acceptodds
Under review as a conference paper at ICLR 2027

scHiDiffEM: Reconstructing Sparse Single-Cell Hi-C Contact Maps with Diffusion Expectation–Maximization

Abstract

Single-cell Hi-C (scHi-C) profiling elucidates three-dimensional (3D) genome architecture at cellular resolution, yet individual contact maps are severely corrupted by extreme count sparsity, cell-to-cell sequencing depth heterogeneity, and power-law genomic distance decay. An observed zero conflates true biological non-contact with technical sampling dropout, making it difficult to reconstruct cell-specific topology without averaging away biological variation. We introduce scHiDiffEM (Single-Cell Hi-C Diffusion Expectation–Maximization), which treats reconstruction as a population-informed inverse problem. A distance-aware low-rank decoder represents chromosome contact rates, and a library-size-normalized negative-binomial model relates these rates directly to raw counts. With clean single-cell targets unavailable, depth-normalized mini-pseudo-bulks from unsupervised subpopulations supply canonical seeds to initialize a population-conditioned diffusion prior. A generalized expectation–maximization (EM) framework then jointly trains the model: the E-step restores individual single-cell latents via raw-count data consistency and bounded diffusion proximal updates, while sequential M-steps iteratively optimize the observation likelihood parameters and refine the diffusion prior. Across 18 independent benchmark datasets spanning five biological collections, scHiDiffEM substantially improves overall performance across cell embeddings, contact-pattern imputation, and topological structure recovery without dataset-specific tuning. These results establish scHiDiffEM as a principled framework for scHi-C reconstruction, enabling faithful recovery of single-cell chromatin architecture from extreme sparsity without requiring unattainable ground truth.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
scHi-C, 3D genome, latent diffusion, expectation-maximization, sparse map imputation, single-cell, geometric decoder, population prior

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_EbQrOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper represents a new approach to generating representations for scHi-C data by formulating it as an inverse problem. Reconstruction of cells is done via a generalized EM-iterative procedure.

Critical strengths

Formulating this problem as a inverse problem is an interesting choice.

Critical weaknesses

The paper introduces an interesting idea, but fails to (1) convincingly describe the method, (2) describe its evaluation procedure carefully and (3) connect its conclusions to the experiment. It should therefore be rejected.

W1: The paper does not provide a well-described method sections. It misses a clear problem formulation, and during the section it remains unclear what are modeling assumptions made by the paper, and what is a new contribution by the author. Examples include:

  • L. 157, 262. What is the role of orthogonal gauge transformations?
  • L. 247. How does this prevent feedback loop error accumuluation
  • L. 258. What is a background threshold?

W2: The evaluation is missing critical details make it hard to assess the performance of the method. Examples:

  • Was a train/test split performed?
  • Why were these specific methods chosen as comparisons. The paper explicitly mentions the Plummer et al. 2025 as a benchmark paper, but does only implement a subset of methods.
  • What is the runtime / time-complexity of the method? (Re)-Training a diffusion model in every iteration of the M-step seems expensive.

W3: Coherent presentation The paper does a bad job of connecting it's experiments to its results, claiming overly broad results. Examples:

  • L483: While it shos results for restoring chromatin architecture, claiming that cell-specific chromatin architecture can be faithfully recovered by the method seems overly broad.
  • L484: It remains unclear why this framework would be considered principled.

Minor comments:

  • L472: The authors should expliclity state the advantages of the method.
Questions for authors

See above.

2 comments

Thank you for the comments. We agree that some parts of the previous draft were dense and unclear, and we have revised the writing throughout the paper to make it clear and accessible to readers from a wider range of fields. We have revised the manuscript accordingly and uploaded the revised PDF. Below we clarify each point.

W1: Method description. We have added a problem setup at the start of Section 3. The 3D chromatin conformation of each cell can be summarized by a dense contact map, a matrix whose entry (i,j)(i,j) is the contact frequency between genomic bins ii and jj. scHi-C sequencing captures only a small, depth-dependent fraction of these contacts, so the observed map is sparse and noisy. Recovering each cell's dense map from this observation is an ill-posed inverse problem, because most entries are zero and no clean single-cell maps exist for supervision. We therefore assume a low-rank contact structure plus a shared distance decay, NB-distributed counts scaled by sequencing depth, and shared structure among similar cells. This directly motivates our geometry-informed generative model, using a generalized EM with a diffusion prior learned from the cohort itself to recover the underlying dense map.

