by single cell datasets. The method constructs simplicial complexes, extracts features using geometry simplicial wavelets, and learns a latent space whose Euclidean distances approximate Geodesic Sinkhorn distances between populations. A separate generation head reconstructs point clouds, while conditional latent flow matching and graph-based interpolation support generation. Experiments evaluate population embeddings and generation on synthetic and biological datasets.

Critical strengths
  1. Learning representations of entire cell populations addresses a practical need to compare biological samples while retaining information beyond average marker expression.
  2. The combination of population encoding, transport-based distance regularization, and point-cloud decoding provides a unified approach to population level analysis and generation.
  3. RhinoNet achieves the best reported generation metrics on PDO and Charville among the evaluated baselines. The paper also includes experiments examining geometric weighting and simplex order.
Critical weaknesses
  1. Intermediate populations are reconstructed from their own latent codes, which does not demonstrate generation of unseen states. The biological meaning of these paths is also unclear.
  2. The unsigned incidence matrices conflict with the claimed graph Laplacian, and Equation (4) contains dimensional inconsistencies.
  3. Patient and batch identifiers are used inconsistently. The data splits and handling of unseen patient identities require clarification.
  4. Important baselines such as Wasserstein Wormhole are missing, and geometry preservation is assessed mainly through visualizations.
Questions for authors
  1. Can intermediate states be generated without accessing the target populations, and how is their biological relevance validated?
  2. Can the authors clarify the incidence-matrix convention and correct Equation (4)?
  3. What are the patient-level data splits, and how does generation handle unseen patient identities?
  4. Can the authors add Wasserstein Wormhole, an encoder trained with the same distance loss, and quantitative geometry preservation metrics? If the authors can satisfactorily address the questions and concerns above, I would consider revising my recommendation to positive.
4 comments
AuthorsOpenReview ↗

Response to Reviewer z7G4 (1/2)

We thank the reviewer for the constructive review and for recognizing the practical need for population-level representations and our unified approach to population analysis and generation. We address the reviewer's comments below.

[W1, Q1] Intermediate states and biological relevance. The reviewer asked for generation of intermediate states without accessing the target populations and the biological relevance.

The intermediate point clouds on each path were held out from the training of both the encoder and the generation head. The submission's Appendix F.3 described this incompletely, mentioning only the generation head, and we have corrected the description in Appendices D and F.6.

We have also added a linear interpolation that never encodes the held-out point clouds: we decode latents spaced evenly on the straight line between the two endpoint latents (Section 3.2). It performs as well as the discrete geodesic interpolation, and its EMD to the held-out point clouds is 20%, 24% and 24% lower than that of Meta Flow Matching on PDO, Charville and UPMC (updated Figure 6, Table A6).

We tested whether the decoded PDO intermediates (Appendix F.1, Figure A1 in the updated manuscript) capture drug-specific biology. The interpolant recovers the culture’s DNA-damage response with its mean γH2AX expression (+1.49 vs. +1.54 control SD). A similar fraction of its cells is strongly damaged with γH2AX above the 99th percentile of DMSO (30.5% vs. 38.8%, against 1% in DMSO). These damaged cells lie in the same region of the embedding as the real damaged cells and show the expected cycling and checkpoint activation (phospho-RB and pCHK1 high). RhinoNet’s interpolation therefore captures the magnitude, location and mechanistic coupling of the SN-38 DNA-damage response at an unmeasured point along the dose axis.

[W2, Q2] Incidence matrix conventions and Eq 4. The reviewer asked for clarification on incidence-matrix convention and Equation 4.

We thank the reviewer for their careful attention to detail. We have corrected the indexing errors in Equation 4. Specifically, we have changed the Mk−1−1M_{k-1}^{-1} to Mk−1M_{k}^{-1} and final MkM_k to Mk+1M_{k+1} in the definitions of P~k+1C\tilde{P}^C_{k+1} and D~k+1B\tilde{D}_{k+1}^B.

The reviewer is correct about the graph Laplacian. Since we are using unsigned incidence matrices, we have Δ0=D+A\Delta_0 = D + A, not D−AD - A. We have removed this sentence.

As in Madhu et al. (2024), which introduced the simplicial scattering transform, we use unsigned incidence matrices so that we can construct generalized degree matrices and, from them, row-normalized transition matrices. With signed incidence matrices, positive and negative entries would cancel when computing these degrees, so a row sum could be zero, the degree matrix would not be invertible, and the row-normalized transition matrix would be undefined.

[W3, Q3] Data splits. The reviewer asked for clarification on data splits and unseen patient identity.

Appendix D.2 has been updated to reflect the fact that c001-c004 in Charville and c001-c007 in UPMC refer to batches of patients, not patients. Patient-level splits are listed in Appendix D.2: on PDO, patients 75 and 99 are held out for validation and testing of the encoder, and on Charville and UPMC whole batches are held out. Conditional generation learns an embedding for each label (patient and treatment, or patient and recurrence), so it can condition only on patients seen during training. Its test set therefore consists of held-out cultures of training patients, and every patient and treatment keeps at least one training example. Point clouds of held-out patients can still be embedded in the phenoscape and decoded from their latents.

