Somewhere in the genome of a Neanderthal, there may be a molecule that kills bacteria that modern medicine can’t touch. That’s not science fiction — that’s the working hypothesis of César de la Fuente, a Presidential Assistant Professor at the University of Pennsylvania, whose lab is now using OpenAI Codex and ChatGPT to systematically mine living and extinct genomes for antimicrobial candidates. It’s one of the most concrete examples yet of AI doing genuine scientific heavy lifting — not summarizing papers, not generating citations, but actually accelerating the discovery pipeline for drugs we desperately need.
Why Antimicrobial Resistance Is the Crisis Nobody’s Fixing Fast Enough
Let’s be honest about the stakes here. The World Health Organization has called antimicrobial resistance one of the top global health threats of the 21st century. Drug-resistant infections already kill around 1.27 million people annually — a number that’s expected to climb dramatically by 2050 if new treatments don’t materialize. The pipeline for new antibiotics has been, to put it bluntly, nearly dry for decades.
Part of the problem is economic. Antibiotics are taken for short courses, which makes them less profitable than drugs for chronic conditions. Big pharma largely abandoned the space. What’s left is a patchwork of academic labs, small biotech startups, and government-funded programs trying to fill a gap that’s measured in millions of deaths.
De la Fuente’s lab sits in that gap — and they’ve been pursuing an unusual angle. Rather than synthesizing entirely new compounds from scratch, they’re looking at what evolution already figured out. Living organisms have been fighting bacteria for billions of years. Their genomes contain encoded proteins and peptides that, in many cases, have antimicrobial properties we haven’t catalogued yet. And extinct species — mammoths, Neanderthals, ancient fish — carry sequences that modern organisms no longer express, which means entirely different chemical structures that modern bacteria have never encountered and have no resistance to.
The problem? There are billions of genomic sequences to sift through. Before AI, that was a bottleneck that could swallow entire careers.
What Codex and ChatGPT Actually Do in the Lab
Here’s where it gets specific, because the actual workflow matters. De la Fuente’s team isn’t just asking ChatGPT to brainstorm antibiotic ideas. The integration is more technical than that, and more interesting.
OpenAI Codex — the code-generation model that powers a lot of developer tooling and is embedded in platforms like GitHub Copilot — is being used to write and refine the bioinformatics scripts that search genomic databases. Researchers describe tasks that would have taken days of manual scripting work being turned around in hours. The model writes Python pipelines for scanning sequences, filtering candidates by structural properties, and preparing datasets for further analysis.
ChatGPT plays a different role: it functions more like a knowledgeable research assistant. Lab members — including graduate students and postdocs who aren’t necessarily expert programmers — use it to understand unfamiliar code, debug pipelines, interpret outputs, and think through experimental design. One of the most telling details from the lab’s account is that ChatGPT helped democratize access to computational tools across the team. You don’t need to be a bioinformatics specialist to participate meaningfully in the search anymore.
The practical output of this AI-assisted pipeline includes:
- Automated scanning of genomic databases, including sequences from extinct species like Neanderthals and woolly mammoths
- Identification of candidate antimicrobial peptides based on structural and chemical criteria
- Faster iteration on experimental hypotheses — moving from genomic data to synthesizable candidates in compressed timelines
- Code generation that handles repetitive scripting tasks, freeing researchers for higher-level scientific decisions
- Natural language interfaces that let non-programmers participate directly in computational workflows
The candidates that make it through this computational filter then get synthesized and tested in the lab the old-fashioned way — biological assays, testing against resistant bacterial strains, measuring efficacy and toxicity. AI accelerates the front end; the bench science still closes the deal.
The “Extinct Genome” Angle Is Genuinely Novel
I want to dwell on this for a moment because it’s the part that most reporting glosses over. Mining living genomes for antimicrobial peptides isn’t new — researchers have been doing it for years. What de la Fuente’s lab has pushed further is the idea that extinct organisms represent an untapped pharmacological library.
The logic is sound. Bacteria evolve resistance to drugs they encounter. If a peptide comes from a Neanderthal genome — a sequence that’s been absent from the biosphere for roughly 40,000 years — modern drug-resistant bacteria have had zero exposure to it. Resistance mechanisms that evolved against contemporary compounds simply don’t apply. That’s a real advantage, not a theoretical one.
This approach has already produced results. The lab has previously published on mining the human proteome for antimicrobial candidates, and their work has generated peptides with genuine activity against resistant pathogens. The AI tooling is scaling that work up significantly.
What This Means Beyond One Lab
De la Fuente’s lab is a proof of concept, but the implications reach further. This is what it looks like when AI tools that were built for software developers — Codex was originally designed to help people write code — find second lives in entirely different domains. Nobody at OpenAI designed Codex specifically for genomic pipeline scripting. That’s a researcher figuring out that a general-purpose tool solves a real problem in their field.
