PlasticEnz: New Tool Combines AI and Databases to Find Plastic-Eating Enzymes

Plastic waste has become one of the planet's most persistent environmental problems. Millions of tons end up in landfills, oceans, and soil each year, and only a fraction gets recycled. Scientists have long known that certain microorganisms can break down plastic polymers, but finding those enzymes in the messy reality of environmental samples has been like searching for a needle in a haystack. A team of researchers has now built a tool to make that search far easier.

The tool is called PlasticEnz, and it represents a deliberate attempt to democratize the hunt for plastic-degrading enzymes. The researchers combined two complementary approaches: traditional sequence similarity searching, which looks for genetic matches to known plastic-eating enzymes, and machine learning models trained to recognize patterns in those enzymes. The result is a free, open-source program that can process protein sequences, genome fragments, or entire assembled genomes and flag candidates worth investigating further. It works particularly well for two common enzyme types: PETases, which degrade polyethylene terephthalate (the plastic in soda bottles and clothing), and PHBases, which break down polyhydroxybutyrate, a biodegradable polymer that some bacteria produce naturally.

Building PlasticEnz required careful groundwork. The researchers assembled a database of 213 experimentally validated plastic-degrading enzymes drawn from 176 peer-reviewed studies, each entry annotated with details about where the enzyme comes from, what it targets, and how it works. They then trained machine learning classifiers—specifically XGBoost and a neural network—on this curated data, teaching the models to recognize the signature patterns of real plastic-degrading enzymes versus false positives. The tool offers two modes: a default setting that prioritizes precision and minimizes false alarms, and a sensitive mode that casts a wider net and catches more candidates, even if some turn out to be dead ends.

When the researchers tested PlasticEnz against existing protein annotation tools like KEGG and eggNOG-mapper, the results were striking. For PET-degrading enzymes, PlasticEnz in sensitive mode achieved 98 percent precision and 96 percent recall—meaning it found nearly all the real PETases while making very few false calls. Even in default mode, it matched or exceeded the performance of established tools. For PHB-degrading enzymes, the picture was more complex because these enzymes are more diverse, but PlasticEnz still outperformed the competition.

The real test came when the researchers applied the tool to actual environmental samples. They analyzed biofilm communities from a hypersaline lagoon in Texas that had been exposed to PET and PHB plastics for 28 days, comparing them to control samples grown on ceramic. PlasticEnz correctly identified a strong enrichment of PHB-degrading enzymes in the plastic-exposed biofilms—over 800 putative candidates—while finding few PETases, a pattern that matched what the original researchers had observed. They also screened soil samples from three plastic-contaminated sites in India and Iran against pristine hot spring sediments from a remote volcano in Kamchatka, Russia. The tool consistently detected more and higher-scoring plastic-degrading enzyme candidates in the polluted soils, with the contaminated site of Sewapura showing the strongest signal. The hot spring samples, by contrast, yielded almost nothing—a result that makes biological sense, since those extreme environments are unlikely to harbor enzymes specialized for breaking down synthetic polymers.

The tool is not without limitations. It currently works well only for PET and PHB because those are the only polymers with enough experimentally validated enzymes to train reliable models. Other plastics like polylactic acid and polycaprolactone remain underrepresented in the training data, and the researchers acknowledge that all predictions are still putative until confirmed through laboratory testing. The computational demands also scale with dataset size, though the researchers have built in options to use graphics processors and multiple CPU cores to speed things up. But for researchers without deep bioinformatics expertise, PlasticEnz removes a major barrier to exploring microbial plastic degradation in their own samples. The tool runs in minutes on typical datasets and produces clear, interpretable output. As more plastic-degrading enzymes are discovered and added to the training set, the tool will only improve, expanding to cover additional polymers and making bioremediation strategies more feasible.