PETase ANnotation and Triage System

Live instrumentation. Everything this project has read, produced and is running on, read straight from the database and the machine.

Nothing here is written by hand. Each figure is a query or a reading, and the page refreshes when the data actually changes.

updated 2026-09-20 23:22:27 UTC · database last written 62801.6 min ago

The funnel

How 14,804,920 sequences become a shortlist, and what that shortlist is judged against. Each of the first steps is a filter, and the one that matters most is the last of them: three catalytic residues can be present in a sequence and still not touch each other in the folded protein, and only a structure can tell you which.

Sequences read

14,804,920

every predicted protein in the assemblies

Survived recall

439

matched a polyesterase profile and carry a complete triad

Given a structure

416

folded with ESMFold, superposed onto IsPETase

Triad confirmed in 3D

403

the three residues actually meet in space

Too long to fold

24

over 450 residues; deferred, not discarded

Reference structures

559

the known enzymes a candidate is compared against, not part of the catch

Measured positives

342

reference enzymes whose activity someone actually assayed and published

Activity measurements

2759

individual published values behind those enzymes, each with its citation

Unlabelled sequences

25,041

homologues nobody has assayed, held apart from the catalogue on purpose so they cannot inflate a count that is quoted as evidence

Sequence clusters

170

groups of related sequences carrying their own ecology — where the source organism was found, and at what temperature

Registered sources

7

public databases and published screens this catalogue is built from, each with its retrieval date and licence

Where the sequences came from

Metagenomics reads the DNA of a whole microbial community straight from the environment, without culturing anything. Each row is one such environment, chosen because it has had decades of plastic exposure and every reason to have evolved a use for it. Yield per million is the comparable figure: raw candidate counts only tell you which environment was sequenced most deeply.

Protein structures

A structure is either measured or predicted, and the difference matters more than any single number on this page. An X-ray crystal structure is experimental evidence. AlphaFold and ESMFold produce a hypothesis — usually a good one for the fold, least reliable exactly where this project looks, in the flexible loops that gate the active site.

What the model learns from

The distinction this whole project turns on. Measured means a published experiment stands behind the label — someone ran the assay. Annotated only means a computer assigned it from sequence similarity and nobody measured anything; a model trained on those is largely rediscovering the rule that produced them, which is why the two are never pooled here.

Evidence and citations

Every measured value in this database points at the paper that reported it. 2757 of 2759 measurements carry a DOI or PubMed identifier, drawn from 38 distinct sources.

Models and software

Recorded per pipeline run rather than declared here, so this is what actually produced the data rather than what a README says it did.

External databases

Public resources this project reads from, with the date each was last retrieved. Nothing is scraped continuously: a source is pulled when a stage runs, and the date below is when that last happened.

Pipeline activity

Every stage writes a record of what went in, what came out, which tool versions ran and how long it took. 124 runs and 121 manifests so far, totalling 32.2 hours of recorded compute.

Model evaluations

Every training run ever recorded, including the ones that failed. AUC is the chance the model ranks a true PET enzyme above a non-enzyme: 1.0 is perfect, 0.5 is a coin flip. The composition baseline beside it is what you get from amino-acid counts alone — a score only means something if it clears that.

The machine this runs on

A single DigitalOcean droplet shared with five other applications, which is why the web layer carries no scientific libraries at all: every structure and every measurement was computed offline and is served here as a precomputed file.

Linux x86_64 · Python 3.12.3 · web process up 6.67 h · host up 1635.9 h · data version 0.3.0