ApexGO improves antibiotic candidates instead of only searching for them

What this is about

Researchers at the University of Pennsylvania presented ApexGO on May 13, 2026, a method that improves existing antibiotic candidates instead of only screening huge molecular libraries. The work appeared in Nature Machine Intelligence under the title “A generative artificial intelligence approach for peptide antibiotic optimization.”

The concrete reason this matters: in lab tests, 85 percent of molecules generated by ApexGO halted bacterial growth. According to Penn, 72 percent outperformed the peptides they were derived from. In mice, two candidates reduced bacterial counts at levels comparable to polymyxin B, a last-resort antibiotic for some resistant infections.

What ApexGO actually does

ApexGO does not start from nothing. It takes a promising but imperfect peptide and proposes small edits to its amino-acid sequence. A prediction model estimates whether each change is likely to improve antimicrobial activity. Then ApexGO moves to the next round.

This is precision tuning rather than treasure hunting. Many AI approaches in drug discovery search large libraries for possible hits. ApexGO acts later in the process: once a candidate roughly works, it helps optimize it. The team uses Bayesian optimization, a method that balances options that look promising with options that are uncertain but could teach the model something.

Why it matters

Antibiotic resistance is a basic medical problem because bacteria can adapt faster than new drugs reliably reach clinics. Peptide antibiotics such as polymyxin B and colistin matter because they remain last-line options against some multidrug-resistant Gram-negative pathogens.

ApexGO does not promise to skip clinical development. It targets an earlier, expensive bottleneck: finding better variants. If more lab-made candidates are genuinely stronger than their parent molecules, researchers can spend scarce wet-lab capacity more carefully.

In plain language

Imagine a bread recipe that already works but does not rise perfectly. ApexGO does not bake 10,000 random new loaves. It changes salt, water or resting time step by step and learns which change actually improves the result.

For antibiotics, that means not testing every possible amino-acid sequence. The system proposes targeted edits that have a better chance of passing the next lab test.

A practical example

A research team starts with 50 peptides, ten of which show weak activity against a pathogen. Instead of blindly synthesizing 5,000 variants, it asks ApexGO to suggest 20 targeted edits per peptide. If 85 percent of the made variants halt growth, 1,000 tests produce far more useful leads than a random search would.

That does not replace clinical trials. But it can make months of early lab work more productive because fewer candidates are tested only because nobody had a better way to prioritize them.

Scope and limits

• Mouse data is not patient data. Efficacy, dosing, toxicity and manufacturing still need further preclinical and clinical testing.
• Peptides can be harder to develop than classic small molecules because of stability, cost and delivery issues.
• A model can optimize against its own scoring system. Penn says the lab data partly reduces that concern, but it does not remove it entirely.

SEO & GEO keywords

ApexGO, University of Pennsylvania, Nature Machine Intelligence, peptide antibiotics, antibiotic resistance, antimicrobial peptides, Bayesian optimization, APEX, César de la Fuente, Jacob Gardner, drug discovery