The Pulse
Stanford AI narrows 1.7 million polymers to 10 antibiotic leads
Stanford researchers trained an AI model to search 1.7 million polymer candidates for molecules that mimic bacteria-killing peptides. Ten candidates outperformed expectations against E. coli, with one showing particular promise against biof
AI.info Team ·
Stanford researchers have reduced a library of 1.7 million potential polymers to 10 antibiotic candidates using an AI system trained first on antimicrobial peptides. All 10 candidates performed better than expected against Escherichia coli, and one showed particular effectiveness against biofilms, the protective microbial communities that often withstand conventional treatment.
The work, published September 21, 2026, in Matter, points to a different route for antibiotic discovery: instead of searching for molecules that block a specific bacterial protein or pathway, the researchers sought polymers that physically damage the cell membrane. Stanford described the findings in a Stanford Report article.
Why Stanford looked beyond peptides
Antimicrobial peptides already offer a model for attacking bacteria without relying on a narrow biochemical target. Many can move toward a bacterial membrane, disrupt it and kill the cell from the outside. That physical mechanism may make it harder for bacteria to develop resistance than it is to alter a drug target, disable a compound or block its entry.
“Antimicrobial peptides are chemically able to get very close to and disrupt the cell membrane, killing the bacteria,” Shoshana Williams, a former Stanford Engineering graduate student who recently earned her PhD, said in the Stanford report. “Importantly, they don’t need to get inside the cell to work, like a typical drug would. Nor do they work on one specific protein or pathway, like drugs do.”
Peptides have practical drawbacks, however. They can be short-lived and expensive to synthesize. The Stanford team instead focused on polymers: long, chain-like molecules that are less costly to make, easier to handle and less prone to degradation.
Training AI where polymer data was scarce
The researchers faced a familiar problem in computational drug discovery: the available data did not match the chemical space they wanted to search. Large collections of antimicrobial peptide data exist, while comparable polymer datasets are much smaller.
“The dataset of antimicrobial polymers simply was nowhere near big enough,” Williams said. The team responded by training its models on the chemical properties of antimicrobial peptides and then transferring that information to polymer candidates.
The process did not rely only on the candidates that received the highest average prediction. Researchers also examined where their models disagreed most. They synthesized and tested the 20 candidates that produced the greatest uncertainty, then fed the experimental results back into the system. That approach gave the models additional information about which chemical features were associated with antimicrobial activity.
“The primary intellectual leap was training on the antimicrobial peptide data first and then to use that data to make predictions on the polymers, an entirely different class of molecules,” Eric Appel, a Stanford professor of materials science and senior author of the study, said.
Ten candidates survive the laboratory test
After the uncertainty-driven testing phase, the system selected 10 polymer candidates for synthesis and biological testing. Each performed well above the researchers’ expectations in tests against E. coli, a Gram-negative bacterium with two membranes.
One candidate stood out in tests against biofilms. Such communities surround themselves with a protective matrix and are often difficult to eliminate with standard antibiotics. The study also found that one polymer could work together with an existing clinical drug regimen to improve the eradication of biofilm-associated E. coli, according to the study summary.
“These 10 candidates are among the most potent antimicrobial polymers ever reported, as far as I’m aware,” Williams said.
A physical attack on bacterial membranes
The candidates are designed to mimic the relevant chemical characteristics of antimicrobial peptides rather than reproduce their exact structures. Their proposed mechanism is membrane permeabilization: the polymer reaches the bacterial surface, interacts with the membrane and creates damage that compromises the cell.
“The peptides permeabilize the microbes … they literally rip holes in the cell membrane to kill them,” Appel said. “It’s much harder for a bacterium to change the entire lipid structure of its membrane or the electrical charge of its surface than to learn to reject a chemical drug or turn off its narrow pathway.”
Stanford says the researchers tested the approach primarily against E. coli, but the polymers also showed activity against Staphylococcus aureus, a Gram-positive bacterium. The result does not establish that the candidates are ready for human use; the reported work is an early demonstration that an AI-guided design process can identify molecules for laboratory testing.
From broad-spectrum candidates to targeted drugs
Appel and Williams say the method could eventually support polymers designed for particular bacterial strains. A targeted compound could, in principle, attack an infection while limiting effects on human cells and beneficial bacteria, though the Stanford report does not claim that the current candidates have achieved that selectivity.
“We really do need better drugs,” Appel said. “This new process opens a promising path to identifying novel antibiotics that work in new and different ways to treat serious and complicated infections and combat resistance.”
The immediate result is narrower and more concrete: an AI model searched 1.7 million polymer structures, used laboratory data from its most uncertain predictions to improve its selection process, and produced 10 candidates that exceeded expectations against E. coli. The next questions concern toxicity, performance in animals and whether the membrane-disrupting polymers can retain their activity without harming human tissue.