PreMLS: The undersampling technique based on ClusterCentroids to predict multiple lysine sites

Yun Zuo, Xingze Fang, Jiayong Wan, Wenying He, Xiangrong Liu, Xiangxiang Zeng, Zhaohong DengPLOS Computational Biology 2024 · DOI

TL;DR

This paper introduces PreMLS, a novel deep learning model that accurately predicts multiple post-translational modifications (PTMs) occurring simultaneously on lysine residues. By employing a ClusterCentroids undersampling technique to address class imbalance and a specialized Convolutional Neural Network (CNN) architecture, PreMLS significantly outperforms existing methods in identifying complex lysine modification patterns.

What Problem Does This Paper Address?

Post-translational modifications (PTMs) are crucial for protein function, and lysine modifications, in particular, play vital roles in various biological processes. While many studies focus on predicting single lysine PTMs, understanding how multiple modifications co-occur on the same lysine residue is essential for a comprehensive view of protein regulation and disease mechanisms. However, existing research often overlooks the interplay between different PTMs at a single site, and datasets for predicting multiple lysine PTMs are typically imbalanced, posing a significant challenge for machine learning models. This imbalance can lead to models that are biased towards the majority class, failing to accurately predict less frequent modification combinations.

Methods

The PreMLS model is built upon several key components:

  • Feature Encoding: The study utilizes the Multi-Label Position-specific Triple Amino Acid Propensity (MLPSTAAP) algorithm to encode protein sequences. This method captures local sequence information and relationships between amino acids around the lysine residue, generating a feature matrix.
  • Class Imbalance Mitigation: To address the severe class imbalance in the dataset, PreMLS employs a novel undersampling technique called ClusterCentroids. This method, based on MiniBatchKmeans, identifies representative "centroids" from the majority class samples, effectively reducing their number while preserving essential data diversity. This is a key innovation compared to random undersampling.
  • Deep Learning Architecture: A custom Convolutional Neural Network (CNN) architecture is designed for multi-label prediction. CNNs are well-suited for capturing local patterns and sequential dependencies in biological sequences, making them effective for this task.
  • Multi-Label Classification: Unlike methods that break down the problem into multiple binary classifications, PreMLS directly predicts all possible combinations of lysine modifications simultaneously.

Datasets

The study uses a dataset derived from the CPLM4.0 database, which contains human protein sequences with known lysine PTMs. After rigorous filtering and deduplication using CD-HIT, the dataset comprises 18,978 human protein sequences with four types of PTMs: acetyllysine, crotonyllysine, methyllysine, and succinyllysine. These were categorized into 11 distinct classes representing single and multiple modification combinations. The data was split into a training set (70%) and a test set (30%).

Key Insights and Ideas

  • The Importance of Direct Multi-Label Prediction: Treating multiple lysine PTM prediction as a direct multi-label problem, rather than a series of binary classifications, is crucial for capturing the interplay between modifications.
  • ClusterCentroids for Effective Undersampling: The ClusterCentroids undersampling technique, leveraging MiniBatchKmeans, proves more effective than random undersampling in maintaining data representativeness and improving model generalization for minority classes.
  • CNNs Excel at Capturing Sequential Dependencies: The CNN architecture effectively leverages the sequential nature of protein sequences and the contextual information around lysine residues for accurate multi-label prediction.
  • Model Interpretability Matters: SHAP analysis reveals which features are most influential in predicting specific lysine modifications, offering insights into the underlying biological mechanisms.

Key Results

  • Superior Performance: PreMLS significantly outperforms existing models like iMul-kSite and predML-Site on an independent test dataset, showing approximately 30% improvement in Aiming and Coverage, and over 25% in Accuracy and Absolute-True.
  • Effective Class Imbalance Handling: The ClusterCentroids undersampling method, when applied, significantly improves the model's predictive accuracy across all classes except the dominant first class, demonstrating its effectiveness in balancing the dataset.
  • Ablation Study Findings: The ablation experiments confirm that the ClusterCentroids undersampling strategy leads to better performance compared to random undersampling or no undersampling.
  • High Absolute Accuracy: The model achieves a high absolute accuracy rate for samples with only crotonyllysine (C) and those with both crotonyllysine and methyllysine (C and M), indicating its strength in predicting specific combinations.

Limitations and Future Work

The authors acknowledge several areas for future improvement:

  • Algorithm Refinement: The ClusterCentroids method might omit some vital data. Future work could explore alternative resampling strategies that better preserve feature information.
  • Model Architecture Optimization: The CNN layers show signs of overfitting on the test set. Incorporating regularization techniques, fine-tuning pooling parameters, and using Dropout could enhance generalization.
  • Data Engineering Advancement: The current feature extraction approach is basic. Exploring diverse feature extraction techniques could further improve multi-class prediction.
  • CNN Limitations: CNNs may struggle with capturing long-range dependencies in protein sequences. Incorporating or substituting with architectures like RNNs or Transformers could address this.

Summary updated

BibTeX
@article{zuo2024premls,
  author = {Yun Zuo and Xingze Fang and Jiayong Wan and Wenying He and Xiangrong Liu and Xiangxiang Zeng and Zhaohong Deng},
  title = {{PreMLS: The undersampling technique based on ClusterCentroids to predict multiple lysine sites}},
  journal = {PLOS Computational Biology},
  year = {2024},
  doi = {10.1371/journal.pcbi.1012544},
}
Figure