acceptodds
Under review as a conference paper at ICLR 2027

Multi-Task Graph Learning for Cell–Cell Communication Inference in Spatial Transcriptomics

Abstract

Intercellular communication is fundamental to multicellular life, orchestrating tissue development, maintaining homeostasis, and driving the pathogenesis of complex diseases. Recent advances in spatial transcriptomics have enabled high-resolution, in situ mapping of intricate ligand-receptor (L-R) interactions within native tissue contexts. However, existing computational approaches often underutilize spatial topological information and overlook the distinct regulatory functions of different L-R interaction types, thereby limiting their capacity to capture the dynamic heterogeneity of cellular communication in complex microenvironments. To address these limitations, we propose CellMM, a novel multi-task learning framework powered by multi-relational graph neural networks. CellMM constructs a multi-relational graph that integrates gene expression profiles, spatial coordinates, and L-R interaction types. It employs relational graph convolutional networks to learn expressive node embeddings and leverages a multi-head attention mechanism to adaptively fuse multi-layer graph features, enabling the identification of active L-R signaling events between spatially adjacent cells. Experiments on three simulated datasets and applications to three real spatial transcriptomics datasets show that CellMM outperforms baseline methods in simulated evaluations and recovers biologically plausible communication patterns in real data.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
spatial transcriptomics, cell-cell communication, ligand-receptor interactions, multi-task learning, graph neural networks

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

This paper studies the problem of cell–cell communication inference from spatial transcriptomics data and proposes a multi-task graph learning framework built on multi-relational GNNs. The method constructs a gene-level graph in which ligand–receptor edges are categorized into spatial interaction types using a distance ratio based on optimal transport. It then learns node embeddings with RGCN layers fused by multi-head attention, and jointly optimizes L–R link prediction and auxiliary relation classification. Experiments are conducted on three simulated datasets and three real spatial transcriptomics datasets.

Critical strengths
  1. The studied application is meaningful. Inferring cell-cell communication from spatial transcriptomics is an important and active problem.
  2. Introducing a heterogeneous graph modality is interesting, and it offers a different modeling perspective from existing methods.
  3. The method is evaluated on both simulated and real datasets spanning different tissues and species.
Critical weaknesses
  1. Regarding the presentation, Figure 1 is confusing; the intuition behind using the ratio d_ratio is not explained; the notation in Eq.2 and Eq.3 is not aligned; and the basic statistics of the constructed graph are missing.
  2. An important ablation study on L-R interactions is missing. The paper claims its main contribution is leveraging the relationships among different types of L-R interactions to predict L-R pairs. To validate this, the authors should include an ablation that collapses all relation types into a single edge type (i.e., a standard GCN over the same graph) and check whether the multi-relational design brings any improvement.
  3. The method lacks major algorithmic novelty. The overall contribution reads as an application-level combination rather than a new learning method.
  4. Many experimental details and analyses are missing, including how the gene pairs are split into training and test sets, whether the target edges remain in the graph during message passing, and no sensitivity study is provided.
  5. Another major concern I have is that the proposed modeling does not truly infer cell–cell communication, which is inconsistent with the paper’s claims. The paper claims to identify “active L-R signaling events between spatially adjacent cells.” However, the constructed graph is defined at the gene level: nodes are ligand and receptor genes, and cells only appear as the dimensions of the node features. The model never represents individual cells or their spatial adjacency, and its output is a dataset-level list of L-R pairs.
Questions for authors

See Weaknesses.

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

This paper considers inferring cell-cell communication patterns from spatial transcriptomics data using ligand-receptor pairs; the authors present an intuitive illustration of how such communication patterns might look in Figure 4. For their approach, they make use of gene-expression profiles as well as spatial information. They calculate Pearson correlation coefficients between gene pairs, and select the highest and lowest correlated ones for training. Based on the selected genes, they construct a matrix of pairwise Euclidean distances from the locations of the sampled cells. The authors define a pipeline for learning representations, building on the idea that their predictor should learn both the relationships between the cells and their types. The authors evaluate their method on a range of synthetic and empirical datasets. In the empirical datasets, their method outperforms the chosen baselines, however, because the empirical datasets come without any ground truth, they cannot estimate their method's performance on real-world data. In an ablation study, the authors show that their method's performance would degrade if they dropped their multi-task learning approach or changed the way they constructed the features.

Overall, I find the paper is well written, and the authors' goal can be understood clearly. Their approach seems to make sense and follows standard strategies in deep learning. However, I find that the interesting questions have been deferred to future work: exploring the learnt representations and analysing how and why they work.

The authors provide their appendices as part of their supplementary material, which is a little uncommon in the ML community, but, as far as I am aware, is the standard in medical fields.

Critical strengths

I believe the most outstanding strength is the authors' realisation that a multi-task approach is beneficial for learning cell-cell interaction patterns. As far as I am aware, this reflects the biological reality better than other, more naive approaches.

Critical weaknesses

To me, the central message of the paper remains unclear: the authors present a method that outperforms three baselines on three synthetic datasets, but it is unclear what performance we can expect on empirical data due to the lack of ground truth. I believe the authors should have invested some effort into analysing the learnt representations, ideally with the outcome of explaining why their method would be expected to work well on empirical data. Because of this, I believe the present work is a nice piece of engineering, but in the lack of further insights, it is difficult to recommend for publication.

