Executive Summary

A paper posted to arXiv reports 46,335 candidate enzymes that break antibiotics, gathered from the genetic material in wastewater and seawater. A team at Beijing Normal University tied five classifiers to the protein language model ESM2 and ran the ensemble across a public database of metagenome proteins. The number of beta-lactamases characterized in the laboratory so far stands at 1,778, so the order of magnitude changes.

The distance matters more than the count. Two candidates sharing only about 30% of their sequence with any known enzyme went into E. coli, and with one of them the bacteria grew even past 512 μg/mL of penicillin, the top of the measured range. A working resistance enzyme sat in a place that searching for similar sequences could not reach. Only two candidates, both class C, were tested on the bench. The rest remain predictions.

Sections 1 through 4 are what the paper reports. Section 5 is our reading.

Key figures

Four numbers carry this study: the scale of the candidate set, the distance from the known enzymes, the evidence that an enzyme at that distance still worked, and what changing a single letter did.

Source: Ge et al., arXiv:2611.14568, main text and figure data.

46,335

Beta-lactamase candidates mined

Spread across all four Ambler classes. The enzymes characterized in the laboratory numbered 1,778

30.6%

Sequence Prot22 shares with a known protein

Prot94 sits at 27.4%. That is a distance a search for similar sequences struggles to reach

>512 μg/mL

Penicillin tolerated by E. coli carrying Prot94

The empty-vector control stood at 8. The value ran past the top of the measured range, so the ceiling is unknown

16 μg/mL

Imipenem resistance after one substitution

Wild-type Prot22 read 4. The same change lifted meropenem from under 0.5 to 2

1

The enzymes a similarity search kept missing

From penicillin to carbapenem, the family of antibiotics used most widely in the clinic goes by the name beta-lactam. Every one of these drugs carries a four-cornered ring at its center, and that ring halts the building of the bacterial cell wall. A beta-lactamase cuts the ring open with water. Once the ring is broken the drug is just another molecule, and the bacterium lives on. That is antibiotic resistance at the level of a single molecule.

These enzymes fall into four Ambler classes according to how closely their sequences resemble one another. Classes A, C and D use the amino acid serine to cut the ring, while class B recruits a zinc ion for the job. The enzymes that break carbapenems as well sit mostly in classes B and D, and the class studied most in the clinic is A.

The trouble lay in how new enzymes were found. Until now candidates were picked by how closely they resembled the sequences already on file, and a method like that returns only the lookalikes. A protein that does the same job from a sequence far away slips straight through the mesh. That is where the team set its starting point. A protein language model learns sequences the way it would learn sentences, so it reads a protein by the regularity of which amino acid sits at which position rather than by how many letters overlap.

The team collected 1,778 experimentally characterized beta-lactamases from BLDB and Swiss-Prot to serve as the reference. ESM2-t33-650M produces a 1,280-dimension vector for every amino acid in a sequence, and five classifiers sit on top of it. A multilayer perceptron, a convolutional neural network, a bidirectional LSTM, a support vector machine and a random forest each judge on their own, and their probabilities are averaged by weight into a final answer. Across five-fold cross-validation the individual classifiers moved around from fold to fold, while the five taken together classified every sample correctly in all folds.

The team also looked at those same training sequences from another angle. A phylogenetic tree built from the 1,066 that came from nature separated the four classes cleanly, yet the branches of the tree matched the taxonomy of the bacteria carrying the enzymes only in part. These enzymes have not simply been handed down the generations; they have also moved from one bacterium to another. And the sequences drawn from environmental samples did not gather on one branch. They scattered across the whole tree. The clue that the environment might be a reservoir of new beta-lactamases was already sitting inside the known list.

The model then went to MGnify, a public database of metagenome proteins. A metagenome is not the genome of one cultured strain. It is every microbial gene in a scoop of seawater or wastewater, read all at once. A hidden Markov model profile for each class passed 57,799 sequences on the first cut, only those between 200 and 400 amino acids long were kept, and the ensemble classifier narrowed the set to 46,335.

The three steps that narrowed the pool to 46,335 MGnify Metagenome proteins 2024_04 release → Hidden Markov model Per-class profile search 57,799 → Ensemble classifier ESM2 + five classifiers 46,335 Training set: 1,778 experimentally characterized beta-lactamases (BLDB, Swiss-Prot)
▲ Pebblous original diagram. The count passing each step comes from §2.3 of the paper

By class, B is the largest at 19,428, followed by C at 14,132, A at 9,092 and D at 3,683. In the reference set of 1,778, class A was the largest at 803, which inverts the order. The team reads the mismatch less as a real difference in nature than as a residue of clinical attention concentrating on class A. The known list is a scale model of where people have been studying rather than a scale model of nature.