Though Negative-binomial count models, denoising diffusion, and EM algorithms are established tools, our contributions lie in (i) a geometry-informed generative model that separates the distance decay shared by all cells from cell-specific low-rank contact structure, and links inferred latents to raw counts through a library-size-aware Negative-binomial likelihood; (ii) a diffusion prior trained without clean single-cell maps, but using mini-pseudo-bulks and sparse counts downsampled from them; and (iii) a generalized EM solver that iteratively performs bounded MAP updates in the E-step while optimizing model parameters in the M-step, to recover each cell's contact map from its own observed counts under this prior.

Here are some details we want to clarify.

  • L157, L262 (orthogonal gauge transformations, now Sections 3.1, 3.2, and 3.4). Since in our contact rate parameterization, ZnZn⊤Z_nZ_n^\top is unchanged under any orthogonal rotation of ZnZ_n, we need to align the latent factors before training the prior (Section 3.2) and derive embeddings from the decoded maps (Section 3.4). Specifically, we align all mini-pseudo-bulk latent factors to a common orientation with Generalized Procrustes Analysis, so that the prior does not have to learn arbitrary rotations. We now use the term "orthogonal rotation" throughout the paper.
  • L247 (feedback-loop error accumulation, now Section 3.3). The prior is not trained on reconstructed single-cell latents. Instead, it is trained on mini-pseudo-bulk latent factors, and hence reconstruction errors are not fed back as training targets. We have clarified this in Section 3.3.
  • L258 (background threshold, now Section 3.4). This is τ=0.01\tau = 0.01 in Eq. 9. It sets near-zero decoded values to zero to suppress background noise. It is applied only when exporting the final dense map, not during training.

W2: Evaluation procedure.

  • Train/test split. scHi-C imputation and embedding are unsupervised, transductive (cohort-level) tasks (similar to matrix completion, clustering, and graph node embedding), and are hence conventionally evaluated without a separate train/test split. The goal is to recover the structure of the observed cells themselves, for which no clean ground truth exists. Existing methods (including scHiCluster, Higashi, Fast-Higashi, scVI-3D, see also Plummer et al., 2025) are likewise applied to the whole cohort, and some have no training stage at all (e.g., scHiCluster directly smooths each cell's map). We apply the same protocol to all methods, and cell-type annotations are used only for evaluation. To avoid circular evaluation, contact and structure recovery use a split-half design: each cell type is split into two halves, and predictions for one half are compared with the raw pseudo-bulk of the other, so query cells never enter their own reference (Section 4.1).

  • Comparison methods. Plummer et al. (2025) benchmark embedding tools instead of dense-map recovery, and we only adopted their embedding metrics. Most tools in that benchmark output only embeddings (e.g., scGAD converts maps into gene-level scores), and several methods integrate scHi-C with other omics modalities such as scRNA-seq. Hence, they are not suitable to be evaluated on contact imputation or structure recovery. We therefore chose four widely used methods (all published in top-tier journals), covering the main method families: smoothing (scHiCluster), hypergraph learning (Higashi), tensor factorization (Fast-Higashi), and generative modeling (scVI-3D).

  • Runtime. The diffusion network is not retrained from scratch. Before EM, it is trained for 20 epochs as a warm-up. In each EM iteration, the E-step updates each cell's latent factors 20 times, where each update involves one likelihood gradient step followed by one denoising step (there is no full diffusion sampling). The M-step then trains the observation parameters for 1 epoch and the diffusion network for 5 epochs (Appendix C). On HiRes-200 kb, scHiDiffEM takes 43 min for training and imputation, versus 90 min for Higashi and 142 min for scVI-3D (see our response to Reviewer vNS4, W1).

W3: Connecting conclusions to experiments.