Questions for authors

There are a couple of points that remained unclear to me

  1. You mentioned in the introduction that "CellMM [...] identifies biologically plausible communication patterns in real-data analyses". Could you elaborate on which of your results support this claim? I may have missed something, but as far as I understand it, the Jaccard similarities you computed between the predictions of different methods only say something about the extent to which they agree, but not about biological plausibility. That is, Fig. 2 shows the agreement between the methods, but not agreement with biological reality. Similarly, if I understand it correctly, Fig. 5 shows inferred interactions between ligands and receptors, but the results are not compared to any ground truth. What am I missing?
  2. I was wondering how sensitive the results are to the "correlation cutoff" of gene pairs? Say, for example, you would choose the top and bottom 5% quantiles instead of 1%?
  3. I feel like the details of constructing the "Euclidean distance matrix" remained a little vague, could you elaborate? Specifically, I understand that you construct the distance matrix based on the location of the cells. Then, you extract "the expression vectors of the corresponding ligand and receptor genes". Does that mean that all gene counts that belong to a certain cell receive the same location, that is, the location of the cell they belong to?
  4. I also didn't quite understand what you mean by that "redundant entries [are] removed through merging". Could you give an example of when an entry would be discarded?
  5. Where do the synthetic datasets come from, did you devise them yourselves? Are they taken from another study? What patterns are "hidden" in the synthetic data?
  6. I couldn't figure out the message of Figure 5, could you help me out? What is the reader supposed to learn from this heatmap? Is there anything expected or unexpected the reader should note?
  7. You mention in l. 324 that "its relatively lower F1-score suggests a higher false-positive rate". I suppose you should be able to check the false-positive rate directly?

Minor points

  • In l. 180, it looks like a backslash might be missing in the LaTeX formula
  • I suggest typesetting things like "input_dim", "hidden_dim", etc. with \texttt or similar
  • The typical ICLR/ML audience usually expects the appendix in the same document as the main text, after the references. I suggest considering moving the supplementary material there.
Reviewer_w5LYOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces CellMM, a multi-task learning framework that integrates multi-relational graph neural networks to infer ligand–receptor (L–R) interactions from spatial transcriptomics data. The central idea is to treat L–R link prediction and L–R spatial-pattern classification as two jointly optimized tasks, on the premise that the spatial category of an L–R interaction (short-, medium-, or long-range) carries information that regularizes and improves link prediction.

The stated contributions are: (1) the first framework to exploit interdependencies among distinct L–R interactions for L–R link prediction; (2) a multi-task RGCN model with adaptive feature fusion and adaptive loss weighting; and (3) empirical validation on three simulated and three real spatial transcriptomics datasets.

Validation consists of accuracy/recall/F1/AUROC comparisons against CellChat, SpaTalk, and spaCI on three simulated datasets (e.g., Simulated Dataset 3: ACC 92.12%, AUROC 96.19%, with t-tests); inter-method Jaccard consistency on melanoma, mouse brain, and human breast cancer data; ablations removing the multi-task objective (OLP) and the multi-layer fusion (NFV), which reduce F1 by 3.66–9.57% and 1.68–6.62% respectively; and a melanoma case study matching the top eight predicted L–R pairs to literature.

Critical strengths
  1. Framing spatial-pattern classification as an auxiliary task that shares the encoder but not the target-relation input is a reasonable way to inject structure without leaking the label. The paper is explicit that the relation decoder does not receive the target relation index or its embedding (Section 3.5), and the adaptive α is detached from the gradient so that it balances loss magnitudes rather than being co-optimized.

  2. Unlike many spatial-communication papers that report only real-data overlap, the paper constructs three simulated datasets with different density and noise regimes, compares against three representative baselines, reports four metrics, and applies independent-samples t-tests (Table 2).

Critical weaknesses
  1. The paper explicitly states there is no gold standard and therefore evaluates by inter-method Jaccard overlap (Eq. 12). This measures agreement, not correctness. All compared methods draw on overlapping L–R databases and all use spatial co-expression or proximity in some form, so high Jaccard can reflect shared priors rather than shared accuracy. The reported values are also modest: the best agreement in melanoma is 0.471 with SpaTalk, and 0.381 with spaCI; in mouse brain the best is 0.454 with spaCI. A method that is “most consistent with SpaTalk in melanoma but with spaCI in mouse brain and breast cancer” does not establish a stable notion of correctness. The melanoma case study (Fig. 5, Section 4.5) selects the top eight predicted pairs and matches them to literature post hoc; this is subject to selection and confirmation bias and cannot distinguish genuine discovery from plausible-sounding retrieval. Since the paper’s headline claim is improved L–R inference, the absence of any held-out, curated, or experimentally anchored evaluation is the single most consequential weakness.

  2. The baseline set is thin relative to the spatial-communication literature, and the training distribution is narrow. The paper cites COMMOT, SpatialDM, NICHES, SpaCcLink, and SpaTalk, but only compares against CellChat, SpaTalk, and spaCI. COMMOT and SpatialDM are directly comparable spatial-aware L–R inference methods and are omitted without justification. Separately, restricting training to the 1% tails of the PCC distribution discards the vast majority of gene pairs and may not match the distribution of true L–R pairs, which are not defined by co-expression extremes; no analysis shows that curated L–R pairs fall predominantly in these tails. Combined with the absence of runtime/memory reporting and of RGCN depth/hidden-width sensitivity, this limits both the strength of the empirical claim and reproducibility.

Questions for authors
  1. Can you evaluate CellMM on a setting with known interactions — e.g., a simulated dataset whose generative process is disclosed and independent of your label construction, or a curated held-out L–R benchmark with positives and matched negatives? Concretely, what is the AUROC of CellMM on distinguishing true curated L–R pairs from co-expression-matched non-pairs?

  2. Can you report a variant in which the auxiliary relation labels are randomly shuffled across pairs (preserving class balance) while keeping the multi-task architecture and loss weighting fixed? If the gain over OLP persists under shuffled labels, the auxiliary task is acting as a regularizer; if it disappears, the spatial categories carry genuine signal.

  3. Can you add COMMOT and SpatialDM to the simulated and real comparisons?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.