Summary: An OpenAI account describes how César de la Fuente’s lab uses Codex and ChatGPT to bridge biology, chemistry, computer science, and engineering while searching biological sequences for antimicrobial candidates. The lab’s own deep-learning models perform sequence search, with the source claiming the initial candidate search can fall from years to hours. The article stresses that wet-lab validation, toxicity testing, manufacturing, regulation, and trials remain unavoidable.
Hunting for a molecule that could become an antibiotic is a needle-in-a-haystack problem, and the haystack keeps getting bigger. In an OpenAI account of the work, bioengineer César de la Fuente describes a lab that treats biology as an information system: DNA nucleotides and the amino acids of proteins work like an alphabet, and deep-learning models trained on biological sequences can scan genome and protein datasets for the patterns that suggest antimicrobial activity. The supplied material says this can cut the initial candidate search from years to hours.
Compute is not the constraint; translation between fields is
The lab de la Fuente runs is transdisciplinary by design, mixing biologists, chemists, computer scientists, and engineers. Writing code comes easily to some members while biology comes less so; for others it runs the other way. Codex and ChatGPT live inside that gap. According to the source, lab members lean on them to get oriented in unfamiliar topics, settle terminology, weigh methods drawn from other fields, and give structure to drug-discovery ideas. Biologists end up able to build programs; programmers end up able to take on biological problems. Notice how far this sits from the pitch that “AI finds drugs.” Here the tools serve as a translation layer among specialists who would otherwise need years to overlap.
Locating the tools inside the pipeline
Ranged alongside the lab’s own deep-learning models, ChatGPT and Codex get used to brainstorm hypotheses, write and refine code, process datasets, analyze results, and link ideas across disciplines. A subset of lab members leans on AI to download, organize, and pre-process large genome datasets. ChatGPT also lets members work in their native languages, which the source frames as lowering barriers to a faster workflow.
De la Fuente treats ChatGPT as a brainstorming partner, and describes the shared workspace as a collaborative sounding board fed by people who approach the same problems from different directions — bad ideas included. His caution is that accuracy always needs double-checking. That caveat matters for anyone building on model output: the sequence search belongs to the lab’s own deep-learning models, while the general-purpose tools carry the reasoning and the code around it.
Prediction is the inexpensive half
The source is explicit that identifying a promising candidate is not the same thing as producing a medicine. Candidates must be confirmed to kill the target microbe, dosed effectively, and tested against human cells. Chemists may optimize for potency, safety, or stability. Further tests cover toxicity thresholds, how readily microbes develop resistance, how the molecule moves through the body, and whether it can be manufactured reliably. Approved drugs still face regulatory review and clinical trials.
That is the reasoning behind de la Fuente’s argument that AI and laboratory biology have to advance together. Ground-truth experiments validate predictions, and he calls that critical for the life sciences in the years ahead. For builders, the practical read is that what AI compresses is the search and the coordination overhead, not the validation. If you are designing agent workflows for research, plan for the handoff into wet-lab confirmation — it is the part you cannot skip, and the part that decides whether a speedup is real.
Takeaways for anyone building research tools
Two points stand out. First, the highest-leverage use here is not autonomous discovery but lowering the cost of crossing disciplines, which is closer to the capability-building pattern in how FDEs build a team’s capability instead of its dependency. Second, the source does not specify which Codex or ChatGPT versions the lab uses, how many candidates have been validated, or what the hit rate looks like. None of those details appear in the supplied material, so the hours-versus-years framing should be read as the lab’s characterization of its search stage rather than an end-to-end result. The honest next step for a builder is to instrument precisely where model output gets checked against an experiment — the value either holds there or evaporates.