AuthorsOpenReview ↗

Response to Reviewer z7G4 (2/2)

[W4, Q4] Baselines and metrics. The reviewer asked for additional baseline comparison such as Wasserstein wormhole trained with the same distance loss.

Please see [PSM] in the global response for phenoscape metrics.

Set encoders trained with the same loss. The PointNet++ and PointTransformer baselines in our phenoscape experiments (Figures 4, 5, A2 and A3) were trained with the same objective as RhinoNet: the PhenoGS distance-matching loss and the reconstruction loss, with the same decoder. The encoder is the only difference between these baselines and RhinoNet, so these comparisons already separate the effect of the encoder from the effect of the PhenoGS targets. The submission did not state this clearly, and we have added it to Section 4.

Wasserstein Wormhole. Wormhole is the closest prior method, and we benchmarked its released implementation on an NVIDIA H200 (141 GB). Wormhole pads every point cloud to the largest cloud in the cohort, and each training step runs attention over all points and solves an optimal transport problem between every pair of clouds in the batch, so its memory grows with the square of the padded cloud size. With the default batch size of 16, a training step on Charville, where clouds are padded to 3,496 cells, uses 33.7 GB and takes 6.5 s, so the default 10,000 steps take about 18 hours. On PDO (padded to 6,533 cells) and UPMC (19,826 cells), a single training step runs out of memory. Encoding is cheaper but still memory-heavy: one cloud takes about 34–37 ms, with peak memory of 1.5 GB on Charville and 19 GB on UPMC, compared with 19–378 ms and under 240 MiB for one RhinoNet view (Appendix E). For comparison, the high-dimensional cohorts in the Wormhole paper have a median of 11 to 69 points per cloud, and its largest clouds (2,048 points) are 3-D shapes (Table A8).

CohortPadded cloud sizeTraining step, batch 16Peak memory, training stepEncode one cloudPeak memory, encode
Charville3,4966,463.8 ± 2.2 ms33.7 GB36.99 ± 7.44 ms1.5 GB
PDO6,533OOMOOM34.47 ± 0.93 ms3.0 GB
UPMC19,826OOMOOM36.90 ± 6.93 ms19.3 GB

We now report quantitative geometry preservation on the test point clouds of every cohort: Spearman correlation between latent and PhenoGS distances, normalized stress, nearest-neighbor overlap and trustworthiness (Table A2). Our PointNet++ and PointTransformer baselines use the same PhenoGS and reconstruction losses and the same decoder as RhinoNet, so the encoder is the only difference. On PDO, RhinoNet reaches ρ = 0.570, against 0.117 and −0.039. On Charville it matches PointNet++ in rank correlation with lower stress, and on UPMC it outperforms PointTransformer on every metric. We also adapted Deep Wasserstein Embedding to point clouds with PhenoGS targets (EMD 4.23 for RhinoNet vs 4.27 on PDO). For Wasserstein Wormhole, we benchmarked the released implementation on an NVIDIA H200 (141 GB). It pads every cloud to the largest in the cohort, so a single training step runs out of memory on PDO and UPMC, while Charville needs 33.7 GB per step and about 18 hours of training.

We thank the reviewer again for the constructive feedback. In summary, we have (1) shown that intermediate states can be generated without encoding the target populations, using linear interpolation between endpoint latents (Table A6), and examined their biological content on PDO (Figure A1); (2) corrected Eq. (4) and clarified the unsigned incidence convention; (3) clarified that c001–c007 denote batches, stated the patient-level splits (Appendix D.2) and described how generation handles unseen patients; and (4) added a Deep Wasserstein Embedding comparison, a Wasserstein Wormhole benchmark, same-loss encoder comparisons and quantitative geometry-preservation metrics (Tables A2, A7 and A8). As the reviewer indicated they would consider a positive recommendation if these points were addressed, we would be grateful if they would revisit their score, and we are happy to answer any remaining questions.

Thank you for your detailed reply, it has solved my main question. I've improved the score to positive.

We thank the reviewer for reading our rebuttal carefully, and we appreciate that they have raised their assessment of the manuscript. If any concerns remain that prevent a stronger recommendation, we would be grateful if the reviewer could let us know, and we would be happy to address them before the discussion period closes.

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

This paper presents RhinoNet, an unsupervised autoencoder for learning population-level representations (phenoscapes) and generating high-dimensional point clouds. The method models each point cloud as a simplicial complex and extracts multi-scale features via a geometric Hodge Laplacian, which incorporates metric mass matrices into simplicial scattering wavelets. The latent space is regularized by pairwise geodesic Sinkhorn distances to preserve optimal transport geometry across point clouds. For generation, RhinoNet employs latent flow matching for label-conditioned sampling and discrete geodesic interpolation on a latent k-NN graph to reconstruct intermediate states. Evaluated on a synthetic Swiss-roll benchmark and three biological cohorts (PDO, Charville CODEX, UPMC CODEX), RhinoNet improves latent manifold organization, outperforms generative baselines (PVD, PointFlow, LION) in EMD, W2, and MMD, and achieves 19%–25% lower interpolation EMD than Meta Flow Matching.