  • L483 (now Section 5). We agree that this statement was overstated and have revised it. The Discussion now states only what the experiments show: superior cell representations, superior contact imputation with the largest gains in cell-type-specific metrics, and structure recovery comparable to the strongest baseline. As stated in the Figure 3 caption, the opposite-half references validate cell-type-level patterns, not individual-cell ground truth (like many other problems in single-cell genomics, we have no access to the ground truth due to technical limitations).
  • L484 (now Section 5). We call the framework principled because the reconstruction follows from an explicit statistical model instead of heuristic smoothing. It is formulated as an inverse problem: raw counts are linked to latent contact rates through a library-size-aware negative-binomial likelihood, the latent space has a geometric low-rank structure whose rotational ambiguity is explicitly handled, and the population prior and single-cell restoration are coupled through generalized EM. Each component therefore has a clear probabilistic role, in contrast to methods that smooth transformed maps without an observation model.

Minor. We have removed "Despite the obvious advantages" and now state the advantages explicitly in the abstract and Section 4.2. Compared with existing methods, scHiDiffEM (i) avoids the over-smoothing of scHiCluster and Higashi, which blurs local contact patterns (Figure 3), and outperforms both in structure recovery on all 18 datasets; (ii) preserves cell-to-cell heterogeneity, whereas scVI-3D pulls predictions toward the population average, as reflected in its much lower type-specific Pearson and top-1 retrieval; (iii) produces better cell embeddings, with the highest embedding aggregate against every baseline in at least 16 of 18 datasets, while Higashi and scVI-3D embeddings degrade sharply on HiRes brain; and (iv) models raw counts and sequencing depth directly, uses one set of hyperparameters for all datasets, and runs faster than Higashi and scVI-3D.

Reviewer_vNS4OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors have proposed an approach for single-cell Hi-C contact map imputation due to sampling dropouts or other experimental artefacts.

Critical strengths
  1. The authors combined EM with diffusion which could be interesting to ICLR audience.
  2. The single-cell Hi-C inference has broad implications in genomics.
  3. The authors have provided fairly comprehensive evaluations on his proposed method in 3 angles (cell embeddings, contact imputation, and chromatin structure)
  4. It is impressive that the authors have fixed K=10 and D=50 across 18 benchmarks while still showing competitive performance results.
Critical weaknesses
  1. Both EM and diffusion are computationally intensive. It is not clear to me with its running costs.
  2. The proposed method relies on population references. If we have population references in Figure 1, why do we still bother with single-cell Hi-C reconstruction ?
  3. The original EM algorithm has rigorous proofs on its convergence. Combined the E-step with diffusion, I am not sure if convergence still exists.
Questions for authors

See weaknesses.

1 comment

Thank you for the positive assessment. We have revised the manuscript accordingly and uploaded the revised PDF. Here is the point-by-point response on your comments.

W1: Running cost. We report the training time of our original runs on all 18 datasets in Table 16 of the revised PDF. We also report the running time of all methods in the multi-seed comparison, including both training and imputation, in minutes (Table 15, mean±SD over five seeds). scHiCluster is deterministic and was run once. Fast-Higashi is a tensor decomposition method that outputs only embeddings, and its running time covers embedding learning only (without imputation).

MethodCell cycle-200 kbHiRes-200 kbHiRes-100 kb
scHiDiffEM116.0±1.343.0±0.458.1±0.1
scHiCluster41.019.444.4
Fast-Higashi27.4±0.210.9±0.115.6±0.1
Higashi142.1±2.690.0±2.4153.4±12.9
scVI-3D377.9±16.9142.0±6.6148.2±3.9

scHiDiffEM is faster than other deep learning methods (Higashi and scVI-3D) on all three datasets. This holds even though scVI-3D was already run in a simplified configuration (latent dimension 4 per distance band, 1 Mb embedding range, default 400 epochs per model). With its default full distance range, a single HiRes run did not finish within two days. Across all 18 datasets, training takes from 0.56 h (HiRes-500 kb) to 13.4 h (PFC-50 kb, which has the most bins; Appendix J).

