The urgency behind that claim is hard to overstate. Antimicrobial resistance is now linked to millions of deaths worldwide each year, and the World Health Organization has placed carbapenem-resistant Acinetobacter baumannii at the top of its list of critical priority pathogens — a Gram-negative bacterium notorious for shrugging off nearly every drug clinicians can deploy against it. The deeper problem is architectural: most antibiotics introduced over the past half-century are chemical variations on a handful of molecular scaffolds, so bacteria that acquire resistance to one member of a family frequently resist its relatives as well. Because established drugs converge on the same targets and the same binding sites, a single resistance mechanism — an altered ribosomal protein, a methylated RNA residue, an overactive efflux pump — can neutralize entire classes at once. New chemical classes that hit old targets in genuinely new ways are therefore among the most valuable commodities in infectious-disease research, and lasso peptides, the review argues, may be exactly that.

Lasso peptides belong to the ribosomally synthesized and post-translationally modified peptides, or RiPPs — a vast family of natural products in which the starting material is made by the ribosome as a linear precursor and then sculpted by dedicated enzymes into something far more elaborate. In lasso peptides, that sculpting produces a mechanical marvel: the N-terminal segment of the chain is stitched into a closed ring through which the C-terminal tail is threaded and then trapped, creating an interlocked lasso fold. The tail is typically cinched in place by bulky amino acids acting as plugs, and in many family members an additional cross-link prevents the tail from slipping back out. The result is a compact, rigid structure with unusual resilience; lasso peptides routinely survive proteases that would dismantle ordinary peptides, as well as temperatures and pH extremes that destroy most proteins. That mechanical constraint also confers pharmacological advantages: the molecule arrives at its target pre-folded, which sharpens binding selectivity and slows chemical degradation inside living systems.

Lariocidin, produced by Paenibacillus sp. M2, is a recently discovered member of this family with an ambition that sets it apart: it is a peptide antibiotic that goes after the ribosome, the molecular factory every bacterium needs to build its proteins. Ribosome-targeting is old territory for antibiotics — aminoglycosides, tetracyclines and macrolides all work there — but lariocidin arrives with a completely different chemical body and, crucially, a different docking address. The review’s authors situate LAR within the emerging subgroup of ribosome-targeting lasso peptides and argue that the class should be treated as a platform of microbial peptide antibiotic scaffolds for future discovery rather than a one-off curiosity. Their synthesis of the evidence gathered so far describes a compound with broad antibacterial activity, an unusual binding mode and a resistance profile that, in early testing at least, looks remarkably clean. If further work confirms that profile, the implications for drug discovery could be considerable.

To understand why that matters, it helps to picture the bacterial ribosome. In bacteria it is built from two subunits: the larger 50S subunit, which catalyzes the formation of peptide bonds, and the smaller 30S subunit, which reads the genetic message. The 30S subunit is organized around 16S ribosomal RNA, the molecular scaffold of the decoding center where messenger RNA codons are checked against the anticodons of incoming aminoacyl-tRNAs, the adaptor molecules that ferry amino acids into place. Classical inhibitors exploit this machinery in well-charted ways: aminoglycosides wedge into the decoding region of 16S rRNA and trick the ribosome into misreading the genetic code, while tetracyclines physically block aminoacyl-tRNA from entering its binding site. Lariocidin does neither. According to the review, it binds a distinct site within the 30S subunit, making contacts with both 16S rRNA and aminoacyl-tRNA — a binding mode that overlaps in function but not in geography with the established drugs. Because the docking site differs, resistance mechanisms tailored to aminoglycosides or tetracyclines do not automatically extinguish lariocidin’s activity.

Functionally, the peptide delivers a double blow. First, it inhibits translocation — the ratcheting step in which the ribosome, having linked one amino acid to the growing chain, must shift the entire messenger RNA and its paired transfer RNAs by exactly one codon to make room for the next. Jam that step and protein synthesis stalls. Second, lariocidin promotes miscoding, causing incorrect amino acids to be inserted into nascent proteins. A bacterium under lariocidin attack therefore faces a pincer movement: whatever proteins it manages to finish are increasingly likely to be garbled and nonfunctional, while the overall production line grinds toward a halt. This combination of translocation inhibition and miscoding, achieved through a binding site distinct from those used by aminoglycosides and tetracyclines, forms the mechanistic heart of the review’s argument and the primary reason the authors believe cross-resistance with established ribosome-targeting antibiotics may be reduced. In an era when resistance genes circulate freely between bacterial species, an unclaimed binding site is a strategic asset.

