Antibiotic resistance is one of the most pressing threats in modern medicine, and few pathogens embody the danger better than Staphylococcus aureus. In 2019, nearly five million deaths worldwide were linked to drug-resistant bacterial infections, and S. aureus alone accounted for more than 700,000 of those deaths. Methicillin-resistant S. aureus, or MRSA, has seen its attributable death toll double between 1990 and 2021, rising from 57,200 to 130,000 per year. Now, a team of researchers at Stanford University and McMaster University reports a major step forward in the fight against this pathogen, using an artificial intelligence system that can design entirely new drug molecules from a chemical space of 46 billion possibilities and then prove, in the laboratory and in living animals, that those molecules actually work.
The new system, called SyntheMol-RL, is described in the journal Molecular Systems Biology. It is a significantly upgraded version of an earlier generative model, SyntheMol, which the same group used to design candidate antibiotics against Acinetobacter baumannii. The core idea behind both systems is the same: rather than screening existing compounds one by one, the AI builds new molecules from purchasable chemical building blocks using well-validated chemical reactions. This constraint guarantees that every molecule the model proposes can, in principle, be synthesized quickly and cheaply, a property that has long been the Achilles heel of generative drug design, where AI models frequently propose compounds that look good on paper but are practically impossible to make.
The key technical innovation in SyntheMol-RL is the replacement of the original Monte Carlo tree search algorithm with a reinforcement learning framework. In the earlier system, the tree search treated every possible combination of building blocks independently, scoring each node based on statistics about how often it had been explored and how well molecules built from it performed. That approach worked, but it could not learn patterns shared across chemically similar building blocks, which limited its search efficiency across enormous combinatorial spaces. The new reinforcement learning model instead uses a deep neural network as a value function that takes the chemical structures of building blocks as input and predicts the expected property score of molecules that could be assembled from them. Because the network generalizes across similar structures, it can navigate the chemical space far more intelligently, sampling promising nodes in proportion to their predicted value while still maintaining exploration.
A second major advance is multi-parameter optimization. Real drug discovery demands molecules that satisfy many properties simultaneously, not just one. In the team’s previous work, only two of six potent antibacterial compounds were soluble enough to be tested in animals, a bottleneck that motivated the new design. SyntheMol-RL combines separate reinforcement learning models, one for each property of interest, into a weighted value function whose weights adjust dynamically during generation. The system also dynamically tunes a temperature parameter that controls the balance between exploiting high-scoring building blocks and exploring diverse regions of chemical space, targeting a user-specified level of similarity among generated molecules. Ablation experiments showed that these dynamic mechanisms achieve optimal or near-optimal performance without manual fine-tuning, which would otherwise be highly sensitive and inconsistent between model variants.
To apply the system against S. aureus, the researchers first trained property prediction models on two datasets. For antibacterial activity, they physically screened an in-house library of 10,716 bioactive compounds against S. aureus RN4220, identifying 1,137 active molecules after statistical thresholding. For aqueous solubility, they used the AqSolDB dataset of nearly 10,000 molecules curated by the Therapeutics Data Commons. Two architectures were compared: Chemprop-RDKit, a graph neural network augmented with 200 molecular descriptors, and a simpler MLP-RDKit model using only the descriptors. The Chemprop-RDKit models performed slightly better and were selected as the fixed property predictors that score every molecule the generative model produces.
The researchers then ran two versions of SyntheMol-RL, one using the Chemprop architecture as the reinforcement learning value function and one using the MLP, alongside the older tree search version and a virtual screening baseline that evaluated 21 million randomly sampled molecules. Defining a hit as a molecule with a predicted antibacterial score of at least 0.5 and high predicted solubility, the RL-Chemprop variant generated hits at a rate of 11.6 percent and RL-MLP at 5.3 percent, compared with 3.0 percent for the tree search and a mere 0.006 percent for virtual screening. The comparison was deliberately fair: virtual screening was given roughly seven days of compute time, matching the slowest generative model. The team also benchmarked against two state-of-the-art generative systems, GFlowNet and REINVENT 4, which are not constrained to synthesizable chemical spaces. Although those models produced molecules with good predicted properties, medicinal chemists at the synthesis partners Enamine and WuXi offered limited synthesis options, with costs up to 62 times higher and timelines up to 5.7 times longer than the SyntheMol-RL compounds.
After filtering for novelty against known antibiotics, diversity, availability, and predicted toxicity, the team submitted 250 compounds for synthesis and obtained 196 unique molecules for laboratory testing. The results were striking. Eleven of 38 tested RL-Chemprop compounds and two of 41 RL-MLP compounds completely inhibited bacterial growth at concentrations of 8 micrograms per milliliter or lower, a hit rate of roughly 16 percent across the 79 SyntheMol-RL compounds synthesized. By contrast, none of the 44 randomly selected control molecules showed any activity up to the detection limit of 128 micrograms per milliliter, suggesting that ultra-large make-on-demand chemical libraries are not naturally rich in antibiotics and that conventional screening benchmarks may not apply to them. After a manual literature search for structural novelty, seven compounds remained as genuinely new antibacterial scaffolds.