The cost is moderate. Before EM, the diffusion network is trained for 20 epochs as a warm-up. In each EM iteration, the E-step updates each cell's latent factors 20 times, where each update involves one likelihood gradient step followed by one denoising step, and there is no full diffusion sampling. The M-step then trains the observation parameters for 1 epoch and the diffusion network for 5 epochs. One latent-gradient pass costs O(NMDK)O(NMDK) for NN cells, MM bins, band width DD, and rank KK.

W2: Why reconstruct single cells when population references are used? Population pseudo-bulks average out cell-to-cell variation, which is exactly what scHi-C is designed to measure; single-cell reconstruction preserves this variation for cell-level analyses. The population information we use is built internally from the input cohort, and no external reference maps are used (clean, dense single-cell-level reference maps do not exist with current experimental techniques). This information trains a shared prior, but each cell's contact map is inferred from its own raw counts through the negative-binomial likelihood, so cells from the same population receive different maps. Figure 3a shows that the scHiDiffEM prediction preserves the individual cell's observed contacts without collapsing to the population average. In a nutshell, the prior regularizes sparse regions, while the cell's own data determine the reconstruction.

W3: Convergence. We agree that classical EM guarantees do not carry over here because of the approximate update we use. As stated in Sections 2 and 3.3, the procedure is a generalized EM scheme without a monotonic convergence guarantee. Appendix K (referenced in Section 3.3) now shows optimization trajectories for all 18 datasets: the observation pseudo-NLL drops within about 10 iterations and then plateaus, and the DSM and recovery losses decrease steadily. Although we do not provide a theoretical convergence proof, these curves show that training is stable in practice.

Reviewer_J7PuOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Single-cell Hi-C (scHi-C) data allows the capture of contacts in individual cells, whereas bulk Hi-C captures population-averaged frequencies. Although the advent of scHi-C allows researchers to look at cell-type heterogeneity in 3D genome organization, it comes with many challenges related to noise, sparsity, and dropout. Existing methods have the problems of dampening true biological signals and masking cell-type-specific contacts.

The paper presents scHiDiffEM, a method that reconstructs scHi-C maps using an expectation-maximization (EM) framework. There are several components to this method: first, a decoder represents chromosome contact rates that are aware of genomic distance decay. A negative binomial distribution is used to relate raw counts to these chromosome contact rates. A method called “mini pseudo-bulking” is used to generate scHi-C map targets (due to the lack of ground-truth maps for training and testing sets). Then, the model is trained using an EM framework, in which the E-step consists of restoring contacts in the latent embedding. The “mini-pseudo-bulks” are used for initialization. The M-step consists of optimizing model parameters. The method is benchmarked using 18 different datasets, and performance is compared with existing methods on three main tasks: cell embedding, contact imputation, and recovery of 3D structures (such as TADs).

Critical strengths

(1) The “mini-pseudo-bulking” approach to deal with the lack of ground-truth scHi-C maps in this context is unique and compelling, since the number of cells used for each pool is small enough that we don’t expect to see too much dampening of biological signals. Additionally, the decision to split each cell type into halves, to prevent data leakage during the mini-pseudo-bulking operation, is sensible and a creative way to handle the lack of ground truth.

(2) Figure 3 does a great job demonstrating how scHiDiffEM avoids the “over-smoothing” problem; however, the labels are inconsistent (panel A says “Ours” and panels B and C say “Diffusion-EM”).

(3) The method itself is explained thoroughly.

Critical weaknesses

(1) The big concern is that there are no variance or significance tests in Tables 1 and 2. Perhaps the metrics should be reported across random seeds to ensure that scHiDiffEM is truly outperforming baselines? This is especially important in instances where reported improvements are very minor- i.e. in Table 2, for the Structure Recovery task, scHiDiffEM reports a mean aggregate of 3.5124 while scVI-3D reports a mean aggregate of 3.5117, which is a very small difference. Significance tests are definitely needed.

