Basecamp Research's metagenomic foundation model — trained on uncultured microbial DNA that most labs cannot grow — now embedded in Anthropic's Claude Science platform, reports a 97% simulation success rate against WHO critical priority drug resistant pathogens.
A clinician types a question in plain language, attaches the morning's microbiology report, and minutes later gets back a shortlist of peptide candidates predicted to attack the drug-resistant pathogen on the slide. That is the loop Basecamp Research is now plugging into Anthropic's new Claude Science platform, via a metagenomic foundation model called EDEN, a roughly 28-billion-parameter AI trained on microbial DNA drawn largely from organisms most labs cannot grow (per a Basecamp-aligned technical write-up). Both EDEN and Claude Science are prediction tools, not finished drugs, and every candidate they propose still has to be synthesized and tested in a wet lab before it can reach a patient.
The human stakes are blunt. Drug-resistant infections contribute to roughly five million deaths a year in Basecamp's framing, against a backdrop where the World Health Organization maintains a "critical-priority pathogens" list of bacteria for which existing antibiotics are failing or absent, the agency's highest concern tier. EDEN's reported 97% in-silico success rate designing peptides active against those organisms is the kind of number that has to be read carefully. It reflects simulation-side predictions, not clinical outcomes, and according to Genetic Engineering & Biotechnology News, Basecamp co-founder Oliver Vince offered a candid caveat: "human-ready antibiotics at the click of a button is still a step away."
The structural shift is what got built underneath the chat. Claude Science bundles more than 60 scientific databases and connectors (genomics, proteomics, structural biology, chemistry tooling) behind a single reasoning layer aimed at life-science researchers (GEN). EDEN was trained, per Rewire, on metagenomic sequences drawn from uncultured microbes, organisms most labs cannot grow and most sequence libraries underrepresent. The combination matters because the historical bottleneck in antimicrobial peptide design has been data. The same drug-resistant families that appear in the WHO's critical list are also the ones with the fewest known sequences to learn from. A model that claims to have seen them anyway, then ranks candidates for a clinician who has never run a bioinformatics pipeline, is the actual story underneath the announcement.
The equity argument is the one Basecamp and its collaborators are leaning on hardest. Vince and EDEN program lead Phil Lorenz frame the work as an explicit agency transfer away from a handful of specialized model-building labs and toward the clinicians who actually know local resistance patterns. The thesis, that the people with the most agency over antimicrobial resistance are the ones who should be running candidate design, is the closest thing in the announcement to an actual argument (GEN). The technical track record behind the agency claim is still thin. According to Basecamp, the same model family has separately been used for programmable gene insertion in mammalian cells, a more speculative claim that has not been independently replicated in peer-reviewed form.
What to watch next is the verification chain that this announcement does not provide. In-silico peptide candidates from any foundation model still have to be synthesized, screened in a microbiology lab, tested for toxicity, and pushed into animal studies before any human exposure. None of that work disappears when the design step moves upstream. It just relocates. The open question the story earns is whether wet-lab groups, from academic core facilities to public-health labs in low-resource settings, can absorb a much higher throughput of AI-generated candidates, or whether the bottleneck simply shifts from sequence design to bench capacity.