Prediction of liquid-liquid phase separating proteins using machine learning

Xiaoquan Chu, Tanlin Sun, Qian Li, Youjun Xu, Zhuqing Zhang, Luhua Lai, Jianfeng PeiBMC Bioinformatics 2022 · OpenReview

TL;DR

This paper introduces PSPredictor, a novel machine learning tool that accurately predicts proteins capable of liquid-liquid phase separation (LLPS) based solely on their amino acid sequences. By integrating both sequence composition and order, PSPredictor achieves high accuracy and outperforms existing methods, offering a valuable resource for identifying potential LLPS proteins in the human genome and understanding their roles in cellular function and disease.

What Problem Does This Paper Address?

Cells organize their internal space using membrane-less organelles, which are formed through a process called liquid-liquid phase separation (LLPS). This phenomenon is crucial for various cellular functions, and its dysregulation is linked to diseases like ALS and Alzheimer's. While LLPS is gaining attention, our understanding of which proteins can undergo this process (Phase Separation Proteins or PSPs) and their prevalence is still limited. Existing computational tools for predicting PSPs often rely on specific protein features or small datasets, limiting their general applicability. The authors aim to address this gap by developing a more general and accurate sequence-based prediction tool for PSPs.

Methods

The core of this work is the development of PSPredictor, a machine learning model. The key methodological innovations include:

  • Sequence Embedding with Word2Vec (w2v): Instead of relying on hand-crafted features, the authors use a natural language processing technique, word2vec, to embed protein sequences into vectors. This method captures both the "words" (residues or k-mers) and their contextual relationships within the sequence, effectively encoding both compositional and sequential information.
  • Positional Encoding: To further enhance the capture of sequential information, they incorporate positional encoding using sine and cosine functions, a technique commonly used in transformer networks. This allows the model to understand the order of amino acids.
  • Gradient Boosting Decision Tree (GBDT): For the classification task (predicting whether a protein is a PSP or not), the authors found that GBDT, a powerful ensemble learning algorithm, performed best among several tested machine learning algorithms.
  • Dataset Construction and Balancing: They carefully curated positive (known PSPs) and negative (non-PSPs) datasets. Due to the imbalance in the number of positive and negative samples, they employed undersampling techniques for the negative dataset to ensure fair training.

Datasets

  • Positive Training Dataset (P1): 293 protein sequences from the LLPSDB database, representing proteins known to undergo LLPS.
  • External Test Dataset (T1+): 60 protein sequences from the LLPSDB, held out from the initial training set.
  • Full Positive Training Dataset (P): All 353 protein sequences from LLPSDB, used for training the final PSPredictor model.
  • Negative Training Dataset (N1): 5258 protein sequences, selected from the PDB databank, specifically single-domain proteins with solved 3D structures, to represent non-LLPS proteins.

Key Insights and Ideas

  • Sequence Order Matters: The success of PSPredictor, particularly its ability to outperform models relying solely on amino acid composition (like Com-Len), highlights the critical role of sequence order and patterns in determining LLPS propensity.
  • Generalizability is Key: By avoiding feature engineering and relying on a powerful sequence embedding technique, PSPredictor demonstrates a more general approach to PSP prediction, applicable across different protein types and domains.
  • LLPS is Widespread: The analysis of human proteins suggests that a significant portion of the proteome might be capable of LLPS under specific conditions, underscoring the fundamental importance of this process.
  • Scaffolds vs. Clients: The tool can help distinguish between proteins that drive LLPS (scaffolds) and those that are recruited into condensates (clients), offering insights into the mechanisms of condensate formation.

Key Results

  • High Accuracy: PSPredictor achieved a tenfold cross-validation training accuracy of 94.71% and a prediction accuracy of 92.50% on an external test set.
  • Superior Performance: PSPredictor outperformed several previously reported first-generation PSP prediction tools (PScore, CatGranule, PSPer) in predicting PSPs on the test dataset.
  • Web Server Availability: A user-friendly web server for PSPredictor is available at http://www.pkumdl.cn/PSPredictor, allowing researchers to easily predict LLPS propensity for their proteins of interest.
  • Identification of Novel PSPs: The tool identified novel scaffold proteins for stress granules and predicted potential PSP candidates within the human genome, providing valuable leads for further experimental investigation.

Limitations and Future Work

  • Environmental Conditions: The current model does not account for specific environmental conditions (temperature, salt concentration, pH, crowding) that can influence LLPS. Future work could integrate this information into training.
  • PTMs and Mutants: PSPredictor does not currently consider post-translational modifications (PTMs) or protein mutants, which can significantly impact LLPS. Developing models that incorporate these factors is a future challenge.
  • Data Size: The authors acknowledge that the limited size of the current dataset might restrict the application of more complex deep learning algorithms. As more experimental data becomes available, these advanced methods could be explored.
  • Driver vs. Client Proteins: While the tool can predict PSPs, further investigation is needed to definitively distinguish between driver (scaffold) and client proteins, and whether specialized tools are required for each.

Summary updated

Abstract

The liquid–liquid phase separation (LLPS) of biomolecules in cell underpins the formation of membraneless organelles, which are the condensates of protein, nucleic acid, or both, and play critical roles in cellular function. Dysregulation of LLPS is implicated in a number of diseases. Although the LLPS of biomolecules has been investigated intensively in recent years, the knowledge of the prevalence and distribution of phase separation proteins (PSPs) is still lag behind. Development of computational methods to predict PSPs is therefore of great importance for comprehensive understanding of the biological function of LLPS. Based on the PSPs collected in LLPSDB, we developed a sequence-based prediction tool for LLPS proteins (PSPredictor), which is an attempt at general purpose of PSP prediction that does not depend on specific protein types. Our method combines the componential and sequential information during the protein embedding stage, and, adopts the machine learning algorithm for final predicting. The proposed method achieves a tenfold cross-validation accuracy of 94.71%, and outperforms previously reported PSPs prediction tools. For further applications, we built a user-friendly PSPredictor web server ( http://www.pkumdl.cn/PSPredictor ), which is accessible for prediction of potential PSPs. PSPredictor could identifie novel scaffold proteins for stress granules and predict PSPs candidates in the human genome for further study. For further applications, we built a user-friendly PSPredictor web server ( http://www.pkumdl.cn/PSPredictor ), which provides valuable information for potential PSPs recognition.
BibTeX
@article{chu2022prediction,
  author = {Xiaoquan Chu and Tanlin Sun and Qian Li and Youjun Xu and Zhuqing Zhang and Luhua Lai and Jianfeng Pei},
  title = {{Prediction of liquid-liquid phase separating proteins using machine learning}},
  journal = {BMC Bioinformatics},
  year = {2022},
  url = {https://openreview.net/forum?id=wU2qXTLZtS},
}
Figure