Codex can help researchers find molecules, but it does not create a finished drug
Published: · Sergey Zhadov
A headline saying that AI found a new medicine is easy to click. The reality is much less magical. In César de la Fuente's laboratory, specialist neural models accelerate the first screening of biological sequences, while Codex and ChatGPT help with code, data, and hypotheses. They do not replace test tubes, safety checks, clinical trials, or regulators.
The main biological work is not done by a chat window
The group searches genomes from living and extinct organisms for short molecules. Its specialist deep-learning models look for useful patterns in protein and peptide sequences. A huge digital archive becomes a shorter list of candidates that may be worth taking into the laboratory.
ChatGPT and Codex work alongside that pipeline in a different role. Team members use them to discuss hypotheses, write and revise code, load and clean datasets, and explain terminology across biology, chemistry, and computing. Calling this an autonomous invention of an antibiotic would be false.
Only the first search shrinks from years to hours
An algorithm can review a volume of data that no person could inspect manually. It finds related patterns and raises candidates with the desired features. This saves time because researchers spend fewer reagents and less attention on obviously weak options.
The ranking is still a hypothesis. It does not show whether a molecule kills the target microbe, what concentration is required, or what it will do to human cells. At that point the screen ends and wet-lab work begins.
Every candidate faces a long series of gates
Researchers first test the molecule against a real pathogen and healthy cells. They then examine toxicity, microbial resistance, stability, behavior in the body, and whether production is practical. Failure at any stage can remove an attractive candidate from the programme.
Even encouraging early experiments do not make a medicine available to patients. Further studies, regulatory review, and clinical trials are still required. This is why the honest word is candidate, not drug, and why this article cannot be used as treatment advice.
Useful AI leaves an auditable trail
The valuable part of scientific AI is not the conversation itself but a reproducible process. Researchers need to preserve the code version, data source, filtering parameters, and reasons a molecule advanced. Then the assistant accelerates work without hiding uncertainty.
The same rule applies to ordinary analytics: a machine sorts and suggests, while a specialist verifies and owns the conclusion. The higher the cost of an error, the less room there is for confident prose without an experiment.