Dataset Card for "ProstT5Dataset"

• Contributors: Michael Heinzinger and Konstantin Weissenow, Joaquin Gomez Sanchez and Adrian Henkel, Martin Steinegger and Burkhard Rost
• Licence: MIT

Overview

The ProstT5Dataset is a curated collection of tokenized protein sequences and their corresponding structure sequences (3Di). It is derived from the AlphaFold Protein Structure Database and includes various steps of clustering and quality filtering. To capture 3D information of the sequence, the 3Di structure string representation is leveraged. This format captures the spatial relationship of each residue to its neighbors in 3D space, effectively translating the 3D information of the sequence. The sequence tokens are generated using the ProstT5 Tokenizer.

Data Fields

• input_id_x (3Di Tokens): Corresponding tokenized 3Di structure representation sequences derived from the proteins.
• input_id_y (Amino Acid Tokens): Tokenized amino acid sequences of proteins.

Dataset Description

We compare basic protein properties (sequence length, amino acid composition, 3Di-distribution) between our dataset (training, validation, test sets) and proteins obtained from the Protein Data Bank (PDB). Key findings include similar amino acid distributions across datasets, an overrepresentation of certain 3Di-tokens (d, v, p) and helical structures in AlphaFold2 predictions compared to PDB, and a tendency for shorter protein lengths in this dataset (average 206-238) relative to PDB proteins (average 255). The analysis also highlights the relationship between 3Di states and secondary structures, with a notable distinction in strand-related tokens between datasets.

Data Collection and Annotation

The dataset began with the AlphaFold Protein Structure Database , undergoing a two-step clustering process and one step of quality filtering:

1. First Clustering: 214M UniprotKB protein sequences were clustered using MMseqs2, resulting in 52M clusters based on pairwise sequence identity.
2. Second Clustering: Foldseek further clustered these proteins into 18.8M clusters, expanded to 18.6M proteins by adding diverse members.
3. Quality Filtering: Removed proteins with low pLDDT scores, short lengths, and highly repetitive 3Di-strings. The final training split contains 17M proteins.

Data Splits

Data splits into train, test, and, validation were created by moving whole clusters (after quality filtering - see above), to either of the sets. For validation and test, we only kept representatives to avoid bias towards large families. This resulted in 474 proteins for test, 474 proteins for validation and around 17M proteins for training.

Citation

@article{heinzinger2023prostt5,
  title={ProstT5: Bilingual language model for protein sequence and structure},
  author={Heinzinger, Michael and Weissenow, Konstantin and Sanchez, Joaquin Gomez and Henkel, Adrian and Steinegger, Martin and Rost, Burkhard},
  journal={bioRxiv},
  pages={2023--07},
  year={2023},
  publisher={Cold Spring Harbor Laboratory}
}

Tokens to Character Mapping

Amino Acid Representation	3DI	Special Tokens
3: A	128: a	0: <pad>
4: L	129: l	1: </s>
5: G	130: g	2: <unk>
6: V	131: v	148: <fold2AA>
7: S	132: s	149: <AA2fold>
8: R	133: r
9: E	134: e
10: D	135: d
11: T	136: t
12: I	137: i
13: P	138: p
14: K	139: k
15: F	140: f
16: Q	141: q
17: N	142: n
18: Y	143: y
19: M	144: m
20: H	145: h
21: W	146: w
22: C	147: c
23: X
24: B
25: O
26: U
27: Z