Critical strengths

S1. The geometric Hodge Laplacian formulation is a principled and well-motivated architectural choice. Prior simplicial networks for point clouds (e.g., HiPoNet) rely on the combinatorial Hodge Laplacian, which considers only topological incidence (0/1 boundary matrices) and ignores the metric information embedded in simplex volumes. RhinoNet's discrete exterior calculus (DEC)-inspired formulation incorporates diagonal mass matrices MkM_k into the transition matrices, providing a metric-aware operator that captures continuous geometric variation. The ablation study (Fig. 7, Appendix F.2) empirically validates this formulation: the geometric Hodge Laplacian recovers smooth marker gradients across a 2D manifold, whereas the combinatorial counterpart collapses 74% of the latent variance onto a degenerate 1D arc.

S2. The phenoscaping application is novel and biologically meaningful. RhinoNet lifts representation learning to the entire point cloud (population) level without requiring marker-level supervision. The learned phenoscape captures biologically meaningful axes of variation—such as patient identity in organoids and continuous tumour–stroma–immune transitions in spatial proteomics—outperforming baselines that either collapse or shatter the cohort structure (Figs. 4, 5, A1, A2). Furthermore, the latent geodesic interpolation experiment (Fig. 6) demonstrates that discrete shortest paths traverse plausible intermediate cellular states, offering practical utility for modelling disease progression and drug perturbation trajectories.

Critical weaknesses

W1. Modality limitations, scalability bottlenecks, and lack of biological context. The evaluation is restricted to targeted multiplexed proteomics (CyTOF/CODEX, d≈40–44d \approx 40\text{--}44). It remains unproven whether the method scales to high-dimensional, sparse modalities central to single-cell biology, such as scRNA-seq (thousands of genes), where building Vietoris–Rips complexes and computing mass matrices face severe curse-of-dimensionality and density challenges. Moreover, precomputing the pairwise Geodesic Sinkhorn matrix scales quadratically (O(M2)O(M^2)), posing a practical bottleneck for cohorts with thousands of samples. Lastly, the reported generation errors (e.g., EMD ≈4.23\approx 4.23 on PDO) lack practical biological context, such as a baseline comparison against the natural variation between technical or biological replicates.

Questions for authors

The current benchmarks are restricted to targeted proteomics with around 40 protein markers. How does RhinoNet handle the curse of dimensionality and sparsity in whole-transcriptome single-cell data (e.g., scRNA-seq with thousands of genes), both in terms of simplicial complex construction (Vietoris–Rips filtration) and the quadratic precomputation cost (O(M2)O(M^2)) of the pairwise Geodesic Sinkhorn distance matrix for large-scale patient cohorts?

3 comments
AuthorsOpenReview ↗

Response to Reviewer NRoc

We thank the reviewer for their positive assessment, and in particular for recognizing the geometric Hodge Laplacian as a principled and well-motivated architectural choice that the ablation supports, and the phenoscaping application as novel and biologically meaningful. We address the reviewer's comments below.

scRNA-seq results. The reviewer asked if RhinoNet can be used for sparse, high-dimensional point clouds.

We have added a whole-transcriptome experiment on the Perturb-seq screen of Schnitzler et al., in which candidate coronary artery disease (CAD) genes were knocked down by CRISPR interference in human arterial endothelial cells and the cells were profiled by scRNA-seq (Appendix F.10, Figure A8). Each cell is described by 2,541 genes of which 294 are perturbed genes. We kept experiments with at least 100 cells per population, giving 296 point clouds. RhinoNet orders these point clouds along a gradient of the proliferation genes NUSAP1 and CENPF (relative Dirichlet energy E = 0.62 and 0.65, where E = 1 means no structure), while PointNet++, PointTransformer and DiffusionEMD show none (E = 0.94–1.02). Because the Vietoris–Rips complex depends only on pairwise distances, moving from 40 protein markers to 2,541 genes required no change to the method. This confirms empirically that the curse-of-dimensionality and density concerns do not prevent RhinoNet from learning a structured phenoscape from thousands of genes.

Vietoris-Rips scalability. The reviewer asked how well the Vietoris-Rips construction performs in high dimensions.

Please see [CC] in the global response for the full table. The Vietoris–Rips complex is built from pairwise distances between cells, so the ambient dimension enters only through computing those distances, which grows linearly with the number of features. The cost is set mainly by the number of cells and the number of simplices at the learned filtration scale. For one point cloud of 2,000 cells, construction and scattering take about 20 ms per view with K = 1 and 111–378 ms with K = 2, with peak memory below 240 MiB (Appendix E, Table A1).

