HydraProt: A New Deep Learning Tool for Fast and Accurate Prediction of Water Molecule Positions for Protein Structures

Andreas Zamanos, George Ioannakis, Ioannis Z. EmirisJournal of Chemical Information and Modeling 2024 · DOI · OpenReview

TL;DR

This paper introduces HydraProt, a novel deep learning tool that accurately and rapidly predicts the positions of water molecules around protein structures. By combining a 3D U-Net for initial sampling and a Multi-Layer Perceptron for refinement, HydraProt significantly outperforms existing methods in both accuracy and speed, offering a valuable resource for drug discovery and structural biology.

What Problem Does This Paper Address?

Water molecules are not just passive bystanders in biological systems; they are integral to protein structure, stability, and function. They play crucial roles in protein-protein interactions, enzyme catalysis, and, importantly, protein-ligand binding – a cornerstone of drug design. Accurately predicting where these water molecules are located around a protein is vital for understanding these interactions and for designing effective drugs. However, precisely mapping these water molecules is a significant computational challenge. Traditional methods like molecular dynamics are computationally intensive, and even advanced techniques struggle with accuracy, scalability, and speed, especially for large protein complexes. This gap highlights the need for a faster, more accurate, and scalable method for explicit water prediction.

Methods

HydraProt employs a two-stage deep learning approach:

  1. 3D U-Net for Coarse Sampling: The process begins by converting protein structures into a 3D grid representation. A 3D U-Net, a well-established deep learning architecture for image segmentation, is then used to rapidly identify regions with a high probability of containing water molecules. This "coarse sampling" significantly reduces the search space. The output is a set of candidate water positions.
  2. Multi-Layer Perceptron (MLP) for Refinement: For each candidate water position, an "embedding" is generated. This embedding captures the water molecule's positional relationship to nearby protein atoms, including atom types, distances, and angles. This rich feature set is then fed into an MLP, which predicts a confidence score for each candidate water position.
  3. Post-processing Refinement: A final refinement step is applied to prune low-confidence predictions, eliminate steric clashes between water molecules and protein atoms, and ensure that the predicted water positions adhere to physically realistic distances observed in experimental data. This multi-step process allows HydraProt to achieve both speed and accuracy.

Datasets

The authors curated a large dataset for training and evaluation:

  • Initial Dataset: 17,988 protein structures from the RCSB-PDB database.
  • High-Resolution Selection: Structures resolved at 1.5 Å or higher were selected for data reliability.
  • Nonredundancy Analysis: CD-HIT was used to reduce sequence similarity to 40%, resulting in a nonredundant set of 4968 PDB entries.
  • Training/Validation/Test Split: A total of 4000 protein structures were used, with 1000 for 3D U-net training, 2000 for MLP training, and 600 for testing.
  • Protein-Protein Complex Set: A separate test set of 151 complexes from 149 unique PDB structures was used to evaluate performance on inter-chain water interactions.

Key Insights and Ideas

  • Bridging Coarse and Fine Prediction: The core innovation lies in effectively combining a fast, coarse sampling method (3D U-Net) with a more detailed, discriminative model (MLP). This hybrid approach tackles the speed-accuracy trade-off that plagues many existing methods.
  • Embedding for Positional Context: The use of a carefully designed embedding that captures the geometric and chemical context of a water molecule relative to its protein environment is key to the MLP's success.
  • Post-processing for Physical Realism: The final refinement steps are crucial for ensuring that the predicted water positions are not only accurate but also physically plausible, adhering to known hydration shell distributions and avoiding steric clashes.
  • Speed as a Major Advantage: The paper strongly emphasizes the computational efficiency of HydraProt, making it suitable for large-scale studies and applications where rapid inference is critical.

Key Results

  • Superior Accuracy: HydraProt consistently outperformed state-of-the-art methods like Accutar and Galaxy-CNN across various metrics (precision, recall, F1 score) on both single protein structures and protein-protein complexes.
  • High Recall and Precision: Notably, HydraProt achieved high recall (identifying a large percentage of experimental waters) while maintaining competitive precision (minimizing false positives), particularly within the first hydration shell.
  • Low RMSD: The predicted water positions by HydraProt showed the lowest Root-Mean-Square Deviation (RMSD) compared to experimental waters, indicating high positional accuracy.
  • Significant Speedup: HydraProt achieved an average prediction time of 3.82 seconds per protein structure, a dramatic improvement over Galaxy-CNN's average of 137.89 seconds. This speed advantage is attributed to the efficient deep learning architecture.
  • Case Study Validation: The method demonstrated strong performance in predicting conserved water molecules in specific protein cavities (AGO4 and Trypsin), highlighting its utility in understanding functionally important water sites.

Limitations and Future Work

  • "Failed" Prediction Cases: The authors acknowledge that two specific PDB structures (6RXN and 3PQH) showed lower recall, suggesting potential issues with the experimental data itself or limitations in handling extremely low-density hydration.
  • Orientation of Water Molecules: The current representation of water molecules by their oxygen atom does not capture the orientation of hydrogen atoms, which can be important for specific hydrogen bonding interactions. The authors note that tools like UCSF-Chimera can add hydrogen atoms later.
  • Future Work: The authors suggest that HydraProt's speed and accuracy make it a valuable tool for various applications, including structural studies, drug discovery, and molecular docking, implying further exploration in these areas. The availability of the code, pretrained models, and data sets also facilitates future research and development.

Summary updated

BibTeX
@article{zamanos2024hydraprot,
  author = {Andreas Zamanos and George Ioannakis and Ioannis Z. Emiris},
  title = {{HydraProt: A New Deep Learning Tool for Fast and Accurate Prediction of Water Molecule Positions for Protein Structures}},
  journal = {Journal of Chemical Information and Modeling},
  year = {2024},
  doi = {10.1021/acs.jcim.3c01559},
}
Figure