New Pharmaceuticals Through AI

An article from carl 03|2025

by Carolin Sage

For many years, scientists have been using computer-aided tools in the research, development and approval of new medicines. Now AI methods are also helping to develop pharmaceutical products with new active ingredients.

When Özlem Türeci and Ugur Sahin, the duo who founded BioNTech, decided in January 2020 to develop a vaccine against coronavirus, they called their project “Project Lightspeed”. And they kept their word: the vaccine was approved within the year [1]. When we are not in the special situation of a pandemic, however, it takes several years for a new active ingredient to be developed [2]. Only every eighth newly developed medicine is actually approved [3] and it takes on average ten to fifteen years for a new pharmaceutical product to be launched on the market.

Back in 2012, an analysis already showed that the prospects of successfully bringing development of a medicine to completion have continuously deteriorated in the course of recent decades [4]. This is indicated by the fact that fewer and fewer new medicines have been approved on the US market since the 1950s, largely due to the complexity of the diseases for which there are not yet any effective medicines.

For example, researchers are still searching for treatment options for many forms of dementia. The symptoms are known, but there is not yet any precise knowledge of the molecular processes that cause them. Yet if you want to cure a disease or slow its progress, you must first understand the mechanisms behind the disease and they key molecules that play a role. Only then can you seek out targets that active ingredients could attach to. For this, researchers must analyse a breathtaking volume of data on possible genetic factors, complex biological networks and metabolic processes.

That’s where AI comes into play, because it is extremely helpful in analysing large data volumes. AI doesn’t only save time, but it can also find connections where people (still) see none. How AI can help with data analysis in these cases is not only part of current research [5], but has already been in use by some pharmaceutical companies for several years. For example, Boehringer Ingelheim is collaborating with Phenomic AI, a Canadian company that has specialised in new targets for forms of cancer that are hard to treat [2]. 

But it’s not always just about a new and improved understanding of diseases. Even if the cause of an illness and its progression are known, it’s not always possible to get to the root of the problem: potential active ingredients, so-called hits, must first be identified, even if the target is known. From a now long list of active ingredient candidates, the most promising are selected and optimised (lead selection and lead optimisation) – a very difficult and lengthy process.

In the case of structurally simple active ingredients, such as  small molecules , AI helps in a very pragmatic way: you just have to ask. In October 2024, researchers from the University of Bonn developed an application that suggests active ingredient candidates based on a voice model, similar to ChatGPT [6]. The trick: these hits are designed to bond with two proteins that play a role in the disease incidence, and, according to the plan, can replace a combined treatment with several active ingredients.

Once optimisation has succeeded, the new active ingredient can be tested in cell studies and later in animal studies. Only..then are its effectiveness and safety demonstrated under real conditions. The substances that come off best are then considered for clinical studies on people.

The pharmaceutical giant Pfizer has said that it used AI in the analysis of clinical data during the development of the mRNA vaccine against coronavirus. The analysis included around ten million data points and, using conventional means, would have taken about a month, according to Pfizer in a press release [7]. 

Using the Smart Data Query software (SDQ), the time required was shortened to just 22 hours.

Despite all the euphoria regarding the smart solutions that AI is capable of delivering, it is still not possible to state conclusively whether pharmaceutical research and development will actually bring more new or improved active ingredients onto the market in the near future, because there are still obstacles. For example, the training data required for AI developers is not always available in adequate quality. The possibility of fake publications from AI applications [8] and the risk of dual use should also be given consideration, because a powerful tool that can discover the best active ingredient can equally well bring the strongest poison to light. 

This was shown by Collaborations Pharmaceuticals Inc. from the USA, which usually uses AI to identify highly effective substances with low toxicity. In an experiment, they deliberately reversed the requirements of the active ingredient. In less than six hours, the AI algorithm delivered 40,000 toxic compounds, including chemical weapons that actually already exist in reality [9]. 
 

[1] K. Leisinger, 2024, Research Ethics 20, 847-856
[2] S. Paul et al., 2010, Nat. Rev. Drug Discov. 9, 203-214
[3] vfa Biotech-Report 2024, 26
[4] J. Scannell et al., 2012, Nat. Rev. Drug Discov. 11, 191-200
[5] F. Gao et al., 2022, GBP 20, 811-813 
[6] S. Srinivasan und J. Bajorath, 2024, Cell Reports Physical Science 5, 102255 
[7] mRNA and Artificial Intelligence for Advanced Vaccine Innovation. www.pfizer.com/stories/articles?field_fs_tags_target_id[7046]=7046
[8] J. Haider et al., 2024, Misinformation Review 5, doi.org/10.37016/mr-2020-156
[9] F. Urbina, 2022, Nat. Mach. Intell. 4, 189-191

Image credits: freepik / Tatjana Pospiech, Carl ROTH / Tatjana Pospiech, Carl ROTH