🧬 biology

scShapeBench: Discovering geometry from high dimensional scRNAseq data

This paper introduces scShapeBench, a benchmark dataset and evaluation framework for automated shape detection in high-dimensional single-cell RNA sequencing data, along with a baseline method called scReebTower that leverages diffusion geometry to outperform existing tools in identifying underlying topological structures like clusters, trajectories, and archetypes.

Original authors: Andrew J Steindl, João Felipe Rocha, Brian Tshilengi Di Bassinga, Zachary Warren, Matthew Scicluna, César Miguel Valdez Córdova, Shabarni Gupta, Leire Torices, Daniel Neumann, Timothy J. Mann, Ihuan G

Published 2026-05-14

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Andrew J Steindl, João Felipe Rocha, Brian Tshilengi Di Bassinga, Zachary Warren, Matthew Scicluna, César Miguel Valdez Córdova, Shabarni Gupta, Leire Torices, Daniel Neumann, Timothy J. Mann, Ihuan Gunawan, Dhananjay Bhaskar, John G Lock, Christine L Chaffer, Guy Wolf, Smita Krishnaswamy

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you have a giant, messy box of thousands of tiny, glowing marbles floating in 3D space. These marbles represent individual cells from a biological sample (like blood or tissue). In the real world, scientists use special cameras to take pictures of these marbles, but the pictures are so complex and high-dimensional that they look like a chaotic cloud of dust.

The big question scientists face is: What is the actual shape of this cloud?

Is it a bunch of separate, distinct islands (like clusters of different cell types)? Is it a long, winding river where marbles slowly flow from one end to the other (a continuous process like cell growth)? Is it a tree with branches splitting off? Or is it a hollow triangle where the marbles are spread out between three corners?

The Problem: Guessing the Shape

Currently, figuring out the shape is a manual job. A scientist has to look at a 2D picture of the cloud and use their intuition to guess, "Oh, that looks like a tree," or "That looks like a cluster." Based on that guess, they pick a specific software tool to analyze the data.

If they guess clusters, they use a tool like Seurat.
If they guess a tree, they use Monocle.
If they guess a river, they use TrajectoryNet.

The problem is that if they guess the wrong shape, they pick the wrong tool, and the important biological signals get lost. It's like trying to fix a bicycle with a hammer because you thought it was a toaster.

The Solution: scShapeBench

The authors of this paper created a new "training ground" called scShapeBench. Think of it as a video game level designed specifically to teach computers how to recognize shapes without human help.

This benchmark has two main parts:

The "Fake" Clouds (Synthetic Data): They built computer-generated clouds where they know the exact shape beforehand (like a perfect circle or a perfect tree). This is like a teacher giving a student a math problem where they already know the answer, so they can check if the student's method is correct.
The "Real" Clouds (Expert Data): They gathered 102 real biological datasets. Since we don't know the "true" shape of real cells, they asked 9 expert biologists to look at the data and vote on what shape it looks like. This creates a "gold standard" based on human expertise.

The New Tool: scReebTower

To test this benchmark, the authors built a new method called scReebTower.

Imagine the cloud of marbles is a mountain range.

scReebTower doesn't just look at the marbles; it simulates a "heat wave" spreading across the mountain.
It then looks at the "contour lines" (like on a hiking map) created by this heat.
As the heat rises, it sees how the islands of land merge or split.
By tracking these merges and splits, it builds a simplified skeleton (a graph) that represents the true shape of the mountain, whether it's a single peak, a ridge, or a complex network of valleys.

The paper claims this method is "topology-agnostic," meaning it doesn't assume the shape is a cluster or a tree beforehand. It just looks at the data and says, "Oh, I see a branch here," or "I see a loop there."

The Results

The authors tested scReebTower against other popular tools (like PAGA and Mapper) using their new benchmark:

On the "Fake" Clouds: scReebTower was the best at guessing the correct shape. It matched the "ground truth" better than the others.
On the "Real" Clouds: When the experts looked at the shapes recovered by scReebTower, they agreed with the experts' original labels more often than the other tools did.

The Bottom Line

This paper introduces a way to automatically figure out the "shape" of complex biological data and a new tool (scReebTower) that does this job better than current methods. The goal is to stop scientists from having to guess which tool to use and instead let the data tell them the answer, ensuring that the right analysis is applied to the right biological story.

Technical Summary: scShapeBench

Problem Definition

High-dimensional single-cell RNA sequencing (scRNAseq) data and other scientific point clouds possess intrinsic low-dimensional structures (shapes) that dictate the appropriate downstream analysis pipeline. Current analysis workflows often presuppose a specific data topology: standard pipelines like Seurat assume clustered data; tools like Monocle and SPADE assume tree-like branching trajectories; and flow models assume continuous transitions. However, selecting the correct pipeline typically relies on manual visual inspection by bioinformaticians, a process that is subjective, non-scalable, and prone to missing signals encoded in trajectories, cycles, or hybrid geometries if the default clustering assumption is applied.

