SidechainNet

Colab Walkthrough, Paper

SidechainNet is a protein structure prediction dataset that directly extends ProteinNet¹ by Mohammed AlQuraishi.

Specifically, SidechainNet adds measurements for protein angles and coordinates that describe the complete, all-atom protein structure (backbone and sidechain, excluding hydrogens) instead of the protein backbone alone.

This repository provides the following:

1. SidechainNet datasets stored as pickled Python dictionaries.
2. Methods for loading and batching SidechainNet data efficiently in PyTorch.
3. Methods for generating protein structure visualizations (.pdb, 3Dmol, .gltf) from model predictions.
4. Methods for augmenting SidechainNet to include new proteins and specify dataset organization.

Summary of SidechainNet data

*L reperesents the length of any given protein in the dataset.

^**Secondary structure is acquired from ProteinNet for training sets only. Blank characters are added by SidechainNet to refer to unmatched amino acids after alignment.(Added January 2021)

^***SidechainNet explicitly includes oxygen atoms as part of the backbone coordinate data in contrast to ProteinNet, which only includes the primary N, C_alpha, C atoms.

^§Stored as string values in the underlying SidechainNet data dictionary.

Other included data

^§Stored as string values in the underlying SidechainNet data dictionary.

^†Includes a 1 for each residue that has been modified to a standard residue according to the mapping in sidechainnet.utils.measure.ALLOWED_NONSTD_RESIDUES (e.g., selenomethionine -> methionine).

Installation

pip install sidechainnet

Usage Examples

Loading SidechainNet as a Python dictionary

>>> import sidechainnet as scn
>>> data = scn.load(casp_version=12, thinning=30)

In its most basic form, SidechainNet is stored as a Python dictionary organized by train, validation, and test splits. These splits are identical to those described in ProteinNet.

Within each train/validation/test split in SidechainNet is another dictionary mapping data entry types (seq, ang, etc.) to a list containing this data type for every protein. In the example below, seq{i}, ang{i}, ... all refer to the i^th protein in the dataset.

data = {"train": {"seq": [seq1, seq2, ...],  # Sequences, 1-letter codes
                  "ang": [ang1, ang2, ...],  # Angles
                  "crd": [crd1, crd2, ...],  # Coordinates
                  "evo": [evo1, evo2, ...],  # PSSMs and Information Content
                  "sec": [sec1, sec2, ...],  # Secondary structure labels (DSSP)
                  "res": [res1, res2, ...],  # X-ray crystallographic resolution
                  "ids": [id1, id2,   ...],  # Corresponding ProteinNet IDs
                  "mod": [mod1, mod2, ...],  # Modified residue annotations
                  "ums": [ums1, ums2, ...]   # Unmodified sequences, 3-letter codes
                  },
        "valid-10": {...},
            ...
        "valid-90": {...},
        "test":     {...},
        "settings": {...},
        "description" : "SidechainNet for CASP 12."
        "date": "January 20, 2020"
        }

By default, the load function downloads the data from the web into the current directory and loads it as a Python dictionary. If the data already exists locally, it reads it from disk. Other than the requirement that the data must be loaded using Python, this method of data loading is agnostic to any downstream analysis.

Loading SidechainNet as an interactive SCNDataset object

The easiest way to interact with SidechainNet is most likely by using the SCNDataset and SCNProtein objects.

>>> data = scn.load("debug", scn_dataset=True)
>>> data
SCNDataset(n=461)
>>> data["1HD1_1_A"]
SCNProtein(1HD1_1_A, len=75, missing=0, split='train')
>>> data[0]
SCNProtein(2CMY_d2cmyb1, len=23, missing=2, split='train')

Available features:

• SCNDataset is iterable,
• proteins (SCNProteins) can selected from the dataset by name or index,
• proteins can be visualized with .to_3Dmol() and writable to PDBs with .to_pdb().
• non-terminal hydrogens can be added with SCNProtein.add_hydrogens(),

Additionally, all of the attributes below are available directly from the SCNProtein object:

• coords, angles, seq, unmodified_seq, mask, evolutionary, secondary_structure, resolution, is_modified, id, split

Loading SidechainNet with PyTorch DataLoaders

The load function can also be used to load SidechainNet data as a dictionary of torch.utils.data.DataLoader objects. PyTorch DataLoaders make it simple to iterate over dataset items for training machine learning models. This method is recommended for using SidechainNet data with PyTorch.

