A Penn Lab Is Running Antibiotic Discovery Through Codex and ChatGPT

OpenAI published a case study on Sept. 10 describing how César de la Fuente, a professor at the University of Pennsylvania, uses Codex and ChatGPT alongside his lab's own machine-learning models to search genomes for new antimicrobial compounds.

De la Fuente, whose title is Presidential Associate Professor at Penn, leads the university's Machine Biology Group, according to insideai.news, which reviewed OpenAI's post. The lab's approach treats genetic sequences as a kind of language: "The nucleotides that make up DNA, and the amino acids that make up proteins and peptides are sort of like an alphabet," de la Fuente said, according to the case study. "Thinking about biology as information enabled us to develop methods that can begin to decipher the organizing principles of life."

What the AI tools actually do

The lab's own deep-learning models handle the core scientific task: recognizing patterns in biological sequences across genome and protein databases to flag candidate antimicrobial peptides. Codex and ChatGPT sit around that process, according to the case study, used for writing and refining the code that processes datasets, brainstorming hypotheses, and translating findings between the lab's biologists and programmers. "Our ChatGPT workspace is receiving input from all these different people that think differently about the problems," de la Fuente said.

The case study credits the combination with cutting the initial search for candidate molecules from years to hours of computation, though it does not specify how many candidates a given run screens or how that time was measured. De la Fuente is explicit that computation is not the last step: "Ground-truth experiments are essential to validate AI predictions."

The track record behind the case study

De la Fuente's group has been doing this kind of AI-driven search since before Codex existed in its current form. A February profile in MIT Technology Review describes his team mining DNA from extinct organisms, including Neanderthals, woolly mammoths and giant sloths, for antimicrobial peptides, work done with collaborators James Collins at MIT, Jonathan Stokes at McMaster University and James Zou at Stanford. That reporting put his peptide library at more than 1 million genetic sequences and named several resurrected compounds, including mammuthusin-2 and mylodonin-2, drawn from mammoth and sloth genomes respectively.

Figure	Source
5 million deaths linked to bacterial resistance, 2021	insideai.news, citing OpenAI case study
Projected to roughly double by 2050	insideai.news, citing OpenAI case study
50 years since the last new antibiotic class reached patients	insideai.news, citing OpenAI case study
Over 1 million peptide sequences in de la Fuente's library	MIT Technology Review

Why the setting matters

The case study is a vendor-published account of one customer's workflow, not an independent evaluation of Codex or ChatGPT's contribution to the results. OpenAI does not disclose in the post how much of the lab's output depends on its models specifically versus the lab's own deep-learning tools, which predate the collaboration and are credited separately in de la Fuente's published work. The antimicrobial-resistance death toll and timeline figures in the case study are OpenAI's own citations, not independently verified in this report. That attribution question, how much credit an AI tool gets versus the researcher directing it, is the same one 25 Fields Medalists raised about AI-assisted mathematics proofs this month.

The broader pattern, an AI tool handling code and cross-disciplinary translation rather than the core scientific model, echoes how GitHub has framed the tooling around coding agents as doing work distinct from the underlying model itself.