acceptodds
Under review as a conference paper at ICLR 2027

Large-scale analysis of mouse and human dendritic spines reveals no evidence of local spatial clustering and weak morphological similarity between neighbouring synaptic contacts

Abstract

Dendritic spines receive most excitatory synaptic inputs in the cortex, and their morphology and arrangement influence how neurons integrate signals. Whether spines form spatial clusters and whether neighbouring synaptic contacts resemble one another remain important questions for understanding synaptic organization and constructing computational representations of neurons. We investigated these questions using large-scale reconstructions of mouse (Minnie65) and human (H01) , with morphological analyses covering more than 600,000 spines. Analyses of spine placement in the mouse dataset did not support local spatial clustering. Instead, spines were less closely spaced within dendritic branches than expected under independent placement; models incorporating short-range suppression improved, but did not fully reproduce, the observed spacing distributions. Across both datasets, neighbouring spines showed only weak excess morphological similarity after accounting for differences between branches. In the mouse dataset, repeated contacts from the same presynaptic axon showed correlated morphology, and their measurements improved spine-volume prediction beyond geometry and partner distances, although limited partner availability restricted the population-wide benefit. We compared our findings with previous studies and examined possible reasons for discrepant conclusions, including differences in tissue, morphological measures and statistical assumptions. The weak morphological resemblance between neighbouring spines challenges the use of spatial proximity as a basis for aggregating morphological information in spine-level graph models. Consistent with this finding, a graph-based LeJEPA model showed limited performance, illustrating the difficulty of exploiting spatial neighbourhoods to learn representations of spine morphology.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Dendritic spines, Spine morphology, Spatial organization, Morphological similarity, Graph representation learning, Self-supervised learning

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_2hsrOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper performs an analysis-oriented battery of experiments on the dendritic spines of human and mouse neurons. It holds a lot of statistical testing along with a single application concerning Graph Neural Networks and Cross-Attention. The paper presents a negative result on finding no significant evidence on spatial clustering, weak evidence on synaptic morphological similarities between two cases: neighbouring spines and same-spine synaptic target (shared-axon).

Critical strengths

The authors did a great job trying to statistically cover the main questions that affect how spines cluster and morphological similarity. Each part was well founded and the propagation of the idea through the parts is sound. This shows a great deal of understanding how the authors tried to locate a significant finding out of a dendritic study but rather showed that we do not have sufficient evidence.

The authors carried out a lot of experiments with statistical analysis, traditional machine learning and deep learning. Ablations and different settings were present covering additional aspects of the experiments.

Critical weaknesses

Due to the substantially lots of experiments, the paper organization is not easy at all to track. The main paper is very dependent on the supplementary appendix. I found it extremely difficult to track a single experiment, for example Part 1, where you will need to leave the section and go to the start of the appendix to know the model specifics, then go down in the appendix to know the results. This was a prominent issue in every section of the manuscript. I find it also extremely difficult to align the organization of the figures, tables with the appendix organization. Tracing a single piece of information requires great effort.

Some important details are underspecified, for instance:

  • In part 1, the held-out evaluation set lacks: how many folds are there, what is the inclusion criteria, the numbers of neurons, branches, spines in each fold.
  • In part 1, the analysis lacks a motivation behind the usage of Poison count.
  • In part 1, the objective behind spine spatial distribution matters is underdeveloped.
  • In part 2, the geometry-only six-knot spline model is explained very shallowly.
  • In part 2, a further analysis on the neighbourhood connectivity is crucial in order to determine if the graph architecture is actually a fit choice for the data The lack of these details highly affect the motivation behind main experiments done in the paper.

The paper is highly numeric which is very clear by the big number of tables present. On the other hand, figures do not convey any value except for two figures (Figure 1 and figure S1). Any other figure is actually another view of a table present.

Questions for authors

In part 1, the authors called the two models complementary, I have only seen the branch occupancy as the link between two models. So why is predicting occupancy scientifically important to the objective? Can’t a simple question like predicting the number of spines suffice? And if that is the case, what is the significance of predicting the number of spines to the spatial distribution?

