OpenAI on September 10 published a case study describing how the laboratory of University of Pennsylvania bioengineer César de la Fuente has wired ChatGPT and Codex into its antimicrobial discovery pipeline: using the models to brainstorm hypotheses, write and refine code, process genomic datasets and connect ideas across disciplines, while scanning the genomes of living and extinct organisms for molecules that might fight drug-resistant infections.
The background numbers deserve to go first. Bacterial antimicrobial resistance was associated with about 5 million deaths in 2021, a toll projected to roughly double by 2050. De la Fuente notes in the piece that humanity has not found a new class of antibiotics in 50 years, and that the traditional paths — modifying existing drugs, working familiar chemical classes — yield diminishing returns. His lab starts elsewhere: treat biology as an information system, where nucleotides and amino acids form an alphabet, then search the entire tree of life, including extinct genomes, for a needle in a genomic haystack.
By the article's account, the pipeline compresses the initial candidate search from years to hours. The lab's own deep-learning models recognize patterns in biological sequences; ChatGPT and Codex act as the glue — letting biologists who cannot code write analysis scripts, letting engineers who do not know biology wrangle genomic data, and letting team members work in their native languages. De la Fuente calls ChatGPT the lab's "communal brain."
[1]What the blog says — and what it doesn't
What it says: the workflow, the motivation, the efficiency narrative, and one crucial act of self-restraint — de la Fuente himself stresses that "ground-truth experiments are essential to validate AI predictions." A candidate must first prove in a dish that it kills the target microbe, then survive dosing, toxicity, resistance, pharmacokinetics, manufacturing, regulatory review and clinical trials. Any step can kill a candidate.
What it does not say is equally clear: the article names no molecule discovered by this pipeline, attaches no peer-reviewed paper, and offers no in vivo or clinical data. Independent commentators were quick to point this out — the current evidence is a corporate blog post. That does not make the direction wrong; the de la Fuente lab has a published track record on antimicrobial peptides from extinct organisms. But "AI discovers antibiotics" and "AI speeds up the manual labor of candidate search" are different claims, and only the latter is what this blog actually demonstrates.
[1][2]The product signal worth noticing
For OpenAI, the value of this piece is not the scientific conclusion but the scene-setting. Codex is a terminal agent built to write code; it is now shown inside a wet-lab workflow. In the same week, OpenAI also launched a financial-services edition of ChatGPT and an enterprise data agent. All three announcements share one narrative template: a general tool plus domain data plus a compressed unit of time — years to hours, days to minutes. The template sells well. In science it needs separate accounting: the time compression happens in candidate screening, while the cost centers of drug development — validation, toxicology, clinical trials — remain entirely untouched.
Drug-resistant bacteria do not read press releases. Whether this pipeline holds up has exactly one test: whether a molecule emerges from this search and survives the dish, the animal studies and the clinic to reach a prescription. Until then, the accurate sentence is that AI shrank the needle-hunt from years to hours. Whether the needle works is still the slow part.
[1]