SmoPSI: Analysis and Prediction of Small Molecule Binding Sites Based on Protein Sequence Information

Wei Wang, Keliang Li, Hehe Lv, Hongjun Zhang, Shixun Wang, Junwei HuangComput. Math. Methods Medicine 2019 · DOI

TL;DR

This paper introduces SmoPSI, a novel method for predicting small molecule binding sites on proteins using only their amino acid sequences. By combining evolutionary information (PSSM) with physicochemical properties and employing an XGBoost classifier, SmoPSI achieves high accuracy, outperforming existing sequence-based methods and offering a faster, more cost-effective alternative to experimental approaches for drug discovery.

What Problem Does This Paper Address?

Identifying where small molecules bind to proteins is crucial for understanding biological functions and designing new drugs. Traditional experimental methods for this are time-consuming and expensive. While computational methods exist, many rely on 3D structural information, which isn't always available. Existing sequence-based methods often focus on limited features. This paper addresses the need for a robust, sequence-only method that can accurately predict these binding sites, thereby accelerating the drug discovery pipeline.

Methods

The core of SmoPSI lies in its feature engineering and classification approach.

  • Feature Engineering: The method extracts several key features from protein sequences:
  • PSSM (Position-Specific Scoring Matrix): Captures evolutionary conservation information, indicating how likely each amino acid is to mutate.
  • Physicochemical Properties: This includes hydrophilicity, hydrophobicity, electrostatic charge, and hydrogen bonding potential of amino acids. These properties are hypothesized to influence a residue's role in binding.
  • Amino Acid Type: Represented using one-hot encoding.
  • Sliding Window Sampling: To account for the local environment influencing binding, a sliding window of a fixed size (determined to be 15 residues) is applied to the protein sequence. This window captures the central residue and its neighbors.
  • XGBoost Classifier: An ensemble learning algorithm based on gradient boosting is used to classify residues as either binding or non-binding. XGBoost is chosen for its efficiency and high performance in machine learning tasks.

Datasets

The study utilized the SC-PDB database, a comprehensive resource for protein-small molecule complexes.

  • Data Preprocessing: Redundant protein-small molecule entries were removed using the PISCES program, and sequences with less than 30% identity and a minimum chain length of 40 amino acids were filtered.
  • Binding Site Definition: A residue was classified as a "binding residue" if its central atom was within 3.5 Å of any atom in a ligand molecule; otherwise, it was considered a "non-binding residue."
  • Training/Testing: The model was trained and validated on 14 different types of protein-small molecule complexes, including various nucleotides and other small molecules. The paper notes a significant data imbalance between binding and non-binding residues, which was addressed using a downsampling method during training.

Key Insights and Ideas

  • Multifaceted Feature Importance: The study highlights that a combination of evolutionary information (PSSM) and physicochemical properties is more effective than relying on a single feature type. PSSM emerged as the most significant predictor, followed by hydrophilicity, hydrophobicity, and charge.
  • Physicochemical Differences: Binding regions tend to exhibit a more hydrophilic character and stronger electrostatic charge compared to non-binding regions. This suggests that these properties play a direct role in molecular recognition.
  • Environmental Context Matters: The use of a sliding window (specifically, a window size of 15) demonstrates that the local sequence environment around a residue is crucial for accurate binding site prediction, as it captures the broader context of the binding pocket.
  • XGBoost Efficacy: The XGBoost algorithm proved to be a powerful tool for this classification task, achieving high accuracy and outperforming simpler models.

Key Results

  • High Performance: The SmoPSI model achieved an overall AUC (Area Under the ROC Curve) of 0.918 and an ACC (Accuracy) of 0.913 across the 14 tested datasets.
  • Specific Molecule Performance: For the ATP type of protein-small molecules, SmoPSI achieved particularly high results with an AUC of 0.935 and an ACC of 0.927.
  • Outperforms Existing Methods: When compared to existing sequence-based predictors like TargetS and EC-RUS on specific ligand types (ADP, ATP, GDP), SmoPSI demonstrated superior recall and AUC values. For instance, SmoPSI's average recall was 0.938, significantly higher than TargetS (0.561) and EC-RUS (0.627).

Limitations and Future Work

  • Data Imbalance: While addressed with downsampling, the inherent imbalance between binding and non-binding residues in real-world datasets remains a challenge.
  • Window Size Optimization: The optimal window size (w=15) was determined empirically. While effective, further exploration into adaptive window sizes or more sophisticated contextual modeling could be beneficial.
  • Future Directions: The authors suggest that their method can aid in understanding protein-small molecule binding mechanisms and exploring the characteristics of these interactions. Future work could involve expanding the range of molecules tested and exploring more advanced feature extraction techniques.

Summary updated

BibTeX
@article{wang2019smopsi,
  author = {Wei Wang and Keliang Li and Hehe Lv and Hongjun Zhang and Shixun Wang and Junwei Huang},
  title = {{SmoPSI: Analysis and Prediction of Small Molecule Binding Sites Based on Protein Sequence Information}},
  journal = {Comput. Math. Methods Medicine},
  year = {2019},
  doi = {10.1155/2019/1926156},
}
Figure