The paper defines the single-cell shape detection problem as recovering a reduced graph $S$ from a high-dimensional point cloud $X$ such that $S$ reflects the underlying latent topology $G$ . The goal is to automate the identification of whether data organizes into discrete clusters, continuous trajectories, branching structures, or archetypal shapes (simplexes), without prior assumptions about the data's geometry.

Methodology

1. SCSHAPEBENCH Benchmark

To address the lack of standardized evaluation for shape detection, the authors introduce SCSHAPEBENCH, a comprehensive benchmarking framework comprising five components:

Synthetic Datasets: A collection of point clouds generated from ground-truth "skeleton graphs" with known topology. These graphs span four shape classes: clusters, single trajectories, multi-branching trees, and archetypes. The dataset includes controlled variations in noise scale, feature separation, sampling density, and feature thickness to test robustness across difficulty regimes.
Expert-Annotated scRNAseq Corpus: A curated collection of 102 real-world scRNAseq datasets from public repositories (CELLxGENE, 10x Genomics, Broad Institute, EMBL-EBI). Nine expert biologists and bioinformaticians annotated the shape of each dataset based on 2D PHATE and UMAP embeddings. The annotation scheme uses a union aggregation strategy, allowing datasets to be labeled with multiple shape classes (e.g., clusters and trajectories) to reflect biological complexity.
Topology-Aware Evaluation Metrics:
- For Synthetic Data: The authors use Wasserstein persistence similarity (comparing persistence diagrams of recovered vs. ground-truth graphs) and Graph Edit Distance (GED) to measure combinatorial agreement.
- For Real Data: Since ground-truth graphs are unavailable, evaluation is indirect. Recovered graphs are fed into two readout models: a Topological readout (PI-MLP) using persistence images as features, and a Connectivity readout (GNN) using graph neural networks. These models predict the expert-annotated shape labels.
Standardized Protocol: A unified interface accepting point clouds and outputting graphs, with uniform preprocessing (Scanpy pipeline, fixed cell count of 50,000) to ensure comparability.

2. The scReebTower Baseline Method

The authors propose scReebTower, a novel baseline method for automated shape detection that constructs Reeb graphs from diffusion geometry. The algorithm proceeds as follows:

Graph Construction: A $k$ -nearest-neighbor graph is built from the input point cloud.
Diffusion Operator: An adaptive Gaussian kernel is used to compute a diffusion operator $P$ .
Morse Function: The Fiedler vector (the first non-trivial eigenvector of $P$ ) is used as a discrete Morse function $f$ .
Reeb Graph Construction: The algorithm tracks the connected components of the level sets of $f$ across the graph. As the function value increases, components merge or split. These components are contracted into nodes, and edges track their evolution, forming a Reeb graph.
Reduction: The graph is reduced by contracting maximal paths of degree-two vertices, preserving cycles and branch points.
Multiscale Extension: The method can be applied iteratively using diffusion condensation, where data points are replaced by weighted averages of their diffusion neighbors, creating a hierarchy of coarse-grained Reeb graphs.

Key Results

Synthetic Performance: On the synthetic track, scReebTower achieved the highest Wasserstein persistence similarity (0.3199) and the lowest Graph Edit Distance (13.3216), outperforming baselines including PAGA and Mapper. The multiscale version (with diffusion condensation) slightly improved upon the base method. PAGA remained competitive in persistence similarity but suffered from higher edit distances, while Mapper performed substantially worse on both metrics.
Real-World Performance: On the scRNAseq track, scReebTower demonstrated the best accuracy and F1 scores across the different shape classes (Clusters, Single Trajectory, Multi-Branching, Archetypal) when evaluated via the GNN and MLP readouts. The results validated the benchmark's ability to resolve performance differences even under varying architectural constraints and annotation complexities.

Significance and Claims

The paper claims that SCSHAPEBENCH provides the first formal mathematical framework and standardized evaluation protocol for the problem of single-cell shape detection. Its significance lies in:

Automating Pipeline Selection: By providing a method to detect intrinsic topology, the work aims to remove the manual bottleneck of selecting analysis pipelines, which is critical for the deployment of "agentic AI scientists" in biology.
Diagnostic and Developmental Utility: The benchmark is not merely diagnostic; the authors demonstrate that it enables the development of better methods, as evidenced by scReebTower's state-of-the-art performance.
Topological Agnosticism: Unlike existing tools that assume specific shapes (e.g., trees or clusters), scReebTower is topology-agnostic, capable of recovering clusters, cycles, and branches without prior bias.
Scalability: The framework is designed to grow, with a dedicated web interface for expanding the expert-annotated corpus over time to ensure long-term coverage of diverse biological datasets.

The authors maintain modesty regarding limitations, noting that the current shape vocabulary is restricted to common scRNAseq structures and does not yet cover higher-dimensional features or nested structures. They also acknowledge that the correlation between benchmark performance and improved biological discovery remains an open empirical question.