scDNM-VAE enables directly inspectable deep clustering of single-cell RNA-seq data through signed dendritic gating

arXiv:2608.17228 · q-bio.QM, q-bio.GN · Submitted 2026-08-18 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.

Ines: We're looking at the paper: "scDNM-VAE enables directly inspectable deep clustering of single-cell RNA-seq data through signed dendritic gating".

Marcus: Here's the gist of the paper. Deep clustering models for single-cell RNA sequencing often assign cells through latent or centroid-based mechanisms that are difficult to inspect scDNM-VAE enables directly inspectable deep clustering of single-cell RNA-seq data through signed dendritic gating.

Ines: Model Introduction and Architecture scDNM-VAE is a deep clustering framework that combines a variational autoencoder with a dendritic neuron-inspired head Cluster assignments are governed by learnable signed synaptic weights and thresholds The weight sign determines the direction of a gate’s response to a latent coordinate, its magnitude controls steepness, and the weight–threshold pair determines the transition location.

Marcus: This architecture is inspired by the dendritic neuron model (DNM) which has properties that support direct parameter-level inspection Stage 1 uses signed weight-threshold gating for cluster assignment, whereas Stage 2 uses a modified dendritic branch aggregation based on the geometric mean for gene-module decomposition.

Ines: Clustering Performance and Benchmarking The model was benchmarked on four datasets spanning immune, cortical, cardiac, and hematopoietic cells against scVI followed by KMeans and an MLP-DEC ablation scDNM-VAE performed better than scVI on PBMC3k, comparably on the Human Heart Cell Atlas and Paul15, and worse on Zeisel Ablating each cluster’s three highest-magnitude synaptic dimensions causes numerically greater reassignment than random-dimension ablation across all datasets, but the margins are modest and negligible on Zeisel.

Paper summary: Marcus: So looking at this paper "scDNM-VAE enables directly inspectable deep clustering of single-cell RNA-seq data through signed dendritic gating" and the authors Melih Ağraza, Deniz Karapınarc, Aysel Topsırd, Qianying Caoe, Erol Eğrioğluc, Gaurav Choudharya <ref:2608.17228#pg1>, what do we actually take away from this work?

Ines: The main thing is that they've shown a way to get directly inspectable clustering decisions by using signed dendritic gating instead of the black box methods we usually rely on <ref:2608.17228#pg3>

Yuki: They prove that you can examine the model's own signed weights and thresholds for every cluster and latent dimension rather than relying on a separate approximate explanation step <ref:2608.17228#pg3>

Marcus: It's a benchmark against scVI plus KMeans and an MLP-DEC ablation, but it shows that geometrically tight partitions can still be misaligned with expert curated labels <ref:2608.17228#pg3>

Ines: So the implication is that we can examine clustering decisions from the model’s internal parameters, which is a big step forward for high-stakes biological analysis <ref:2608.17228#pg3>

Conclusion: Ines: So, we've seen how scDNM-VAE uses these signed dendritic gates to make clustering directly inspectable by looking at weights and thresholds instead of just seeing a final cluster assignment.

Marcus: Yeah, and that's what the authors are really pointing toward with this paper titled "scDNM-VAE enables directly inspectable deep clustering of single-cell RNA-seq data through signed dendritic gating."

Yuki: It seems like the core idea is moving away from those black box methods where you just get a cluster label and then try to guess how that decision was made later.

Ines: Exactly. They show that by looking at the learned weights and thresholds, you can see exactly which latent dimensions are pushing a cell into one cluster versus another.

Marcus: It’s a benchmark against scVI plus KMeans and even an MLP-DEC ablation, and they’re claiming it performs better on some datasets than those methods.

Yuki: That's interesting because the paper mentions that while it does look promising, there are still some limitations, like how tight geometric partitions can still be misaligned with what experts label.

Ines: Right. And the faithfulness ablation they did showed that even when you focus on a small subset of parameters, the effect is bigger than random noise suggests, but only by a factor of one point one to one point three times larger.

