spARC recovers human glioma spatial signaling networks with graph filtering

Manik Kuchroo, Danielle Miyagishima, Holly R. Steach, Abhinav Godavarthi, Yutaka Takeo, Phan Q. Duy, Tanyeri Barak, E. Zeynep Erson‐Omay, Scott E. Youlten, Ketu Mishra-Gorur, Jennifer Moliterno, Declan McGuone, Murat Günel, Smita KrishnaswamybioRxiv 2022 · DOI

TL;DR

This paper introduces spARC, a novel method that leverages graph filtering and data diffusion to denoise spatial transcriptomics data and uncover cell-cell signaling networks. By integrating spatial location and gene expression, spARC effectively recovers gene dynamics and simulates paracrine interactions, offering a more comprehensive view of biological systems, particularly in complex tissues like gliomas.

What Problem Does This Paper Address?

Spatial transcriptomics technologies have revolutionized our ability to study gene expression within its tissue context. However, these technologies are plagued by technical artifacts like "dropout" (where a gene is detected but its expression is zero due to technical limitations) and noise, which obscure true biological signals and hinder discovery. Existing methods for denoising or imputing gene expression often overlook the crucial information provided by a cell's spatial location. Cells in close proximity are influenced by shared microenvironments and direct cell-cell communication, factors that are not captured by analyzing gene expression in isolation. This paper addresses the challenge of effectively integrating both spatial and gene expression information to accurately recover gene dynamics and model complex signaling interactions within tissues.

Methods

The core of spARC is a diffusion-based filtration method that operates on graphs. It builds upon concepts from manifold learning, graph signal processing, and data diffusion.

  1. Graph Construction: spARC constructs two types of graphs:
  2. Gene Expression Graph (Gexp): This graph captures relationships between cells based on their gene expression profiles, using an alpha-decay affinity kernel.
  3. Spatial Graph (Gspatial): This graph is more novel. It uses a "shared nearest neighbor" approach, where an edge exists between two cells only if they are both spatially proximate and transcriptomically similar. This aims to overcome the "noise" of spatial proximity alone, where neighboring cells might be of different types.
  4. Diffusion Operators: From these graphs, spARC computes diffusion operators (Pexp and Pspatial) that represent transition probabilities for random walks on the graph. These operators are powered to a certain time step (texp, tspatial) to capture longer-range relationships.
  5. Integrated Diffusion Operator: The key innovation is the integration of these two diffusion operators into a single Pintegrated operator. This operator captures the spatio-transcriptomic relationships between cells.
  6. Graph Filtering for Denoising: The integrated diffusion operator is then used as a diffusion filter to denoise the gene expression data. This filter effectively shares expression information between cells that are both spatially close and transcriptomically similar, recovering true gene expression dynamics.
  7. Paracrine Signaling Simulation: spARC can also simulate ligand diffusion through tissue by applying a spatial diffusion operator to ligand expression. By combining diffused ligand concentrations with cognate receptor expression, it can compute interaction scores and model paracrine signaling.

Datasets

  • Mouse Brain Data: Used for initial validation of spARC's ability to recover layer-specific gene expression patterns. This data was obtained from the 10X Visium platform and compared against Allen Mouse Brain Atlas in situ hybridization (ISH) data.
  • Human Glioma Data: 10 IDH-mutant surgically resected human gliomas (WHO grades II-IV) were analyzed to study signaling networks across disease progression.
  • MERFISH Data: Used for quantitative benchmarking of spARC against other denoising methods. This dataset was chosen for its high fidelity and resolution. Simulated noise and dropout were added to this data.
  • Visium Human Lymph Node Data: Used to test spARC's ability to recover known cell-cell interactions and to compare its signaling analysis performance against other algorithms.
  • DBIT-seq Data: Used to demonstrate the framework's extensibility to multiple data modalities (protein and gene expression in situ).
  • Various Spatial Technologies: spARC was tested on slide-seq, seqFISH, MERFISH, MIBI-TOF, and 4i data to show its generalizability.