>>> dataloaders = scn.load(casp_version=12, with_pytorch="dataloaders")
>>> dataloaders.keys()
['train', 'train_eval', 'valid-10', ..., 'valid-90', 'test']
>>> dataloaders['train'].dataset
ProteinDataset(casp_version=12, split='train', n_proteins=81454,
               created='Sep 20, 2020')
>>> for batch in dataloaders['train']:
....    predicted_angles = model(batch.seqs)
....    sb = scn.BatchedStructureBuilder(batch.int_seqs, predicted_angles)
....    predicted_coords = sb.build()
....    loss = compute_loss(batch.angs, batch.crds,               # True values
....                        predicted_angles, predicted_coords)   # Predicted values
....    ...

For more information on the batch variable, see the section Using SidechainNet to train an all-atom protein structure prediction model below.

By default, the provided DataLoaders use a custom batching method that randomly generates batches of proteins of similar length. For faster training, it generates larger batches when the average length of proteins in the batch is small, and smaller batches when the proteins are large. The probability of selecting small-length batches is decreased so that each protein in SidechainNet is included in a batch with equal probability. See dynamic_batching and collate_fn arguments for more information on modifying this behavior. In the example below, model_input is a collated Tensor containing sequence and PSSM information.

Converting Angle Representations into All-Atom Structures

An important component of this work is the inclusion of both angular and 3D coordinate representations of each protein. Researchers who develop methods that rely on angular representations may be interested in converting this information into 3D coordinates. For this reason, SidechainNet provides a method to convert the angles it provides into Cartesian coordinates.

In the below example, angles is a NumPy matrix or Torch Tensor following the same organization as the NumPy angle matrices provided in SidechainNet. sequence is a string representing the protein's amino acid sequence.

>>> (len(batch.seqs), batch.angs.shape)  # 12 angles per residue
(128, (128, 12))
>>> sb = scn.StructureBuilder(batch.seqs, batch.angs)
>>> coords = sb.build()
>>> coords.shape  # 14 atoms per residue (128*14 = 1792)
(1792, 3)

Visualizing & Interacting with All-Atom Structures (PDB, py3Dmol, and gLTF)

SidechainNet also makes it easy to visualize both existing and predicted all-atom protein structures. These visualizations are available as PDB files, py3Dmol.view objects, and Graphics Library Transmission Format (gLTF) files. Examples of each are included below.

The PDB format is a typical format for representing protein structures and can be opened in software tools like PyMOL. py3Dmol (built on 3Dmol.js²) enables users to visualize and interact with protein structures on the web and in Jupyter Notebooks via an open-source, object-oriented, and hardware-accelerated Javascript library. Finally, gLTF files, despite their larger size, can be convenient for visualizing proteins on the web or in contexts where other protein visualization tools are not supported.

>>> sb.to_pdb("example.pdb")
>>> sb.to_gltf("example.gltf")

Using SidechainNet to train an all-atom protein structure prediction model

Below is an outline of how to use this repository for machine learning model training. Assuming you have a predictive model variable model and a loss function loss_fn used to evaluate your model, you can load SidechainNet using our DataLoaders and begin training.

import sidechainnet as scn

data = scn.load(casp_version=12,thinning=30, with_pytorch="dataloaders")

for epoch in range(100):
    # Training epoch
    for batch in data['train']:
        predictions = model(batch.seqs)
        loss = loss_fn(predictions, batch.angs, batch.crds)
        loss.backwards()
        ...
    
    # Evaluate performance on down-sampled training set for efficiency
    for batch in data['train-eval']:
        predictions = model(batch.seqs)
        loss = loss_fn(predictions, batch.angs, batch.crds)
        loss.backwards()
        ...