Marcus: So the implication for us is that we might not need to rely solely on those complex black box methods if we can access these internal parameters directly.

Yuki: It suggests that understanding *why* a cluster forms is just as important as knowing *what* the final clusters are, especially when you're trying to characterize biological differences in populations.

Ines: And it sets up a new kind of benchmark for deep clustering methods—one where inspectability isn't just an afterthought, but built into the architecture itself.

Department of Medicine, Brown University · Division of Cardiology, Brown University Health · Department of Data Science and Analytics, Faculty of Arts and Sciences, Giresun University · Department of Industrial Engineering, Faculty of Mechanical Engineering, Yildiz Technical University · Department of Mechanical Engineering, Rochester Institute of Technology · Alpert Medical School of Brown University and Brown Health Cardiovascular Institute · VA Providence Healthcare System

q-bio.QM, q-bio.GN

Submitted: 2026-08-18

Updated: 2026-11-05

Comments: 39 pages, 7 figures, 4 tables. Supplementary information available from the authors

License: http://creativecommons.org/licenses/by-nc-nd/4.0/

Importance score: 83/100

The gist: Deep clustering models for single-cell RNA sequencing often assign cells through latent or centroid-based mechanisms that are difficult to inspect scDNM-VAE enables directly inspectable deep…

Key concepts

Signed Dendritic Gating
This is the core mechanism where cluster assignments are made using learnable signed synaptic weights and thresholds. The sign of the weight dictates whether a neuron's response increases or decreases based on a latent coordinate, allowing for direct inspection of the clustering decision process.
Dendritic Neuron Model (DNM)
The architecture is inspired by biological dendritic neurons. This inspiration allows the model to have properties that support parameter-level inspection, meaning researchers can examine the learned weights and thresholds to understand how the model makes its decisions.
Gene-Module Decoder
After clustering, this stage uses a decoder with twenty branches. One branch is specifically trained to reconstruct the differential expression signature genes for each cluster, helping to biologically characterize what each identified cluster represents.

Terminology

Summary

Deep clustering models for single-cell RNA sequencing often assign cells through latent or centroid-based mechanisms that are difficult to inspect scDNM-VAE enables directly inspectable deep clustering of single-cell RNA-seq data through signed dendritic gating. p. 0

Where the paper says it

scDNM-VAE enables directly inspectable deep clustering of single-cell RNA-seq data through signed dendritic gating Melih Ağraza,b,c,∗ , Deniz Karapınarc , Aysel Topsırd , Qianying Caoe , Erol Eğrioğluc , Gaurav Choudharya,b,f,g aDepartment of Medicine, Brown University, Providence, RI 02903, USA bDivision of Cardiology, Brown University Health, Providence, RI 02903, USA cDepartment of Data Science and Analytics, Faculty of Arts and Sciences, Giresun University, Giresun, Türkiye dDepartment of Industrial Engineering, Faculty of Mechanical Engineering, Yildiz Technical University, Istanbul, Türkiye eDepartment of Mechanical Engineering, Rochester Institute of Technology, Rochester, NY 14623, USA fAlpert Medical School of Brown University and Brown Health Cardiovascular Institute, Providence, RI 02903, USA gVA Providence Healthcare System, Providence, RI 02908, USA Abstract Deep clustering models for single-cell RNA sequencing often assign cells through latent or centroid-based mechanisms that are difficult to inspect. We introduce scDNM-VAE (single-cell Dendritic Neuron Model Variational Autoencoder), a deep clustering framework that combines a variational autoencoder with a dendritic neuron-inspired head.

