Confirmed 3 Nov 2026 · Science A one-minute read

A model designed 603 plastic-eating enzymes, and more than 400 of them worked.

Nature Communications published a method that redesigned PET-digesting enzymes from evolutionary sequence models. The best designs beat the field's benchmark enzyme in 15 of 16 test conditions.

Until now

AI protein design produced single wins, checked one molecule at a time.

Now

It produces hundreds of working variants at once, most beating their parent.

Why it matters

PET is the plastic in bottles. Recycling it with enzymes has been held back by one or two fragile candidates. 400 working variants, carrying up to 68 substitutions each, give engineers a library to choose from for heat, acid and crystallinity. The bottleneck moves from design to plant.

Established

603 designs were made and tested; more than 400 depolymerised PET.
Top designs averaged 34.5 times the parent enzyme's activity.
4 rational mutations raised one scaffold's melting point 18.3 degrees.

Not established

That industrial recycling improves. No plant data is reported.
Cost. The paper reports no production or enzyme cost.
That the method generalises beyond plastic-digesting enzymes.

The five-minute readThe record, the timeline, the weight and 2 sources

The record

Iterative computational and rational design generates diverse PET hydrolases for plastic depolymerization was published open access in Nature Communications, volume 17, article 11371, on November 3, 2026, and we read it at nature.com. First authors are Brenna Norton-Baker, Benjamin Fram and Nicole N. Thadani; the listed authors include Debora S. Marks, John E. McGeehan, Chris Sander, Gregg T. Beckham and Nicholas P. Gauthier of the Dana-Farber Cancer Institute. The method, EVdesign, builds an evolutionary Hamiltonian from EVcouplings and direct coupling analysis, generates sequences by Gibbs sampling, and uses HMMER for alignments and AlphaFold 3 for structures. The numbers. 603 designs were made and tested from 8 parent enzymes; more than 400 variants, carrying 12 to 68 substitutions each, depolymerised PET, and nearly all outperformed their parents. 399 of 513 designs built on the TfCut2 scaffold stayed soluble and active, with a mean of 41.0 mutations. 5 of the 8 parents yielded active designs. TfCut2 designs averaged about 3.1 times parent activity, to a maximum of 14 times, while designs from LCC never beat the LCC parent, peaking at 0.54 times. In the diversity experiment, 184 designs were made and 98 purified, grouped Low 95, Med 30, High 30 and Max 29, with average activity against wild type of 2.33, 1.57, 1.02 and 0.19 times. 4 rational mutations raised TfCut2's thermostability by 18.3 degrees Celsius, its melting point going from 72.1 to 90.4, with a mean rise of 15.2 degrees across 22 designs and a 16-fold activity gain on TfCut2. Of 80 designs built on that stabilised scaffold, all 74 tested were active in at least one condition. Iterative rounds produced 201 designs from 19 parents, 137 in generation 3 and 64 in generation 4, adding 7 to 9 substitutions each, with melting points falling 4.3 degrees on average and about 15% matching their parent's stability. Against the benchmark LCC-ICCG, which melts near 93 degrees, at least one design matched or beat 48-hour product formation in 15 of 16 conditions; the top design per condition averaged 1.48 times LCC-ICCG, reaching 2.2 times on crystalline powder and 3.3 on amorphous film at pH 4.5 and 70 degrees. More than 36 designs beat LCC-ICCG on powder and more than 24 on film. Top designs averaged 34.5 times TfCut2, from 3.7 to 206.5. Two crystal structures matched the original enzyme to 0.30 and 0.35 angstroms. Not established: cost, scale yield, and mixed waste.

How it got here

Oct 22
Hugging Face publishes an open model that reads DNA and answers questions in English. Edition 059. Biological sequence models reached the public the same month this work went to press.
Oct 24
Britain funds a world-first AI biosecurity laboratory, with OpenAI and Anthropic supplying model credits. Edition 060. The same design power, aimed at what it should not build.
Nov 3
Nature Communications publishes 603 tested enzyme designs. More than 400 depolymerise PET; the best beat the LCC-ICCG benchmark in 15 of 16 conditions, and 2 crystal structures confirm the predicted fold.

Who has to react

Enzymatic recycling companies, whose plants are built around single enzymes; Carbios and the holders of LCC-ICCG patents, now benchmarked against and beaten; bottle and packaging producers under recycled-content mandates; the United States Department of Energy, whose renewable energy laboratory co-authored this; protein design groups that validate one molecule at a time; and waste regulators writing recycling targets.

What to watch

Would strengthen itAn independent group reproduces the hit rate on a different enzyme family, or an enzymatic recycling operator adopts one of these variants, by December 31, 2027.
Would weaken itA replication reports an activity hit rate below 20% for EVdesign on any scaffold, or the top designs fail on post-consumer waste, by June 30, 2028.
Would strengthen itThe method is applied to an enzyme family with sparse sequence data and yields fewer than 10% active designs, by December 31, 2027.

Why it weighs this much

Signal strength

CostHigh

CommitmentMedium

ReversibilityHard

PersistenceStructural

Cost is high: 603 designed genes were synthesised, expressed, purified and assayed, with 2 crystal structures solved, which is years of laboratory work and not a computational claim. Commitment is medium, because publishing open access in Nature Communications stakes reputations on numbers others will try to reproduce, but commits no money to a plant. Reversibility is hard: the designs and data are published and cannot be withdrawn. Persistence is structural, because a validated library of 400 working variants becomes the baseline every later PET enzyme is measured against.

The words

PET: Polyethylene terephthalate, the plastic in drinks bottles and polyester fibre. It resists water, which is why breaking it down needs a strong enzyme.
Hydrolase: An enzyme that splits a chemical bond using water. A PET hydrolase cuts the polymer back into the units it was made from.
Thermostability: How much heat an enzyme survives. It matters because PET softens when hot, and a hot reaction digests it far faster.
Melting temperature: The temperature at which half a protein sample unfolds. Raising it by 18.3 degrees turns a fragile enzyme into a usable one.

Sources

Your call

Which explanation do you believe?

A call is final. Other people's calls show after yours.