HiPoNet: A Topology-Preserving Multi-View Neural Network For High Dimensional Point Cloud and Single-Cell Data

Siddharth Viswanath, Hiren Madhu, Dhananjay Bhaskar, Jake Kovalic, David R. Johnson, Rex Ying, Christopher J. Tape, Ian Adelstein, Michael Perlmutter, Smita Krishnaswamy

Abstract

HiPoNet is a novel neural network architecture designed to process high-dimensional point clouds, with a primary focus on single-cell biological data. Unlike traditional models that rely on 3D spatial heuristics or single-graph representations, HiPoNet learns multiple "views" of the data and models them as higher-order simplicial complexes. By employing multiscale simplicial wavelets, the network captures complex topological patterns and multi-way interactions that standard graph neural networks (GNNs) often miss. We demonstrate HiPoNet's superior performance across several cancer cohorts and spatial transcriptomics datasets, highlighting its stability and capacity for intrinsic geometry learning.

Introduction

The advent of single-cell RNA sequencing and spatial transcriptomics has produced massive datasets where each data point represents a cell characterized by high-dimensional gene or protein expression profiles. Analyzing these cohorts to predict clinical outcomes is challenging because standard point cloud architectures (e.g., PointNet) are typically hardcoded for 3D spatial data. Furthermore, conventional dimensionality reduction and GNN approaches often collapse features into a single graph, conflating distinct biological processes and ignoring higher-order geometric relationships. HiPoNet addresses these issues by disentangling biological signals through a multi-view framework and preserving data topology using simplicial complexes.

Methods

HiPoNet utilizes a three-stage pipeline to extract robust features from high-dimensional biological manifolds:

  1. Multi-View Creation: The model learns multiple feature-weighting vectors to generate distinct "views" of the point cloud, dynamically highlighting relevant biomarkers while suppressing noise.
  2. Simplicial Complex Modeling: For each view, a Vietoris-Rips simplicial complex is constructed. This allows the model to capture higher-order interactions (e.g., triangles and tetrahedrons) beyond simple pairwise edges.
  3. Multi-Scale Simplicial Wavelets: To prevent feature oversmoothing, HiPoNet replaces standard message passing with simplicial wavelet transforms based on heat diffusion. These transforms act as filters that preserve both local and global topological signals.
HiPoNet architecture (A) and multi-view feature importance (B).

Results

The authors evaluated HiPoNet on several single-cell and spatial transcriptomics cohorts, including Melanoma (MIBI), Patient-Derived Organoids (PDOs), and CODEX spatial data. HiPoNet consistently outperformed baseline models in classification tasks, achieving 90.90% accuracy on the Melanoma dataset and a 0.916 AUC on the DFCI spatial transcriptomics outcome task.

A key strength of HiPoNet is its ability to learn intrinsic geometry without explicit feature engineering. When tested on predicting persistence homology, HiPoNet achieved the lowest Mean Squared Error (MSE) across all benchmarks, significantly outperforming both GNNs and 3D point cloud models.

Performance comparison (MSE) for geometry prediction tasks across different models and datasets.

Discussion and Conclusion

HiPoNet offers a powerful and interpretable framework for high-dimensional data analysis. By examining learned feature weights, researchers can identify specific biological markers driving the model's predictions. While the computational cost of higher-order simplices remains a consideration for very large datasets, HiPoNet’s stability and topological fidelity represent a significant step forward in biological machine learning. Future research will focus on scaling these methods and exploring their utility in non-biological high-dimensional domains.

Summary updated

Abstract

In this paper, we propose HiPoNet, an end-to-end differentiable neural network for regression, classification, and representation learning on high-dimensional point clouds. Single-cell data can have high dimensionality exceeding the capabilities of existing methods point cloud tailored for 3D data. Moreover, modern single-cell and spatial experiments now yield entire cohorts of datasets (i.e. one on every patient), necessitating models that can process large, high-dimensional point clouds at scale. Most current approaches build a single nearest-neighbor graph, discarding important geometric information. In contrast, HiPoNet forms higher-order simplicial complexes through learnable feature reweighting, generating multiple data views that disentangle distinct biological processes. It then employs simplicial wavelet transforms to extract multi-scale features - capturing both local and global topology. We empirically show that these components preserve topological information in the learned representations, and that HiPoNet significantly outperforms state-of-the-art point-cloud and graph-based models on single cell. We also show an application of HiPoNet on spatial transcriptomics datasets using spatial co-ordinates as one of the views. Overall, HiPoNet offers a robust and scalable solution for high-dimensional data analysis.
BibTeX
@inproceedings{viswanath2025hiponet,
  author = {Siddharth Viswanath and Hiren Madhu and Dhananjay Bhaskar and Jake Kovalic and David R. Johnson and Rex Ying and Christopher J. Tape and Ian Adelstein and Michael Perlmutter and Smita Krishnaswamy},
  title = {{HiPoNet: A Topology-Preserving Multi-View Neural Network For High Dimensional Point Cloud and Single-Cell Data}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2025},
  url = {https://openreview.net/forum?id=UoKt9B1aY8},
}
Figure