PhenoGS scalability. The reviewer asked how well the PhenoGS computation performs in high dimensions.

One of the key factors in our use of PhenoGS is that it scales well with both the size of the point cloud and its ambient dimension, as reported in (Wilkinson et al., 2026). We have added a more explicit description of PhenoGS in Appendix E.2.

For the PDO cohort (M=1,678M = 1{,}678, N=2×106N = 2 \times 10^6 cells, nnz(L)≈9.8×107\mathrm{nnz}(\mathbf{L}) \approx 9.8 \times 10^7), the full 1,678×1,6781{,}678 \times 1{,}678 PhenoGS distance matrix took 301 seconds of wall-clock time on 32 CPU cores, including construction of the shared graph, with about 4 GB of memory. The PhenoGS distance is computed once before training. No PhenoGS distance is computed at test time.

Biological Interpretability. The reviewer asked about the biological interpretability of the generated point clouds and metrics.

We tested whether the decoded PDO intermediates (Appendix F.1, Figure A1 in the updated manuscript) capture drug-specific biology. The interpolant recovers the culture’s DNA-damage response with its mean γH2AX expression (+1.49 vs. +1.54 control SD). A similar fraction of its cells is strongly damaged with γH2AX above the 99th percentile of DMSO (30.5% vs. 38.8%, against 1% in DMSO). These damaged cells lie in the same region of the embedding as the real damaged cells and show the expected cycling and checkpoint activation (phospho-RB and pCHK1 high). RhinoNet’s interpolation therefore captures the magnitude, location and mechanistic coupling of the SN-38 DNA-damage response at an unmeasured point along the dose axis.

As a reference for natural variation, we computed the EMD between pairs of different real point clouds from the same cohort: 4.7 on PDO, 5.2 on Charville and 3.8 on UPMC. RhinoNet’s generation errors (4.23, 3.74 and 2.96) fall below these values, because each generated point cloud is compared with a real point cloud of the same condition, while the reference pairs are drawn from different conditions. These are now in the Interpolation paragraph of Section 4.

We thank the reviewer again for their positive assessment and constructive questions. In summary, we have (1) added a whole-transcriptome Perturb-seq experiment with 2,541 genes (Appendix F.10, Figure A8); (2) reported the runtime and memory of Vietoris–Rips construction and scattering (Appendix E, Table A1) and of the PhenoGS precomputation (Appendix E.2); and (3) added biological validation of decoded intermediate states (Figure A1 in the updated manuscript) and a comparison of generation errors with natural variation between real point clouds. We hope these revisions address the reviewer’s concerns, and we would be grateful if they would raise their score. We are happy to answer any remaining questions.

Thank you for the comprehensive response. My score remains unchanged.

We thank the reviewer for taking the time to go through the rebuttal! We are glad all of your concerns are addressed. What further clarification or modifications could we make to improve your score?

Reviewer_AiC6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces RhinoNet, an autoencoder intended to represent each point cloud in a cohort with one latent vector. The goal is for distances in that latent space to reflect optimal-transport distances between point clouds, creating what the authors call a phenoscape. The method builds on HiPoNet: it creates feature-reweighted views, constructs differentiable Vietoris–Rips complexes, and extracts simplicial-scattering features. RhinoNet replaces the usual incidence-based random-walk operators with operators based on a weighted Hodge Laplacian. An MLP maps the resulting features to latent vectors, trained with a reconstruction loss and a loss that matches latent distances to the precomputed geodesic-Sinkhorn distance (PhenoGS). A separate decoder then generates point clouds from latents, and the paper also explores conditional latent flow matching and latent-space interpolation.

The experiments include a Swiss-roll dataset and three biological cohorts: colorectal-cancer organoids (PDO) and the Charville and UPMC spatial cohorts. The paper shows PHATE visualizations of the phenoscape, compares conditional generation using W1, W2, and multi-bandwidth MMD, evaluates held-out intermediate point clouds with EMD, and presents qualitative ablations of the mass matrices and simplex order.

Critical strengths

The problem is worthwhile. Learning a representation for each point cloud whose geometry follows transport distances, while also supporting decoding, could be useful for cohort-level single-cell and spatial data. The paper brings together a set-level encoder, a distance-aware latent representation, and a decoder, and demonstrates the pipeline on three real cohorts. The discussion in Appendix C also suggests other settings where this setup may be useful.

The motivation for adding metric information to simplicial scattering is also reasonable. The usual HiPoNet and simplicial-scattering operators rely on 0/1 incidence matrices and therefore do not directly use the point-cloud geometry. The ablation in Fig. 7 and Fig. A3 suggests that the weighting changes the learned phenoscape: under the combinatorial Laplacian, 74% of the latent variance lies along one direction. However, the particular weighting and its justification raise important questions.

