TL;DR
This paper introduces MUSE, a novel deep learning framework that integrates single-cell morphology and gene expression data to provide a more comprehensive understanding of tissue composition. MUSE effectively identifies cell subpopulations and spatial organization that are missed by analyzing either data type alone, offering a powerful new tool for biological discovery.
What Problem Does This Paper Address?
Understanding the intricate organization and diversity of cells within tissues is crucial for deciphering their function and how diseases arise. Traditionally, researchers have relied on microscopy to study cell shapes (morphology) and gene sequencing to understand cell activity (transcriptional state). While each method provides valuable insights, they capture only a partial picture. Recent technological advancements allow for simultaneous measurement of both morphology and transcription from the same cells, but effectively integrating these diverse data types into a cohesive understanding remains a significant challenge. This paper tackles this challenge by developing a method to combine these modalities, aiming to overcome the limitations of each individual approach and reveal a more complete view of cellular heterogeneity.
Methods
The core of this work is the development of MUSE (Multi-modal Structured Embedding), a deep learning framework built upon a multi-view autoencoder architecture. MUSE operates in three key stages:
- Modality-Specific Transformations: The raw morphological and transcriptional data for each cell are independently processed by separate neural networks (encoders) to extract initial latent representations.
- Pseudo-Label Learning: Clustering is performed on these individual latent representations to generate "pseudo-labels" for each cell within its respective modality. This helps capture initial population structures.
- Joint Feature Learning: The modality-specific latent representations are then merged and further processed by a joint encoder to create a unified latent representation for each cell.
MUSE is trained using a combination of two loss functions:
- Self-Reconstruction Loss: This loss ensures that the joint latent representation retains as much information as possible from the original input data of each modality. For transcriptional data, a "zero-inflated" reconstruction loss is used to specifically address the issue of dropout (missing gene expression data).
- Structured Self-Supervised Loss (Triplet Loss): This loss encourages cells that are clustered together within a single modality (based on the pseudo-labels) to remain close in the joint latent space, while pushing apart cells from different clusters. This helps preserve the biological relationships learned from each modality in the combined representation.
After training, graph clustering (using PhenoGraph) is applied to the final joint latent features to identify distinct cell subpopulations.
Datasets
The paper utilizes two primary datasets for evaluating MUSE:
- seqFISH+ mouse cortex dataset: This dataset comprises transcriptional data (10,000 genes) and DAPI/Nissl stained images for 523 cells from a mouse cortex.
- STARmap mouse cortex dataset: This dataset includes transcriptional data (1,020 genes) and morphological information (segmentation masks) for 973 cells from a mouse visual cortex.
Additionally, the authors generated simulated data with known ground truth subpopulation assignments to rigorously benchmark MUSE against other methods under controlled conditions, varying factors like dropout rates and the number of subpopulations.
Key Insights and Ideas
- Synergy of Modalities: The most significant insight is that combining morphology and transcriptional data provides a richer understanding of cell heterogeneity than either modality alone. MUSE can identify subpopulations that are not discernible from a single data type.
- Robustness to Data Quality: MUSE demonstrates a remarkable ability to compensate for modality-specific noise and data quality issues, such as sequencing dropouts in transcript data. It can maintain performance even when one modality is of lower quality.
- Preservation of Biological Structure: The structured self-supervised loss (triplet loss) is key to MUSE's success. It ensures that the biologically meaningful relationships and structures learned from individual modalities are faithfully integrated into the joint representation.
- Refining and Discovering Subpopulations: MUSE can both "refine" existing subpopulations (by providing finer distinctions within them) and "discover" entirely new ones that were previously hidden.
Key Results
- Improved Subpopulation Identification: On simulated data, MUSE consistently outperformed single-modality approaches and other multi-modal integration methods in accurately identifying ground-truth cell subpopulations, achieving higher Adjusted Rand Index (ARI) scores.
- Robustness to Dropout: Even with significant dropout rates in the transcriptional data, MUSE was able to maintain accuracy comparable to morphology-alone analysis, outperforming other methods that degraded more severely.
- Biological Validation in Cortex Data: When applied to mouse cortex datasets (seqFISH+ and STARmap), MUSE successfully identified distinct cortical layers and revealed finer distinctions within these layers, leveraging morphological diversity to refine transcript-based classifications. For instance, in the seqFISH+ data, MUSE refined distinctions within the L2/3 layer by identifying distinct morphological features.
- Efficiency: MUSE is reasonably fast, processing 1000 samples in 1.5 minutes on a standard desktop, although it is noted to be slower than some other methods due to its graph clustering component.
Limitations and Future Work
- Computational Cost: While reasonably fast, MUSE is slower than some other methods due to the inclusion of graph clustering during its structured self-supervised training. Future work could explore more efficient clustering strategies.
- Extensibility: The authors highlight that MUSE's architecture is designed to be extensible to other measurement modalities, suggesting future work could involve integrating additional data types (e.g., proteomics, epigenomics) to further enhance the understanding of cellular states and tissue organization.
- Generalizability: While demonstrated on cortex data, further validation on diverse tissue types and biological contexts would be beneficial to fully establish MUSE's generalizability.
Summary updated