Model Introduction and Architecture

scDNM-VAE is a deep clustering framework that combines a variational autoencoder with a dendritic neuron-inspired head Cluster assignments are governed by learnable signed synaptic weights and thresholds The weight sign determines the direction of a gate’s response to a latent coordinate, its magnitude controls steepness, and the weight–threshold pair determines the transition location. This architecture is inspired by the dendritic neuron model (DNM) which has properties that support direct parameter-level inspection Stage 1 uses signed weight-threshold gating for cluster assignment, whereas Stage 2 uses a modified dendritic branch aggregation based on the geometric mean for gene-module decomposition. p. 3

Where the paper says it

These quantities are part of the clustering decision function itself and can therefore be inspected without fitting a separate post-hoc model. In scDNM-VAE, we adapt the DNM principle differently across the two stages: Stage 1 uses signed weight-threshold gating for cluster assignment, whereas Stage 2 uses a modified dendritic branch aggregation based on the geometric mean for gene-module decomposition.

Clustering Performance and Benchmarking

The model was benchmarked on four datasets spanning immune, cortical, cardiac, and hematopoietic cells against scVI followed by KMeans and an MLP-DEC ablation scDNM-VAE performed better than scVI on PBMC3k, comparably on the Human Heart Cell Atlas and Paul15, and worse on Zeisel Ablating each cluster’s three highest-magnitude synaptic dimensions causes numerically greater reassignment than random-dimension ablation across all datasets, but the margins are modest and negligible on Zeisel.

Interpretability of Decision Logic

Unlike post-hoc explanation methods, scDNM-VAE exposes the parameters of the clustering rule itself For each cluster k and latent dimension d, the gate Akd(zd) = σ(ksharp(zdWkd − Qkd)) is determined by a learned weight Wkd and threshold Qkd Its derivative, ∂Akd/∂zd = ksharpWkdAkd(1 − Akd), shows that the sign of Wkd determines whether the gate increases or decreases with the latent coordinate. The complete matrices Wsyn and Qsyn therefore provide a compact, model-internal description of how every cluster reads every latent dimension, without fitting a separate post-hoc explanation model. p. 3

Where the paper says it

Crucially, each signed weight-threshold pair specifies how a cluster reads one latent coordinate: the weight sign determines whether the gate increases or decreases, the magnitude controls the steepness of this dependence, and the weight–threshold pair determines the transition location. These quantities are part of the clustering decision function itself and can therefore be inspected without fitting a separate post-hoc model.

Biological Characterization via Gene Modules

Stage 2 freezes the encoder and feeds the latent representations through a dendritic gene-module decoder containing twenty branches One branch was assigned to the differential-expression signature of each populated Stage-1 cluster and masked to reconstruct that cluster’s differential-expression signature genes. This supervised activation term encouraged each module to become active in its target cluster The resulting gene modules generally showed their strongest activation in their assigned target clusters. p. 3

Where the paper says it

The threshold-gated formulation carves the latent space into regions defined by distinct activation patterns, which is a stronger inductive bias than Euclidean distance to a centroid; note that our Stage-1 head aggregates the per-dimension gates additively, as distk = P d (1 − Akd) 2 , and reserves the DNM’s multiplicative branch aggregation for the Stage-2 gene-module decoder. Crucially, each signed weight-threshold pair specifies how a cluster reads one latent coordinate: the weight sign determines whether the gate increases or decreases, the magnitude controls the steepness of this dependence, and the weight–threshold pair determines the transition location.

Key Findings and Limitations

The MLP-DEC ablation achieves dramatically higher Silhouette scores than scDNM-VAE on every dataset (0.88–0.96 vs. 0.37–0.84) and substantially better Davies–Bouldin and Calinski–Harabasz indices This is not a marginal effect Architectural inspectability does not imply that a small subset of parameters dominates the decision. p. 8

Where the paper says it

The MLP-DEC ablation achieves dramatically higher Silhouette scores than scDNM-VAE on every dataset (0. 84) and substantially better Davies–Bouldin and Calinski–Harabasz indices (Supplementary Figure S5); on the Heart Atlas its Calinski–Harabasz index exceeds scDNM-VAE’s by an order of magnitude (1.

The faithfulness ablation shows that the own-dimension effect exceeds the random control by 1. p. 1
Where the paper says it

scDNM-VAE performs better than scVI on PBMC3k, comparably on the Human Heart Cell Atlas and Paul15, and worse on Zeisel, while producing biologically coherent marker-gene signatures. Ablating each cluster’s three highest-magnitude synaptic dimensions causes numerically greater reassignment than random-dimension ablation across all datasets, but the margins are modest and negligible on Zeisel.