Critical weaknesses

1. The paper needs a clearer account of what is new and stronger comparisons to the closest prior work

Much of the framework appears to be adapted from prior work. Section 2.1 closely follows Madhu et al. (ICML 2024), including the simplicial-scattering construction, transition matrices, diffusion recursion, aggregation, wavelets, and higher-order coefficients. The feature-reweighted views and differentiable Vietoris–Rips construction follow HiPoNet. The weighted Hodge Laplacian is a standard construction based on weighted inner products on cochains; the proposed contribution is the particular choice of weights. The latent distance-matching objective is closely related to the stress objectives used in Deep Wasserstein Embedding and Wasserstein Wormhole, while the decoder loss and decoded interpolation are also closely related to Wormhole.

Appendix A explicitly states that Deep Wasserstein Embedding and Wormhole use the same general encoder–decoder idea, with the difference being RhinoNet’s geodesic rather than Euclidean ground cost. In that framing, the loss is not a new regularizer; it is an existing idea applied to a different target distance. These methods, ideally trained with the same geodesic-Sinkhorn targets, are therefore essential baselines, yet neither is included. A second important baseline would be a standard set encoder trained with the same distance-matching objective and decoder. Without these comparisons, it is difficult to tell whether the gains come from the proposed encoder or simply from the target distance. On PDO, the main phenoscape setting uses K = 1, so the encoder is a weighted graph model and does not use higher-order simplicial structure.

2. The Hodge-Laplacian motivation and convergence claim are not convincing as written

The paper motivates its geometric Laplacian through discrete exterior calculus (DEC), but the chosen weights do not appear to be the DEC Hodge-star weights. In DEC, the diagonal Hodge star on k-cochains uses the ratio of dual to primal volume. The paper instead uses primal quantities: a uniform M0, edge length for M1, and triangle area for M2. Relative to DEC, the edge and triangle weights are inverted.

This has a concrete effect. At the vertex level, B1 M1 B1ᵀ is a weighted graph Laplacian with edge weights equal to edge length. The resulting up-walk places more probability on longer edges, which is opposite to the usual intuition for diffusion and to common DEC-style weights. This matters particularly on PDO, where K = 1 and this vertex walk is essentially the full geometric modification. The cited convergence results concern DEC stars on simplicial meshes that approximate a manifold. A Vietoris–Rips complex built from points in 40–44 dimensions is not such a mesh: it need not triangulate the data manifold and may contain simplices of dimension higher than the manifold. I therefore do not see support for the claim that this construction approximates the smooth Hodge Laplacian. At present, the evidence for the weighting is mainly qualitative, aside from the reported variance concentration.

3. The experiments do not yet test the central claims in a sufficiently direct way

Phenoscape quality is evaluated primarily through two-dimensional PHATE plots. Since the Swiss roll has known structure and PhenoGS distances are already computed for every cohort, the paper should report direct measures of distance preservation: correlation between latent and PhenoGS distances, normalized stress, neighborhood overlap, and trustworthiness. The appendix says that the geometric Laplacian and K = 2 fit the distances more closely, but does not provide values. Because RhinoNet alone is trained to match those distances, the visual advantage could arise from the loss rather than the architecture. The appendix also notes that DiffusionEMD gives the smoothest marker gradients on UPMC, which is difficult to reconcile with the stronger claim in the main text that the other methods fail to represent this structure.

The interpolation experiment appears to be reconstruction rather than interpolation. Appendix F.3 indicates that the intermediate point clouds are held out only from decoder training and are reconstructed from their own encoded latents. The 20-nearest-neighbor graph is built on the full cohort, and the encoder was trained on these samples. In the PDO example, the path includes a training patient; in Charville, it includes a training batch. This tests whether the decoder can reconstruct samples unseen by that decoder, not whether the method can generate an intermediate population from two endpoints. Meta Flow Matching is asked to generate the target population from a source population, so the comparison is not like-for-like. The reported EMD reduction therefore does not establish the interpolation claim.

The generation comparison in Table 1 also needs more detail. PVD, PointFlow, and LION are unconditional 3-D shape generators, whereas RhinoNet is conditioned on patient and treatment labels. The paper says the baselines were adapted to high-dimensional marker spaces but does not explain how or whether they received the same conditioning. The very large LION values on Charville suggest that this adaptation may not be working reliably. Finally, the experiments use one seed unless stated otherwise and Table 1 does not report variability. The Charville and UPMC descriptions also need to clarify whether c001–c007 denote patients or batches, since the text uses both descriptions.