(2) It is stated that “scHiDiffEM achieves superior overall performance across cell-type clustering agreement, contact pattern recovery, and insulation and boundary metrics”; however, upon looking at Tables 1 and 2, there are several instances where other methods outperform scHiDiffEM. The paper would benefit from directly acknowledging these instances and perhaps adding a paragraph in the discussion.

(3) Cross-seed stability is only reported for one dataset.

Minor points:

(1) Replace the model label “Ours” with the name of the method, scHiDiffEM, throughout the figures/tables

(2) The method is limited to a distance of 50 genomic bins from the diagonal, and it is not explored how the performance of the method is affected as you increase the number of genomic bins, and how this compares with existing methods (as acknowledged in the limitation). A figure on this (with heatmaps showing results) or a table showing how the metrics change as you move further away from the diagonal would be useful to include.

Questions for authors

(1) What about the scVI-3D in the Figure 3 comparison? How does scHiDiffEM perform in comparison to that method, which has so far been reported in all the benchmarking tables/figures? Same with Figures 9-13.

2 comments

Thank you for the constructive suggestions. We have revised the manuscript accordingly and uploaded the revised PDF. Here is the point-by-point response on your comments.

W1: Variance and significance tests. We compared scHiDiffEM with each baseline using two-sided paired Wilcoxon signed-rank tests on the task aggregates of all 18 datasets in Table 2 (Appendix H, Table 12). Each entry reads "wins, pp", where wins is the number of the 18 datasets on which scHiDiffEM has the higher task aggregate:

ComparisonEmbeddingImputationStructure
scHiDiffEM vs scHiCluster17/18, p=7.6×10−5p=7.6\times10^{-5}18/18, p=7.6×10−6p=7.6\times10^{-6}18/18, p=7.6×10−6p=7.6\times10^{-6}
scHiDiffEM vs Higashi16/18, p=3.8×10−5p=3.8\times10^{-5}18/18, p=7.6×10−6p=7.6\times10^{-6}18/18, p=7.6×10−6p=7.6\times10^{-6}
scHiDiffEM vs Fast-Higashi17/18, p=3.8×10−5p=3.8\times10^{-5}––
scHiDiffEM vs scVI-3D18/18, p=7.6×10−6p=7.6\times10^{-6}13/18, p=0.010p=0.01010/18, p=0.83p=0.83

scHiDiffEM is significantly better than all baselines in cell embedding and imputation. In structure recovery, scHiDiffEM significantly outperforms scHiCluster and Higashi, while the difference between scHiDiffEM and scVI-3D is not significant.

We also reran all methods with five seeds on three datasets, keeping the evaluation cells and split-half assignments fixed (Appendix H, Table 11). The table below reports the mean±SD of the task aggregates, defined as in Table 2. scHiCluster is deterministic, so its values are taken from the main benchmark.

TaskDatasetscHiDiffEMscHiClusterHigashiFast-HigashiscVI-3D
EmbeddingCell cycle-200 kb1.9183±0.05021.69091.8550±0.04541.6010±0.02451.5110±0.0300
HiRes-200 kb2.0153±0.11691.48770.7180±0.04141.5077±0.02660.1853±0.0274
HiRes-100 kb1.8342±0.04131.66540.6001±0.02101.3960±0.07360.0560±0.0378
ImputationCell cycle-200 kb4.3585±0.00212.98303.0155±0.1005–4.1769±0.0107
HiRes-200 kb3.4861±0.00132.82082.4506±0.0960–2.8582±0.0139
HiRes-100 kb2.9573±0.00692.54642.1710±0.1009–2.6367±0.0133
StructureCell cycle-200 kb3.5206±0.00032.90192.8832±0.0698–3.5209±0.0023
HiRes-200 kb3.5628±0.00133.10482.7792±0.1551–3.4826±0.0031
HiRes-100 kb3.4489±0.00063.01712.7768±0.0780–3.3437±0.0030

scHiDiffEM has the highest mean in 8 of the 9 cases. In the remaining case (structure on Cell cycle-200 kb), it is within one SD of scVI-3D. Both analyses are consistent with the revised claim.