One of those compounds, named synthecin, emerged as the standout. Although it shares some superficial features with salicylanilides, a class of known antibacterial agents, the researchers showed that it lacks the rigid hydroxybenzamide core and characteristic intramolecular hydrogen bonding of that class, instead incorporating a flexible aliphatic linker with a tertiary alcohol and trifluoromethyl group. All seven hit compounds retained potency against USA300, the most prevalent community-associated MRSA strain for skin infections, and against a panel of vancomycin-intermediate S. aureus isolates from the CDC database covering all major resistance mechanisms. Notably, the compounds showed no meaningful activity against other ESKAPE pathogens, indicating a narrow-spectrum profile that could spare beneficial microbiota and potentially slow the spread of resistance.
The team then moved synthecin into animal testing. Because S. aureus causes most skin and soft tissue infections, they first developed a simple assay called C.R.E.A.M., in which compounds formulated in a topical cream base are applied to MRSA-inoculated agar plates to evaluate chemical release and antibacterial effect. Synthecin produced the largest zone of inhibition among the seven candidates, and all seven compounds dissolved readily in the cream base, a practical payoff of optimizing solubility in silico. In a mouse wound infection model, mice pre-treated with cyclophosphamide were infected with roughly 21.5 million colony-forming units of MRSA USA300 and then treated topically with either vehicle or a 2 percent synthecin formulation at five time points over 20 hours. Vehicle-treated mice carried about 6.39 billion bacteria per gram of wound tissue at the endpoint, with visible inflammation, while synthecin-treated mice carried only about 51.4 million bacteria per gram, a burden comparable to untreated infection controls at the start of treatment, and their wounds showed no signs of inflammation.
The authors emphasize that SyntheMol-RL is not limited to antibiotics. Because it is compatible with any property predictor and any combinatorial chemical space, it can be applied across therapeutic domains, and it can even expand beyond single-reaction make-on-demand libraries to larger, multi-step synthetic spaces at the cost of higher expense. The work also highlights remaining challenges, particularly the need for more accurate property prediction models so that high-scoring computational designs translate more reliably into laboratory hits. Still, the journey from a 46-billion-molecule search space to a structurally novel compound that suppresses an MRSA wound infection in mice, accomplished with a 16 percent experimental hit rate, marks a compelling demonstration that reinforcement learning guided generative design can bridge the long-standing gap between computational molecular design and real-world drug discovery.
Subject of Research: Reinforcement learning-based generative AI for designing synthesizable antibiotics against drug-resistant Staphylococcus aureus
Article Title: SyntheMol-RL: a flexible reinforcement learning framework for designing easily synthesizable antibiotics
Article References: SyntheMol-RL: a flexible reinforcement learning framework for designing easily synthesizable antibiotics. (n.d.). https://doi.org/10.1038/s44320-026-00206-9
Image Credits: AI Generated
DOI: 10.1038/s44320-026-00206-9
Keywords: SyntheMol-RL, reinforcement learning, generative AI, antibiotic resistance, MRSA, Staphylococcus aureus, drug discovery, molecular design, synthesizability, synthecin, chemical space, preclinical validation
Cite Scienmag News
APA MLA Chicago
Drew Townsend. (October 1, 2026). AI Designs a New Antibiotic From 46 Billion Possibilities and Beats MRSA in Mice. Scienmag. https://scienmag.com/ai-designs-a-new-antibiotic-from-46-billion-possibilities-and-beats-mrsa-in-mice/
Drew Townsend. “AI Designs a New Antibiotic From 46 Billion Possibilities and Beats MRSA in Mice.” Scienmag, 1 October 2026, https://scienmag.com/ai-designs-a-new-antibiotic-from-46-billion-possibilities-and-beats-mrsa-in-mice/. Accessed 1 October 2026.
Drew Townsend. “AI Designs a New Antibiotic From 46 Billion Possibilities and Beats MRSA in Mice.” Scienmag. October 1, 2026. https://scienmag.com/ai-designs-a-new-antibiotic-from-46-billion-possibilities-and-beats-mrsa-in-mice/
Copy citation Download RIS
Tags: AI in pharmaceutical developmentAI-designed antibioticsAI-driven drug molecule synthesisAntibiotic resistancechemical spacechemical space exploration for drug designcombating antibiotic resistance with AIcombating MRSA with artificial intelligencedrug discoverydrug-resistant bacterial infectionsfighting multidrug-resistant bacteria with AIgenerative AIinnovative approaches to infectious disease treatmentlaboratory validation of AI-designed antibioticsmolecular designmolecular design for new antibioticsMRSApreclinical validationreinforcement learningStaphylococcus aureussynthecinSyntheMol-RLSyntheMol-RL antibiotic discoverysynthesizability
