Life sciencesPreprintData analysis4 min read

46,000 ENZYMES THAT COULD BREAK OUR ANTIBIOTICS

Penicillin and its relatives, the β-lactam antibiotics, all carry the same chemical ring. Bacteria fight back with enzymes, β-lactamases, that snip that ring open and leave the drug useless. Nearly 1,800 of these enzymes have been characterised in the laboratory and gathered in reference databases.

But bacteria live everywhere, and so do their enzymes. Liangjun Ge, Zhiwei Qin, Heqian Zhang and colleagues, at Beijing Normal University’s Zhuhai campus, went looking for the ones nobody has described.

Teaching a model what the family looks like

The problem is similarity. The usual search tools find sequences that resemble known enzymes. An enzyme that does the same job with a very different sequence slips through.

So the team turned to a protein language model, a network trained on vast numbers of protein sequences, which represents a protein by what it has learned about proteins in general rather than by direct resemblance. They fed its output to five different classifiers — three neural networks and two classic machine-learning methods — and had them vote. Trained on 1,778 experimentally characterised β-lactamases, the combination classified every validation sequence correctly.

They then let it loose on MGnify, a database of proteins predicted from environmental DNA, after a first filter by classic profile searches and by length. More than 46,000 candidate β-lactamases came out, spanning all four classes of the standard classification.

A map that does not match the clinic

Where the samples carry coordinates, the candidates span the world, though unevenly: 14,099 from North America, 11,248 from Asia, 10,665 from Europe, with far fewer from Africa, South America and Oceania — partly real, partly a matter of who collects samples.

The composition is the surprise. Among the 1,778 characterised reference enzymes, class A dominates. In the environment, class B — the ones that need zinc to work — make up 34.0 to 44.5% everywhere. Habitats differ too: wastewater and activated sludge are rich in classes B and C, while human faecal samples are dominated by class A (48.4%) — a composition the authors link to the clinical attention that class has received.

Two candidates, tested for real

Predictions are cheap; the test is not. The team picked two class C candidates from environments under different antibiotic pressure, sharing only about 30% of their sequence with their closest characterised relatives. Both were put into Escherichia coli, alongside a known enzyme as a positive control and an empty vector as a blank.

The blank strain showed no appreciable resistance. Both candidates did protect the bacterium. The first pushed the minimum concentration needed to stop the bacteria above 256 micrograms per millilitre for penicillin, oxacillin and ampicillin; the second reached 512 micrograms per millilitre for penicillin. These are real resistance enzymes, not statistical artefacts.

One letter away from the carbapenems

The team then asked a different question: how far could these enzymes still evolve? Two more models — one that proposes sequences compatible with a known 3D structure, one that scores how natural a substitution looks — jointly suggested 31 promising sites in the first enzyme and 42 in the second. Eight were chosen for testing; seven were successfully built.

All seven changed the resistance profile. Some lost ground. But one substitution in the first enzyme, a single amino acid replaced at position 106, broadened its resistance and added a new one: against carbapenems, a family the known reference enzyme in the same experiment barely resisted. In the second enzyme, another single change pushed resistance to a cephalosporin and to a carbapenem to 512 micrograms per millilitre. Simulations suggest these substitutions hold the antibiotic closer to the enzyme’s cutting site.

Sewage versus ocean

Finally, the team ran the same site prediction on four representative environmental enzymes, one per class, and compared the 97 sites it flagged with natural sequences. At those sites, natural variation was higher than at other positions: these are tolerant spots, not the untouchable core. And in all four classes, enzymes from wastewater varied more than those from the ocean — consistent with sewage mixing drugs, pollutants and microbes from homes, hospitals and industry.

The authors are careful about what this does not prove. Only two enzymes were tested, by resistance measurements alone; the mutations were tested one at a time; and more sequence variation is not evidence of selection, nor does a mutation that works in a lab strain tell us it would spread in the wild. What the work does show is that the catalogue of resistance enzymes is far from complete, and that one letter can be enough to turn one of them against a whole new family of antibiotics.