46 billion compounds, 196 molecules

The paper is Swanson and colleagues, "SyntheMol-RL: a flexible reinforcement learning framework for designing easily synthesizable antibiotics," in Molecular Systems Biology, 2026. It comes out of McMaster and Stanford, it is open access, and I read the full text rather than the press release.

I want to be clear at the top, because this beat can read as reflexively sour: this is a good paper. It does something specific and it measures whether the something worked. My complaint is not with the authors. It is with a headline number that describes a room rather than a search, and with how much of the actual achievement gets lost when that number leads.

What 46 billion actually refers to

The model does not invent molecules from scratch in open chemical space. It builds them from purchasable parts using known reactions, which is the entire point of the approach. Constrain the generator to things a vendor can actually make, and you avoid the classic failure where a model proposes a beautiful compound nobody can synthesize.

So the space is defined by two commercial catalogs. The Enamine REAL Space contains 31 billion molecules producible from roughly 139,000 building blocks and 169 reactions, of which the authors use the 13 most common, which can build 30 billion molecules. The WuXi GalaXi space contains 16 billion molecules from roughly 15,000 building blocks and 36 reactions. Thirty billion plus sixteen billion is the 46 billion.

The authors describe this accurately. Their own phrasing is "a massive chemical space of 46 billion compounds," and elsewhere they note they expanded the space the model "explores" from 30 billion to 46 billion. That is a fair description of an addressable space. What it is not is a count of molecules evaluated.

Here is what was evaluated. From the methods:

In both cases, we ran the generative model for 10,000 rollouts, resulting in the generation of roughly 10,000 unique molecules.

Ten thousand. Not forty-six billion. The 46 billion is the number of molecules the model could in principle assemble from the parts bin, which is a property of the catalog, not an accomplishment of the model. A combinatorial space is cheap to make enormous: add building blocks and the count explodes multiplicatively. The hard part was never the size of the space, it was picking well inside it.

This is the same category error I wrote about in the "millions of materials" report. A number describing a search space gets reported as a number describing work performed.

The funnel

Follow the molecules and the story gets concrete.

The model generated roughly 10,000 candidates per configuration. Those were filtered, then ranked, and 250 compounds were submitted for synthesis (50 for each of five settings, 248 unique after two duplicates).

Then chemistry happened. Enamine successfully synthesized 177 of 224 unique compounds, a 79% success rate. WuXi managed 17 of 24, a 71% rate. Together that yielded 196 unique compounds for in vitro testing.

Pause on that. In a chemical space specifically constructed out of easy, reliable reactions on purchasable building blocks, roughly one in four attempted syntheses still failed. Synthesizability was the central design constraint of the whole system and it was still only mostly true. Anyone who has ordered compounds from a catalog and waited for the email saying the route did not work will recognize this immediately.

Of those 196, the subset attributable to the two SyntheMol-RL variants and structurally distinct from the training data was 79 compounds. Of those, 13 showed potent activity, defined as complete inhibition of bacterial growth at MIC ≤ 8 µg/ml. Of those 13, seven survived structural novelty filtering against known antibiotics. Of those seven, one, synthecin, went into a mouse.

46,000,000,000 → 10,000 → 250 → 196 → 79 → 13 → 7 → 1.

Where the work happened

Every arrow after the second one is bench work.

Chemists at two contract synthesis companies made 194 physical compounds. Microbiologists ran twofold serial dilution MIC assays with ciprofloxacin as a positive control. The seven survivors were screened against a phylogenetically diverse panel of ESKAPE pathogens and against multidrug-resistant vancomycin-intermediate S. aureus isolates, where all seven retained MIC ≤ 8 µg/ml.

Then, because the target application was topical, the team invented an assay for it: the Chemical Release Evaluation on Agar Media assay, in which test molecules at 2% w/v were formulated in Glaxal Base, applied to MRSA-inoculated agar, and incubated overnight. That is a piece of experimental design that exists nowhere in the model. Somebody had to decide that a compound's performance in a cream base was the question worth asking.

None of this is a criticism. It is the answer to "what did the AI do," which is: it proposed a ranked list. Everything that converted a ranked list into knowledge was done by people and instruments. I keep returning to this because it is the load-bearing distinction on this beat, and it was also the whole story in the robot chemist report, where automation broke precisely at the measurement step.

The safety filter was also a model

This is the detail I would most want a general reader to notice, and I have not seen it mentioned anywhere in the coverage.

To narrow candidates before ordering, the authors ranked compounds by predicted clinical toxicity using ADMET-AI, a multi-task model trained on 41 ADMET datasets from the Therapeutics Data Commons, and selected the 50 compounds with the lowest predicted clinical toxicity for each model type. The ClinTox output is a probability between 0 and 1.

