petadex/petai-labels-human

PETai Labels, Human Curated

6,655 activity and stability labels on 458 plastic-degrading enzyme sequences, read out of 59 published papers by hand by 19 iGEM Toronto 2026 curators. This is the activity source of petadex/petai-activity-labels, published separately because the merged schema throws away most of what the curators recorded.

Every source_row_id in the unified release that begins with a DOI resolves to exactly one row of labels.csv here.

What makes this set worth the trouble

It is the only one of the four label sources where a person read the paper. Assay conditions survive: pH on 90.4% of labels, temperature on 42.8%, incubation time on 16.7%, plus buffer, substrate pretreatment, crystallinity, expression host and enzyme loading in the native tables. It also carries the only honest negatives of literature origin in the whole release, and 26 of the 65 papers were curated independently by two or more people, so inter-curator agreement is measurable.

Files

File	Rows	What it is
labels.csv	6,655	Long format, the shared schema the unified release merges
measurements.csv	9,357	One row per measurement, every condition column the curators filled
experiments.csv	520	One row per (paper, curator, experiment): design, methodology, results
enzymes.csv	783	Enzyme as named in a paper, with accessions and mutant status
sequences.csv	534	One row per distinct sequence, keyed by seq_md5
papers.csv	93	One row per (paper, curator). 65 distinct papers, 4 curation rounds
sequences.fasta	534	The sequences, named by MD5
summary.json		Build counts, link statistics and dropped rows

Join everything on seq_md5 (MD5 of the uppercase amino acid sequence) and doi.

labels.csv schema

Column	Type	Notes
source_row_id	string	doi|scraper_id|experiment_id|measurement_id
seq_md5	string	Join key
task	string	quantitative on all 6,655 rows
variable	string	What was measured, 79 distinct strings
unit	string	Not harmonised, see below
substrate	string	PET 2,647, pNPB 1,067, BHET 331, blank 1,027
value	float	Target, mixed units
is_negative	bool	32 labels on 16 sequences
ph	float	90.4% present
temperature_c	float	42.8% present
time_h	float	16.7% present
ref	string	DOI, 59 distinct
curator	string	Pseudonym S01 to S19

From 9,357 measurements to 6,655 labels

A measurement becomes a label when it has a sequence and a usable value, meaning value_numeric is present or value_not_detected is true. 2,386 measurements carry no sequence, and a further 316 have a sequence but no numeric value and no not-detected flag. Use measurements.csv directly if you want the unusable rows, for example to see which papers still need work.

seq_link in measurements.csv records how each row was tied to a sequence: expressed_sequence 5,053, name matched against that curator's own enzyme rows 1,918 including 454 matched to a co-curator's sequence, 177 left unlinked because the name matched more than one sequence, and unlinked 2,209. The 32 rows tagged exact_cross_curator_ambiguous are curators disagreeing about what an enzyme name means, which is a curation question and is not settled by picking one.

Human only, on purpose

The curation effort existed to get human-read data, so LLM-submitted rows are excluded: 61 measurement, 33 experiment, 30 enzyme and 8 paper rows. One paper, 10.3389/fmicb.2019.02187, was only ever machine-extracted and drops out entirely. Machine extractions over a much wider corpus live in petadex/petai-labels-llm.

Read before training

• Units are not harmonised. variable mixes released monomer concentration, specific activity, relative activity in percent, weight loss, kcat and melting temperature, and 79 variable strings and 59 unit strings carry no shared vocabulary: relative activity alone is spelled % (1,146) and % relative activity (369). Filter to one variable plus unit pair, or model them as separate tasks.
• Negatives are nearly absent. 32 not-detected labels against 6,623 numeric ones. A binary task built on this table alone is effectively single class. The halo screen in petadex/petai-halo-screen supplies 168 more.
• Collapse the duplicate sequences before splitting. 33 of 534 sequences are a strict substring of another, usually the same protein entered once with and once without its signal peptide, and 36 are flagged is_fragment. contained_in names the containing sequence.
• A sequence is not a unit of independent evidence. 14 sequences carry measurements from more than one paper and 32 come from more than one curator.
• Curation is uneven. The most productive curator contributed 1,475 measurement rows, the least 25, median 435.
• Substrate is free text. pNPB and pNP-C2 through pNP-C16 are soluble surrogate esters, not PET. Keep them apart from PET hydrolysis.

Recovered sequences

148 of the 534 sequences were never typed in by a curator, and 19 more arrived both ways, entered by one curator and recovered for another. 367 are purely curator typed. Filter sequences.csv on sequence_source == 'as_entered' to keep those 367 alone.

sequence_source records how each was recovered, so the recovery can be audited or switched off. Counted over the 783 rows of enzymes.csv: derived_from_wildtype 91 and derived_from_fetched_wildtype 18 apply the mutation string in an enzyme name to a wild type, searching for the numbering offset that makes every stated wild-type residue match and refusing to guess when several offsets fit; fetched_* 57 resolve an accession the curator recorded against EBI Dbfetch, UniProt or RCSB; translated_nt 17 translate a nucleotide entry that is a complete ORF.

One caution: an accession usually holds the precursor while curators often typed the mature protein, so fetched entries sit next to mature-form entries under a different MD5. On the 45 rows with both, 34 fetched sequences are byte identical to what the curator typed; LCC is 266 aa as entered against 293 from HQ704839.

Biggest remaining gap

Panel papers whose enzymes were entered as internal IDs. 10.1038/s41467-022-35237-x has 1,377 measurements and 19 named enzymes with no sequence against 1 with, because the panel was entered as bare numbers with N/A for the sequence. The sequences are in that paper's supplementary tables and need transcribing. 10.1007/s00253-018-9374-x, 10.1021/acs.est.4c01495, 10.1016/j.ijbiomac.2026.150252 and 10.1111/1462-2920.16516 are the same problem.

Curator identity

scraper_id is a stable pseudonym. Real names are not published. IDs are stable across builds: a curator keeps the ID they had and new curators are appended. source_round is the curation round, R1 to R3 or rescrape.

Licensing

CC BY 4.0, which covers what is ours to license: the compilation, the curation, the derived and recovered columns, and the summary text written by the curators in summary, methodology, results and comments. CC BY 4.0 also licenses the database right in the compilation where that right exists.

What it does not cover:

• The measured values are facts reported in the cited papers. Attribution to this dataset is not attribution to them. Cite the DOI in ref when you use a value.
• The papers' own text, tables and figures stay with their publishers. No source PDF is redistributed here.
• Sequences fetched from EBI, UniProt, RefSeq and RCSB carry those services' own terms. UniProt is CC BY 4.0.

Citation

@misc{petai_labels_human_2026,
  title  = {PETai Labels, Human Curated},
  author = {{iGEM Toronto 2026}},
  year   = {2026},
  url    = {https://huggingface.co/datasets/petadex/petai-labels-human}
}

Built from the Human-scraped CSVs-20261013 snapshot by build_activity_labels.py and staged by build_source_repos.py.