Can the Machine Beat the Microbes?
By: Shumin Wang
The Superbug Problem, and How AI Could Change the Fight
In recent years, bacteria have become increasingly resistant to drugs that once worked against them, causing a surge of multidrug-resistant ‘superbugs’, posing a major threat to modern medicine. Through natural selection, accelerated by the misuse and overuse of antibiotics, resistant bacteria can survive treatments and reproduce, potentially making even common infections increasingly difficult to cure.
Malaysia faces the same threat. A WHO projection, cited by the Malaysian Ministry of Health, estimated that the country could experience 87,000 antimicrobial resistance (AMR) deaths between 2020 and 2030 without intervention. It is vital to realise that superbugs are no longer confined to laboratories, and have quietly seeped into everyday healthcare. AMR can affect anyone who develops a bacterial infection caused by a resistant bacterium, particularly people who are more vulnerable to severe illnesses.
Nonetheless, scientists are turning to artificial intelligence to search through millions of molecules for potential new antibiotics, accelerating one of the earliest stages of drug discovery. The technology could offer a new weapon against superbugs, but the misconceptions about antibiotic use cannot remain unaddressed.
Malaysia's National Antimicrobial Guideline recognises skin, soft-tissue, and wound infections as conditions in which antimicrobial treatment may sometimes be required. From dental infections to complex surgery, modern medicine continues to depend on effective antibiotics to prevent or treat bacterial infections when they are clinically necessary, yet misconceptions about antibiotics still remain widespread. A 2026 nationwide Malaysian study found that 72% of surveyed adults reported using antibiotics for respiratory illnesses. The problem is that many respiratory illnesses are caused by viruses, which renders antibiotics ineffective.
Sadly, inappropriate use extends beyond misunderstanding. Malaysians may also self-medicate, keep leftover antibiotics or stop treatment without professional advice. The 2026 study found that 34% of respondents kept leftover antibiotics and 45% reported stopping treatment early. These behaviours can contribute to the misuse of antibiotics and increase the biological pressure that drives antibiotic resistance as bacteria are not fully eliminated.
The consequences are not only medical, but also economic. A 2024 national study found that the total cost of treating AMR patients in Malaysian public hospitals sharply increased, from RM3.7 million in 2017, to RM9.7 million in 2019. When antibiotics fail, patients may require alternative treatments, additional monitoring, or longer hospital stays. This places further pressure on patients, hospitals, and the wider healthcare system in Malaysia.
Moving back to antibiotics, the problems researchers face are enormous. There could potentially be millions of compounds that, in theory, could inhibit bacteria. However, testing them individually would be overly time-consuming. Conventional drug discovery begins with researchers identifying and testing likely compounds before moving the strongest candidates into further studies. AI accelerates the first part of the process by allowing scientists to examine huge numbers of molecules digitally, and identify the ones most worth testing.
In context, the Graph Neural Networks (GNN) is one method researchers can use to achieve this. Instead of treating a molecule as a solitary compound, a GNN can represent chemical bonds as connections between atoms — learning patterns that link a molecule's structure to its potential biological activity.
Another approach resembles a recommendation algorithm. Just like how social media platforms can examine millions of user habits and trends to predict which content a viewer might enjoy, machine-learning models can analyse molecular structures and previous experimental data to predict which compounds are most likely to inhibit bacterial growth. This allows researchers to prioritise promising candidates before spending time and resources testing them in the laboratory.
Beyond existing molecules, researchers used the GNN in 2025 to screen more than 45 million chemical fragments; generative algorithms then assembled the most promising fragments into new molecular structures. More than seven million candidate molecules were generated. One compound, NG1, demonstrated promising antibacterial activity against Neisseria gonorrhoeae (a bacterium that causes gonorrhea, an STI) in laboratory testing, reducing bacterial numbers in a mouse infection model.
This marks an important shift in the role of AI. Instead of just asking, “Which molecules should we test?” researchers can now ask, "What new molecules could we design?”
