acceptodds
Under review as a conference paper at ICLR 2027

GO-Based Clustering for Learning Cluster-Level Causal Gene Regulatory Networks

Abstract

Discovery of causal relationships in high-dimensional Gene Regulatory Networks (GRN) is computationally challenging and often difficult to interpret due to dense connections. Therefore, grouping genes together into functional modules can improve tractability and biological interpretability. However, existing cluster level causal discovery methods assume access to a predefined admissible partitions, requiring the graph over clusters to be acyclic. Constructing such partitions is therefore challenging. In this work, we introduce GO-based Clustering for Causal Discovery (GO4CD), an algorithm that uses Gene Ontology (GO) to construct biologically meaningful gene partitions at multiple levels of granularity, while favoring those more likely to be admissible for causal discovery. GO4CD groups together genes participating in a shared biological process, and propagates gene annotations through the ontology hierarchy to achieve different granularity of partitions. Furthermore, we integrate GO4CD with Causal Learning over Clusters (CLOC) algorithm and evaluate recovery of true Markov equivalence class both with an oracle of conditional independencies and on simulated gene expression data using multivariate conditional independence tests. We evaluate GO4CD on multiple E.coli regulatory subnetworks and find that it is inadmissible in 18.1% of the cases, compared with 65.3-82.3% for the semantic-similarity baselines. Our results indicate that GO4CD is substantially better suited to learning causal GRNs defined over biologically meaningful gene clusters.

Primary area:
causal reasoning
Keywords:
causal discovery, causal abstraction, genes regulatory networks, ontology, graph clustering, semantic similarity

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

Some of the current methods for learning DAGs over clustered groups of variables need the variable clusters to be specified a priori. This partition must be admissible, meaning no cycles should exist between the variable groups. Specifying such a partition can be quite challenging; for instance, in the context of gene regulatory networks, feedback loops are common between genes.

The authors present GO4CD, a method for obtaining partitions over genes that (a) is based on biological knowledge, namely Gene Ontology Biological Processes, and (b) shold in principle lead to partition that are likely to be admissible.

The authors couple GO4CD with the causal discovery algorithm CLOC, which can provably reconstruct the DAG over clustered variables under specific assumptions. Quantitative analysis on simulated gene expression data provide evidence for the effectiveness of the proposed method.

Critical strengths

The GO4CD algorithm is a straigthfoward solution for the complex problem of identifying functional modules in gene regulatory networks. This type of modelling is useful and relevant, as it can lead to more intepretable models for the biological machinery underlying the cell operation.

Critical weaknesses

My main concern relates to the actual interpretability of the produced partitions. Table 1 indicates that most of the genes congregates in a single giant cluster, while the remaining genes mostly remain as singletons. A second concern is the absence of evaluation on actual gene expression data. The authors retrieve E coli known regulatory structures from Regulon DB, however no transcriptomics data from E. coli experiments are used. Thus, it is not really possible to evaluate how GO4CD would perform in a real study.

Questions for authors

I would like the author to answer my concerns described above. In particular:

  1. Can they elaborate on how a useful biological interpretation can be obtained when most of the genes are grouped in a single cluster? It would be interesting if the authors could apply GO4CD to Human Gene Ontology Biological Processes, and characterize the clusters obtained at different resolutions. Would you obtain a more or less balanced partition depending on the chosen resolution? Under which conditions Biological Processes grouped in the same cluster would actually form a meaningful regulatory module?
  2. While I realize that time constraints might be tight, I nevertheless trongly encourage the author to retrieve and analyze E coli actual gene expression data for their analysis on the E coli regulatory network
Reviewer_xvDoOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a clustering algorithm for building partitions in gene networks, so that the clusters could be used later to learn efficiently relations between the gene groups.

The proposed algorithm is validated in a simple E. coli gene network and shows improvement over classical algorithms.

The proposal seems to be highly specific to the particular application and not very generalizable.

Critical strengths

Paper was able to show improvements over classical algorithms

Critical weaknesses

The paper proposal seems to be very specific for this particular application that is not really the focus of ICLR. It seems to be the benefit for the average ICLR participant would be practically zero because I can't really see this clustering algorithm being directly applied to other applications.

The manuscript could be significantly improved, Too much space has been dedicated to the fundamentals and still it wasn't clear to me why ICLR would be a good venue for this paper.

Questions for authors

No specific questions

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

This paper considers causal discovery over clusters of variables, focused on gene regulatory networks. The paper uses Gene Ontology (GO) as a resource, which groups genes by the biological processes in which they participate and records regulatory relationships; this prior knowledge can be represented as a DAG. From this graph, the authors propose a procedure for partitioning the genes into clusters (based on a choice of granularity represented by a depth parameter dd, and contracting strongly connecting components to ensure an admissible partition), these clusters are then used to form a Cluster DAG (C-DAG) from the original DAG; here, admissible means that the resulting cluster-level graph is indeed acyclic.

In synthetic experiments, the authors use a curated E. coli gene regulatory network from RegulonDB, edited to be acyclic. The experiments investigate whether different clustering methods return admissible partitions, and whether a cluster-level causal discovery algorithm returns the correct structure. As expected from the method design, their method performs better than acyclicity-agnostic clustering baselines, both in "oracle mode" (using d-separation for conditional independence) and in the finite-sample setting (where data is generated using CausalRegNet and conditional independence tests use the Generalised Covariance Measure).

Critical strengths

The overall direction studied by the paper is an interesting one, combining two threads in causal discovery: (1) cluster-based causal discovery, related to causal abstraction, and (2) integration of prior knowledge. Especially in high-dimensional settings, as common in gene regulatory network inference, these two threads are promising for greater statistical/computational scalability and stability; hence, it is good to see work at their intersection.

Critical weaknesses

The paper is somewhat poorly written, with nearly two pages devoted to preliminaries, many of which could be abstracted / relegated to the appendix. Overall, the main weakness is the lack of depth: the main novelty of the paper is essentially a preprocessing algorithm that takes a gene ontology graph (already assumed to be a DAG) and coarsens it. In particular, this algorithm is separate from the downstream causal structure learning step; the partitioning is not informed by the statistical data (Algorithm 1 does not take a dataset as input), and there is nothing particularly clever about clustering algorithm.

I'd encourage the authors to continue pursuing the direction, but greater depth is needed for a conference like ICLR/ICML/NeurIPS, e.g., integration between the clustering and structure learning stage, insightful theoretical analysis, and/or real-data results with a careful comparison to converging evidence from other analyses.

Questions for authors
  1. Please respond to the weaknesses listed above, especially regarding any potential misconceptions about whether statistical data (i.e., samples and conditional independence tests) are used to construct the partition, or if they are only used for structure learning.

  2. By design, it seems that Algorithm 1 should never return an inadmissible partition (since the while loop always contracts SCCs), so why do the experimental results show that it sometimes returns an inadmissible partition? I see the statement that "This is a limitation of GO’s knowledge, merge can only remove the cycles GO knows about", but it seems that the experiments use the stripped-down RegulonDB network as ground truth; some detail is either missing or was not emphasized enough for me to notice.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.