    # Evaluate performance on each of the 7 validation sets
    for valid_set in [data[f'valid-{split}'] for split in scn.utils.download.VALID_SPLITS]:
        for batch in valid_set:
            predictions = model(batch.seqs)
            loss = loss_fn(predictions, batch.angs, batch.crds)
            loss.backwards()
            ...

# Evaluate performance on test set
for batch in data['test']:
    predictions = model(batch.seqs)
    loss = loss_fn(predictions, batch.angs, batch.crds)
    ...

The batch variable above is a collections.namedtuple that has the following attributes:

Reproducing or Extending SidechainNet

If you would like to reproduce our work or make modifications/additions to the dataset, please see the example we provide in our Colab Walkthrough. In simple terms, you will need to call scn.create with the desired CASP/ProteinNet information or provide a list of ProteinNet-formatted IDs to scn.create_custom. Please note that since some data is acquired from ProteinNet directly (e.g., Position Specific Scoring Matrices), protein entries will exclude this data if it was not previously available in ProteinNet.

# Reproduce SidechainNet
scn.create(casp_version=12, training_set=30)

# Create a custom version of SidechainNet
custom_ids = scn.get_proteinnet_ids(casp_version=12, split="train", thinning=30)
# Include a protein released in April 2021 (not included in SidechainNet)
custom_ids += ['7C3K_1_A'] 
scn.create_custom(pnids=custom_ids,
                  output_filename="custom.pkl",
                  short_description="Custom SidechainNet.")

Package Requirements

• Python 3
• ProDy (pip install ProDy)
  • Biopython
  • numpy
  • scipy
• PyTorch
• tqdm
• py3Dmol (pip install py3Dmol)
• pymol (optional, for gltf support, repo, linux install requires libxml2)

Acknowledgements

Thanks to Mohammed AlQuraishi for his inspiring work on protein structure prediction. Thanks, also, to Jeppe Hallgren for his development of a ProteinNet text record parser, which I have used in part here.

This work is supported by R01GM108340 from the National Institute of General Medical Sciences, is supported in part by the University of Pittsburgh Center for Research Computing through the resources provided, and by NIH T32 training grant T32 EB009403 as part of the HHMI-NIBIB Interfaces Initiative.

Project structure (continuous integration, docs, testing) based on the Computational Molecular Science Python Cookiecutter version 1.1.

References

1. SidechainNet: An All-Atom Protein Structure Dataset for Machine Learning. J.E. King, D. Koes. Proteins: Structure, Function, and Bioinformatics (Vol. 89, Issue 11, pp. 1489–1496). Wiley. (2021).
2. ProteinNet: a standardized data set for machine learning of protein structure.. M. AlQuraishi. BMC Bioinformatics 20, 311 (2019).
3. 3dmol.js: molecular visualization with WebGL. N. Rego and D. Koes. Bioinformatics, 31(8):1322–1324, (2014).

Other Resources

Computational Biology Skills Seminar, U.C. Berkeley, May 13, 2021

• Slides
• Notebook
• Recording

Citation

Please cite our paper if you find SidechainNet useful in your work.

@article{https://doi.org/10.1002/prot.26169,
	author = {King, Jonathan Edward and Koes, David Ryan},
	title = {SidechainNet: An all-atom protein structure dataset for machine learning},
	journal = {Proteins: Structure, Function, and Bioinformatics},
	volume = {89},
	number = {11},
	pages = {1489-1496},
	keywords = {dataset, deep learning, machine learning, protein structure, proteins, software},
	doi = {https://doi.org/10.1002/prot.26169},
	url = {https://onlinelibrary.wiley.com/doi/abs/10.1002/prot.26169},
	eprint = {https://onlinelibrary.wiley.com/doi/pdf/10.1002/prot.26169},
	abstract = {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 and can be extended by users to include new protein structures as they are released. In this article, 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.},
	year = {2021}
}

Copyright

Copyright (c) 2023, Jonathan King