In part 2, for the graph study, does the choice of the graph really captures the essence of the connectivity? Since proximity is the edge contributor in this relationship, we risk having separate disconnected components along the chain. Message passing would stop propagation if the disconnected components are among the chain. A further analysis on the neighbourhood connectivity is crucial.

Why didn’t the authors do multiple morphometric descriptors correction in the correlation shuffling experiment?

In the geometry-only six-knot spline, what anatomical and geometric features were included? A model equation is necessary. Can the authors present a justification on why only 3 out of the 17 morphology descriptors were reported in the study?

Reviewer_X3rQOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper uses connectomic reconstructions from mouse and human cortex data to analyze organization of synapses in dendritic spines. Specifically, three questions are considered:

  • distribution of spines along a dendritic branch and whether this distribution is consistent with independent placement or local clustering
  • whether spines share morphological similarities based on spatial proximity
  • and whether spines share morphological similarities based on input from a shared axon

Overall, the analysis seems weakly conclusive, that:

  • spines do not exhibit local clustering, but without a model that well-fits the observed distribution (with soft suppression being the best of the three models tested)
  • weak evidence that neighboring measurements can improve morphology prediction, at least for surface area and skeletal length
  • weak evidence for morphological correlation between contacts of the same axon, over geometry-and-distance baseline, but not different-axon neighbor baselines.
Critical strengths

authors are careful to give confidence intervals for their experimental studies and put their results into the appropriate context as far as statistical significance

Critical weaknesses

in my view, there are two primary weaknesses of the paper as far as being an ICLR submission:

  • significance - it seems the experimental findings do not offer an especially strong or surprising result - that there is weak morphological correlation for some measurements but not others, and that same axon connections may offer some information but not statistically distinct from different axon connections.

  • relevance / fit for ICLR - the paper does not seem the best fit for ICLR, given that the methods used for analysis do not require state-of-the-art methods or advancements in machine learning (ie suitability as an application paper), and the results, at least as-is, do not have much as far as implications or take-aways for ML/representation learning

Questions for authors

with respect to the above listed weaknesses:

  • is there an element to the experimental findings that would be of strong importance for a researcher working in the field to be aware of?

  • Is there a specific reason/angle in which this paper is a better fit for ICLR as opposed to a more traditional biology/computational neuroscience venue?

Reviewer_UYs6OpenReview ↗
Rating
1
Confidence
1
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents a large-scale analysis of dendritic spine organization using 3D reconstructions of mouse and human cortical neurons that includes more than 600,000 spines. The paper investigates if dendritic spines form spatial clusters, neighbouring spines have similar morphological characteristics, and spines receiving input from the same axon show greater morphological similarity.

The authors use statistical models, correlation analyses, and prediction experiments to investigate these questions. The results do not show any evidence of local spatial clustering in the mouse dataset. Additionally, there is only weak morphological similarity between neighbouring spines in both datasets. Spines receiving input from the same axon show modest positive morphological correlations, although the additional predictive information is limited.

The paper also performs experiments using a graph-based LeJEPA model to investigate whether information from neighbouring spines can help predict missing morphological measurements. However, the paper reports that the model collapsed.

Critical strengths
  • The paper uses more than 600,000 dendritic spines from mouse and human cortical datasets which is quite larger for studying spatial organization and morphological similarity.

  • The results provide evidence against local spatial clustering of dendritic spines and reveal only weak morphological similarity between neighbouring spines. These findings may contribute to our understanding of synaptic organization.

  • The paper uses multiple statistical approaches to investigate their hypotheses.

  • The paper acknowledges the limitations of their experiments.

Critical weaknesses

The paper primarily presents a large-scale statistical analysis of dendritic spine morphology using existing methods. Although the neuroscience findings are potentially interesting, I do not have sufficient expertise in this area of neuroscience to assess their novelty and significance relative to existing literature. My evaluation therefore focuses primarily on the machine learning aspects of the paper, where I find the technical contribution and experimental insights limited. From the representation learning perspective, the graph-based LeJEPA experiments do not provide additional insight, as the model collapsed.

Questions for authors

No specific questions for the authors.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.