How researchers are pairing Codex and ChatGPT with AI-driven antibiotic discovery
Every year, resistant bacteria pick up new tricks — a gene here, a mutation there — and slip past whatever antibiotic we last threw at them. It's not a niche worry anymore.
Public health researchers have been saying for years that drug-resistant infections could become one of the biggest killers on the planet, and part of the reason is embarrassingly simple: we've barely found any new classes of antibiotics in decades. The old way of hunting for one — dig up soil, isolate compounds, test them one at a time — can eat years before a single candidate is even ready for the next stage.
Computational biology has been quietly eating into that timeline for a while. What's new is who's doing the computing. In 2026, general-purpose tools like ChatGPT and Codex, OpenAI's coding agent, started showing up in that process too.
The lab that's already proving it works
César de la Fuente runs the Machine Biology Group at Penn, and his approach starts from a strange but useful idea: biology is just code. DNA has four letters. Proteins and peptides are strings of twenty. If you can search text, you can search a genome.
It's paid off. His lab's first big project mined the human proteome — every protein your body makes — and found antimicrobial peptides buried inside proteins that had nothing to do with the immune system.
A much bigger follow-up scanned genomes and metagenomes from tens of thousands of organisms worldwide, reportedly the largest antibiotic discovery effort ever attempted, and turned up close to a million candidates. About 100 of the strongest got synthesized and tested against live bacteria. Seventy-nine percent killed at least one microbe.
Since then the group has mined the proteomes of Neanderthals and Denisovans for antibiotic leads, and pulled a new antimicrobial, prevotellin-2, out of an ordinary gut microbe. De la Fuente talks about compressing searches that used to take years down to hours.
None of that ran on ChatGPT. His lab built its own system for the actual prediction work — a family of models called APEX, ApexGO, ApexDuo, and ApexOracle, trained to guess a peptide's function straight from its sequence.
Codex and ChatGPT don't touch that part. What they're doing more and more is everything around it: writing the code that moves data through the pipeline, catching researchers up on literature, drafting the paper at the end. The biology stays with the specialized model.
Why everyone's suddenly talking about this
In late July 2026, OpenAI gave away access to its best tools — Codex included — to as many as 100,000 scientists, mathematicians, and engineers, through a program called ChatGPT for Academic Researchers. It launched with around 10,000 seats at places like the Institute for Advanced Study and France's École normale supérieure, with more to come through 2027.
The package ships with over 75 built-in life-science skills — genetics, genomics, sequencing, protein modeling, drug discovery — and OpenAI says researchers are already using both tools across most of a project: framing hypotheses, reading literature, writing code, working through data.
Antimicrobial discovery fits that mold almost perfectly. A little biological insight, a lot of grinding software work — parsing files, fixing mismatched data formats, wiring a sequence database into a prediction model, running a search that used to be done by hand. It's the kind of grunt work a coding agent handles well, and you don't need Penn's resources to get the benefit.
What a search like this actually looks like
None of what follows is a documented case — no named scientist, no specific paper. It's a composite, built around a target that already has real precedent: multidrug-resistant Klebsiella pneumoniae, one of the WHO's critical-priority resistant bacteria.
The choice isn't arbitrary. In 2021, IBM researchers ran exactly this kind of search on K. pneumoniae: a generative model designed candidate peptides, and within 48 days the team had synthesized and tested 20 of them.
Two worked, showing real potency against several pathogens including drug-resistant K. pneumoniae, with low toxicity in mice. Two months from computer to validated hit, for a target that normally takes years — that's the kind of speed smaller labs are chasing when they don't have IBM's machine-learning team behind them.
Picture a researcher with a hunch: peptides from a specific family of soil bacteria might be potent against Gram-negative pathogens like K. pneumoniae without being toxic. Before writing a line of code, they'll spend an hour or two in ChatGPT catching up on recent papers, comparing classification methods, poking holes in the idea. Nothing here invents new biology — it just compresses a day of reading into one sitting. The researcher still decides what's worth chasing.
Then the actual pipeline. This is Codex's job. The researcher describes it in plain English: parse this batch of genome files, translate the reading frames, keep only peptides between 8 and 50 amino acids with a positive charge and a mix of water-loving and water-repelling regions, then score the survivors with an activity model. Codex writes it.
It rarely works the first time — real genome data is full of ambiguous residues, spots where the sequencer couldn't confidently call an amino acid, and a naive filter chokes on them. Telling Codex the thing crashed is usually enough:
def is_candidate(peptide):
if not peptide or "X" in peptide:
return False # skip sequences with ambiguous residues
length_ok = 8 <= len(peptide) <= 50
charge_ok = net_charge(peptide) > 0
has_hydrophobic = any(aa in HYDROPHOBIC for aa in peptide)
has_hydrophilic = any(aa in HYDROPHILIC for aa in peptide)
return length_ok and charge_ok and has_hydrophobic and has_hydrophilicWrite it, run it against messy data, hit a wall, fix it, run it again. That loop is where a coding agent actually earns its keep, not in whatever clean draft it hands you first.
Once it works, point it at something bigger — public databases covering tens of thousands of organisms — because Codex can run several tasks in parallel and doesn't get tired. A search that once needed a dedicated engineering team now needs one person and a few overnight runs.
What comes out is a long, ugly spreadsheet: tens of thousands of scored candidates, most of them junk or duplicates. Back to ChatGPT to sort through it — group by predicted mechanism, flag the obvious artifacts, boil it down to a shortlist short enough to bring to a lab meeting.
Then the part no AI touches. Out of that list, maybe a few dozen to a hundred peptides, going by published examples, actually get synthesized and thrown at real bacteria in a dish. Slow, expensive, and there's no way around it.
Whatever survives that gets written up, and that's the last place these tools show up: a methods section, a README so someone else can rerun the pipeline in six months, a first pass at a manuscript. A human rewrites all of it before it goes anywhere near a journal.
The catch
Call this "AI discovers antibiotics" and you'd be wrong. What's changed is that a search which used to need a small team of programmers and months of scripting can now get built and rerun by one biologist in a few days.
The researcher spends less time on plumbing and more time on the stuff that actually needs a scientist: which hypothesis to chase, which candidates deserve scarce lab time, what the data is actually saying.
Codex and ChatGPT are still just assistants. They'll state a wrong mechanism with total confidence if you let them, and any code they write needs a human to actually read it, not just run it.
The real prediction work — will this peptide be active, will it be toxic, is it worth synthesizing — still runs on specialized systems like APEX or IBM's generative models. A candidate isn't an antibiotic until someone's tested it in a dish, and that part hasn't sped up at all.
What has sped up is everything before that: how fast a researcher can get from a hunch to a list of things actually worth testing.
Further reading
OpenAI launches ChatGPT for Academic Researchers — eWeek, July 30, 2026
Exclusive: OpenAI offers 100,000 academics free ChatGPT access — Axios, July 29, 2026
Penn researchers show encrypted peptides could be wellspring of natural antibiotics — Penn Engineering
The scientist using AI to hunt for antibiotics just about everywhere — MIT Technology Review, February 2026
AI discovers nearly 1 million new antibiotics — Warp News
Accelerated antimicrobial discovery via deep generative models and molecular dynamics simulations — Das et al., Nature Biomedical Engineering, March 2021
Comments (0)
Login to post a comment.