AI Antibiotics: How MIT and McMaster University are Searching for New Drugs Against Superbugs

Bacterial resistance to antibiotics directly impacts the capabilities of modern medicine. According to the latest data from the World Health Organization, bacterial antimicrobial resistance was associated with more than 4.7 million deaths worldwide in 2021. In 2023, approximately one in six laboratory-confirmed bacterial infections was resistant to antibiotics. From 2018 to 2023, resistance increased by more than 40% in the “pathogen-antibiotic” combinations tracked by the WHO.

The traditional search for new drugs requires screening a vast number of chemical substances. Machine learning is changing this process. An algorithm can rapidly evaluate millions of structures and identify molecules that are more likely to inhibit a specific bacterium.

Primarily, AI is used for such tasks:

• searching for antibacterial properties in already known molecules;
• predicting the activity of substances against resistant bacteria;
• filtering out potentially toxic compounds;
• searching for new mechanisms of antibiotic action;
• generating molecules that were previously absent in chemical libraries;

This does not mean that artificial intelligence alone creates ready-made drugs. After computational selection, each molecule still needs to be synthesized, tested in the laboratory, and subjected to preclinical and clinical trials.

Halicin – How MIT Found a Promising Antibiotic Among Millions of Molecules

One of the best-known examples is halicin. In 2020, MIT researchers applied a deep learning model to search for substances with antibacterial activity. As reported by MIT News in their article about the halicin discovery, this molecule was previously considered a potential treatment for diabetes, but the model uncovered a completely different property.

After training the system, researchers used it to analyze over 100 million chemical compounds from the ZINC15 database. The computational screening took three days.

The result was illustrative:

1. The algorithm analyzed more than 100 million potential compounds.
2. Researchers selected 23 promising molecules for laboratory testing.
3. Eight substances demonstrated antibacterial activity.
4. Halicin became one of the most interesting candidates due to its unusual mechanism of action.

In laboratory tests, halicin was effective against several bacteria, including Acinetobacter baumannii. In a mouse infection model, ointment containing this substance eliminated A. baumannii within 24 hours. However, this was a preclinical result, not proof of the drug’s efficacy in humans.

Abaucin – The Discovery of McMaster University and MIT Against Acinetobacter baumannii

In 2023, the team from McMaster University together with MIT researchers presented another result – abaucin. The detailed scientific work was published by the journal Nature Chemical Biology. Scientists first tested about 7,500 molecules for the ability to inhibit Acinetobacter baumannii, then used the obtained data to train a neural network.

Acinetobacter baumannii is of special interest due to its ability to acquire multiple drug resistance and cause complex hospital-acquired infections.

AI helped identify abaucin – a substance with a relatively narrow spectrum of activity specifically against A. baumannii.

Why Narrow Spectrum Activity Matters

A typical broad-spectrum antibiotic can affect not only the disease-causing agent but also other bacteria. A narrow-spectrum drug potentially allows for more targeted action.

Researchers established several important properties of abaucin:

• The molecule showed activity against A. baumannii;
• Its action is associated with disrupting the transport of bacterial lipoproteins;
• The mechanism involves the protein LolE;
• The substance controlled A. baumannii in a mouse infected wound model;

The authors of the study explicitly call abaucin a promising lead molecule for further development, not a finished drug.

Generative AI Already Creates New Molecules from Scratch

The next step is considerably more complex. AI can not only search for the desired structure among known substances but generate new molecules.

In 2025, MIT researchers used generative AI to design antibiotics against resistant Neisseria gonorrhoeae and multi-resistant Staphylococcus aureus. According to MIT’s report on generative antibiotic research, the system created over 36 million potential compounds.

The scientific results were published in the journal Cell. The original paper, “A generative deep learning approach to de novo antibiotic design,” is available on PubMed. Of the 24 synthesized compounds, seven demonstrated selective antibacterial activity.

To combat N. gonorrhoeae, researchers initially worked with approximately 45 million chemical fragments. After multi-stage selection, the promising fragment F1 became the basis for generating about 7 million new candidates. One was named NG1.

For Staphylococcus aureus, the model acted even more freely and generated more than 29 million potential molecules. Researchers managed to synthesize 22 selected substances, six of which showed strong activity. The candidate DN1 reduced bacterial load in a mouse MRSA skin infection model.

Enterololin – AI Helped Identify the Mechanism of Antibiotic Action

In 2025, MIT and McMaster University demonstrated another possibility of AI. Algorithms can not only help find molecules but also explain which bacterial structure they target.

According to MIT CSAIL on the enterololin study, scientists worked with over 10,000 molecules and used the DiffDock model to analyze the mechanism of action of the narrow-spectrum substance enterololin.

Within minutes, DiffDock predicted the interaction of enterololin with the LolCDE protein complex, which is involved in lipoprotein transport in bacterial cells. The scientists did not accept the AI prediction as definitive evidence. They experimentally verified it using resistant mutants, RNA sequencing, and CRISPR.

A task that could traditionally take approximately 18 months to two years or longer took about six months in this case.

When AI Antibiotics Might Become Real Medicines

Search queries like “new antibiotics 2026,” “AI antibiotics against superbugs,” or “drugs created by artificial intelligence” may give the false impression that such drugs are already ready for use. In reality, halicin, abaucin, NG1, DN1, and enterololin remain research candidates.

After discovery, a molecule must go through several stages:

1. Optimization of structure and synthesis method.
2. Toxicity testing.
3. Pharmacokinetics and dosage studies.
4. Preclinical trials.
5. Clinical trials involving humans.
6. Regulatory evaluation before market release.

AI primarily accelerates the early stages – search, selection, design, and mechanism analysis. Without biological validation, an algorithm’s prediction cannot be considered proof of safety or effectiveness.

The technology also demonstrates how much broader the application of artificial intelligence is becoming. On Poshuk.info, there is a separate look at practical AI use in 2026, comparing the capabilities of modern AI systems for different tasks. And when future technologies transition from research to practical healthcare services in Ukraine, requirements for medical practice licensing become particularly important.

Research by MIT and McMaster University shows a major shift: instead of manually screening thousands of substances, scientists gain a tool to work with tens of millions of structures. But AI here does not replace the laboratory; it identifies precisely where to look.