TL;DR
This paper introduces DEGAS, a novel deep transfer learning framework that bridges the gap between patient-level disease information and individual cell transcriptomics. By learning "impressions" of disease attributes, DEGAS can prioritize cells for disease association, offering a flexible and powerful tool for understanding disease mechanisms at a cellular level.
What Problem Does This Paper Address?
Analyzing single-cell RNA sequencing (scRNA-seq) data has revolutionized our understanding of cellular heterogeneity in tissues and diseases. However, a significant challenge remains in directly associating individual cells with specific disease attributes like diagnosis, prognosis, or treatment response. Existing methods often rely on cell type identification or deconvolution, which can be limited by clustering resolution and may fail to capture nuanced disease associations within cell types. This paper addresses the need for a method that can directly infer disease attributes from patient data and transfer this information to individual cells, enabling a more granular understanding of disease at the cellular level.
Methods
The core of DEGAS is a deep transfer learning framework that integrates scRNA-seq data (representing cells) with patient-level bulk transcriptomic data (representing patients). The key innovation lies in using domain adaptation to align the feature spaces of these two disparate data types. This allows the model to learn a shared latent representation where both patients and cells occupy a comparable space.
The framework employs a multitask learning approach to simultaneously:
- Label cells with their cellular subtypes.
- Label patients with their clinical attributes (e.g., disease status, prognosis).
- Generate a latent space where patients and cells are comparable.
This is achieved through a deep neural network architecture, typically a DenseNet, with a shared hidden layer for feature extraction and separate output layers for patient and cell predictions. The training process minimizes a combined loss function that includes terms for patient classification, cell type classification, and Maximum Mean Discrepancy (MMD) to minimize the distribution differences between patient and cell data representations. The output of the model, termed "impressions," represents the transferrable disease associations.
Datasets
DEGAS was trained and validated on a diverse set of datasets, including:
- Simulated data: Generated to mimic real-world scenarios with known cell types and disease associations, serving as a controlled benchmark.
- Glioblastoma (GBM) datasets: Including scRNA-seq data from five GBM tumors and TCGA microarray data for 111 GBM patients, with known GBM subtype labels.
- Alzheimer's Disease (AD) datasets: Combining AIBS human scRNA-seq data and MSBB patient RNA-seq data, with known AD diagnosis labels.
- Multiple Myeloma (MM) datasets: Including IUSM scRNA-seq data from four MM patients and MMRF bulk tissue RNA-seq data for 647 patients, with progression-free survival (PFS) and overall survival (OS) information.
These datasets represent a significant effort to validate the framework across different diseases and data modalities.
Key Insights and Ideas
- Direct cell-level disease association: DEGAS moves beyond cell-type level analysis to directly assign disease attributes to individual cells, overcoming limitations of traditional clustering-based approaches.
- "Impressions" as a novel concept: The framework introduces "impressions" – transferrable disease associations – as a way to quantify the relationship between cells and patient outcomes.
- Domain adaptation is crucial: The success of DEGAS hinges on domain adaptation techniques to bridge the inherent biases and differences between bulk and single-cell transcriptomic data.
- Flexibility in data types and labels: DEGAS can handle various data types and label combinations (classification, survival, or even no labels), making it broadly applicable.
- Identification of novel subtypes: In MM, DEGAS identified a CD138+/PHF19high subtype associated with aggressive disease, suggesting potential therapeutic targets.
Key Results
- Accurate prediction: DEGAS demonstrated high performance in predicting patient characteristics and cell types across simulated data and real-world disease datasets (GBM, AD, MM). For instance, it achieved high PR-AUCs (0.79-0.97) for GBM subtypes and AD diagnosis.
- Robustness: The framework showed robustness to hyper-parameter choices, with most parameters not significantly affecting performance, suggesting a stable and reliable model.
- Identification of disease-associated cells: DEGAS successfully identified disease-associated cells and subtypes, such as the PHF19high myeloma cells linked to progression and specific astrocyte and microglia subtypes in AD.
- Comparison with existing tools: DEGAS outperformed Augur in identifying disease associations in certain simulation scenarios and demonstrated its ability to capture intra-cell type differences that other methods missed.
Limitations and Future Work
- Data requirements: DEGAS requires both patient-level transcriptomic data and single-cell transcriptomic data. For scenarios where only case-control scRNA-seq is available, tools like Augur might be more suitable.
- Imbalanced data: While DEGAS showed promise, the authors noted that performance could be further improved for highly imbalanced datasets.
- Future work: The authors suggest that DEGAS can be expanded to accommodate other data modalities beyond transcriptomics, provided proper normalization. They also intend to provide more options for output label combinations in future versions. The identified CD138+/PHF19high subtype in MM presents a promising avenue for targeted immunotherapies.
Summary updated