Introduction

Protein secondary structure can be calculated based on its atoms' 3D coordinates once the protein's 3D structure is solved using X-ray crystallography or NMR. Commonly, DSSP is the tool used for calculating the secondary structure and assigns one of the following secondary structure types (https://swift.cmbi.umcn.nl/gv/dssp/index.html) to every amino acid in a protein:

1. C: Loops and irregular elements (corresponding to the blank characters output by DSSP)
2. E: β-strand
3. H: α-helix
4. B: β-bridge
5. G: 3-helix
6. I: π-helix
7. T: Turn
8. S: Bend

However, X-ray or NMR is expensive. Ideally, we would like to predict the secondary structure of a protein based on its primary sequence directly, which has had a long history. A review on this topic is published recently, Sixty-five years of the long march in protein secondary structure prediction: the final stretch?.

For the purpose of secondary structure prediction, it is common to simplify the aforementioned eight states (Q8) into three (Q3) by merging (E, B) into E, (H, G, I) into E, and (C, S, T) into C. The current accuracy for three-state (Q3) secondary structure prediction is about ~85% while that for eight-state (Q8) prediction is <70%. The exact number depends on the particular test dataset used.

Dataset

The main dataset lists peptide sequences and their corresponding secondary structures. It is a transformation of https://cdn.rcsb.org/etl/kabschSander/ss.txt.gz downloaded at 2018-06-06 from RSCB PDB into a tabular structure. If you download the file at a later time, the number of sequences in it will probably increase.

Description of columns:

1. pdb_id: the id used to locate its entry on https://www.rcsb.org/
2. chain_code: when a protein consists of multiple peptides (chains), the chain code is needed to locate a particular one.
3. seq: the sequence of the peptide
4. sst8: the eight-state (Q8) secondary structure
5. sst3: the three-state (Q3) secondary structure
6. len: the length of the peptide
7. has_nonstd_aa: whether the peptide contains nonstandard amino acids (B, O, U, X, or Z).

Key steps in the transformation:

• Both Q3 and Q8 secondary structure sequences are listed.
• All nonstandard amino acids, which includes B, O, U, X, and Z, (see here for their meanings) are masked with "*" character.
• An additional column (has_nonstd_aa) is added to indicate whether the protein sequence contains nonstandard amino acids.
• A subset of the sequences with low sequence identity and high resolution, ready for training, is also provided

For details of curation, please see https://github.com/zyxue/pdb-secondary-structure.

A subset (9079 sequences) based on sequences culled by PISCES with more strict quality control is also provided. This dataset is considered ready for training models.

The culled subset generated on 2018-05-31 with cutoffs of 25%, 2Å, and 0.25 for sequence identity, resolution and R-factor respectively, is used. The URL to the original culled list is http://dunbrack.fccc.edu/Guoli/culledpdb_hh/cullpdb_pc25_res2.0_R0.25_d180531_chains9099.gz, but it may not be permanently available. This dataset contains more columns from cullpdb_pc25_res2.0_R0.25_d180531_chains9099.gz with self-explanatory names.

For more about PISCES, please see https://academic.oup.com/bioinformatics/article/19/12/1589/258419.

Acknowledgements

The peptide sequence and secondary structure are downloaded from https://cdn.rcsb.org/etl/kabschSander/ss.txt.gz. The culled subset is downloaded from http://dunbrack.fccc.edu/PISCES.php.

Inspiration

Kaggle provides a great platform for sharing ideas and solving data science problem. Sharing a cleaned dataset help prevent others from duplicated work and also provides a common dataset for more comparable benchmark among different methods.

Early attempts on this (or related) problem:

1. Baldi, Pierre, Søren Brunak, Paolo Frasconi, Gianluca Pollastri and Giovanni Soda. “Bidirectional Dynamics for Protein Secondary Structure Prediction.” Sequence Learning (2001). http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.104.7092&rep=rep1&type=pdf
2. Chen, J. and Chaudhari, N. S.. "Protein Secondary Structure Prediction with bidirectional LSTM networks." Paper presented at the meeting of the Post-Conference Workshop on Computational Intelligence Approaches for the Analysis of Bio-data (CI-BIO), Montreal, Canada, 2005. http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.104.7092&rep=rep1&type=pdf (Couldn't find a pdf)
3. Sepp Hochreiter, Martin Heusel, Klaus Obermayer; Fast model-based protein homology detection without alignment, Bioinformatics, Volume 23, Issue 14, 15 July 2007, Pages 1728–1736, https://doi.org/10.1093/bioinformatics/btm247

Context

Protein secondary structure prediction dataset. Used by 2015 NAR paper* from Barton group. There are a total of 1507 protein sequences, each represented by an integer identifier (e.g. 24695). 1348 in the training folder, and the rest in the blind test folder.

