petadex/petai-activity-labels

PETai Activity Labels

12,888 measured activity and stability labels on 1,219 plastic-degrading enzyme sequences, curated from 649 published references plus one in-house halo screen. Built for iGEM Toronto 2026 (project PetaBite) as the supervised label set for the PETadex corpus of 307,155,746 catalytic ORFs.

PETadex supplies candidates. This supplies the targets. No activity label exists anywhere in the PETadex release, so the two are published separately and joined on sequence MD5.

v2, and what changed

This version replaces the first release (13,092 labels, published 2026-10-15). Two changes, one of which matters a great deal:

• The split is re-stratified on negatives. In v1 the test fold held 7 negative labels on 5 sequences and the in-house screen contributed 0 of its 168 negatives to it, which made every classification metric computed on that fold uninterpretable. Negatives are now allocated 124 / 39 / 33 labels across train / val / test, on 109 / 39 / 32 sequences. Label provenance is also no longer inverted: the LLM share of labels is 39.5% in train against 39.9% in test, where v1 ran 19.6% against 62.9%.
• 204 duplicate LLM rows were removed, so the label count falls from 13,092 to 12,888. The sequence set, the clustering and the cluster assignments are byte identical to v1.

Every label now also carries source_repo and source_row_id, which resolve to a single row in the source dataset it came from.

The four sources, published separately

Source	Labels	Sequences	What it is
petai-labels-human	6,655	458	Hand-read from 59 papers by 19 curators, with assay conditions
petai-labels-llm	4,694	455	LLM extraction from 572 papers, evidence sentence on every row
petai-labels-kinetics	1,342	206	kcat, Km and kcat/Km compiled from 74 papers
petai-halo-screen	197	197	In-house BHET plate screen, 168 of the 200 negatives

Each source repo ships its own native tables rather than a slice of the merged schema, so anything the merge drops, including assay conditions, evidence sentences, raw value strings and plate-level data, is recoverable there.

Configs

Config	Rows	Description
labels	12,888	One measurement per row, keyed by seq_md5
sequences	1,219	One row per sequence, with cluster and split assignment
split	1,137	The 30%-identity fold assignment for the clean sequences
clusters	1,219	Cluster membership at 90%, 50% and 30% identity

sequences.fasta holds all 1,219 sequences, named by MD5. summary.json is the release summary emitted by the build. The four CSVs are published beside the parquet for anyone who would rather not use datasets.

labels schema

Column	Type	Notes
seq_md5	string	Join key, MD5 of the uppercase sequence
source	string	activity 6,655, llm 4,694, kinetics 1,342, halo 197
task	string	quantitative 12,691, binary 197
variable	string	What was measured, 99.95% present
unit	string	See the unit warning below
substrate	string	92.03% present. PET 4,182, pNPB 1,078, BHET 817
value	float	Not one target. See below
is_negative	bool	200 of 12,888 (1.55%)
ph	float	54.10% present
temperature_c	float	30.21% present
time_h	float	10.15% present
ref	string	650 values: 649 DOIs plus halo:BHET25@120h
curator	string	89.59% present
source_repo	string	Which of the four datasets above the row came from
source_row_id	string	That dataset's own row id

value is not a single regression target

The column mixes incompatible units and spans -372.91 to 1.21e9 as a result. Do not regress on it directly. The three collisions that matter:

• Relative activity is split across % (1,352), percent (865) and % relative activity (369). Same quantity, three spellings.
• Melting data, 888 labels in all, mixes fraction of folded protein (751), which are CD spectroscopy curve points rather than a melting temperature, with temperature units on 137. Only those 137 are Tm values.
• Kinetics are three separate targets, not one: kcat (556), Km (551) and kcat/Km (473).

Clean target choices, both well populated: binary halo activity for classification, and PET monomer release for regression.

Read this before training on the split

The split is now usable, with two limits worth knowing.

