petadex/petadex-corpus-full

PETadex corpus: full

Every ORF in the PETadex v1.1 catalytic corpus, 307,155,746 putative plastic-degrading enzyme ORFs (PAZy 211, NCBI nr 4,723,128, Logan/SRA 302,432,407), with every annotation that could be keyed to them. Built 2026-10-15.

The companion petadex/petadex-corpus-centroids90 has one row per 90%-identity cluster (18,172,960 rows) and is the place to start for most uses.

Not included, by design: protein-language-model embeddings (and anything derived from them, such as UMAP coordinates), and activity data (kinetics, curated activity labels). Activity labels are a separate PETai dataset.

Layout

All Parquet (zstd). Every table is a config: load_dataset("petadex/petadex-corpus-full", "signalp6"). For joins over the full corpus, querying the files directly with DuckDB or Polars is much faster than datasets.

orfs/: the spine, one row per ORF. Rows are not globally sorted by orf_id: orf_id ranges overlap across row groups and files, so sort explicitly if you need order. orf_id, origin (PAZy / NR / Logan), accession (NR/PAZy), library_id (SRA run, Logan), contig, orf_start, orf_end, orf_type, length, has_stop, md5, c90_id, c60_id, c30_id, is_c90_centroid, is_c60_centroid, structure_path, sequence.

is_c60_centroid is true for the 1,814,296 centroids of the 60% clusters (dims/clusters60.centroid_orf_id); every 60% centroid is also a 90% centroid. structure_path is the S3 URI of the predicted structure for 1,813,242 of them (gzip mmCIF, pLDDT in the B-factor column), from s3://petadex-protein-structures/petadex-v1.1-60pid-centroids-structures/: ESMFold2 (Full under 150 aa, Fast 150-2,500 aa) or, where the centroid's exact sequence is in AlphaFold DB, the AFDB v4 model. It is null for every other ORF and for the 1,054 centroids longer than 2,500 aa. That folder's README.md and structure_index.tsv give the model and metrics per structure (iGEM Toronto 2026, issue #6).

md5 is the MD5 of the uppercase sequence with a trailing stop (*) removed; has_stop records the stop. md5(sequence) holds row by row.

layers/: annotations keyed on orf_id (an ORF can have several rows).

Layer	Rows	ORFs covered	Contents
afdb_structures	308,394	0.1%	AlphaFold DB structure pointers (S3 keys, UniProt) for ORFs whose NR accession matched an AFDB entry. This is an accession match, not a sequence match: the AFDB model can be of a related sequence (e.g. 27 of the 19,430 60% centroids with a key differ from their AFDB entry)
biochem_full	307,155,746	100%	MW, pI, GRAVY, instability index, aromaticity of the full ORF
biochem_mature	42,522,458	13.8%	The same properties for the mature sequence, signal-peptide ORFs only
catalytic_domains	307,155,746	100%	PETadex HMM catalytic domain call per ORF: HMM, coordinates, catalytic residues
catalytic_domains_v260701	8,246,431	2.15%	Domain calls from the July 2026 rebuild: E-value, component, complete_catalytic_hit
family_v1	61,900	0.02%	Legacy 64,730-family membership (family representative = trimmed domain)
interproscan_family_reps	262,260	0.02%	InterProScan hits on the family representatives (coordinates are domain-relative)
nr_taxonomy	2,653,237	0.86%	Organism and lineage for NR-origin ORFs
plasticome	522	0%	Plasticome (known plastic-degrading enzymes) membership
signalp6	307,155,746	100%	SignalP 6 class probabilities and cleavage site

A missing row in a subset layer means not annotated, not absent.

dims/: tables keyed on something else, reached from the spine.

Table	Rows	Contents
biosamples	6,796,722	BioSample annotation: metadata profile, tier, environment, geography, temperature, pH
cluster90_profile_distribution	24,236,640	Profile counts per 90% cluster
cluster90_profiles	18,172,960	Dominant metadata profile per 90% cluster
clusters30	22,235	30% cluster table
clusters60	1,814,296	60% cluster table
clusters90	18,172,960	90% cluster table: centroid, size, composition
component_cath	42	Component to CATH lookup
hmms	45	PETadex HMMs and their catalytic positions
plasticome_reference	57	Plasticome reference residues per component
profiles	22	Metadata profile summary
runs	6,213,292	SRA run metadata for the runs that contribute Logan ORFs

Joins: orfs.library_id → runs.acc, runs.biosample → biosamples.biosample, orfs.c90_id → clusters90.cluster_id (and cluster90_profiles.pid90), c60_id → clusters60.cluster_id, c30_id → clusters30.cluster_id.

import duckdb
duckdb.sql("""
  select o.origin, s.prediction, count(*)
  from 'orfs/*.parquet' o join 'layers/signalp6.parquet' s using (orf_id)
  group by all""")

Things to know

• Redundancy. Only 123,966,734 ORFs (40.4%) are distinct sequences. Deduplicate on md5 before per-sequence work.
• Logan ORFs are called stop-to-stop: only ~3% start with M. N-terminal annotations (SignalP, biochem_mature) are unreliable for Logan rows.
• orf_type (Logan) is inherited and undocumented: types 0/2 end in a stop; 1/3 never do and are cut off by the contig end.
• HMM versions: catalytic_domains names _260407 models; dims/hmms lists _260507 models.
• Family layer / InterProScan coordinates are relative to the trimmed family-representative domain, not the ORF.
• centroid_domain_name in the cluster tables is known to be wrong for some representatives; prefer the HMM calls.
• Layers outside the orf_id key space were joined by sequence MD5, then accession (bridge_method). 44.5% of the July 2026 rebuild's NR proteins postdate this corpus and are not here.

Provenance

Corpus FASTA s3://petadex/logan/petadex.catalytic_orfs.v1.1.fa; PETadex RDS tables (clustering, catalytic domains, origins, metadata); PETadex v260701 rebuild; SignalP 6 and biochemical-property runs. Build code and checks: petadex-corpus-dataset (stages 01–06, notebook PETadex_Corpus_Dataset.ipynb).