For each example, there are the following files: .fasta -> amino acid sequence for that domain .dssp -> ground truth 3-state secondary structures, obtained from PDB 3D crystal structures using the DSSP algorithm .pssm -> PSI-BLAST matrices, obtained from running the PSI-BLAST algorithm on the sequence, which returns both the matrix and a multiple-sequence alignment (MSA) .hmm -> profile HMM matrices, obtained by running the HMMer3 algorithm on the MSA generated from PSI-BLAST

The suggested k for cross validation is 7, such that each fold will have 193 (the last will have 190) protein sequences.

This leads on to the purpose of the third file in this dataset - shuffle.pkl. This file contains the suggested 7-fold split for cross-validation, in the form of a nested list. Random splits were generated until the 3-state secondary structure contents were within 1% of each other, to balance the prediction labels across the 7 folds.

*Alexey Drozdetskiy, Christian Cole, James Procter, Geoffrey J. Barton, JPred4: a protein secondary structure prediction server, Nucleic Acids Research, Volume 43, Issue W1, 1 July 2015, Pages W389–W394, https://doi.org/10.1093/nar/gkv332

ProteinNet is a standardized data set for machine learning of protein structure. It provides protein sequences, structures (secondary and tertiary), multiple sequence alignments (MSAs), position-specific scoring matrices (PSSMs), and standardized training / validation / test splits. ProteinNet builds on the biennial CASP assessments, which carry out blind predictions of recently solved but publicly unavailable protein structures, to provide test sets that push the frontiers of computational methodology. It is organized as a series of data sets, spanning CASP 7 through 12 (covering a ten-year period), to provide a range of data set sizes that enable assessment of new methods in relatively data poor and data rich regimes.

To use this dataset:

import tensorflow_datasets as tfds

ds = tfds.load('protein_net', split='train')
for ex in ds.take(4):
 print(ex)

See the guide for more informations on tensorflow_datasets.

Context and Content

The data is for the paper titled "DeepHelicon: accurate prediction of inter-helical residue contacts in transmembrane protein by residual neural networks". It contains four sub-folders as follows: 1. Fasta: the protein sequences in the TRAIN, PREVIOUS, and TEST datasets, respectively. 2. PDB: the protein native structures in the TRAIN, PREVIOUS, and TEST datasets, respectively. 3. Predictions: the contact predictions on the PREVIOUS and TEST datasets, which are predicted by the contact prediction methods mentioned in the DeepHelicon paper. 4. 3D modelling: the 3D models, which are guided by the secondary structures predicted by SCRATCH1D and guided by the residue contacts predicted by DeepHelicon and DeepMetaPSICOV, respectively, are finally generated by CONFOLD2.

Acknowledgements

sunmt, jianfeng; Frishman, Dmitrij (2020), “Experiment data used in DeepHelicon”, Mendeley Data, V2, doi: 10.17632/k8tfvgftv3.2

Inspiration

Predict the intern-helical residue contacts in transmembrane protein

Context

This dataset contains TM-scores, a common metric for evaluating the structural similarity of proteins or RNA molecules. These scores likely correspond to predictions submitted to the Stanford Ribonanza RNA Folding competition held on Kaggle.

Content

The dataset file(s) presumably contain identifiers (e.g., submission IDs or model identifiers) and their corresponding TM-scores against the ground truth structures from the competition.

(Please refer to the data file(s) within the dataset for exact column names and details.)

Potential Use Cases

• Competition Analysis: Analyze the distribution of scores achieved in the Ribonanza competition.
• Model Performance Comparison: Compare the performance of different RNA folding models based on their TM-scores.
• Benchmarking: Use these scores as a benchmark for evaluating new RNA structure prediction methods.

License

Setting license to CC0-1.0 (Public Domain Dedication) for maximum usability, assuming the underlying competition data allows this. If derived from specific competition data, the original competition rules might apply.

Context

This dataset combines monoclonal antibody (mAB) information from a variety of sources into a more concise and convenient format.

Here is a quick introduction to monoclonal antibodies in context with the COVID-19 pandemic.

Sources

• RCSB PDB - 3D protein models
• SAbDab - pairs of antigens and antibodies from RCSB
• Thera-SAbDab - therapeutic monoclonal antibodies
• CoV-AbDab - COVID-19 related antibodies
• ANARCI - CDR predictions
• DSSP - secondary structure and solubility predictions