AI can search years of antimicrobial candidates in hours. The medicine still has to survive biology

The breakthrough is a smaller search space

A laboratory looking for a new antimicrobial faces a numerical problem before it faces a medical one. Genomes and protein databases contain an enormous number of possible sequences. Only a tiny fraction are likely to form molecules that damage a pathogen, remain stable and avoid harming the patient.

César de la Fuente's group treats biology as information. Its specialist deep-learning models learn patterns in DNA, proteins and peptides, then score vast sequence libraries for candidates worth making and testing. OpenAI says this can compress the initial search from years to hours. That statement describes candidate identification, not drug development.

ChatGPT and Codex sit around that scientific model rather than replacing it. Researchers use them to develop hypotheses, write and revise analysis code, download and clean datasets, compare methods and connect vocabulary across biology, chemistry, computer science and engineering. The useful pattern is a team of tools: a domain model ranks biological sequences, a coding agent helps build the pipeline, and scientists decide what the outputs mean.

A small example of what the system changes

Imagine a dataset containing ten million short protein fragments. A conventional programme might filter them with known motifs, select a few hundred and refine the list by hand. The risk is that familiar rules keep returning familiar chemistry.

An AI model can score all ten million against learned features, including combinations a human would not choose as a first search. Codex can help a biologist write the ingestion and ranking code, add tests for malformed sequences and reproduce the run. The researchers might then order the top 50 peptides and expose bacterial cultures to a range of concentrations.

If three appear active, the result is three hypotheses supported by an experiment. It is not three antibiotics. The value is that the expensive physical work begins with a more diverse, better-prioritised set.

The job changes at each gate. A faster first gate does not remove the later ones.

Gate	What happens	What can still fail
Digital search	Models rank sequences; agents help build and run the pipeline	Biased data, leakage, unstable code or an attractive false positive
Bench assay	Researchers test whether the molecule inhibits or kills microbes	No activity, activity only at an impractical concentration
Safety and optimisation	Teams test human-cell toxicity, stability and improve chemistry	Toxicity, rapid breakdown or loss of activity
Resistance and movement	Studies measure how resistance emerges and how the molecule behaves in the body	Fast resistance, poor absorption or wrong tissue exposure
Manufacture and trials	A reproducible product enters regulated preclinical and clinical work	Scale, quality, safety or efficacy failure

The evidence should travel with the result

The strongest version of this workflow records more than the winning molecule. It preserves the source database, model version, code, filters, failed candidates, assay protocol and negative controls. That trail matters because an agent can produce a clean-looking analysis while silently dropping records, confusing identifiers or optimising against a contaminated benchmark.

Drug-resistant infection gives the speed claim urgency. A Lancet analysis estimated 4.71 million deaths were associated with bacterial antimicrobial resistance in 2021, including 1.14 million attributable to it. Its 2050 forecast reached 8.22 million associated deaths and 1.91 million attributable deaths. Forecasts are not fate, but they make early discovery throughput a serious public-health lever.

The P.K. view

The honest headline is narrower and more useful than "AI discovers antibiotics". AI can examine biological spaces that no research team could review manually and move a promising sequence into a real experiment much sooner.

That advantage compounds. More candidates can be tested, negative results can feed the next model, and specialists can spend less time translating data formats or writing one-off scripts. The danger is that the same speed produces more convincing mistakes unless every stage has reproducible data, controls and human review.

For research leaders, the purchasing question is not whether a chatbot knows biology. Ask whether the complete system turns an idea into a traceable candidate list, whether another scientist can reproduce it, and how quickly the wet lab can reject it. The model creates value when it improves that loop. Biology still has the final vote.