OpenAI is using a new Applied AI case study to show how general-purpose models can fit into a scientific discovery workflow. The September 10 profile follows César de la Fuente, a Presidential Associate Professor at the University of Pennsylvania, and his lab’s search for antimicrobial molecules hidden in the genomes of living and extinct organisms.
The development is a workflow story rather than a drug announcement. OpenAI does not report a newly approved antibiotic, a human study, or a clinical result in the post. It describes how the lab combines its own deep-learning systems with ChatGPT and Codex to narrow a large computational search and help a multidisciplinary team move from questions to testable candidates.
Search the code of life
De la Fuente’s approach starts from the idea that biology is an information system. DNA nucleotides and the amino acids that make up proteins can be read as sequences, and some sequence patterns encode molecules with useful biological activity. The lab trains deep-learning models to recognize patterns across genome and protein datasets, then uses those models to prioritize sequences that may have antimicrobial potential.
That search can reach well beyond the organisms researchers typically sample in a laboratory. Digital databases make it possible to scan biological information from plants, animals, microbes, insects, water, soil, and extinct organisms. The practical challenge is not a lack of candidates but an excess of them: most sequences will not produce a molecule that is useful, safe, stable, or manufacturable.
OpenAI says the lab’s computational approach can reduce the initial search from years to hours. That is a claim about finding and ranking promising starting points. It does not mean an AI-selected sequence becomes a medicine without further work; the candidates still have to survive biological experiments and a long development process.
Codex as a bridge between disciplines
Alongside the lab’s specialized models, the team uses ChatGPT and Codex for the connective work around research. OpenAI says the tools help researchers brainstorm hypotheses, write and refine code, download and preprocess large genome datasets, analyze results, review unfamiliar concepts, and compare methods across fields.
That division of labor matters because de la Fuente’s group spans biology, chemistry, computer science, and engineering. A biologist may need help building a data pipeline, while a programmer may need a fast explanation of a biological method. OpenAI also says ChatGPT lets team members work in their native languages, lowering a practical barrier in a collaboration that depends on ideas moving between specialties.
The lab’s use of a shared ChatGPT workspace is presented as a sounding board rather than an autonomous scientist. Team members feed in both good and bad ideas, and de la Fuente continues to work with colleagues. He also cautions that AI output must be checked for accuracy—an important qualification when a plausible explanation or a working script can still encode a scientific mistake.
Experiments remain the gate
A computational hit is only the beginning. Researchers need to confirm that a candidate can kill the target microbe, measure how much is required, and test how it affects human cells. Later studies examine toxicity, the likelihood that microbes will develop resistance, how the molecule behaves in the body, and whether it can be made reliably. Candidates that clear those hurdles still face regulatory review and clinical trials.
Newsroom analysis
The useful signal is the handoff between model types and human expertise. Specialized sequence models do the high-volume biological screening; ChatGPT and Codex reduce the friction of coding, explanation, and collaboration around that screening; laboratory experiments decide whether a prediction survives contact with biology. The pattern is less about replacing a scientist than about making a small transdisciplinary team faster at the front end of a very long pipeline.
The limits are equally important. This is an OpenAI account of one lab’s process, not an independent evaluation of the models or evidence that a new antimicrobial has reached patients. The strongest near-term takeaway is that AI can help search biological data and organize research work at a scale that would be difficult to manage manually, while the evidentiary burden for a real treatment remains experimental, clinical, and regulatory.