Questions for authors
  • How does RhinoNet compare with Wasserstein Wormhole and Deep Wasserstein Embedding when all methods use the same geodesic-Sinkhorn targets? How does a standard set encoder, such as PointNet++ or a Set Transformer, perform when trained with the same distance-matching loss and decoder?
  • For interpolation, can you evaluate latents sampled directly along a path between the two endpoint latents, without encoding the held-out intermediate point clouds? Please also evaluate a split where those intermediate point clouds are excluded from encoder training.
  • Why use primal volumes (edge length and triangle area) rather than the dual-to-primal ratios used in DEC? On PDO, is the preference for longer-edge transitions intentional? What happens with inverse-length or Gaussian-kernel edge weights?
  • Please report quantitative phenoscape results for every cohort and method: latent–PhenoGS correlation, normalized stress, and neighborhood preservation. Please also report the numbers behind the claims in Appendix F.2 and F.4.
  • How were PVD, PointFlow, and LION adapted to 40–44-dimensional point clouds with roughly 2,000 points? Were they conditioned on the same labels as RhinoNet? How many seeds were used and what is the variance?
  • Please define PhenoGS in the paper, including its ground cost, heat-kernel scale, and entropic regularization. What is the wall-clock time to compute it for the 1,678 PDO samples?
  • As written, Eq. (4) seems to pair B(k+1) with M(k−1)⁻¹ and Mk, which is dimensionally inconsistent. Is the intended term Mk⁻¹ B(k+1) M(k+1) B(k+1)ᵀ?
  • The generation head is described as mapping a latent vector to one point, while Appendix E suggests one forward pass per generated point. What additional input does it use, and how is the number of output points selected?
  • What is the computational cost of constructing the Vietoris–Rips complexes and computing scattering features for roughly 2,000 points in 40 dimensions, per point cloud and per view?
  • For Charville and UPMC, are c001–c007 patients or batches?
7 comments
AuthorsOpenReview ↗

Response to Reviewer AiC6 (1/6)

We thank the reviewer for their careful consideration of our manuscript. We are glad the reviewer found the problem worthwhile and the motivation for adding metric information to simplicial scattering reasonable, and noted that the ablation shows the weighting changes the learned phenoscape. We address the reviewer's comments below.

[W1] Novelty. The reviewer asked for clarification on what components are new.

RhinoNet is novel because it is the first neural network method to both faithfully embed and conditionally generate high-dimensional point clouds with many points. Existing methods each cover only part of this: MFM doesn’t produce interpretable embeddings, PointNet++ is designed for low dimensions, and Wasserstein Wormhole cannot handle point clouds that are both high dimensional and large. RhinoNet models high dimensional simplicial complexes like HiPoNet, but HiPoNet was a classification model; RhinoNet adds generative capabilities, a latent space that is organized by biologically interpretable optimal transport geometry, and metric information directly into the simplicial wavelets. The incorporation of metric information into simplicial wavelets is itself a novel contribution.

RhinoNet is indeed based on prior work. However, this is the case for most advances in computational research. Notable examples include Transformers, which combined attention and residual connections; Latent Diffusion, which paired a pretrained autoencoder with a diffusion model; and conditional flow matching, which combined continuous normalizing flows with a simulation-free regression objective adapted from score matching.

We have clarified the novel components in the introduction of the revised manuscript.

  1. Metric information in simplicial scattering. Prior simplicial scattering uses only incidence structure. What is new is not the weighted Hodge Laplacian, but putting metric volumes into the wavelets.
  2. OT distances have been used as latent regularizers before, but always with a Euclidean ground cost. PhenoGS computes a heat-diffusion distance over a graph of the cells, which approximates geodesic distance along the data manifold, so RhinoNet’s latent geometry follows the manifold. The use of heat-diffusion OT as a regularizer is the novel addition.
  3. Conditional generation of entire high-dimensional point clouds. Prior work focuses on low dimensions, unconditional generation, or smaller point clouds.

Section 2 now opens by further clarifying that the preliminaries are a review of ideas from the existing literature (Madhu et al., 2024; Viswanath et al., 2025).

AuthorsOpenReview ↗

Response to Reviewer AiC6 (2/6)

[W1, Q1] Comparisons to prior work. The reviewer asked about baselines including encoders trained with the same loss as RhinoNet, Deep Wasserstein Embedding, and Wasserstein Wormhole.

We have added comparisons to set encoders trained with PhenoGS loss, DWE, and Wasserstein Wormhole.

Set encoders trained with the same loss. The PointNet++ and PointTransformer baselines in our phenoscape experiments (Figures 4, 5, A2 and A3) were trained with the same objective as RhinoNet: the PhenoGS distance-matching loss and the reconstruction loss, with the same decoder. The encoder is the only difference between these baselines and RhinoNet, so these comparisons already separate the effect of the encoder from the effect of the PhenoGS targets. The submission did not state this clearly, and we have added it to Section 4.