Fold	Sequences	Labels	30% clusters	Mean cluster size	Largest cluster	Negative labels
train	795	6,888	141	5.64	124	124
val	171	1,445	27	6.33	113	39
test	171	1,418	28	6.11	104	33

• One cluster dominates each held-out fold. 113 of val's 171 sequences and 104 of test's 171 sit in a single 30%-identity cluster, so the top three clusters hold 76.6% of val and 73.7% of test. Effective evaluation diversity is nearer 27 families than 171 sequences, and bootstrap confidence intervals must resample clusters rather than sequences. A per-cluster breakdown of any headline metric is worth more than the headline.
• The split is homology clean. It is a 30%-identity partition and no cluster straddles folds. The leakage check passes.
• 82 sequences carry no fold assignment, described under Fragments.

Conditions are trainable but unavailable at inference

pH is on 54.10% of labels and temperature on 30.21%, with both on 3,398 (26.37%). This is far better coverage than the PETadex corpus, which has melting_point and thermostability null on all 307M rows. The conditions come almost entirely from the human-curated and kinetics sources; the LLM source has none.

The asymmetry is the catch. A PETadex candidate mined from Logan has no measured pH or temperature, so a model conditioned on them at training time has nothing to condition on at inference. Condition dropout during training is the standard fix.

Overlap with the PETadex corpus is small

Joining on exact sequence MD5 against all 131 orfs shards of petadex/petadex-corpus-full, 401 of 1,219 labelled sequences (32.90%) appear in the corpus at all. Of those, 327 come from NCBI nr, 109 from PAZy and 21 from Logan. The sequence set is unchanged from v1, so these numbers still hold.

Two consequences. Corpus-derived columns cannot be model features, since they exist for only about a quarter of the train sequences. And the 21 Logan hits are the only direct evidence of how a model trained here behaves on the population it is meant to screen.

This is exact-MD5 overlap and the label set is full of engineered point variants, so it understates true homology coverage. An MMseqs2 search against the 22,235 30%-identity centroids is the number that actually bears on transfer, and is not yet run.

Fragments

82 of the 1,219 sequences are fragments or contained duplicates and carry no split assignment. They hold 3,137 labels (24.3% of the release).

73 of those 82 name a contained_in parent that is already in a fold, so 3,101 of the 3,137 labels can be recovered by remapping onto the parent. Only 9 sequences with 36 labels are genuinely unrecoverable. The remap is not applied in this version.

Provenance and limitations

• 8,194 labels are human-curated, 4,694 are LLM-extracted from published PDFs. The LLM subset is unvalidated against a human-labelled gold set and contains zero negatives, which is evidence of extraction bias rather than of biology. Filter on source != "llm" for an evaluation set that rests only on human reading.
• The halo screen is in-house iGEM Toronto wet lab data under the condition BHET25 @ 120h, active at z>3.0. It is a BHET assay, not a PET polymer assay, so its negatives mean "no BHET hydrolysis detected" rather than "not a PETase", and heterologous expression failure is indistinguishable from catalytic failure.
• Values are transcribed from assays run under widely differing protocols. Cross-reference comparisons are not controlled.
• value has no uncertainty estimate. The human-curated source carries standard deviations on 252 of its measurements; the merge drops them.

Licensing

CC BY 4.0, which covers what is ours to license: the compilation, the curation, the merge, the clustering, the split, and the in-house halo measurements. 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 649 cited references. Attribution to this dataset is not attribution to them. Cite the DOI in ref when you use a value.
• Those papers' text, tables and figures stay with their publishers. No source PDF is redistributed here or in any of the four source datasets.
• petai-labels-llm carries one verbatim sentence per row, quoted from the source paper so the extraction can be checked. Those quotations are not covered by this grant. They are not merged into this dataset; see that dataset's card.
• Sequences resolved against EBI, UniProt, RefSeq, GenBank and RCSB carry those services' own terms. UniProt is CC BY 4.0.

If you only need the parts that are unambiguously ours to give, that is the halo screen plus the compilation, clustering and split.

Citation

@misc{petai_activity_labels_2026,
  title  = {PETai Activity Labels},
  author = {{iGEM Toronto 2026}},
  year   = {2026},
  version = {2},
  url    = {https://huggingface.co/datasets/petadex/petai-activity-labels}
}

v1 built 2026-10-14, published 2026-10-15. v2 built 2026-10-18. See igem-toronto issue 116.