PDRLGB: precise DNA-binding residue prediction using a light gradient boosting machine

Lei Deng, Lei Deng 0002, Juan Pan, Xiaojie Xu, Wenyi Yang, Chuyao Liu, Hui LiuBMC Bioinformatics 2018 · DOI

TL;DR

This paper introduces PDRLGB, a novel computational method for accurately predicting which amino acid residues in a protein are involved in binding to DNA. By combining a wide array of sequence and structural features with an efficient LightGBM model and a robust feature selection process, PDRLGB achieves state-of-the-art performance, offering a valuable tool for understanding protein-DNA interactions.

What Problem Does This Paper Address?

Understanding how proteins interact with DNA is fundamental to many biological processes, including DNA replication, repair, and gene regulation. Identifying the specific protein residues that directly bind to DNA (the "binding interface residues") is crucial for deciphering these complex mechanisms. While experimental methods exist, they are often costly and time-consuming. Computational approaches offer a more accessible alternative, but existing methods still face challenges in achieving high accuracy and efficiently handling the imbalanced nature of DNA-binding residue data (where non-binding residues vastly outnumber binding ones). This imbalance can lead to models that overfit and perform poorly. This paper aims to address these limitations by developing a more accurate and efficient computational predictor.

Methods

The PDRLGB method employs a multi-step pipeline:

  • Feature Extraction: A comprehensive set of 913 features is extracted, encompassing both sequence-based information (like Position-Specific Scoring Matrices - PSSMs, physicochemical properties, and evolutionary information) and structure-based information (derived from DSSP and NACCESS, including secondary structure, solvent accessibility, and B-factors). These features are generated using a sliding window of size 11 around each residue.
  • Feature Selection: To manage the large feature set and identify the most informative ones, a two-stage approach is used. First, a Random Forest algorithm ranks features by their importance (using the Mean Decrease Gini index). Second, an Incremental Feature Selection strategy iteratively adds features from the ranked list and evaluates the performance of a LightGBM classifier. This process selects an optimal subset of features that maximizes predictive performance.
  • Prediction Model: A Light Gradient Boosting Machine (LightGBM) is used to build the final prediction model. LightGBM is chosen for its efficiency and ability to handle large datasets, incorporating techniques like Gradient-based One-Side Sampling (GOSS) and Exclusive Feature Bundling (EFB) to speed up training and improve accuracy.

Datasets

The study utilizes several established datasets for training and evaluation:

  • PDNA-62: A benchmark dataset of 67 sequences from 62 protein-DNA complexes, with sequence identity between any two sequences ≤ 25%.
  • PDNA-224: A larger benchmark dataset containing 224 proteins, with redundant sequences removed using a 25% sequence identity cutoff.
  • TS-72: An independent test dataset comprising 72 protein chains.
  • TS-61: Another independent test dataset of 61 non-redundant DNA-binding protein sequences, with ≤30% sequence identity to sequences in the other datasets.

A residue is defined as a binding amino acid if its closest atom is within 3.5Å of a binding DNA atom.

Key Insights and Ideas

  • The Power of Hybrid Features: Combining a broad spectrum of sequence and structural features is crucial for accurately capturing the nuances of protein-DNA interactions.
  • Feature Selection is Key: Simply using all extracted features is not optimal. A rigorous feature selection process, like the one employed using Random Forest and incremental selection, is essential for building a high-performing and efficient model.
  • LightGBM's Efficiency: LightGBM proves to be a highly effective and computationally efficient algorithm for this task, outperforming other standard machine learning methods like SVM, Random Forest, and AdaBoost in terms of both accuracy and training speed.
  • Handling Imbalance: The study acknowledges the significant imbalance between binding and non-binding residues and implicitly addresses it through robust feature selection and the choice of LightGBM, which are known to handle such issues well.

Key Results

  • State-of-the-Art Performance: PDRLGB achieves superior performance compared to existing state-of-the-art methods on benchmark datasets (PDNA-62 and PDNA-224) and independent test datasets (TS-72 and TS-61). For instance, on PDNA-62, PDRLGB shows significant improvements in ST, F1, MCC, and AUC scores compared to SVM, RF, and AdaBoost. It also outperforms other specialized predictors like PreDNA and EL_PSSM-RT.
  • Optimal Window Size: The study found that a sliding window size of 11 provided the best performance for feature extraction.
  • Optimal Feature Subset: The optimal feature subset consisted of the top 800 features, which contributed most to classification accuracy.
  • Computational Efficiency: LightGBM demonstrated a substantial speedup in training time, being up to 20 times faster than SVM and also faster than AdaBoost, making it a practical choice for large-scale predictions.
  • Case Study Validation: Applying PDRLGB to the 2XMA complex showed highly accurate prediction of DNA-binding residues, with a high percentage of true positives and a low number of false positives and negatives, validating its real-world applicability.

Limitations and Future Work

  • Requirement for Structural Information: A noted limitation is that PDRLGB requires protein structural information. While this can be a constraint, the authors suggest that the increasing availability of solved protein structures and advancements in homology modeling will mitigate this in the future.
  • Future Directions: The authors believe PDRLGB can be a powerful tool for identifying DNA-binding residues, especially with the growing availability of high-quality protein-DNA complex structures. Further development could explore integrating more advanced feature extraction techniques or exploring different ensemble methods.

Summary updated

BibTeX
@article{deng2018pdrlgb,
  author = {Lei Deng and Lei Deng and Juan Pan and Xiaojie Xu and Wenyi Yang and Chuyao Liu and Hui Liu},
  title = {{PDRLGB: precise DNA-binding residue prediction using a light gradient boosting machine}},
  journal = {BMC Bioinformatics},
  year = {2018},
  doi = {10.1186/s12859-018-2527-1},
}
Figure