TL;DR
This paper introduces CoDropleT, a novel deep learning method that leverages AlphaFold2's protein structure predictions to accurately forecast whether pairs of proteins will co-condense into liquid-like cellular structures called membraneless organelles (MLOs). This tool promises to accelerate research into MLO function and dysfunction in diseases like neurodegeneration.
What Problem Does This Paper Address?
Cells organize themselves using membraneless organelles (MLOs), which are dynamic compartments formed by the liquid-liquid phase separation (LLPS) of proteins and other biomolecules. These MLOs are crucial for cellular processes, but their formation and composition are not fully understood. Experimentally identifying which proteins co-condense to form these structures is challenging. While previous computational methods predicted individual protein phase separation propensity, predicting co-condensation of protein pairs remained a significant hurdle. This gap hinders our ability to understand MLO assembly and to develop therapies for diseases linked to aberrant protein condensation.
Methods
The core of this work is the development of CoDropleT, a deep learning model built using the Haiku framework. CoDropleT takes pairs of protein sequences as input and utilizes AlphaFold2 to generate structural representations (single and pair representations) for each protein. These representations are then fed into a series of transformer blocks, which employ self-attention mechanisms to capture complex sequence and structural features. Finally, a "PPS head" analyzes these features to predict the probability of co-condensation between the protein pair. The model's architecture is designed to handle long-range dependencies within protein sequences, which are critical for understanding protein interactions.
Datasets
The CoDropleT model was trained and evaluated using datasets derived from the CD-CODE database, which contains experimental data on co-condensing proteins.
- PPS-positive dataset: Formed by creating pairs from proteins known to co-condense within specific MLOs.
- PPS-negative dataset: Created by pairing proteins that undergo phase separation individually but not together, or by excluding known co-condensing pairs.
To address potential biases and improve robustness, the dataset was expanded to include proteins that self-condense. The authors also curated non-redundant datasets filtered by 50% sequence identity, containing 1,791 PPS-positive and 1,579 PPS-negative human protein pairs for final training and testing.
Key Insights and Ideas
- Leveraging AlphaFold2 for Co-condensation: The key innovation is integrating AlphaFold2's highly accurate protein structure predictions as input features for predicting co-condensation. This moves beyond sequence-based predictions by incorporating structural information, which is intuitively important for molecular interactions.
- Transformer Architecture for Protein Interactions: The use of transformer blocks with self-attention is well-suited for capturing the complex, non-linear relationships between protein sequences and their propensity to interact and co-condense.
- Predicting MLO Composition: CoDropleT demonstrates a strong ability to predict relevant co-condensing protein pairs within known MLOs like stress granules, P-bodies, and nucleoli, suggesting its utility in understanding MLO assembly.
- Identifying Scaffold-Client Relationships: The model can differentiate between a protein scaffold's co-condensation propensity with its known partners versus unrelated proteins, offering insights into how MLOs are built around specific core components.
- Proteome-Wide Predictions Reveal Amino Acid Preferences: Analyzing the entire human cytosolic proteome revealed that proteins with a high propensity for co-condensation tend to be enriched in charged residues and depleted in disorder-promoting residues, offering a glimpse into the molecular grammar of co-condensation.
Key Results
- The initial CoDropleT model achieved an AUC of 0.874.
- After expanding the training dataset to include self-condensing proteins, the final model showed improved performance with an AUC of 0.923, accuracy of 0.884, precision of 0.749, and recall of 0.908.
- CoDropleT consistently predicted relevant co-condensing protein pairs for three well-characterized MLOs (stress granules, P-bodies, and nucleoli) when compared to random protein sets.
- Protein scaffolds of stress granules showed significantly higher co-condensation scores with known SG components compared to non-SG proteins.
- Approximately 25% of all possible cytosolic human protein pairs were predicted to have a co-condensation score above 0.8.
Limitations and Future Work
- Dataset Limitations: The authors acknowledge the challenge of creating large, standardized datasets for training and validation, particularly for PPS-negative examples. This limitation can lead to overfitting and highlights the need for more robust, curated datasets.
- Extrinsic Factors: CoDropleT currently relies primarily on intrinsic protein sequence and structural properties. Future work could incorporate extrinsic factors, such as cellular conditions or interactions with nucleic acids, to provide a more comprehensive prediction of phase separation.
- Experimental Validation: The authors emphasize the crucial need for experimental validation of their computational predictions to confirm the model's accuracy and utility.
- Model Refinement: There is scope for further refinement of the datasets used for training and validation to improve model performance.
Summary updated