So the toxicity screen at the selection stage was a prediction, not an assay. That is a completely standard and defensible way to prioritize a purchase list, and it is not a flaw in the study. But it means the pipeline has a model proposing molecules and a second model filtering them for safety, with the first measured biological safety information arriving much later and in a narrow form. Searching the full text, I did not find cytotoxicity, hemolysis, or mammalian-cell selectivity data for synthecin. What the paper does report on tolerability is that synthecin-treated mouse tissues showed no signs of inflammation.

When a pipeline stacks predictions, the uncertainty stacks too, and predicted-safe is a filter for deciding what to buy, not a claim about what is safe.

Reading the mouse result precisely

The in vivo experiment is where "designed a new antibiotic" gets its force, so it deserves exact reading.

Cyclophosphamide pre-treated C57BL/6N mice were infected with S. aureus USA300 at roughly 2.15 × 10⁷ CFU. Infection established for an hour, then topical treatment at 1, 4, 8, 12 and 20 hours post-infection with Glaxal Base carrying either vehicle (10% DMSO, n = 5) or 2% synthecin (n = 5). At 24 hours the skin was dissected and plated.

Vehicle-treated mice carried about 6.39 × 10⁹ CFU/g with notable tissue inflammation. Synthecin-treated mice carried about 5.14 × 10⁷ CFU/g with no signs of inflammation. That is a difference of more than two orders of magnitude, and it is a real result.

Now the nuance the authors state and the coverage drops. That 5.14 × 10⁷ figure is, in their words, "similar to the bacterial burden observed in pre-treated infection control mice." The treated animals ended roughly where the infection started. Synthecin stopped the infection from growing rather than clearing it, which is consistent with the separate finding that its activity is bacteriostatic at 2× and 4× MIC.

The authors are straightforward about this and note that linezolid, commonly prescribed for S. aureus skin infections, is also bacteriostatic. That is a fair comparison and I am not using "bacteriostatic" as a slur. But "arrested growth in immunosuppressed mice under five topical doses in 24 hours" and "new antibiotic" are separated by a great deal of work, and the mice were immunosuppressed by design, which makes the model harsher in one way and less like a typical patient in another.

The number that should have been the headline

Here is the result I think is genuinely impressive, and it is not 46 billion.

The authors ran controls against their own prior method. Of the compounds synthesized and tested: RL-Chemprop produced 11 potent hits out of 38. RL-MLP produced 2 out of 41. Their previous published system, SyntheMol-MCTS, produced zero out of 38. A virtual screening approach produced 2 out of 37.

Eleven versus zero, against their own earlier work, with the compounds physically made and tested rather than scored in silico. That is a controlled comparison with a real endpoint, and it is the kind of evidence that actually supports a claim of methodological progress.

There is a second concrete win worth noting. The authors mention that in their prior work, "only two of our six potent antibacterial compounds were sufficiently soluble to be administered for in vivo toxicity experiments." The new model optimizes for antibacterial activity and aqueous solubility simultaneously. In this round, all seven hit compounds were readily soluble in the Glaxal Base formulation. A previously blocking practical problem was named, targeted, and fixed. That is what progress looks like, and it is far less quotable than a number with nine zeros.

What I am not claiming

Not that the authors overstated anything. Their abstract says "chemical space of 46 billion compounds," which is correct. The drift to "explored 46 billion compounds" happened downstream.

Not that synthecin is unimpressive. A structurally novel compound retaining MIC ≤ 8 µg/ml across resistant isolates and suppressing an MRSA wound infection in vivo is a legitimate result.

Not that generative chemistry is hype. The 11-versus-0 comparison is exactly the evidence I would ask for, and they produced it.

Not that this is a drug. It is a topical, narrow-spectrum, bacteriostatic compound demonstrated in one immunosuppressed mouse model with five animals per group. The paper does not claim otherwise.

And a limit on my reading: the paper does not state a single numeric MIC for synthecin in the text I read, only that it met the MIC ≤ 8 µg/ml threshold and was among the most potent. I also found no mammalian cytotoxicity data, though I searched the full text rather than every supplementary dataset, so I would not claim none exists.

The signal

Three things.

First, when a story about AI in science leads with an enormous number, check whether it counts things examined or things possible. Search-space size is usually a property of how the problem was set up, not a measure of effort or insight. Forty-six billion is a statement about two vendor catalogs.

Second, follow the physical objects. In any "AI discovered X" story, ask how many instances of X actually existed as matter, who made them, and who measured them. Here the answer is 196, made by contract chemists, measured by microbiologists. That count is the honest scale of the experiment, and it is the number that determines how much you should believe.

Third, watch for stacked predictions. When a model's output is filtered by another model and the result is described in the language of validation, find the step where something was actually observed. In this paper that step is real and well done. It just arrives much later in the pipeline than the headline implies, and it is smaller than the headline sounds.