Can AI really create the right drug formula? What scientists have already achieved may surprise you

One of the biggest myths around artificial intelligence is that it can simply create a new medicine on its own. That is not how drug discovery works today. AI can suggest ideas, spot patterns and predict which drug candidates look promising, but every prediction still has to be checked in a laboratory. According to Let’s Talk Science, this is why modern drug development is becoming a partnership between computers and scientists rather than a competition between the two.
The organisation explains that researchers now work through a continuous cycle involving “dry labs” and “wet labs.” In dry labs, scientists use computer models to study proteins and ask AI to predict new protein designs. In wet labs, researchers create those proteins and test how they actually behave. The findings from these experiments are then fed back into the computer models, allowing AI to learn from real-world results. The cycle repeats again and again, helping researchers improve their predictions with every round of testing.

This approach is also changing how scientists spend their time. Instead of carrying out hundreds of repetitive experiments to eliminate unsuitable drug candidates, AI can narrow down the list much earlier. That allows researchers to focus on the ideas that have a better chance of succeeding. The final decision, however, still rests with scientists who carefully analyse every result before moving a medicine towards clinical testing.

Teaching AI to understand proteins

Proteins are extremely complex molecules. Even a tiny change in their structure can completely change the way they behave inside the human body. For years, analysing these complicated structures was one of the slowest parts of drug development because the amount of data was simply too large for older computer systems to handle efficiently.

According to Let’s Talk Science, advances in machine learning and faster computing have changed that situation. Modern AI models are now able to recognise patterns hidden inside huge collections of protein data. In some cases, they can even suggest completely new protein designs that scientists may not have considered on their own.

This does not mean every new design automatically becomes a medicine. It simply means researchers have more promising options to investigate. Instead of searching blindly, scientists can begin with candidates that AI believes are more likely to work. That saves valuable research time while still keeping scientific testing at the centre of the process.

Predicting problems before human trials begin

Finding a possible drug is only one part of the challenge. Researchers also need to know whether that drug will actually be practical to use in patients. One important factor is viscosity, or how thick a protein solution becomes.
A protein that is too thick could become difficult to inject into a patient. Discovering that problem after years of research would waste both time and resources. To avoid this, researchers at Amgen developed a machine learning model using data collected from 83 antibody proteins. The scientists compared the amino acid sequences of those proteins with laboratory measurements of their viscosity.

Once the model learned those patterns, it could predict whether newly designed antibody proteins were likely to have high or low viscosity before extensive laboratory work even began. The company says this allows researchers to identify unsuitable candidates much earlier and concentrate on proteins that appear more practical for development.

According to Chadwick King, Associate Vice President at Amgen British Columbia, generative biology is already helping patients by changing the way medicines are developed. He says AI and machine learning are allowing researchers to create new protein-based treatments more quickly while giving scientists more time to focus on innovative ideas instead of repetitive laboratory tasks that computers can help simplify.

Read More: Why is Instagram facing backlash over Meta’s AI glasses ads? User claims her profile picture was used in a targeted ad on her boyfriend’s feed amid privacy concerns

Why better data matters

Artificial intelligence cannot make reliable predictions without learning from reliable information. Every experiment carried out in a laboratory becomes another lesson that helps improve future predictions. That is why scientists say high-quality data remains one of the most valuable parts of modern drug discovery.

According to Let’s Talk Science, one challenge is that individual pharmaceutical companies develop only a small number of protein drugs each year. That limits the amount of information available to train AI models. To overcome this, researchers are exploring a system known as federated learning.

Under this approach, companies do not need to hand over their confidential research data. Instead, they improve a shared machine learning model using their own information. Those improvements are then combined to build a stronger global model while sensitive company data stays protected. Researchers believe this could help AI learn from a much wider range of scientific knowledge without compromising privacy or commercial secrets.

AI could help solve one of medicine’s biggest problems

The need for faster drug discovery has become even more urgent because many existing medicines are slowly losing their power. According to the BBC, antibiotic resistance has become one of the biggest challenges facing modern healthcare. Bacteria are becoming harder to kill because they continue to evolve and develop resistance to medicines that once worked well. As a result, infections that were once easy to treat are becoming far more difficult to control.

The BBC reports that developing completely new antibiotics has also been painfully slow. Between 2017 and 2022, only 12 new antibiotics were approved for use, and many of them were similar to medicines that bacteria are already learning to resist. Drug development is expensive, takes many years and often carries a high risk of failure. This is one reason why researchers are increasingly looking at AI to speed up the search for new medicines.

One of the scientists leading this effort is James Collins, Professor of Medical Engineering and Science at the Massachusetts Institute of Technology (MIT). According to the BBC, Collins and his team have shown how AI can search through massive collections of chemical compounds in a fraction of the time it would normally take researchers. Instead of manually studying millions of possibilities, AI can quickly identify compounds that appear to have the characteristics needed to fight harmful bacteria.

The BBC explains that Collins’ team trained a generative AI system using the chemical structures of existing antibiotics. By learning what successful antibiotics have in common, the AI could begin searching for entirely new compounds with similar disease-fighting properties. The researchers then used the system to examine more than 45 million chemical structures while looking for compounds that could target bacteria responsible for infections such as gonorrhoea and MRSA, both of which have become increasingly resistant to available medicines.

Rather than simply searching for existing drugs, the AI also helped create completely new chemical compounds. According to the BBC, researchers allowed the system to build new molecular structures by adding different atoms, bonds and chemical substructures while continuously checking whether each design looked more like a potential antibiotic. The process generated millions of possible compounds, giving scientists a much larger pool of ideas than they could realistically create by hand.

Of the millions of AI-designed compounds, researchers selected just 24 for laboratory testing. Seven showed antimicrobial activity, while two proved especially effective against drug-resistant bacteria. Just as importantly, these compounds appeared to attack bacteria differently from many existing antibiotics. That raises hopes they could eventually become part of an entirely new class of medicines, although they still have to undergo extensive testing before they can be used in patients.

The same research group has also used AI to identify promising compounds against other dangerous bacteria, including those responsible for tuberculosis and Clostridium difficile infections. Scientists are now exploring whether similar AI tools could help speed up research into diseases that currently have few or no effective treatments. The technology is also being studied for conditions such as Parkinson’s disease and thousands of rare diseases where finding new medicines has always been especially difficult.

Read More: Will AI ever feel happiness, see colours or know it exists? The consciousness puzzle explained

The final verdict

So, can AI make drugs faster than humans? The answer appears to be yes at least during the early stages of discovery. AI can search enormous datasets, recognise hidden patterns and suggest promising drug candidates far more quickly than people working alone. That can save months or even years of research and help scientists focus their efforts where they matter most.
But can AI produce the correct drug formula on its own? Not yet. Every prediction made by AI still needs careful laboratory testing, detailed scientific analysis and strict safety checks before it can become a real medicine. AI is not replacing researchers; it is becoming one of the most powerful tools they have ever had.

That may be the biggest lesson from the latest research. The future of medicine is unlikely to belong to either humans or machines alone. Instead, it will depend on scientists and artificial intelligence working together, combining human knowledge with the speed and computing power of AI to discover safer and better medicines for patients around the world.