SPACE-GM: geometric deep learning of disease-associated microenvironments from multiplex spatial protein profiles

Zhenqin Wu, Alexandro E. Trevino, Eric Q. Wu, Kyle Swanson, Honesty J. Kim, H. Blaize D’Angio, Ryan Preska, Gregory W. Charville, Piero Dalerba, Ann Marie Egloff, Ravindra Uppaluri, Umamaheswar Duvvuri, Aaron T. Mayer, James ZoubioRxiv 2022 · DOI

TL;DR

This paper introduces SPACE-GM, a novel geometric deep learning framework that models tumor microenvironments (TMEs) as graphs to predict patient outcomes. SPACE-GM significantly outperforms existing methods by capturing spatial relationships between cells, offering deeper insights into disease mechanisms.

What Problem Does This Paper Address?

Understanding the tumor microenvironment (TME) is crucial for cancer research and treatment. TMEs are complex ecosystems of various cells and molecules, and their spatial organization plays a significant role in disease progression and patient response to therapy. While advanced imaging techniques like multiplexed immunofluorescence can capture high-resolution spatial data, effectively learning disease-relevant information from this data remains a challenge. Previous methods often rely on simple cell-type compositions of neighborhoods, potentially missing crucial spatial interactions. This paper aims to bridge this gap by developing a method that can learn from the intricate spatial arrangements of cells within TMEs to predict clinical outcomes.

Methods

The core of this work is SPACE-GM (SPAtial CEllular Graphical Modeling), a geometric deep learning framework built upon Graph Neural Networks (GNNs), specifically a Graph Isomorphism Network (GIN).

Here's a breakdown of the key methodological steps:

  1. Data Preprocessing: Multiplexed immunofluorescence images are processed to segment and classify individual cells. The spatial relationships between these cells are then represented as a graph, where each cell is a node and edges connect neighboring cells.
  2. Microenvironment Definition: A "microenvironment" is defined as a 3-hop neighborhood around a central cell within the graph, with an additional constraint of physical distance (within 75 µm). This captures a local cellular community.
  3. Graph Neural Network Architecture: A 3-layer GIN is used to learn embeddings for individual cells and their microenvironments. These embeddings capture both cell features (like type) and graph connectivity.
  4. Model Training: SPACE-GM is trained in two stages:
  5. Pre-training: The model is first trained on predicting cellular properties (e.g., expression profiles) to learn robust cell embeddings.
  6. Fine-tuning: The pre-trained GIN backbone is then connected to prediction heads and fine-tuned on patient-level phenotypic data (e.g., survival, recurrence) using the microenvironment embeddings.
  7. Prediction and Aggregation: For inference, SPACE-GM generates predictions for each microenvironment in a sample. These predictions are then mean-aggregated to derive a single patient-level prediction.
  8. Baseline Comparison: The authors compare SPACE-GM against composition-based methods. These baselines use cell-type composition vectors (counts of cell types) as input, either for the whole sample or for microenvironments, losing detailed spatial information.

Datasets

The study utilizes three large-scale multiplexed immunofluorescence datasets:

  • UPMC-HNC: Head and neck cancer samples (308 samples, 2,061,102 cells).
  • Stanford-CRC: Colorectal cancer samples (292 samples, 632,280 cells).
  • DFCI-HNC: Head and neck cancer samples (58 samples, 131,748 cells), collected before and after neoadjuvant therapy.

These datasets were annotated with clinical data, including patient survival, disease recurrence, and response to therapy. The study also mentions using a subset of cells for manual annotation and training cell type classifiers.

