TL;DR
This paper introduces SubCell, a novel suite of self-supervised deep learning models that can accurately interpret single-cell fluorescence microscopy images to capture morphology, protein localization, and cellular organization. Trained on a massive proteome-wide dataset, SubCell generalizes exceptionally well to diverse biological contexts and outperforms existing methods, paving the way for a more comprehensive understanding of cellular function.
What Problem Does This Paper Address?
Understanding cellular function and behavior is crucial for biology and medicine. This understanding relies heavily on analyzing cell morphology and the precise location of proteins within cells. While fluorescence microscopy generates vast amounts of data, manually interpreting these images is time-consuming and limited by human perception. Existing machine learning approaches often require task-specific training and struggle to generalize across different datasets and imaging modalities. This creates a need for a robust, general-purpose model that can extract meaningful biological insights from microscopy images without extensive fine-tuning.
Methods
The core of SubCell is a multi-task, self-supervised learning framework built upon Vision Transformers (ViTs). Unlike standard self-supervised methods that focus solely on image reconstruction, SubCell incorporates novel "cell-specific" and "protein-specific" objectives. These objectives guide the model to learn features related to cell identity and protein localization, respectively. The framework leverages a proteome-aware learning objective trained on the Human Protein Atlas (HPA) dataset, which covers a vast number of proteins and cell lines. Additionally, the authors employ attention-based pooling to refine feature extraction and improve interpretability.
Datasets
- Training Dataset: Human Protein Atlas (HPA) v23 dataset, containing images of over 1.13 million single cells across 37 human cell lines, encoding the expression and subcellular distribution of 13,141 proteins.
- Evaluation Datasets:
- OpenCell dataset (HEK293T cells with CRISPR-based protein tagging)
- AllenCell dataset (human iPSC cells)
- Yeast dataset (Saccharomyces cerevisiae cells)
- U2OS FUCCI dataset (cell cycle monitoring)
- Bridge2AI dataset (drug-treated triple-negative breast cancer cells)
- JUMP1 dataset (Cell Painting images for perturbation screens)
The authors also mention using the Kaggle challenge dataset for protein localization.
Key Insights and Ideas
- Proteome-aware learning is key: Training on the entire proteome's spatial patterns, not just individual proteins, significantly improves the model's ability to capture biologically relevant features.
- Beyond reconstruction: Simply reconstructing images isn't enough for complex biological tasks like protein localization. Explicitly guiding the model with cell and protein identity information is crucial.
- Generalization is paramount: SubCell's ability to perform well across diverse datasets (different cell types, species, imaging resolutions, and even channel combinations) without fine-tuning is a major breakthrough.
- Vision-based maps are powerful: SubCell enables the creation of the first vision-only, proteome-wide hierarchical map of protein organization, revealing new insights into cellular subsystems and protein interactions.
- Multimodal integration unlocks synergy: Combining SubCell's image embeddings with protein sequence embeddings (from models like ESM2) creates a more robust gene representation than either modality alone, particularly for understanding protein interactions and functional divergence.
Key Results
- SubCell outperforms state-of-the-art methods across a variety of tasks, including protein localization classification, cell cycle modeling, drug response prediction, and mechanism-of-action identification.
- On the HPA test set, SubCell achieved a micro AP of 0.878 for protein localization, surpassing existing models.
- SubCell accurately classifies drug treatments on triple-negative breast cancer cells, outperforming previous models.
- The model demonstrates strong performance on the JUMP1 dataset for replicate retrieval (mAP of 0.387) and MoA prediction, even without being trained on Cell Painting data.
- The multimodal embedding significantly outperformed unimodal embeddings in recovering known interacting protein pairs.
Limitations and Future Work
- Distinguishing rare classes: SubCell, like other models, struggled to distinguish rare protein localization classes (e.g., mitotic spindle, centrosome) compared to specialized models.
- Interphase/prophase distinction: In cell cycle analysis, SubCell had difficulty distinguishing between interphase and prophase stages in some datasets.
- Uncertainty in pseudotime: The model showed higher uncertainty in predicting cell cycle pseudotime at the beginning and end of the cycle.
- Future Work: The authors envision integrating SubCell embeddings with transcriptome-based models for improved biological context and causal reasoning. They also suggest fine-tuning and training models with lower-resolution images to further enhance performance. The creation of a universal virtual cell model is a long-term goal.
Summary updated