Researchers Are Using Codex and ChatGPT to Search for New Antimicrobials: What AI Actually Does

*This analysis covers OpenAI's September 10, 2026 profile of César de la Fuente's University of Pennsylvania lab and supporting scientific literature. OpenAI's article is a vendor-published case study, not a controlled evaluation of productivity or evidence that ChatGPT discovered an approved drug.*

César de la Fuente's lab uses ChatGPT and Codex alongside specialist deep-learning systems to brainstorm hypotheses, write and refine code, prepare genome datasets, analyze results, and connect concepts across biology, chemistry, engineering, and computer science. The specialist models scan biological sequences and prioritize molecules that might have antimicrobial activity.

AI can shrink the initial computational search from years to hours, according to the lab and OpenAI. But a predicted molecule is only a candidate. Researchers must synthesize it, test whether it kills the target microbe, measure toxicity and resistance, study how it behaves in the body, establish manufacturability, and eventually complete regulatory review and clinical trials.

They are general-purpose collaborators around the scientific pipeline, not substitutes for the lab's antimicrobial prediction models. Team members use them to explain unfamiliar terminology, compare methods, organize ideas, write data-processing programs, and work across disciplinary and language boundaries.

That distinction is essential. A coding assistant can accelerate a pipeline or help a biologist express an analysis in code. It does not independently establish that the input data is representative, the method is biologically valid, the implementation is correct, or a candidate is safe.

The lab treats genomes and proteomes as searchable information. Specialist models learn patterns associated with antimicrobial peptides, then screen very large collections of biological sequences from humans, microbes, living species, and extinct organisms. The output is a smaller candidate list that researchers can afford to investigate experimentally.

Published work from the group includes machine-learning searches across ancient proteins and generative methods for optimizing peptide antibiotics. These studies support the broader method, but they should not be read as independent validation of every claim in OpenAI's customer story or of the incremental contribution made by ChatGPT and Codex.

Not in the sense most patients mean. OpenAI's new article describes how tools are used and points to a research program that has reported promising antimicrobial candidates. It does not announce an approved medicine, a completed clinical trial, or a controlled comparison showing how much Codex and ChatGPT improved scientific outcomes.

Discovery has many gates. A candidate must show activity at a practical concentration, avoid damaging human cells, remain stable, resist rapid microbial adaptation, reach the relevant tissue, be manufacturable, and outperform available options. Most early candidates do not become drugs.

Use general assistants for bounded work whose output can be checked: drafting scripts, documenting a pipeline, translating terminology, creating test cases, or generating hypotheses to rank. Preserve data provenance, pin software and model versions, review generated code, test against known controls, and record which claims came from a model.

Keep experimental ground truth outside the assistant. A credible workflow separates ideation, computational prediction, code verification, laboratory validation, and clinical evidence. It also prevents unpublished sequences, patient information, credentials, or partner data from entering a service without appropriate terms and approval.

This case study is a strong illustration of AI's near-term scientific value: reducing search and coordination costs across a multidisciplinary team. It is not proof of autonomous drug discovery, and the vendor profile does not quantify ChatGPT's causal contribution.

The important trend is the combination of general-purpose assistants, specialist biological models, and physical experiments. AI can help scientists explore more candidates and write better tooling; laboratory results must still be allowed to say no.