adrianhenkel commited on
Commit
522317c
1 Parent(s): 3144661

Adds Dataset Card for the ProstT5Dataset

Browse files

Hi

@mheinz

,
when you find the time, please review the content of this card and check which licence is suitable.

Files changed (1) hide show
  1. README.md +95 -1
README.md CHANGED
@@ -17,7 +17,101 @@ dataset_info:
17
  num_examples: 17070828
18
  download_size: 810671738
19
  dataset_size: 65394099456
 
 
 
 
 
 
 
20
  ---
21
  # Dataset Card for "ProstT5Dataset"
22
 
23
- ## ⚠️ This is work in progress! ⚠️
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
17
  num_examples: 17070828
18
  download_size: 810671738
19
  dataset_size: 65394099456
20
+ license: cc
21
+ task_categories:
22
+ - text-generation
23
+ tags:
24
+ - biology
25
+ size_categories:
26
+ - 10M<n<100M
27
  ---
28
  # Dataset Card for "ProstT5Dataset"
29
 
30
+ # Currently under review
31
+
32
+ * **Contributors:** Michael Heinzinger and Konstantin Weissenow, Joaquin Gomez Sanchez and Adrian Henkel, Martin Steinegger and Burkhard Rost
33
+ * **Licence:** TBD
34
+
35
+ ## Table of Contents
36
+ - [Overview](#overview)
37
+ - [Dataset Description](#dataset-description)
38
+ - [Data Collection and Annotation](#data-collection-and-annotation)
39
+ - [Data Splits](#data-splits)
40
+ - [Dataset Structure](#dataset-structure)
41
+ - [Data Fields](#data-fields)
42
+ - [Data Instances](#data-instances)
43
+ - [Data Considerations](#data-considerations)
44
+ - [Social Impact of Dataset](#social-impact-of-dataset)
45
+ - [Discussion of Biases](#discussion-of-biases)
46
+ - [Other Known Limitations](#other-known-limitations)
47
+ - [Licensing Information](#licensing-information)
48
+ - [Citation Information](#citation-information)
49
+ - [Contributions](#contributions)
50
+
51
+ ## Overview
52
+ The ProstT5Dataset is a curated collection of *tokenized* protein sequences and their corresponding structure sequences (3Di).
53
+ It is derived from the [AlphaFold Protein Structure Database](https://alphafold.ebi.ac.uk/) and includes various steps of clustering and quality filtering.
54
+ To capture 3D information of the sequence, the [3Di structure string representation](https://www.nature.com/articles/s41587-023-01773-0#Sec2) is leveraged. This format
55
+ captures the spatial relationship of each residue to its neighbors in 3D space, effectively translating the 3D information of the sequence.
56
+ The sequence tokens are generated using the [ProstT5 Tokenizer](https://huggingface.co/Rostlab/ProstT5).
57
+
58
+ ## Data Fields
59
+ - **input_id_x** (3Di Tokens): Corresponding tokenized 3Di structure representation sequences derived from the proteins.
60
+ - **input_id_y** (Amino Acid Tokens): Tokenized amino acid sequences of proteins.
61
+
62
+ ## Dataset Description
63
+
64
+ ![image/png](https://cdn-uploads.huggingface.co/production/uploads/62c412251f45e8bdb2b05855/BgiKOoFUGjlHDPjbxJWOX.png)
65
+ We compare basic protein properties (sequence length, amino acid composition, 3Di-distribution) between our
66
+ dataset (training, validation, test sets) and proteins obtained from the [Protein Data Bank (PDB)](https://www.rcsb.org/). Key findings include similar amino acid distributions across datasets,
67
+ an overrepresentation of certain 3Di-tokens (d, v, p) and helical structures in AlphaFold2 predictions compared to PDB, and a tendency for shorter protein
68
+ lengths in this dataset (average 206-238) relative to PDB proteins (average 255). The analysis also highlights the relationship between
69
+ 3Di states and secondary structures, with a notable distinction in strand-related tokens between datasets.
70
+
71
+ ## Data Collection and Annotation
72
+ The dataset began with the AlphaFold Protein Structure Database , undergoing a two-step clustering process and one step of quality filtering:
73
+ 1. *First Clustering:* 214M UniprotKB protein sequences were clustered using MMseqs2, resulting in 52M clusters based on pairwise sequence identity.
74
+ 2. *Second Clustering:* Foldseek further clustered these proteins into 18.8M clusters, expanded to 18.6M proteins by adding diverse members.
75
+ 3. *Quality Filtering:* Removed proteins with low pLDDT scores, short lengths, and highly repetitive 3Di-strings. The final training split contains 17M proteins.
76
+
77
+ ## Citation
78
+ ```
79
+ @article{heinzinger2023prostt5,
80
+ title={ProstT5: Bilingual language model for protein sequence and structure},
81
+ author={Heinzinger, Michael and Weissenow, Konstantin and Sanchez, Joaquin Gomez and Henkel, Adrian and Steinegger, Martin and Rost, Burkhard},
82
+ journal={bioRxiv},
83
+ pages={2023--07},
84
+ year={2023},
85
+ publisher={Cold Spring Harbor Laboratory}
86
+ }
87
+ ```
88
+
89
+ ## Tokens to Character Mapping
90
+ | Amino Acid Representation | 3DI | Special Tokens |
91
+ |---------------------------|-----------|--------------------|
92
+ | 3: A | 128: a | 0: \<pad\> |
93
+ | 4: L | 129: l | 1: \</s\> |
94
+ | 5: G | 130: g | 2: \<unk\> |
95
+ | 6: V | 131: v | 148: \<fold2AA\> |
96
+ | 7: S | 132: s | 149: \<AA2fold\> |
97
+ | 8: R | 133: r | |
98
+ | 9: E | 134: e | |
99
+ | 10: D | 135: d | |
100
+ | 11: T | 136: t | |
101
+ | 12: I | 137: i | |
102
+ | 13: P | 138: p | |
103
+ | 14: K | 139: k | |
104
+ | 15: F | 140: f | |
105
+ | 16: Q | 141: q | |
106
+ | 17: N | 142: n | |
107
+ | 18: Y | 143: y | |
108
+ | 19: M | 144: m | |
109
+ | 20: H | 145: h | |
110
+ | 21: W | 146: w | |
111
+ | 22: C | 147: c | |
112
+ | 23: X | | |
113
+ | 24: B | | |
114
+ | 25: O | | |
115
+ | 26: U | | |
116
+ | 27: Z | | |
117
+