Key Insights and Ideas

  • Spatial Proximity Alone Isn't Enough: The paper highlights that simply considering spatial proximity in graph construction can introduce "noise" because neighboring cells are often of different types. spARC's "shared nearest neighbor" approach, requiring both spatial and transcriptomic similarity, is crucial for overcoming this.
  • Integrating Modalities is Key: The success of spARC hinges on its ability to create a unified diffusion operator that effectively combines information from spatial location and gene expression. This integrated approach leads to significantly better denoising and signaling inference.
  • Denoising Enables Deeper Biological Discovery: By cleaning up noisy spatial transcriptomics data, spARC unlocks the ability to identify subtle spatial niches, characterize tumor microenvironments, and uncover complex cell-cell signaling interactions that were previously obscured.
  • Simulating Diffusion for Signaling: The novel application of diffusion operators to model ligand movement in tissue provides a more biologically realistic way to infer paracrine signaling compared to methods that rely solely on expression levels or predefined cell clusters.

Key Results

  • Superior Denoising: spARC significantly outperformed other graph-based denoising techniques (kNN smoothing, MAGIC) in recovering ground truth expression values from simulated noisy MERFISH data, achieving higher Pearson correlations.
  • Accurate Gene Co-expression Recovery: In mouse brain data, spARC recovered known layer-specific gene expression patterns more accurately than other methods, as measured by mutual information.
  • Identification of Glioma Niches and Interactions: Applied to human gliomas, spARC identified distinct spatial niches (e.g., Dense Tumor, Leading Edge, Vasculature) and revealed key signaling interactions, such as the Osteopontin-CD44 axis, which was enriched in higher-grade gliomas and showed clinical relevance.
  • Effective Signaling Simulation: spARC demonstrated higher association between computed interaction scores and known signaling response programs in lymph node data compared to CellPhoneDB and SpaOTsc.
  • Computational Efficiency: The method can analyze a standard Visium 10X sample in under a minute.
  • Platform Generalizability: spARC was shown to be effective across a range of spatial technologies.

Limitations and Future Work

  • Computational Cost: While efficient for analysis, the initial graph construction and diffusion operator calculation can be computationally intensive, especially for very large datasets.
  • Parameter Sensitivity: The performance of spARC can be influenced by parameter choices (e.g., bandwidth of the kernel, diffusion time steps), although default settings are provided and parameters can be user-tuned.
  • Future Work: The authors suggest that the framework can be extended to other data modalities and applied to a wider range of biological questions, including understanding spatial patterning and tissue morphogenesis. The ability to integrate multiple data types (e.g., spatial transcriptomics with proteomics or epigenomics) is a promising avenue. The paper also hints at using the diffusion operator for visualization and clustering, which could be further explored.

Summary updated

Abstract

Abstract Biological networks operate within architectural frameworks that influence the state and function of cells through niche-specific factors such as exposure to nutrients and metabolites, soluble signaling molecules, and direct cognate cell-cell communication. Spatial omics technologies incorporate environmental information into the study of biological systems, where the spatial coordinates of cells may directly or indirectly encode these micro-anatomical features. However, they suffer from technical artifacts, such as dropout, that impede biological discovery. Current methods that attempt to correct for this fail to adequately integrate highly informative spatial information when recovering gene expression and modelling cell-cell dynamics in situ . To address this oversight, we developed spatial Affinity-graph Recovery of Counts (spARC), a data diffusion-based filtration method that shares information between neighboring cells in tissue and related cells in expression space, to recover gene dynamics and simulate signalling interactions in spatial transcriptomics data. Following validation, we applied spARC to 10 IDH-mutant surgically resected human gliomas across WHO grades II-IV in order to study signaling networks across disease progression. This analysis revealed co-expressed genes that border the interface between tumor and tumor-infiltrated brain, allowing us to characterize global and local structure of glioma. By simulating paracrine signaling in silico , we identified an Osteopontin-CD44 interaction enriched in grade IV relative to grade II and grade III astrocytomas, and validated the clinical relevance of this signaling axis using TCGA.
BibTeX
@misc{kuchroo2022sparc,
  author = {Manik Kuchroo and Danielle Miyagishima and Holly R. Steach and Abhinav Godavarthi and Yutaka Takeo and Phan Q. Duy and Tanyeri Barak and E. Zeynep Erson‐Omay and Scott E. Youlten and Ketu Mishra-Gorur and Jennifer Moliterno and Declan McGuone and Murat Günel and Smita Krishnaswamy},
  title = {{spARC recovers human glioma spatial signaling networks with graph filtering}},
  year = {2022},
  doi = {10.1101/2022.08.24.505139},
  howpublished = {bioRxiv},
}
Figure