2

Candidates sharing 30% of their sequence still broke the drugs

A prediction on its own proves nothing. The team picked out the two candidates furthest from the known enzymes in each of the four classes, then put two class C candidates on the bench. Prot22, from groundwater, shares 30.6% of its sequence with the closest known protein. Prot94, from activated sludge, shares 27.4%. The two came from environments under completely different antibiotic pressure.

Both sequences were codon-optimized, cloned into an E. coli expression vector, and measured for minimum inhibitory concentration against 11 beta-lactam antibiotics and two inhibitor combinations. The minimum inhibitory concentration is the lowest drug concentration that keeps bacteria from growing. A larger value means the drug no longer works. AmpC, an already characterized class C enzyme, ran alongside as the positive control, with E. coli carrying an empty vector as the negative control.

Antibiotic Empty vector AmpC Prot22 Prot94
Penicillin 8 >512 256 >512
Oxacillin 64 >512 512 512
Amoxicillin <0.5 >512 256 256
Imipenem <0.5 16 4 <0.5
Amoxicillin + clavulanate 1 512 512 1

▲ Minimum inhibitory concentration (μg/mL). Five rows out of the 13 entries in Fig. 4C of the paper. Measurement ran in 11 two-fold steps from 0.5 to 512

The empty-vector control reads 8 for penicillin and 64 for oxacillin. That is the baseline E. coli carries on its own, and the comparison has to be read on top of that baseline. Above it Prot22 recorded 256 for penicillin and 512 for oxacillin, while Prot94 ran past the top of the measured range for penicillin. A protein overlapping a known enzyme by barely 30% of its sequence does break antibiotics inside E. coli.

AmpC, the positive control, is the yardstick for how far a documented class C enzyme goes. Penicillin, oxacillin and amoxicillin all exceeded the highest concentration tested, while imipenem stopped at 16 and ceftriaxone, left out of the table, at 4. A known class C enzyme does not break carbapenems or ceftriaxone the way it breaks the penicillins. The two candidates came close to that yardstick, and for penicillin Prot94 landed in the same cell.

Something reassuring sits in the same table. Add clavulanate to amoxicillin and Prot94 drops to 1, which means the existing inhibitor still works on this enzyme. Prot22 recorded 512 in the same combination, where the inhibitor did not get through. The two proteins lie a similar distance from the known enzymes, and yet they hold off drugs in different ways.

3

Where you sample decides what you find

Once the 46,335 are split by habitat, the composition shifts noticeably. In human fecal samples class A takes 48.38%, close to half, while in seawater it falls to 11.00% and class B rises to 56.32% in its place. Wastewater sits between the two. The same model with the same criteria returns a different kind of enzyme depending on where the sample was lifted.

Class mix shifts with the place the sample came from Class A Class B Feces n=4,328 48.38% 34.43% Wastewater n=9,530 18.38% 39.51% Ocean n=8,044 11.00% 56.32% Classes C and D are left out. Shares follow Fig. 3B of the paper
▲ Pebblous original diagram. It draws classes A and B only, out of the habitat breakdown in §2.3

The team reads the difference as pressure from the environment. Wastewater and activated sludge are where water from homes, hospitals and factories runs together, so drugs and bacteria pile up side by side, and cells sit densely enough for genes to travel between them. The human gut is an entirely different environment, with host physiology and a history of antibiotic use attached to it. Metagenome data is itself skewed by which places have been sampled and how heavily, so this difference in composition holds a real difference in ecology and a tilt in collection at the same time.

The tilt in sampling is sharper by geography. Among samples with coordinates attached, North America holds 14,099, Asia 11,248 and Europe 10,665, while Africa, South America and Oceania are rare. The global map of resistance is in practice a map of the regions whose metagenomes have been read.

The picture goes quiet when the same data is split by continent rather than habitat. Class B was the largest in every region, and its share stayed between 34.0% and 44.5%. The axis that moved the class composition was habitat rather than continent, seawater against feces rather than North America against Asia.

The part worth more attention is how much room there is to vary. The team took 97 positions in the representative sequences of the four classes where a mutation could plausibly be accepted, and measured how much those positions actually vary across natural sequences, as site entropy. Enzymes from wastewater scored significantly higher than enzymes from the ocean in all four classes. The same enzyme, in other words, carries more letters it can change when it comes from wastewater. The paper states plainly that this is no direct evidence of positive selection, and that differences in sampling depth or phylogenetic composition may be mixed into it.

The team also asked what those positions were to begin with. The median site entropy of the 97 came to 3.20, above the 2.93 of the other 1,053 positions. The positions marked as able to accept a mutation sat away from the tightly conserved catalytic core, in regions where letters changed often to begin with. The rate at which a predicted mutation actually appears as that amino acid in natural sequences, though, came to about 25% for the ocean and about 30% for wastewater, pointing the same way without reaching statistical significance. The width of the positions open to change and the convergence onto a particular amino acid ran separately.

