New Horizons in Biotechnology and AI: Screening Existing Drugs for Resistant Pneumococcus
A new study shows how artificial intelligence can turn a large catalogue of existing medicines into a focused shortlist of possible antibacterial treatments. Researchers screened 6,747 small-molecule drugsĀ against Streptococcus pneumoniae, a pathogen responsible for pneumonia, meningitis, ear infections and other diseases. Eleven compounds were selected for laboratory testing, and nine reduced bacterial growth in vitro. One of the strongest candidates also worked against multidrug-resistant pneumococcal strains. 1Ā 2
Why pneumococcus is a difficult target
Streptococcus pneumoniaeĀ can live in the nose and throat without causing symptoms, yet it can also produce severe invasive disease. When the bacterium reaches the lungs, bloodstream or central nervous system, treatment depends on antibiotics reaching the infection at an effective concentration. Resistance can narrow that therapeutic window.
Ā
The U.S. Centers for Disease Control and Prevention reports that pneumococci not susceptible to at least one antibiotic cause roughly two in five infectionsĀ in the United States. Resistant disease can lead to treatment failure, longer hospital stays and more expensive care. Vaccination has lowered disease caused by several resistant serotypes, but no vaccine covers every pneumococcal strain. 4
Ā
The wider antimicrobial-resistance crisis adds urgency. The World Health Organization estimates that bacterial antimicrobial resistance was associated with more than 4.7 million deaths in 2021, and that about one in six laboratory-confirmed bacterial infections worldwide was resistant to antibiotics in 2023. 5Ā Drug resistance is driven by unnecessary or inappropriate antimicrobial use, weak infection prevention, inadequate diagnostics and a thin pipeline of new medicines.
Ā
What the AI system did
The research team, led by Pedro J. Ballester at Imperial College London, treated drug repurposing as a prediction problem. Drug repurposing means testing a medicine developed for one indication against a different disease target. Existing drugs can offer known manufacturing routes, pharmacology and some safety information, although a new use still requires its own efficacy, dosing and safety assessment.
Ā
The investigators assembled activity data from public bioassay resources, including PubChem and ChEMBL. Their training material contained 1,849 compounds known to be active against pneumococcus and 34,503 inactive compounds associated with another drug-resistant bacterium. Three types of model ensembles then ranked the 6,747 drugs: decision-tree models, graph neural networks and sequence-based transformer models. 1Ā 2
Ā
The design was not based on a single algorithm making an isolated guess. The different models represented molecules in different ways and supplied partly independent signals. The researchers selected candidates with strong model support and chemical dissimilarity from the known training molecules. This helped prioritize compounds that were both promising and not simple repeats of familiar pneumococcal antibiotics.
Ā
Ā
From thousands of predictions to eleven experiments
The computational screen produced a manageable experimental set. After removing molecules that were difficult to synthesize, already known to be active, redundant with other selections or otherwise less suitable for testing, the team retained 11 candidates.
Ā
In culture experiments with the laboratory strain S. pneumoniaeĀ R6, nine of the 11 compounds inhibited growthĀ at half-maximal inhibitory concentrations of 0.4 micrograms per millilitre or lower. This is a strong hit rate for a prospective screen, but it is still a laboratory result rather than evidence of clinical benefit.
Ā
The two most potent compounds were thiostrepton and ceftiofur. Thiostrepton showed an ICā ā of 0.0001 micrograms per millilitre, while ceftiofur showed 0.0004 micrograms per millilitre. Thiostrepton retained high activity against pneumococcal strains resistant to several antibiotic classes. 1
Ā
Ceftiofur is a third-generation cephalosporin used in veterinary medicine. The study found that it was more potent in the tested setting than cefotaxime, a commonly used third-generation cephalosporin. That comparison is scientifically interesting because cephalosporins are important in severe pneumococcal disease, including infections where penicillin susceptibility is reduced. It does not mean that ceftiofur is ready for routine human treatment.
