SidechainNet: An All-Atom Protein Structure Dataset for Machine Learning

Jonathan E. King, David Ryan Koes

TL;DR

This paper introduces SidechainNet, a comprehensive dataset that extends existing protein structure datasets by including detailed information about protein sidechains. This new dataset aims to enable machine learning models to better understand and predict the full, all-atom structure of proteins, which is crucial for understanding their function and for applications like drug discovery.

What Problem Does This Paper Address?

Predicting the 3D structure of proteins from their amino acid sequences is a grand challenge in biology. While deep learning has made significant strides, most current methods focus primarily on predicting the protein's "backbone" – the main chain of atoms. However, the "sidechains," which are the chemically diverse groups attached to the backbone, play a critical role in a protein's function, including enzyme activity, drug binding, and interactions with other molecules. Existing datasets often treat backbone and sidechain prediction as separate problems or lack the detailed sidechain information needed for comprehensive modeling. This gap limits the ability of machine learning models to capture the full atomic detail of proteins, potentially hindering progress in areas like structure-based drug discovery.

Methods

The authors built SidechainNet by extending the existing ProteinNet dataset. They re-downloaded all-atom protein structures from the Protein Data Bank (PDB) and integrated this detailed information, including atomic coordinates and torsional angles for both backbone and sidechains, into the ProteinNet framework. A key methodological insight came from realizing that simply predicting backbone torsional angles wasn't enough for accurate 3D reconstruction; specific backbone bond angles also needed to be considered. SidechainNet includes these crucial bond angles, along with sidechain-specific torsional angles and atomic coordinates, to provide a complete representation. They also developed a Python package to facilitate the loading and manipulation of this data for machine learning tasks.

Datasets

The primary dataset introduced is SidechainNet. It is built upon and extends the ProteinNet dataset. SidechainNet incorporates all the features of ProteinNet (primary sequence, PSSM, missing residue masks, backbone coordinates, and backbone torsion angles) and adds:

  • Backbone bond angles
  • Sidechain torsion angles
  • Sidechain coordinates

The dataset is organized into training, validation (with varying similarity thresholds to the training set), and testing splits, mirroring the structure of ProteinNet to prevent information leakage and ensure robust evaluation.

Key Insights and Ideas

  • Sidechains are Crucial for Function: The paper strongly emphasizes that sidechain structure and orientation are not mere details but are fundamental to protein function, influencing everything from enzyme catalysis to drug interactions.
  • Backbone Bond Angles Matter: A surprising finding was that even with accurate backbone torsional angles, reconstructing precise 3D coordinates was challenging if backbone bond angles were assumed to be fixed. Including these angles significantly improved reconstruction accuracy.
  • Unified Representation is Powerful: By providing a single dataset with comprehensive all-atom information (backbone and sidechain), SidechainNet enables the development of models that can learn from and predict the complete protein structure, leading to more expressive and potentially more accurate representations.
  • Bridging the Gap to Applications: The dataset and associated tools are designed to facilitate research in areas like structure-based drug discovery and molecular dynamics simulations, where atomic-level detail is paramount.

Key Results

The paper doesn't present quantitative results from training specific models on SidechainNet, as its primary contribution is the dataset itself and the framework for using it. However, they demonstrate:

  • Improved Reconstruction Accuracy: By including backbone bond angles in their reconstruction process, they achieved lower Root-Mean-Square Deviation (RMSD) compared to models where these angles were held constant.
  • Comprehensive Data Representation: SidechainNet provides a rich set of features, including up to 10 atomic coordinates and 6 torsional angles for each sidechain, allowing for detailed structural modeling.
  • Accessible Tools: They provide a Python package that simplifies loading and using the dataset with popular machine learning frameworks like PyTorch, including DataLoaders for efficient batching.

Limitations and Future Work

  • Exclusion of Some Structures: A small percentage (0.3%) of ProteinNet structures were excluded due to parsing issues or sequence discrepancies, with plans to address these in future releases.
  • Single-Molecule Chains Only: The dataset currently focuses on single protein chains; multi-chain proteins are split into independent entries.
  • No Energetic Minimization: Structures are used directly from the PDB and are not energetically minimized. The authors plan to incorporate AMBER-based minimization in a future release to improve consistency and reconstruction accuracy.
  • Secondary Structure Exclusion: Secondary structure information was temporarily excluded from the dataset.
  • Future Research Directions: The authors are interested in determining if including sidechain information improves the performance of structure prediction methods. They also suggest that the dataset and construction methods could be adapted for creating similar all-atom datasets using alternative clustering approaches.

Summary updated

Abstract

Despite recent advancements in deep learning methods for protein structure prediction and representation, little focus has been directed at the simultaneous inclusion and prediction of protein backbone and sidechain structure information. We present SidechainNet, a new dataset that directly extends the ProteinNet dataset. SidechainNet includes angle and atomic coordinate information capable of describing all heavy atoms of each protein structure. In this paper, we provide background information on the availability of protein structure data and the significance of ProteinNet. Thereafter, we argue for the potentially beneficial inclusion of sidechain information through SidechainNet, describe the process by which we organize SidechainNet, and provide a software package (https://github.com/jonathanking/sidechainnet) for data manipulation and training with machine learning models.
BibTeX
@inproceedings{king2021sidechainnet,
  author = {Jonathan E. King and David Ryan Koes},
  title = {{SidechainNet: An All-Atom Protein Structure Dataset for Machine Learning}},
  booktitle = {Proteins},
  year = {2021},
  doi = {10.1002/prot.26169},
}
Figure