From Halicin to Abaucin: What Has AI Already Found?
AI has continued to deliver promising results as early as 2020, when researchers at MIT used a deep-learning model to identify halicin, a compound that showed activity against multiple strains of antibiotic-resistant bacteria. The experiment demonstrated that machine learning could identify antibacterial candidates that conventional screening had not previously highlighted. Three years later, researchers reported another new example: abaucin, an antibiotic candidate discovered using a machine learning model against Acinetobacter baumannii. Commonly found in hospitals, A. baumannii can cause multiple serious infections, and has developed resistance to many commonly used antibiotics.
The model was trained on roughly 7,500 tested molecules before analysing 6,680 previously unseen compounds from the Broad Institute's Drug Repurposing Hub. The screening took less than two hours and produced several hundred promising findings, from which researchers selected 240 compounds for laboratory testing. These experiments identified nine potential antibiotics, including abaucin, which had originally been explored as a potential diabetes drug. It proved highly effective against A. baumannii, but showed no effect against several other bacterial species.
By inhibiting the system used to maintain the bacterium’s cell wall and allow nutrient uptake, abaucin effectively treated wound infections in mice. Abaucin is particularly interesting as it demonstrates that AI can bring light on a previously forgotten compound, or a potential compound intended to serve another purpose. Additionally, the AI's ability to select narrow-spectrum antibiotics (antibiotics that only target specific bacteria strains), rather than multiple bacteria, can potentially reduce the unnecessary side effects on beneficial bacteria.
Research in 2025 and 2026 has pushed this approach further, with AI systems moving from screening thousands of compounds to exploring millions and generating molecules that do not solely resemble existing antibiotics. Nevertheless, these compounds remain experimental, and not as approved medicines.
Discovery Doesn’t Mean Approval
Still, only discovering a promising molecule on a computer is very different from developing a medicine that is safe and effective for patients. The American Food and Drug Administration (FDA) describes the development pathway as progressing from discovery and preclinical research, to clinical research, regulatory review, and continued safety monitoring after approval.
The general procedure goes as follows:
A candidate identified by AI must first undergo laboratory and preclinical testing to determine how it behaves, whether it is toxic, and whether it can function as intended. Only candidates that pass those stages can move into human clinical trials.
Human trials are then divided into phases to fulfill different purposes: early trials focus on safety and dosage, while later studies examine its effectiveness over large groups of patients.
Finally, phase three trials involve hundreds — or even thousands of participants, before regulators assess whether the evidence justifies approval.
Therefore, though an AI system can reduce the time researchers spend searching for a promising molecule, it cannot remove the need to prove that the molecule is safe and is able to effectively work in humans. Perfecting such technology requires years in most cases. In other words, AI may now be able to search and select millions of possibilities in hours or days.
Regardless of AI’s ability, the most important question must still be answered without the computer: does the drug help a real patient without causing unacceptable or unforeseeable harm?
What Happens Next
AI has displayed its potential to become a keystone in the fight against antimicrobial resistance, but it cannot solve antibiotics misuse on its own. Malaysia has responded with a national action plan on antimicrobial resistance based on a “One Health” Approach, linking human, animal and environmental health together.
For us, that means the future of antibiotics is partly being shaped in laboratories developing AI models, and partly by how antibiotics are used today. We have to take responsible use of antibiotics by avoiding antibiotics for illnesses they cannot treat, following medical advice, and not sharing or reusing leftover medicines. These practices ensure the effectiveness of existing drugs.
Ultimately, the promise of AI is not that it will produce a miracle cure overnight. Its real potential lies in giving scientists a faster way to find new antibiotics before bacteria develop resistance to the ones we already have. That being said, discovering new antibiotics is only part of the fight. Since antimicrobial resistance is driven in large part by the misuse and overuse of antibiotics, protecting the medicines we have remains equally important.
Artificial intelligence may help us find tomorrow’s treatments, but responsible antibiotic use can help ensure that today’s treatments continue to work.
For sources, click here.
Comments