Ā
Thiostrepton has a different profile. It is also used mainly in veterinary medicine and has poor water solubility, which limits how it can be administered. The authors therefore point to possible value in non-invasive infections, such as selected ear, eye or sinus infections, where a narrow-spectrum approach might reduce collateral pressure on the wider microbiota. Such use would require diagnostic confirmation, formulation work and controlled clinical studies. 1
Ā
Ā
Why the result is scientifically important
The most useful result is not simply the number eleven. It is the demonstration that AI can help connect three stages that are often separated: biological activity data, chemical structure and experimental prioritization.
Ā
Traditional screening can test large libraries, but laboratory resources are limited. A model that ranks compounds before testing can reduce the number of experiments needed to find viable leads. In this study, the models were complementary: a consensus across different representations helped identify molecules that would have been difficult to prioritize by intuition alone.
Ā
Repurposing also changes the starting point. A molecule that has already been manufactured or studied for another use may have a more developed evidence base than a completely new chemical entity. The advantage is time and information, not automatic approval. A repurposed drug can still fail because the required dose is unsafe, the compound does not reach the infected tissue, its formulation is unsuitable or its activity does not translate from culture to patients.
Ā
The work also supports a more precise approach to antibiotic stewardship. Broad-spectrum antibiotics can affect beneficial bacteria and select for resistance beyond the pathogen being treated. If clinicians can identify the causative organism and use a narrow-spectrum option, treatment may reduce this bystander pressure. That possibility is a future clinical hypothesis, not a current prescription recommendation.
Ā
The limits that should shape the next phase
The study tested growth inhibition in vitro. It did not establish how the compounds behave in animals or humans, how they are absorbed and distributed, or whether they are safe at concentrations needed to treat infection. The models also depend on the quality and scope of their training data. A prediction can be useful without revealing the exact mechanism by which a compound harms a bacterium, and mechanism studies remain essential.
Ā
The resistant-strain result is encouraging, but resistance is diverse. Pneumococci carry different combinations of mutations, resistance genes and serotypes in different regions. Activity against the strains tested cannot guarantee activity against every circulating isolate. Surveillance, susceptibility testing and molecular diagnostics will remain necessary.
Ā
The next steps include confirming antibacterial mechanisms, testing additional clinical isolates, evaluating pharmacokinetics and toxicity, developing appropriate formulations and comparing the candidates with current standards of care. Only well-designed clinical trials can establish whether the predicted benefits reach patients.
Ā
Closing Thoughts
The striking feature of this research is its balance between ambition and restraint. AI did not replace microbiologists, pharmacologists or clinicians. It helped them decide which molecules deserved attention first. That is a more credible and useful vision of AI in medicine than the idea of an automated system producing a finished antibiotic on demand.
Ā
The discovery also invites a broader question about the medicines already around us. Thousands of compounds have been developed for human and veterinary health, and many may contain biological effects that were not fully explored for every pathogen. Better data and careful experimentation can reveal some of those hidden possibilities.
Ā
At the same time, responsible progress requires discipline. A promising hit is not a treatment. Patients need evidence about dose, delivery, side effects and outcomes. Antibiotics also need stewardship after approval so that any benefit is not quickly lost to new resistance.
Ā
What is the true implication of this?
The true implication is that antibiotic discovery can become more selective, iterative and evidence-led. Instead of testing every molecule equally, researchers can use AI to concentrate scarce laboratory resources on candidates with converging signals. That may shorten the early discovery stage and widen the search beyond conventional antibiotic families.
Ā
The work also suggests that drug repurposing can complement, rather than replace, the creation of new antibiotics. Existing molecules may offer nearer-term leads, while AI can support the longer process of designing compounds with new targets and mechanisms. The strongest pipeline will use both paths.
Ā
What is the relevance of this?
Its relevance is practical for public health, research funding and clinical preparedness. Drug-resistant pneumococcal infections already cause treatment problems, and the World Health Organization identifies antimicrobial resistance as a major global threat. A nine-in-11 laboratory hit rate shows that computational triage can generate experimentally testable options from approved-drug libraries.
Ā
For now, the finding should encourage investment in validation, diagnostics, surveillance and responsible antibiotic useānot self-medication or off-label treatment. Its value will be measured by whether carefully tested candidates become safe, effective and appropriately targeted therapies for people facing infections that current antibiotics can no longer reliably control.
Ā
References
Comments