Deep Wasserstein Embedding (DWE). DWE was designed for images: it trains a siamese CNN on 2-D pixel grids, so it cannot be applied directly to point clouds of varying size. We replaced its CNN encoder with a GNN and trained it with PhenoGS targets in place of its Wasserstein targets, keeping its KL reconstruction loss. Results on PDO (Appendix F.8, Table A7):

MethodEMD ↓W2 ↓MMD ↓
RhinoNet4.23054.32410.021284
DWE-GNN (PhenoGS)4.27274.35440.02408

Wasserstein Wormhole. Wormhole is the closest prior method, and we benchmarked its released implementation on an NVIDIA H200 (141 GB). Wormhole pads every point cloud to the largest cloud in the cohort, and each training step runs attention over all points and solves an optimal transport problem between every pair of clouds in the batch, so its memory grows with the square of the padded cloud size. With the default batch size of 16, a training step on Charville, where clouds are padded to 3,496 cells, uses 33.7 GB and takes 6.5 s, so the default 10,000 steps take about 18 hours. On PDO (padded to 6,533 cells) and UPMC (19,826 cells), a single training step runs out of memory. Encoding is cheaper but still memory-heavy: one cloud takes about 34–37 ms, with peak memory of 1.5 GB on Charville and 19 GB on UPMC, compared with 19–378 ms and under 240 MiB for one RhinoNet view (Appendix E). For comparison, the high-dimensional cohorts in the Wormhole paper have a median of 11 to 69 points per cloud, and its largest clouds (2,048 points) are 3-D shapes (Table A8).

CohortPadded cloud sizeTraining step, batch 16Peak memory, training stepEncode one cloudPeak memory, encode
Charville3,4966,463.8 ± 2.2 ms33.7 GB36.99 ± 7.44 ms1.5 GB
PDO6,533OOMOOM34.47 ± 0.93 ms3.0 GB
UPMC19,826OOMOOM36.90 ± 6.93 ms19.3 GB

[Q2] Interpolation details. The reviewer asked for additional details on the interpolation capabilities.

The intermediate point clouds on each path were held out from the training of both the encoder and the generation head. The submission's Appendix F.3 described this incompletely, mentioning only the generation head, and we have corrected the description in Appendices D and F.6.

We have also added a linear interpolation that never encodes the held-out point clouds: we decode latents spaced evenly on the straight line between the two endpoint latents (Section 3.2). It performs as well as the discrete geodesic interpolation, and its EMD to the held-out point clouds is 20%, 24% and 24% lower than that of Meta Flow Matching on PDO, Charville and UPMC (updated Figure 6, Table A6).

AuthorsOpenReview ↗

Response to Reviewer AiC6 (3/6)

[W2, Q3] Geometric Hodge Laplacian. The reviewer asked about differences between our usage of DEC and the DEC literature.

Reviewer Comment: “The Hodge-Laplacian motivation and convergence claim are not convincing as written”

We agree that this construction does not approximate a smooth Laplacian and we do not claim that it does. However, this construction is also the suitable path to embed metric information in simplicial wavelets.

Reviewer Comment: “Vietoris–Rips complex built from points in 40–44 dimensions is not such a mesh: it need not triangulate the data manifold and may contain simplices of dimension higher than the manifold.”

The Vietoris–Rips complex suits this because it depends only on pairwise distances, not on the ambient dimension (40–44). Because we stop at triangles, it contains no simplex above the manifold dimension unless the data are intrinsically one-dimensional.

Reviewer Comment: “Why use primal volumes (edge length and triangle area) rather than the dual-to-primal ratios used in DEC? On PDO, is the preference for longer-edge transitions intentional?”

The reason we cannot use the DEC theory in its precise formulation is that not every simplex is guaranteed to have a dual, e.g. an edge many not belong to any triangle. Although it is possible to guarantee a dual exists by artificially imposing a larger filtration or removing simplices without duals, this would either introduce numerical instabilities, such as thin triangles with area near 00, or remove information equivalent to forgetting entire points in the point cloud. Therefore we encode as much metric information as is faithfully possible by only including the primal volumes or inverse primal volumes. On PDO, the reported results use the combinatorial operators (unit masses), so this weighting does not affect them.

Reviewer Comment: “What happens with inverse-length or Gaussian-kernel edge weights?”

To test the choice of weights, we compared primal volumes with inverse primal volumes (Appendix F5, Table A4). The two give nearly the same embedding: every metric differs by 0.021 or less, and neither weighting is better on all of them.

Metric (Charville test point clouds, n = 23)Primal (submission)Inverse primal
Spearman correlation ↑0.28060.3016
Normalized stress ↓0.29060.2933
kNN overlap, k = 5 ↑0.28700.2783
kNN overlap, k = 10 ↑0.55220.5435

Where they differ is in the random walk that the wavelets are built from. Neighboring cells with almost the same neighborhood are joined by edges of very small volume. Primal volumes give these near-duplicate pairs little weight, so each cell's walk spreads over its more varied neighbors. Inverse volumes give them large weights, so the walk concentrates on look-alike cells that add little beyond the cell's own signal, which the lazy step already keeps. We measured this over a dataset as shown below (Table A5):

Measure (typical cell)Charville: primalCharville: inverseUPMC: primalUPMC: inverse
Share going to the single largest neighbor0.460.500.220.33
Number of neighbors the walk effectively uses2.92.78.05.7
Share going to near-duplicate neighbors0.290.340.220.33

These data support this explanation on both cohorts: inverse volumes restrict the walk to fewer neighbors. The shift is small on Charville, where the walk is already concentrated on about three neighbors, and larger on UPMC, where the effective number of neighbors drops from 8.0 to 5.7 and the share sent to near-duplicates rises from 0.22 to 0.33. The walk does not collapse onto a single neighbor under either weighting (the largest share stays at or below 0.50), which is consistent with the modest differences in the phenoscape metrics above. We therefore keep primal volumes: they give a better-spread walk at no cost in embedding quality.