W2: Instances where other methods outperform scHiDiffEM. We have revised the central claim accordingly: scHiDiffEM outperforms existing methods in cell embedding and contact imputation, and is comparable to the strongest baseline in structure recovery. A new paragraph in Section 4.2 analyzes where baselines perform better, and the pattern is systematic. scVI-3D closely fits the smooth distance-decay profile shared by all cells. This profile dominates SCC and the global and stratum-wise correlations, so scVI-3D scores higher on these metrics, particularly at fine resolutions, where single-cell maps are sparsest and distance decay accounts for most of the signal. The same fit, however, pulls predictions toward the population average. After the shared profile is removed, scVI-3D retains much less cell-type-specific signal than scHiDiffEM (type-specific Pearson 0.1670 vs 0.2451, top-1 retrieval 0.4577 vs 0.7213). In structure recovery, scVI-3D is higher in all six datasets at 500 kb or 1 Mb, where insulation profiles are smooth, whereas scHiDiffEM is higher in 10 of the 12 datasets at 50–200 kb, where sharp domain boundaries must be resolved.

W3: Cross-seed stability on one dataset. Table 5 in the revised PDF now covers four datasets from three collections. In all four, the SD is below 0.001 for SCC and below 0.0015 for the structure aggregate. Clustering varies more (ARI SD 0.026–0.089). All of these variations are far smaller than the differences between methods in Table 2.

M1: Label "Ours". All figures and tables now use the label scHiDiffEM.

M2: Distance beyond 50 bins. We added both analyses on all 18 datasets. They are summarized in Section 4.3, with per-dataset results in Appendix I (Tables 13–14, Figures 15–16). The preprocessed PFC-100 kb data contain no contacts beyond 20 bins, so this dataset is excluded beyond that range.

(a) Extending the band from D=50D=50 to D=100D=100. With all other settings fixed, the D=100D=100 model improves SCC on bins 1–50 in 17 of the 18 datasets and type-specific Pearson in 16 of 18, so modeling a wider band does not hurt the original range. On bins 51–100, which the baselines do not cover, it exceeds the raw input in both metrics on all 17 scored datasets (PFC-100 kb is excluded because there is no contacts beyond 20 bins in our downloaded dataset, mean SCC 0.2161 vs 0.1655, mean type-specific Pearson 0.1581 vs 0.1271). Full results are shown in Table 14.

(b) Comparison with existing methods across distance. The advantage of scHiDiffEM in cell-type-specific recovery is largest at bins 1–10, where the contact signal is strongest. There it ranks first in top-1 retrieval on all 18 datasets and in type-specific Pearson and margin on 17. Here is the mean over 18 datasets, with the number of datasets ranked first in parentheses (full results in Table 13):

MethodType-specific PearsonTop-1 retrievalMargin
scHiDiffEM0.2578 (17)0.7294 (18)0.0963 (17)
scHiCluster0.0980 (0)0.2437 (0)−0.0732 (0)
Higashi0.0969 (0)0.2663 (0)−0.1067 (1)
scVI-3D0.1753 (1)0.4643 (0)−0.0117 (0)

At bins 11–25, scHiDiffEM still has the highest mean top-1 retrieval (0.5887) and margin (0.0306). At bins 26–50, cell-type-specific signal is weak for all methods (mean margins between −0.06 and 0.03), and scHiDiffEM no longer leads. scVI-3D has the highest SCC at all distances, consistent with Table 1.

Q1: scVI-3D in Figure 3. Figure 3 and Figures 9–14 now include scVI-3D. scVI-3D recovers the overall domain layout in Figure 3, but its predictions vary little across the DG, Inh, and MGE examples in Figure 13. This is consistent with its lower cell-type-specific scores and negative mean retrieval margin in Table 1, and reflects smoothing toward the shared population profile without preserving heterogeneity among different cell types.

Comments on the paper

3 comments

Many thanks for the follow-up. We address W1 and W3 below.

W1: Mathematical formulation of the inverse problem.

