Will AI improvement revolutionize drug improvement? Researchers say it is determined by how it’s used

Using the potential to use artificial intelligence in the discovery and development of drugs in drug development has triggered both excitement and skepticism among scientists, investors and the public.

“Artificial intelligence takes over the development of drugs,” says some companies and researchers. In recent years, interest in the use of AI to design medication and the optimization of clinical studies has led to an increase in research and investments. AI-controlled platforms such as Alphafold, who have won the Nobel Prize 2024 for its ability to design the structure of proteins and new new ones, show the potential of AI to accelerate the development of pharmaceuticals.

AI in the discovery of drugs is “nonsense”, warns some industry veterans. They demand that “KIS potential to accelerate drug discovery needs a reality test”, since the medicines with AI producers cannot yet demonstrate the ability to address the 90% failure rate of new medicinal products in clinical studies. In contrast to the success of AI in image analysis, its effect on drug development remains unclear.

There are many, many more, many more, which failed.
Nortonrsx/iStock via Getty Images Plus

In our work as a pharmaceutical scientist, we have both in science as well as in the pharmaceutical industry and as a former program manager in the agency for defense research projects (Darpa we argue that AI is not yet a game changer in drug development, and it is also not complete nonsense. Ki is not a black box that can transform every idea into gold.

Most work with AI in drug development to shorten the time and money to bring a medication onto the market -currently 10 to 15 years and $ 1 billion to $ 2 billion. But can AI really revolutionize the development of medicines and improve the success rates?

AI in drug development

Researchers have used AI and machine learning in every phase of the drug development process. This includes the identification of goals in the body, screening potential candidates, the design of pharmaceutical molecules, the prediction of toxicity and the selection of patients who could best react to medicinal products in clinical studies.

Between 2010 and 2022, 20 AI-focused startups discovered 158 pharmaceutical candidates, 15 of whom rose in clinical studies. Some of these drug candidates were able to carry out preclinical tests in the laboratory and enter human experiments in just 30 months, compared to the typical 3 to 6 years. This performance shows the potential of the AI ​​to accelerate drug development.

The development of medicinal products is a long and costly process.

On the other hand, the success of these candidates in clinical studies – where the majority of drug failure occur – remains very uncertain, while AI platforms can quickly identify connections that work on cells in a petri dish or in animal models.

In contrast to other fields that have large, high-quality data records to train AI models such as image analysis and language processing, the AI ​​is restricted in drug development by small data sets with low quality. It is difficult to create medication -related data sets for cells, animals or people for millions to billions. While Alphafold is a breakthrough in the prediction of protein structures, the accuracy of the drug design remains uncertain. Minor changes in the structure of a drug can strongly influence its activity in the body and thus treatment in the treatment of diseases.

Survival priority

As with AI, previous innovations in drug development such as computer-aided drug design, the human genome project and the high-throughout screening have improved the individual steps of the process in the past 40 years, but the installment failure of the drug failure has not improved.

Most AI researchers can tackle certain tasks in the drug development process if they receive high-quality data and certain questions for answering. However, they are often not familiar with the full extent of drug development and reduce the challenges into the pattern recognition problems and the refinement of the individual steps of the process. In the meantime, many scientists with specialist knowledge in drug development lack training in AI and machine learning. These communication barriers can prevent scientists from going beyond the mechanics of current development processes and identifying the basic causes of drug failure.

Current approaches to drug development, including those who use AI, could have fallen into a survival project that concentrates excessively on less critical aspects of the process and at the same time overlook important problems that contribute most to fail. This is analogous to the repair of damage to the wings of aircraft that return from the battlefields in the Second World War and neglect the fatal weaknesses in engines or cockpits of the aircraft that have never made it back. Researchers often concentrate excessively on how the individual properties of a drug and not the basic causes of failure can improve.

While returning aircraft may survive hits for the wings, those that damage the engines or cockpits are less likely back.
Martin Grandjean, McGeddon, US Air Force/Wikimedia Commons, CC BY-SA

The current drug development process acts like a assembly line and is based on a control box approach with extensive tests with every step of the process. While AI may be able to shorten the time and costs of the preclinical laboratory stages of this assembly line, it is unlikely that the success rates in the more expensive clinical phases that contain tests in humans. The persistent 90% failure rate of drugs in clinical studies underlines this restriction.

Add the root causes

Medicines failure in clinical studies are not only due to how these studies were designed. The selection of the wrong drug candidates for examining in clinical studies is also an essential factor. New AI guided strategies could help to deal with these two challenges.

At the moment, three dependent factors are driving most of the drug errors: dosage, security and effectiveness. Some medications fail because they are too toxic or unsafe. Other drugs fail because they are classified as ineffective, often because the dose cannot be increased further without causing damage.

We and our colleagues suggest a mechanical learning system to select drug candidates by predicting dosage, security and effectiveness based on five previously overlooked features of pharmaceuticals. In particular, researchers could use AI models to determine how specific and potent the medicine to known and unknown goals, the degree of these goals in the body, how concentrated the medicine in healthy and sick tissues and the structural properties of the medicine binds.

These characteristics of AI-generated drugs could be tested in the so-called phase-0 studies, with ultra-lowering doses being used in patients with serious and lighter illness. This could help the researchers to identify optimal medication and at the same time reduce the costs of the current “test-and-lake” approach for clinical studies.

While AI alone may revolutionize the development of pharmaceuticals, it can help to fix the basic causes of why medication fails and optimize the lengthy process for admission.

This dux, Associate Dean for Research, Charles Walgreen Jr. Professor of Pharmaceutical and Pharmaceutical Sciences, University of Michigan and Christian Macedonia, extraordinary professor of pharmaceutical sciences, University of Michigan

This article will be released from the conversation under a Creative Commons license. Read the original article.