[W3, Q4] Quantitative phenoscape evaluation. The reviewer asked for metrics to demonstrate the quality of the embeddings.

We have added detailed discussion in the Global Response [PSM] and Table A2 in Appendix F3, Figures A2, A3 in Appendix F2 and F10.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Global response

We thank the reviewers for their careful assessment of our work. Reviewers found the problem worthwhile and of practical importance for comparing biological samples (AiC6, z7G4), and the phenoscaping application novel and biologically meaningful (NRoc). They found the geometric Hodge Laplacian principled and well-motivated (NRoc), saw the motivation for adding metric information to simplicial scattering as reasonable (AiC6), and noted the strong generation results (z7G4, NRoc).

We believe the revisions and new experiments address the concerns raised, and we would be grateful if the reviewers would consider updating their scores accordingly. We are happy to answer any further questions during the discussion period.

Here we include responses to questions that were common amongst the reviewers.

Phenoscape Metrics [PSM]

As reviewers AiC6 and z7G4 suggested, we now report direct measures of distance preservation on the test point clouds of each cohort: Spearman correlation between latent and PhenoGS distances (ρ), normalized stress, k-nearest-neighbor overlap (NN), and trustworthiness (T), for k = 5 and 10. PointNet++ and PointTransformer are trained with the same PhenoGS and reconstruction losses as RhinoNet. These results are now in Table A2 in Appendix F3.

CohortMethodρ ↑Stress ↓NN5 ↑NN10 ↑T5 ↑T10 ↑
PDO (n = 168)RhinoNet0.5700.3880.1790.2210.7880.769
PointNet++0.1170.4730.0580.0880.5890.590
PointTransformer−0.0390.9260.0390.0880.5870.554
Charville (n = 23)RhinoNet0.2810.2910.2870.5520.6380.651
PointNet++0.2830.4030.2870.5170.5650.600
PointTransformer0.0241.0000.2700.4910.5360.511
UPMC (n = 52)RhinoNet0.1470.3020.2380.3480.6270.623
PointNet++------
PointTransformer0.0040.3610.1190.2250.5250.563

On PDO, RhinoNet preserves PhenoGS distances far better than PointNet++ and PointTransformer trained with the same loss (ρ = 0.570 vs 0.117 and −0.039). On Charville, RhinoNet and PointNet++ have similar rank correlation and 5-NN overlap; RhinoNet has lower stress and higher 10-NN overlap and trustworthiness. On UPMC, RhinoNet is better than PointTransformer on every metric and PointNet++ ran out of memory.

Computation Cost [CC]

Reviewers AiC6 and NRoc asked for details on wall time for certain computations.

We measured the cost of Vietoris–Rips construction and simplicial scattering for one point cloud of 2,000 points, per view, for K = 1 and K = 2. These results are now in Appendix E, Table A1.

ViewKVR construction (ms)Operator setup (ms)Scattering (ms)Total (ms)Peak memory, construction / scattering (MiB)
Molecular17.84 ± 0.356.51 ± 0.065.72 ± 0.0418.80 ± 0.28152.8 / 107.9
Molecular288.78 ± 1.9910.97 ± 0.0913.00 ± 0.15111.46 ± 2.21152.8 / 108.2
Spatial18.00 ± 0.187.86 ± 0.166.67 ± 0.0521.34 ± 0.35152.5 / 138.1
Spatial2345.02 ± 4.9215.25 ± 0.1918.91 ± 0.27377.77 ± 4.65152.5 / 236.8

With K = 1, one view of a 2,000-point cloud takes about 20 ms. Adding triangles (K = 2) raises this to 111 ms on the molecular view and 378 ms on the spatial view. Almost all of the increase comes from enumerating triangles in the Vietoris–Rips complex, which accounts for 80% and 91% of the total; operator setup and scattering stay below 20 ms in every setting. Peak memory stays below 240 MiB, and construction memory does not depend on K. The ambient dimension enters only through the pairwise distances, so the cost is driven by the number of points and the number of simplices at the learned filtration scale.

Dear Reviewers,

As the discussion period draws to a close, please read and acknowledge the authors’ rebuttal within the next 24 hours. I also encourage you to engage in the discussion and share any remaining questions or concerns.

Thank you for your time and thoughtful reviews.

Best,

Your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.