Consider one chromosome with MM bins. Each cell nn has an unknown dense contact map Rn={rnid}R_n=\{r_{nid}\}, where rnidr_{nid} is the expected contact count between bins ii and i+di+d at a specific sequencing depth (1≤d≤D1\le d\le D, 1≤i≤M−d1\le i\le M-d). scHi-C sequencing does not observe RnR_n directly. It samples a small, depth-dependent number of contacts from it, producing a sparse count matrix Yn={ynid}Y_n=\{y_{nid}\} in which most entries are zero. The observation is therefore generated from the dense map by a stochastic forward operator, Yn∼Aℓn(Rn),Y_n\sim\mathcal{A}_{\ell_n}(R_n), and recovering RnR_n from YnY_n is an inverse problem. It is analogous to recovering a clean image from a corrupted observation in which most pixels are missing and the remaining ones are noisy (like image inpainting and denoising).

Forward model. The operator AℓnA_{\ell_n} scales the dense map by the relative library size ℓn\ell_n of the cell and samples each entry independently as ynid∣rnid∼NB(ℓnrnid,θd),y_{nid}| r_{nid}\sim\mathrm{NB}(\ell_n r_{nid},\theta_d), where θd\theta_d is a distance-dependent dispersion that accounts for overdispersion beyond Poisson sampling (variance exceeding the mean, which is a common phenomena in single-cell genomics).

Inverse problem. Each entry is observed only once, and most counts are zero (due to both sampling dropouts and biological true zero). The maximum-likelihood solution r^nid=ynid/ℓn\hat r_{nid}=y_{nid}/\ell_n therefore only reproduces the corrupted observation, so the problem is ill-posed. We regularize it in two ways.

  1. Parameterization. We leverage the geometric structures behind sc-HiC contact map: rnid=softplus(zni⊤zn,i+d+bϕ(d))r_{nid}=\mathrm{softplus}(z_{ni}^\top z_{n,i+d}+b_\phi(d)) with Zn∈RM×KZ_n\in\mathbb{R}^{M\times K} (Eq. 1), which is a combination of a cell-specific rank-KK term and a distance decay bϕb_\phi shared by all cells.
  2. Population Prior. Zn∼pψ(Z∣qn)Z_n\sim p_\psi(Z\mid q_n), a diffusion model conditioned on the subpopulation qnq_n (obtained via clustering) of cell nn, so that we can borrow information from cells with similar identities. Here we assume that cells of the same identity have similar true contact rates, so that their sparse counts can be viewed as independent samples from closely related dense maps.

Estimator. Each cell's map is the MAP estimate Z^n=arg⁡max⁡Z ∑i,dwdlog⁡pNB(ynid∣ℓnrnid(Z),θd)+log⁡pψ(Z∣qn),R^n=R(Z^n),\widehat Z_n=\arg\max_{Z}\ \sum_{i,d} w_d\log p_{\mathrm{NB}}\big(y_{nid}\mid \ell_n r_{nid}(Z),\theta_d\big)+\log p_\psi(Z\mid q_n),\qquad \widehat R_n=R(\widehat Z_n), where the first term is the distance-weighted log-likelihood (Eq. 2). Because log⁡pψ\log p_\psi is defined only through the score network, we approximate the maximization in a plug-and-play manner. Each update takes a gradient step on the likelihood and then a bounded denoising step (Eqs. 6–8).

Model training. Since ϕ\phi, θ\theta, and ψ\psi are unknown, and no clean maps exist, we employ the proximal EM updates. Given the current {Z^n}\{\widehat Z_n\}, (ϕ,θ)(\phi,\theta) maximize the same likelihood, and ψ\psi is trained on mini-pseudo-bulk latent factors (Eqs. 3–5). Alternating the MAP step (E-step) with these updates (M-step) gives the recovered map (generalized EM procedure in Section 3.3).

We have updated a more rigorous mathmatical formulation at the begining of section 3 in the revised PDF.

W3: Linking conclusions to experiments.