The preclinical evidence assembled in the review is, by the sober standards of early-stage antibiotic research, encouraging. Lariocidin displays broad antibacterial activity, including against multidrug-resistant pathogens such as Acinetobacter baumannii, one of the organisms for which new therapies are most desperately needed. In laboratory assays, spontaneous resistance arises at low frequency, a sign that the ribosomal binding site cannot be trivially mutated around. Safety signals from cell-based studies are similarly promising: mammalian cytotoxicity is limited, and the peptide causes minimal hemolysis, the rupture of red blood cells that serves as a standard early warning of membrane-damaging toxicity. Most significantly, the compound has shown efficacy in mouse infection models — a hurdle many antibiotic candidates never clear, because activity in a culture dish does not guarantee that a molecule will survive, reach its target and work inside a living host. Taken together, these findings sketch a compound that is simultaneously potent, selective and demonstrably active in vivo, a combination rare enough to turn heads.

Part of the excitement lies in what the lasso scaffold itself makes possible. Because RiPPs are genetically encoded, their sequences can be edited like software: individual amino acids in the ring or the threaded tail can be swapped to tune potency, spectrum or stability, while the knot-tying biosynthetic enzymes can be repurposed for the chemoenzymatic production of analogs that would be difficult to synthesize by chemistry alone. The same genetics that make lasso peptides engineerable also make them findable. Biosynthetic gene clusters for these molecules lie scattered across bacterial genomes and metagenomic datasets, meaning computational mining of sequence databases can point researchers toward previously hidden lasso peptides before anyone has cultured the producing organism. The review frames lariocidin as a flagship for this strategy — proof that searching the microbial pan-genome for ribosome-targeting lasso peptides can surface antibiotic scaffolds with properties that conventional small-molecule screening has struggled to deliver, and a starting point for engineering next-generation derivatives.

The authors are nonetheless explicit about how far the compound still has to travel. Lariocidin remains at an early developmental stage, and several fundamental questions are unresolved. Its pharmacokinetics and pharmacodynamics — how the molecule is absorbed, distributed, broken down and cleared, and how those behaviors translate into safe, effective dosing regimens — remain to be characterized. Formulation is untested at scale, and scalable production is a genuine obstacle for any RiPP, since fermentation yields, the efficiency of the knot-forming enzymes and downstream purification all require optimization before industrial manufacturing becomes realistic. Resistance surveillance will be essential as well: low spontaneous resistance in the laboratory does not guarantee that resistance will not emerge and spread under the intense selective pressure of clinical drug exposure. And like all early findings, the preclinical results await independent validation by research groups beyond the original discoverers. The review is careful to note that this body of work includes no clinical trial; the road from mouse models to medicine runs through years of further development.

Even with those caveats, the arrival of lariocidin marks a meaningful shift in emphasis for the field. For decades, antibiotic development has largely meant redecorating old scaffolds in a running battle to stay ahead of resistance. Lariocidin offers something different: a new chemical architecture — a self-knotting peptide — aimed at the oldest and most validated target in antibacterial therapy, the ribosome, through a binding site the existing armory does not touch. Whether it ultimately survives the gauntlet of pharmacokinetic testing, manufacturing scale-up and clinical trials remains an open question, and the review avoids promising a new medicine on pharmacy shelves. What it does argue is that genome mining, RiPP bioengineering and ribosome structural biology have matured into a discovery engine capable of surfacing candidates that the classical pipeline overlooked. If that engine keeps running, the knotted molecules hiding in the world’s soil may yet supply the next chapter in humanity’s fight against the superbugs.

Lasso peptides are synthesized by bacteria through a fascinating biosynthetic pathway wherein ribosomes assemble linear chains of amino acids, which are subsequently folded into a slipknot-like structure by specialized biosynthetic enzymes. This unique topology bestows lasso peptides with extraordinary stability against enzymatic degradation and environmental stressors. Thousands of different lasso peptides have been identified across microbial species, many exhibiting potent antibacterial, antiviral, and anticancer properties that underscore their therapeutic promise.

Professor Doug Mitchell, co-leader of the study and Director of the Vanderbilt Institute for Chemical Biology, emphasized the untapped potential of these molecules: “The unique structural features of lasso peptides make them ideal candidates for targeting challenging receptors and developing robust oral therapeutics. By creating a dedicated language model tailored for these peptides, we now have a powerful computational tool to accelerate discovery and design in this emerging field.”

