AI Designs New Antibiotics That Reduce MRSA Infections In Mice

By Pooja Toshniwal PahariaReviewed by Lauren HardakerMay 5 2026

A new reinforcement learning–powered platform navigates billions of synthesizable molecules to deliver experimentally validated antibiotic candidates, offering a promising step toward faster, scalable drug discovery.

Study: SyntheMol-RL: a flexible reinforcement learning framework for designing easily synthesizable antibiotics. Image credit: simplystocker/Shutterstock.com

Designing new drugs from billions of possible molecules has long challenged discovery pipelines, but a new study in Molecular Systems Biology may offer a way forward. Researchers developed SyntheMol-RL, a generative artificial intelligence (AI) model powered by reinforcement learning that explores a massive molecular universe of nearly 46 billion synthesizable compounds to identify viable drug candidates.

By shifting from a traditional search strategy to a flexible deep learning framework, the system enables simultaneous optimization of multiple molecular properties. When applied to antibiotic discovery, it generated novel compounds active against Staphylococcus aureus, including a lead candidate that reduced bacterial burden in MRSA infections in a preclinical mouse wound infection model.

Antibiotic Resistance Continues To Outpace Drug Development

Antimicrobial resistance continues to outpace drug development, driving a growing global health crisis. In particular, infections caused by S. aureus, especially methicillin-resistant strains, remain a leading cause of morbidity and mortality, with limited treatment options and rising death rates. Although AI has accelerated antibiotic discovery, current approaches face critical limitations.

Scaling property prediction methods to ultra-large chemical libraries remains challenging, while generative AI approaches frequently produce compounds that are difficult to synthesize. For clinicians, this often translates into potentially limited or failing treatment options for routine infections, particularly in hospital settings, where drug-resistant Staphylococcus aureus infections can quickly escalate and complicate recovery.

Multi-Objective Optimization Improves Drug-Like Property Selection

In the present study, researchers developed SyntheMol-RL, an enhanced generative framework designed to address key limitations of their earlier model. Specifically, they replaced the Monte Carlo tree search (MCTS) approach with a reinforcement learning algorithm. This algorithm can learn shared patterns across chemically similar building blocks, thereby improving the efficiency of chemical space exploration. They also introduced a multi-objective optimization strategy, enabling simultaneous optimization of antibacterial activity alongside aqueous solubility, an important advance over the earlier system, which could only optimize one property at a time.

To guide molecule design, the team trained computational models to predict key chemical properties using in-house screening data from more than 10,000 compounds, alongside publicly available datasets. These models were integrated into an iterative reinforcement learning framework that selects and assembles chemical building blocks from large, commercially accessible libraries, producing compounds expected to be both effective and synthesizable.

The team implemented SyntheMol-RL to design antibiotics targeting Staphylococcus aureus. They prioritized candidates based on predicted performance, novelty, diversity, and safety, and benchmarked results against the previous model and standard virtual screening approaches. The team synthesized a subset of 79 compounds and tested them in vitro, followed by evaluation against drug-resistant strains, including MRSA.

They further evaluated the most promising candidate in a mouse-based experimental model of infection, where infected animals were treated topically with synthecin (2.0 % w/v) or a vehicle control (10 % dimethyl sulfoxide, DMSO). The researchers quantified bacterial burden to assess in vivo efficacy, enabling end-to-end validation of the model’s real-world utility.

AI Model Delivers Higher Hit Rates Than Existing Methods

SyntheMol-RL consistently outperformed both the earlier search-based model and conventional virtual screening across computational and experimental benchmarks. The model achieved markedly higher predicted hit rates, approximately 11.6 % compared to 3.0 % for the previously used MCTS-based approach and just 0.006 % for virtual screening, based on computational predictions rather than experimental validation. The findings highlight improved ability to identify promising candidates within vast chemical spaces.

When applied to antibiotic discovery targeting S. aureus, the model translated these gains into experimental success. Of 79 synthesized compounds, 13 (~16 %) demonstrated potent antibacterial activity in vitro. This exceeded comparator methods, which yielded no viable hits for the earlier model and only two for virtual screening among their respective synthesized and tested compounds. Notably, seven of these compounds met stringent structural novelty criteria and retained activity against clinically relevant drug-resistant strains, including methicillin-resistant S. aureus (MRSA) and vancomycin-intermediate isolates.

Several candidates demonstrated narrow-spectrum activity against S. aureus, potentially reducing off-target effects on beneficial microbiota. Among them, synthecin emerged as the most promising lead. In a mouse study involving MRSA-infected skin, synthecin significantly reduced bacterial burden, lowering counts from approximately 6.4 × 10⁹ colony-forming units per gram (CFU/g) in controls to 5.1 × 10⁷ CFU/g in treated animals.

This reduction was also associated with decreased tissue inflammation, suggesting potential therapeutic benefit in this preclinical setting. Mechanistically, synthecin was shown to act as a bacteriostatic agent rather than killing bacteria outright. Overall, the findings indicate that SyntheMol-RL can generate structurally novel, synthesizable antibiotic candidates with early-stage therapeutic potential.

Platform Offers Scalable Approach For Future Drug Discovery

Together, these findings suggest that SyntheMol-RL represents a meaningful step toward bridging the gap between computational drug design and real-world validation. By integrating reinforcement learning with built-in synthesizability and multi-parameter optimization, the model moves beyond theoretical predictions to deliver experimentally viable candidates, exemplified by synthecin against S. aureus. However, these results remain preclinical and are currently limited to in vitro testing and a mouse wound infection model.

Looking ahead, refining property prediction accuracy, as only a subset of high-scoring generated compounds translated into experimental hits, with many candidates failing during laboratory validation, and expanding optimized drug-like features could improve translation from in silico hits to clinical candidates. The platform’s flexibility also enables applications beyond antibiotics, including other therapeutic areas that require rapid, cost-effective molecule design. As such models evolve, SyntheMol-RL may help reshape early-stage drug discovery by enabling faster, scalable identification of novel compounds in complex chemical spaces.

Journal Reference

Swanson, K., Liu, G., Catacutan, D.B. et al. SyntheMol-RL: a flexible reinforcement learning framework for designing easily synthesizable antibiotics. Mol Syst Biol (2026). DOI: 10.1038/s44320-026-00206-9. https://link.springer.com/article/10.1038/s44320-026-00206-9