01–1.39× on every dataset, but only by 1. p. 8
Where the paper says it

Geometric tightness does not imply biological correctness An important pattern appears when we compare the geometry-based metrics of Supplementary Figure S5 with the label-based metrics of Table 1 and Figure 3. The MLP-DEC ablation achieves dramatically higher Silhouette scores than scDNM-VAE on every dataset (0.

01–1.39×, while random ablation alone already reassigns a large fraction of cells. p. 4
Where the paper says it

We evaluate it against scVI+KMeans and an MLP-DEC ablation on four scRNAseq datasets covering immune cells, cortical neurons, cardiac tissue, and hematopoietic progenitors. Across five random seeds, scDNM-VAE outperformed scVI+KMeans decisively on PBMC3k in mean ARI, fell within seedto-seed variability of the alternatives on Paul15 and on the Human Heart Cell Atlas—where it is nonetheless roughly twice as stable across seeds as scVI—and was outperformed on Zeisel Cortex, where scVI performed clearly better.

The paper notes that the faithfulness ablation was performed on the seed-42 model and is descriptive rather than inferential. p. 4
Where the paper says it

We evaluate it against scVI+KMeans and an MLP-DEC ablation on four scRNAseq datasets covering immune cells, cortical neurons, cardiac tissue, and hematopoietic progenitors. Across five random seeds, scDNM-VAE outperformed scVI+KMeans decisively on PBMC3k in mean ARI, fell within seedto-seed variability of the alternatives on Paul15 and on the Human Heart Cell Atlas—where it is nonetheless roughly twice as stable across seeds as scVI—and was outperformed on Zeisel Cortex, where scVI performed clearly better.

Conclusion

scDNM-VAE shows that a directly inspectable clustering architecture can remain competitive with black-box alternatives on real biological benchmarks that geometrically tight partitions can nevertheless be misaligned with expert-curated labels and that clustering decisions can be examined from the model’s own signed weights and thresholds rather than through a separate approximate explanation step. The authors conclude by stating their contribution is therefore a directly inspectable dendritic clustering function, a benchmark against scVI+KMeans and an MLP-DEC architectural ablation, and a separate supervised gene-module decoder that characterizes the learned clusters. p. 4

Where the paper says it

In this manuscript we report a novel scDNM-VAE pipeline, a method designed from the start to make its clustering rule directly inspectable, building on the DNM [2] for clustering of cells based on scRNA-seq data. We evaluate it against scVI+KMeans and an MLP-DEC ablation on four scRNAseq datasets covering immune cells, cortical neurons, cardiac tissue, and hematopoietic progenitors.

Methods Summary

The training objective combines a reconstruction term, a DEC-style clustering term, and a KL-regularization term L = Lrecon + γ Lcluster + β LKL The KL coefficient was set to β = 10−3 for PBMC3k and Paul15 and β = 5 × 10−4 for Zeisel and Heart. The scVI baseline is trained with the same latent dimensionality as scDNM-VAE (nlatent = 10) for 150 epochs with batch size 128 and a 50-epoch KL warm-up schedule. p. 5

Where the paper says it

Benchmark performance We benchmarked scDNM-VAE, the MLP-DEC ablation, and scVI+KMeans on four datasets with expert-curated labels (Table 1, Figure 3). All three methods, including the scVI baseline, were run with five random seeds (42,

The evaluation metrics include adjusted Rand index (ARI), normalized mutual information (NMI), homogeneity, completeness, and V-measure for label-based metrics and silhouette score, Calinski–Harabasz index, and Davies–Bouldin index for geometry-based metrics. p. 6
Where the paper says it

Label-based metrics compare predicted clusters against expert-curated labels: ARI evaluates how closely two clustering assignments correspond after accounting for agreement expected by chance; Normalized Mutual Information (NMI) captures the extent of information shared between predicted clusters and reference labels; and clustering accuracy (ACC) is the fraction of correctly assigned cells after optimal cluster-tolabel matching. Geometry-based metrics assess cluster compactness and separation directly in the latent space without using labels: the Silhouette score (SIL) and the Calinski–Harabasz index (CH) increase with tighter, betterseparated clusters, whereas the Davies–Bouldin index (DB) decreases with better separation.