This kind of lateral application is going to keep happening. We’ve already seen GPT-5.6 Sol running real quantum experiments at MIT, another case of AI crossing from one domain into hard experimental science. The pattern is the same: AI handles the computational bottleneck, humans handle the experimental design and interpretation, and the combined output exceeds what either could do independently.
For the drug discovery industry more broadly, the question is whether this model scales. Pharma companies have been investing in AI-driven drug discovery for years — companies like Recursion Pharmaceuticals, Insilico Medicine, and Exscientia have raised billions on that promise. What de la Fuente’s work demonstrates is that even a single academic lab, armed with general-purpose AI tools available to anyone with an API key, can run a meaningful discovery pipeline. The barrier to entry has dropped considerably.
The Bottleneck That Remains
None of this eliminates the hard parts of drug development. Identifying a candidate peptide is the beginning of a process that takes years and costs hundreds of millions of dollars before anything reaches a patient. Clinical trials, regulatory approval, manufacturing at scale, pricing and distribution — AI doesn’t touch most of that. And the economic dynamics that drove pharma away from antibiotics in the first place haven’t fundamentally changed.
What’s shifted is the discovery rate. If AI tools help labs like de la Fuente’s identify ten times as many viable candidates in the same timeframe, the probability that some of them make it through the full pipeline improves. It’s a numbers game. More shots on goal. That matters, even if it doesn’t fix everything.
Codex’s Role in Scientific Computing Is Underappreciated
There’s something worth flagging here about Codex specifically. Most of the conversation around AI coding tools focuses on software engineering productivity — 1Password saw a 21% boost in engineering output using Codex, for instance. But the scientific computing use case might ultimately matter more. Researchers across biology, chemistry, physics, and climate science spend enormous amounts of time writing and debugging code for data analysis. They’re often not trained programmers. A tool that closes that skill gap doesn’t just save time — it changes who can do the work.
De la Fuente’s lab is a good example of that. Graduate students running complex bioinformatics pipelines without needing to be fluent in every aspect of the code — that’s a meaningful change in how science gets done.
Key Takeaways
- César de la Fuente’s lab at UPenn is using OpenAI Codex and ChatGPT to search living and extinct genomes for antimicrobial peptides that could fight drug-resistant bacteria
- Codex handles bioinformatics scripting and pipeline automation; ChatGPT serves as a research assistant for interpretation, debugging, and experimental planning
- Mining extinct genomes — from Neanderthals, mammoths, and other vanished species — offers a structural advantage: bacteria have no evolved resistance to compounds they’ve never encountered
- AI accelerates the front end of discovery; wet lab validation and the full drug development pipeline still require traditional methods and timelines
- This is a template for how general-purpose AI tools, designed for one domain, get repurposed by researchers to solve entirely different problems
- The economic problem driving antibiotic underfunding hasn’t been solved — but dramatically higher candidate throughput changes the odds
Frequently Asked Questions
What is César de la Fuente’s lab actually discovering?
The lab is identifying antimicrobial peptides — short protein fragments — encoded in genomic sequences from a wide range of organisms, including extinct species like Neanderthals and woolly mammoths. These candidates are then synthesized and tested in the lab for activity against drug-resistant bacteria. Several have shown genuine efficacy in early testing.
How is Codex different from just using ChatGPT for research?
Codex is specifically optimized for generating and editing code, making it effective for writing the bioinformatics scripts that process and filter genomic data. ChatGPT functions more as a general-purpose research assistant — helping team members understand outputs, plan experiments, and interpret results in plain language. They serve complementary roles in the workflow.
Does this mean AI is solving antibiotic resistance?
Not exactly. AI is accelerating the candidate discovery phase, which is genuinely valuable — but drug development involves years of clinical trials, regulatory review, and massive capital investment before anything reaches patients. The economic and regulatory barriers that have kept antibiotic development underfunded for decades are still very much in place. What changes is the volume and speed of early-stage discovery.
Could other labs replicate this approach?
Yes — and that’s arguably the most significant implication. The tools de la Fuente’s lab is using, including OpenAI Codex and ChatGPT, are commercially available. The specific expertise in antimicrobial peptide biology and genomic analysis is harder to replicate, but the AI infrastructure is accessible to any well-resourced research group. I wouldn’t be surprised if similar pipelines start appearing across biology and chemistry labs in the next two to three years.
The de la Fuente lab’s work points toward something real: AI tools quietly reshaping what small research teams can accomplish, in fields where the stakes are measured not in software releases but in lives. As OpenAI pushes harder on policy frameworks for responsible AI deployment, use cases like this one — concrete, high-impact, and methodologically transparent — are exactly the kind of evidence that shapes how regulators and the public understand what these tools are actually for. The bigger question now is whether the funding and institutional structures around drug development can move fast enough to catch up with what the science is starting to make possible.
