TL;DR
This paper introduces stClinic, a novel dynamic graph model that integrates spatial multi-slice multi-omics (SMSMO) data with clinical information to identify and characterize clinically relevant cellular niches. By learning from evolving neighboring cell profiles and linking niche features to patient outcomes, stClinic offers a powerful new tool for understanding disease progression and developing targeted therapies.
What Problem Does This Paper Address?
Understanding the complex cellular neighborhoods, or "niches," within tissues is crucial for deciphering disease mechanisms, especially in cancer. Spatial multi-omics technologies allow us to study these niches at an unprecedented resolution, but integrating data from multiple slices (spatial multi-slice multi-omics, or SMSMO) and linking it to clinical outcomes remains a significant challenge. Existing methods struggle with the sheer scale and diversity of this data, inter-patient variability, and limited sample sizes, hindering the extraction of actionable clinical insights. This paper aims to bridge this gap by developing a computational framework that can effectively integrate and analyze SMSMO data to uncover clinically relevant niches.
Methods
The core of stClinic is a dynamic graph learning model. Here's a breakdown of its key components:
- Dynamic Graph Construction: stClinic builds a unified graph that connects spots (regions of tissue) not only within the same slice but also across different slices. This is achieved by aggregating information from "evolving neighboring nodes" – cells that have similar feature profiles, even if they are in different slices. This dynamic approach helps learn batch-corrected features that are biologically coherent.
- Mixture-of-Gaussians Prior: A Mixture-of-Gaussians (MOG) prior is used to guide the learning of latent features, helping to classify spots into distinct clusters or niches.
- Attention-Based Clinical Linking: To connect niches to clinical outcomes, stClinic characterizes each slice using a "niche vector." This vector is derived from six geometric statistical measures (mean, variance, max, min of UMAP embeddings, and proportions) of each cluster within the slice. An attention mechanism then learns the importance of these niche features for predicting clinical outcomes like survival time.
- Zero-Shot Learning for Label Transfer: A key innovation is stClinic's ability to transfer labels from a reference dataset to new, unseen samples without retraining. This is enabled by a pre-trained encoder that maps query samples into the same feature space as the reference set.
- Multi-Omics Integration: stClinic is designed to be flexible, allowing it to integrate latent features extracted by other established multi-omics tools (e.g., MultiVI, Seurat), enhancing its applicability across different data types and technologies.
Datasets
stClinic was evaluated on a variety of publicly available datasets, including:
- Human Dorsolateral Prefrontal Cortex (DLPFC) slices: Used for assessing spatial alignment and batch-effect correction capabilities.
- Luminal B breast cancer slices (BAS1, BAS2): Used to demonstrate the identification of intra-tumor niches.
- Triple-negative breast cancer (TNBC) patient slices: Used to evaluate niche malignancy through survival analysis.
- Primary colorectal cancer (CRC) and liver metastasis (LM) slices: Used to validate the prediction of metastasis-relevant niches.
- Mouse embryo and brain datasets: Used to showcase multi-omics integration capabilities.
Key Insights and Ideas
- Dynamic Graphs for Better Alignment: The use of dynamic graphs that consider evolving neighboring cell profiles across slices is crucial for effectively aligning and integrating data from heterogeneous sources, leading to more accurate niche identification.
- Niche-Clinical Linkage is Powerful: Directly linking the statistical properties of identified niches to clinical outcomes (like survival) provides a powerful way to assess niche malignancy and prognostic value.
- Zero-Shot Learning for Scalability: The zero-shot learning capability significantly enhances the model's utility by allowing it to leverage existing knowledge from reference datasets, reducing the need for extensive retraining on new data.
- Uncovering "Missed" Niches: stClinic can identify subtle or complex niches that might be overlooked by other methods, particularly those related to tumor boundaries or specific cellular adaptations.
Key Results
- Superior Spatial Alignment: stClinic demonstrated superior performance in aligning spatial domains across multiple slices and tissues compared to several state-of-the-art methods, as evidenced by higher Adjusted Rand Index (ARI) and Normalized Mutual Information (NMI) scores.
- Accurate Niche Identification in Cancer: The model successfully identified clinically relevant niches in breast, colorectal, and liver metastasis cancer datasets. For instance, it distinguished aggressive niches enriched with tumor-associated macrophages from favorable prognostic niches with B and plasma cells.
- Prognostic Biomarker Potential: In TNBC, stClinic identified niches associated with patient survival, with positive-weight clusters enriched in cancer and macrophage cells (poor prognosis) and negative-weight clusters enriched in plasma and B cells (improved prognosis).
- Metastasis Prediction: For CRC and liver metastasis, stClinic identified niches that accurately predict metastatic potential, highlighting the role of SPP1+ myeloid cells and cancer-associated fibroblasts in supporting CRC adaptation and invasion.
- Efficient Label Transfer: stClinic achieved significantly higher accuracy (~70%) in transferring labels from reference sets compared to Seurat and Geneformer (~16-18%) in both intra- and inter-tissue scenarios.
- Robust Multi-Omics Integration: The model effectively integrated spatial RNA and ATAC-seq data, even from different technologies and resolutions, outperforming other integration tools.
Limitations and Future Work
- Spatiotemporal Relationships: While stClinic excels at identifying static niches, the authors acknowledge the need for future work to infer continuous spatiotemporal relationships between clusters within 3D tissue structures.
- Sample Classification Accuracy: The prediction accuracy for sample classification in supervised tasks is noted as moderate, with suggestions to improve it by integrating additional modalities like histological images.
- Ensemble Learning for Stability: To further enhance robustness and reproducibility, the authors plan to explore ensemble learning and consensus clustering.
- Expanding to More Modalities: The authors suggest that stClinic's framework is adaptable to various spatial multi-omics technologies and could be extended to integrate other data types like spatial proteomics or chromatin accessibility profiling.
Summary updated