4

A single amino acid swap that lifted carbapenem resistance

In the last stage the team changed the question, from which enzymes are out there to how far one of them can change. Two tools did the work. One is ESM1b, which scores how natural a sequence looks once the amino acid at a given position is swapped for another. Run against existing mutation data for the TEM-1 beta-lactamase, it tracked the measured values with a correlation coefficient of 0.667, and separated beneficial mutations from harmful ones with an area under the ROC curve of 0.88. The other is ProteinMPNN, which starts from a structure and works backward to propose sequences that fit it. Overlaying the evolutionary signal with the structural constraint narrowed the candidate positions to 31 in Prot22 and 42 in Prot94.

Eight substitutions at those positions were built, one failed sequence verification, and the remaining seven went into E. coli for another round of minimum inhibitory concentrations. All seven read differently from the wild type. Some values rose and some fell, and either way the positions the AI marked do bear on how the enzyme works.

Antibiotic Prot22 wild type Y106F substitution
Meropenem <0.5 2
Imipenem 4 16

▲ Minimum inhibitory concentrations (μg/mL) for the two carbapenems. Mut2 results in §2.5 of the paper

The substitution that stood out swaps the tyrosine at position 106 of Prot22 for phenylalanine. Meropenem, under 0.5 in the wild type, rose to 2, and imipenem went from 4 to 16. Carbapenems are the drugs held back for the point where the other beta-lactams have stopped working. Changing a single amino acid moved the enzyme in their direction. In Prot94 the strongest substitution swapped the serine at position 73 for threonine, and both cefaclor and imipenem climbed to 512.

The other five went in different directions. M86L and Q65K in Prot22 lifted the values for piperacillin, mezlocillin and aztreonam while lowering some of the penicillins. The list of drugs the enzyme breaks had itself moved. S115T came down across the board, and C376V and G362P in Prot94 read lower than the wild type on the penicillins. Substitutions that raise resistance and substitutions that lower it came out of one prediction. The AI marked positions that bear on function rather than positions that build resistance.

A 100-nanosecond molecular dynamics simulation went after the reason. For S73T the binding pocket moved less than in the wild type, the hydrogen bond network widened, and the distance between the catalytic serine and the substrate closed from 0.69 nm to 0.52 nm. For Y106F the overall structure stayed much as it was while the environment inside the pocket changed. The team records this reading as provisional, since better binding does not have to mean a higher catalytic turnover.

The paper also sets out its own limits. Experimental verification stopped at two class C candidates and seven single substitutions, and the ecological comparison rests on the metagenome sequences accumulated so far rather than on any direct measurement of selective pressure or any tracking over time. That a mutation can change resistance and that the mutation actually occurs in nature are two different statements.

5

Why Pebblous is watching this paper

Antibiotic resistance has long been tracked from the hospital. A strain is isolated from a patient, the drugs it no longer answers to are identified, and the result goes up into a surveillance network. GLASS, run by the World Health Organization, works off reports from clinical laboratories in each country. Resistance gene databases such as CARD and ResFinder are likewise built to annotate and update the genes already observed in the clinic. All of these handle resistance that has already happened.

This paper changed where the samples come from. It gathered data from seawater and wastewater rather than from patients, and came back with more than twenty times the known list. Here a shape surfaces that anyone working with data meets again and again. The site of collection sets what comes back. Fecal samples give half class A, seawater gives half class B. North America, Asia and Europe have been read the most, so the map is densest there. Wastewater yields enzymes with more room to vary than the ocean does. Same model, same criteria, and the input pulls the conclusion along behind it.

That is what makes reading the known list as a scale model of nature dangerous. Class A was the largest among the enzymes characterized so far because the clinic spent its attention there, not because nature is built that way. In any field, accumulated data holds the distribution of observation rather than the distribution of the world. Put a model on top without separating the two, and the model learns the tilt of observation as a law of the world.

The open question is what becomes of the list itself. The 46,335 candidates, together with the 97 positions open to mutation, amount to a list of risk that has not reached the clinic yet. As things stand, no institution holds the job of verifying and updating material of this kind. GLASS takes clinical reports, CARD annotates known genes, and environmental surveillance is only now widening as far as wastewater. Frameworks such as managed access and screening at the input stage are under discussion for AI biodesign tools, but that discussion is aimed at tools that design something, and this paper found something that already exists.

The tension runs both ways. Open the list completely and the path to misuse opens with it. Keep it closed inside a few institutions and the early warning that diagnostics and drug development could have used disappears. The data availability clause in this paper is a single sentence, saying the data is available on reasonable request. No deposit in a public repository is mentioned. That one line is where the absence of an institutional answer shows.

Two things are here for an organization that works with data. One is writing down where the material in hand came from before writing down how the model performed. The other is that output predicting something that has not happened yet also needs rules for storage, verification and release. The candidate list in this paper may turn out wrong and may turn out right. Either way, the fact that nobody has been assigned to find out stays where it is.

R

References