Key Insights and Ideas

  • Spatial Arrangement Matters: The core insight is that the spatial arrangement of cells within a tumor microenvironment, beyond just their types and proportions, holds critical information for predicting patient outcomes. SPACE-GM's graph-based approach effectively captures this.
  • Microenvironments as Predictive Units: Defining and analyzing "microenvironments" (local cellular communities) as the fundamental units for prediction proved highly effective, outperforming whole-sample analysis.
  • Tumor Subtype Mixing is Detrimental: The study found that the spatial mixing of different tumor subtypes within a microenvironment is associated with poorer patient outcomes. This suggests that spatial heterogeneity within the tumor itself can be a negative prognostic factor.
  • Granulocyte Dispersion Impacts Prognosis: The dispersion of granulocytes within the TME also significantly affects SPACE-GM's predictions, with dispersed granulocytes correlating with poorer prognoses. This highlights the complex role of immune cells in cancer progression.
  • Generalizability Across Datasets: SPACE-GM demonstrated robustness and generalizability, performing well even on datasets with different preprocessing pipelines and cell labeling, suggesting its broad applicability.

Key Results

  • Superior Predictive Performance: SPACE-GM consistently outperformed baseline composition-based methods across multiple prediction tasks (primary outcome, recurrence, survival length, HPV infection) on both head-and-neck cancer and colorectal cancer datasets.
  • For UPMC-HNC, SPACE-GM achieved ROC-AUC scores above 0.86 and C-index scores around 0.799 for primary outcome and survival length, respectively, significantly higher than baselines.
  • On the more challenging Stanford-CRC dataset, SPACE-GM achieved ROC-AUC of 0.739 for primary outcome, while composition-based methods performed near random chance.
  • Robustness to Data Variations: SPACE-GM showed consistent performance across different validation folds and demonstrated robustness in cross-study generalization tests.
  • Identification of Disease-Relevant Motifs: Clustering SPACE-GM embeddings identified distinct microenvironment clusters associated with different patient outcomes, providing interpretable biological insights. For example, lymphocyte-rich microenvironments were associated with better prognosis, while mixed tumor cell and myeloid cell clusters were linked to negative outcomes.

Limitations and Future Work

  • Cell Segmentation and Classification: The authors acknowledge that limitations in cell segmentation and classification approaches can affect the generalization of trained models to new datasets or unseen cell types. Future work could incorporate unified image-cell type dictionaries or methods for handling unseen cell types.
  • Aggregation Strategy: While mean aggregation of microenvironment predictions works well, the authors suggest that more sophisticated hierarchical aggregation techniques could reveal deeper insights into tissue architecture at different spatial scales.
  • Extending Modalities: SPACE-GM's framework is versatile and can be extended to other spatial omics data modalities, such as spatial transcriptomics, which is an interesting direction for future research.
  • Downstream Analysis: Coupling SPACE-GM embeddings with downstream analyses like permutation can further elucidate complex relationships between cellular communities and patient phenotypes.

Summary updated

Abstract

Abstract Multiplexed immunofluorescence imaging enables high-dimensional molecular profiling at subcellular resolution. However, learning disease-relevant cellular environments from these rich imaging data is an open challenge. We developed SPAtial CEllular Graphical Modeling (SPACE-GM), a geometric deep learning framework that flexibly models tumor microenvironments (TMEs) as cellular graphs. We applied SPACE-GM to 658 head-and-neck and colorectal human cancer samples assayed with 40-plex immunofluorescence imaging to identify spatial motifs associated with cancer recurrence and patient survival after immunotherapy. SPACE-GM is substantially more accurate in predicting patient outcomes than previous approaches for modeling spatial data using neighborhood cell-type compositions. Computational interpretation of the disease-relevant microenvironments identified by SPACE-GM generates insights into the effect of spatial dispersion of tumor cells and granulocytes on patient prognosis.
BibTeX
@misc{wu2022space,
  author = {Zhenqin Wu and Alexandro E. Trevino and Eric Q. Wu and Kyle Swanson and Honesty J. Kim and H. Blaize D’Angio and Ryan Preska and Gregory W. Charville and Piero Dalerba and Ann Marie Egloff and Ravindra Uppaluri and Umamaheswar Duvvuri and Aaron T. Mayer and James Zou},
  title = {{SPACE-GM: geometric deep learning of disease-associated microenvironments from multiplex spatial protein profiles}},
  year = {2022},
  doi = {10.1101/2022.05.12.491707},
  howpublished = {bioRxiv},
}
Figure