Sharp domain boundaries. Our structure recovery task directly evaluates domain boundaries. Insulation profiles and boundary calls from the predicted maps are compared with held-out references. scHiDiffEM attains the highest insulation-profile correlations (Pearson 0.9141, Spearman 0.8691) and the highest boundary AUPRC (0.2442) among all methods (Table 1). Its structure aggregate is higher than that of scHiCluster and Higashi on all 18 datasets (Wilcoxon p=7.6×10−6p=7.6\times10^{-6}). It is also higher than that of scVI-3D on 10 of the 12 datasets at 50–200 kb, where boundaries are sharpest. Visually, scHiCluster and Higashi blur domain borders, whereas scHiDiffEM effectively keeps them (Figure 3, Figures 9–12).

Not collapsing cells to their population average. Four results support the claim.

  • Type-specific imputation metrics: These metrics first subtract the pseudo-bulk of all cells, which is the population average, and then test whether a prediction still matches its own cell type. A prediction that collapsed to the population average would score near zero. scHiDiffEM reaches a type-specific Pearson of 0.2451, versus at most 0.1670 for the baselines. It retrieves the correct cell type in 72% of cases, versus at most 46% for compared methods, and it is the only map-producing method with a positive margin (Table 1).
  • Cell embeddings: The embeddings are computed from the reconstructed maps (Section 3.4). If those maps collapsed to the population average, cell types could not be separated. Instead, scHiDiffEM gives the highest embedding aggregate (ARI 0.7046, NMI 0.7325) and a much higher silhouette than all baselines (0.1668 vs. at most 0.0840). Its UMAPs also show the clearest cell-type separation (Figure 2, Figures 4–8).
  • Sensitivity analysis on prior strength: The sensitivity analysis shows the risk explicitly: an overly strong prior pulls reconstructions toward the population average and lowers cell-type specificity (Table 4). The bounded prior step in the E-step is designed to avoid this.
  • Reconstructed contact map: Qualitatively, the visualization of reconstructed contact map of scHiDiffEM for DG, Inh, and MGE cells differs from each other, whereas those of scVI-3D are nearly identical (Figure 13).

We agree that these results are measured on cell-type aggregates and embeddings rather than on individual-cell ground truth (which does not exist), and we agree that the original words were used in a slightly absolute way. We are therefore happy to make the wording more precise to a comparative style: "…generalized EM imputes sparse contacts, better preserves domain boundaries, and retains cell-type-specific contact patterns instead of collapsing cells to their population average."

We hope these clarifications address your concerns, and we would be glad to answer any further questions.

Thank you for the follow-up. We agree that scVI-3D is a strong baseline while there are clear advantages of scHiDiffEM over scVI-3D. We summarize below:

  • Cell embeddings. scHiDiffEM has a higher embedding aggregate than scVI-3D on all 18 datasets (2.33 vs 1.75 on average, Wilcoxon p=7.6×10−6p=7.6\times10^{-6}). On the three HiRes brain datasets, scVI-3D nearly fails to separate cell types (aggregate 0.02–0.27), whereas scHiDiffEM remains stable (1.80–1.83).
  • Cell-type-specific imputation. After the contact profile shared by all cells is removed, scHiDiffEM retains much more cell-type-specific signal: type-specific Pearson is 0.2451 vs 0.1670, and top-1 cell-type retrieval is 0.7213 vs 0.4577. The mean retrieval margin of scVI-3D is negative (−0.0119), so its predictions are, on average, closer to another cell type's reference than to their own. The margin of scHiDiffEM is positive (+0.0847). Overall, scHiDiffEM has the higher imputation aggregate on 13 of the 18 datasets (p=0.010p=0.010).
  • Boundary recovery at fine resolutions. At 50–200 kb, where domain boundaries are sharpest, scHiDiffEM has the higher structure aggregate on 10 of the 12 datasets.
  • Efficiency. scHiDiffEM is 2.5–3.3 times faster (43 vs 142 min on HiRes-200 kb, 116 vs 378 min on Cell cycle-200 kb, and 58 vs 148 min on HiRes-100 kb). This holds even though scVI-3D was run in a reduced configuration.

scVI-3D scores higher only on SCC and the global correlations. These metrics are dominated by the distance decay shared by all cells, so a method can score well on them while predicting similar maps for different cells. Since single-cell Hi-C is used to resolve differences between cells, we believe the advantages above make a clear case for scHiDiffEM.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.