While machine learning has become integral in analyzing vast biological datasets, existing AI models such as AlphaFold, despite their revolutionary impact on protein structure prediction, face intrinsic limitations when applied to lasso peptides. The atypical lasso fold deviates significantly from canonical protein structures, rendering traditional prediction algorithms ineffective in accurately modeling their complex topology. This gap motivated the development of LassoESM, a bespoke protein language model specifically trained on the sequences and structural intricacies of lasso peptides.

Unlike generic protein language models that learn from a broad range of protein sequences, LassoESM was meticulously trained on a curated dataset of thousands of experimentally validated lasso peptides. The model uses a masked language modeling technique, wherein fragments of peptide sequences are concealed and predicted, enabling the model to learn the underlying “language” of lasso peptide biosynthesis and folding patterns. This deep understanding allows LassoESM to capture subtle sequence-structure relationships unique to the lasso fold, which conventional models overlook.

A core functionality of LassoESM lies in its ability to predict interactions between lasso peptides and their biosynthetic enzymes, particularly lasso cyclases—the specialized enzymes responsible for catalyzing the knot-tying step of peptide biosynthesis. Since each lasso cyclase recognizes specific peptide substrates much like keys fitting into distinct locks, deciphering these interactions is crucial for engineering novel peptides with designed functionalities. The LassoESM model can infer which lasso cyclase pairs are compatible with a given peptide sequence, a feat that was previously challenging due to sparse experimental data and complex enzyme-substrate specificity.

The collaborative effort harnessed the complementary expertise of the Mitchell and Shukla laboratories, combining bioinformatics, machine learning, and experimental validation. They initially employed bioinformatics approaches to collect a comprehensive catalog of lasso peptides from diverse microorganisms, followed by manual validation to ensure the accuracy of sequence annotations. This high-quality dataset was essential for reliably training the language model. Subsequently, the model was fine-tuned for multiple predictive tasks, including lasso peptide enzymatic compatibility, structural property inference, and functional annotation.

Dr. Diwakar Shukla, co-leader and chemical engineering professor at the University of Illinois Urbana-Champaign, highlighted the transformative impact of this approach: “By decoding the molecular ‘language’ of lasso peptides, LassoESM opens new horizons in predicting properties and functions that have remained elusive. This tool enables us to not only predict structure but also to rationally design peptides with tailored features for specific biomedical applications.”

Despite the limited availability of labeled experimental data—a common bottleneck in peptide research—LassoESM demonstrated robust performance in predicting diverse lasso peptide properties. This capability significantly reduces the empirical trial-and-error burden traditionally associated with discovering and optimizing peptide therapeutics. The model thereby streamlines the development pipeline from sequence to functional candidate, accelerating translational applications in industry and medicine.

Looking ahead, the researchers aspire to extend this AI-driven framework to other classes of peptide natural products beyond lassos. They envision developing specialized language models capable of capturing the nuances of various peptide topologies and their biosynthetic strategies. Additionally, the team aims to leverage LassoESM in engineering peptides that selectively target specific proteins, potentially creating a new generation of peptide-based therapeutics with enhanced efficacy and stability.

The development and application of LassoESM exemplify the power of interdisciplinary collaboration and cutting-edge computational resources. Supported by the National Institutes of Health and facilitated by the robust infrastructure at the Carl R. Woese Institute for Genomic Biology, this research represents a significant advance in peptide engineering. As machine learning continues to evolve, tailored models such as LassoESM are poised to revolutionize how scientists understand, design, and deploy complex biomolecules in real-world therapies.

In summary, LassoESM is an innovative language model that captures the structural and functional essence of lasso peptides, overcoming longstanding challenges in prediction and design. By enabling precise forecasting of peptide properties and enzyme compatibility, it paves the way for rational, AI-driven development of novel therapeutics. This work stands as a compelling testament to the synergy between computational biology and experimental science in transforming drug discovery.

Subject of Research: Lasso peptides and protein language models for biomedical and industrial applications

Article Title: LassoESM a tailored language model for enhanced lasso peptide property prediction

News Publication Date: 29-Sep-2025

Web References:
https://doi.org/10.1038/s41467-025-63412-3

Image Credits: Xuenan Mi, Isaac Mitchell

Keywords: Protein engineering, Machine learning, Artificial intelligence, Peptides, Drug discovery, Bioinformatics