The final stage involves a complementary gene-module decoder trained with a combined objective of reconstruction loss and a supervised activation term. p. 3
Where the paper says it

Two properties make this architecture particularly appealing for clustering. The threshold-gated formulation carves the latent space into regions defined by distinct activation patterns, which is a stronger inductive bias than Euclidean distance to a centroid; note that our Stage-1 head aggregates the per-dimension gates additively, as distk = P d (1 − Akd) 2 , and reserves the DNM’s multiplicative branch aggregation for the Stage-2 gene-module decoder.

The source code implementing scDNM-VAE and the scripts used to reproduce the benchmarking and interpretability analyses will be made available in a public repository upon publication or earlier upon reasonable request for peer-review purposes. The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. p. 3
Where the paper says it

In scDNM-VAE, we adapt the DNM principle differently across the two stages: Stage 1 uses signed weight-threshold gating for cluster assignment, whereas Stage 2 uses a modified dendritic branch aggregation based on the geometric mean for gene-module decomposition. A growing body of work pursues inherent interpretability in single-cell models by constraining the latent space to align with known biology.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improved AI systems can gain direct parameter-level inspectability of their clustering decision function by using a dendritic neuron model head that governs cluster assignments through learnable signed synaptic weights and thresholds. This allows inspection of the weight sign determines the direction of a gate’s response to a latent coordinate, its magnitude controls steepness, and the weight–threshold pair determines the transition location.

  2. The system can be improved by incorporating a two-stage architecture where Stage 1 uses signed weight-threshold gating for cluster assignment and Stage 2 uses a modified dendritic branch aggregation based on the geometric mean for gene-module decomposition. This separates clustering from gene-module interpretation, allowing the model to provide both a direct clustering rule and a supervised gene-level characterization.

  3. The system can be enhanced by performing faithfulness ablation testing to verify that decision-relevant information is distributed across the latent space, as the highest-magnitude synaptic dimensions carry decision-relevant but distributed signal. This prevents overconfidence in interpretations based solely on a few top-ranked weights, showing that random ablation alone already reassigns a large fraction of cells.

  4. The system can be made more biologically informative by implementing a supervised gene-module decoder that is trained to align each module with the differential-expression signature genes of one populated Stage-1 cluster. This ensures that the resulting gene programs are characterized by genes most discriminative for their predicted cluster, rather than just being arbitrary background genes.

Abstract

Deep clustering models for single-cell RNA sequencing often assign cells through latent or centroid-based mechanisms that are difficult to inspect. We introduce scDNM-VAE (single-cell Dendritic Neuron Model Variational Autoencoder), a deep clustering framework that combines a variational autoencoder with a dendritic neuron-inspired head. Cluster assignments are governed by learnable signed synaptic weights and thresholds: the weight sign determines the direction of a gate's response to a latent coordinate, its magnitude controls steepness, and the weight-threshold pair determines the transition location. The trained clustering function can therefore be inspected directly without fitting a post-hoc explanation model. We benchmark scDNM-VAE on four datasets spanning immune, cortical, cardiac, and hematopoietic cells against scVI followed by KMeans and an MLP-DEC ablation. scDNM-VAE performs better than scVI on PBMC3k, comparably on the Human Heart Cell Atlas and Paul15, and worse on Zeisel, while producing biologically coherent marker-gene signatures. Ablating each cluster's three highest-magnitude synaptic dimensions causes numerically greater reassignment than random-dimension ablation across all datasets, but the margins are modest and negligible on Zeisel. These results show that signed dendritic gating supports competitive clustering with a parameter-inspectable decision function, while indicating that decision-relevant information is distributed across the latent space.

Related papers