Artificial intelligence has long been described as one of the most promising technologies for discovering new drugs, and investment in AI biotech has already reached tens of billions of dollars. But so far, real-world results have been noticeably more modest than the promises: AI has learned to identify promising candidates, yet this has not led to the emergence of a large number of new drugs in late-stage clinical trials.

This is the conclusion reached by the authors of a peer-reviewed paper published in Nature Reviews Drug Discovery. They describe AI’s clinical impact as “disappointingly limited.”

Finding a Molecule Turned Out to Be Easier Than Proving Its Effectiveness

As Axios.com reports, the main problem is the gap between the early stages of drug development and testing in humans.

AI has indeed become a useful tool for identifying potential drug compounds and promising biological targets. But after that, development runs into the same problem as traditional drugs: it is necessary to prove that the candidate actually works in humans.

The second stage of clinical trials — Phase 2 — remains especially challenging. At this stage, the drug is tested on patients and researchers try to obtain convincing evidence of its efficacy and safety. Results can vary greatly from person to person, and the biology of the real human body is far more complex than the models on which algorithms are trained.

Therefore, AI’s ability to predict promising molecules does not in itself mean that those molecules will become effective medicines.

The Biggest Shortage Is Human Biology Data

Researchers point to a problem that looks far less impressive than new AI models but may prove more important for the technology’s development: data quality.

Training models requires detailed information about how cells and the body function, yet such datasets often turn out to be insufficiently reliable, incomplete, or simply too complex to standardize.

Biotech journalist Derek Lowe, commenting on the study, noted that scientists still do not know how to convert much of this data into a form suitable for полноценного machine learning, or whether that is even possible at all.

This creates a kind of paradox. Algorithms are getting better and better, but the data they must use to study human biology remains one of the main constraints.

More Than $40 Billion Has Already Been Invested in AI Biotech

According to PitchBook, venture investors have invested more than $40 billion in AI biotech companies this decade.

At the same time, there are still very few drugs developed with substantial use of AI that have already advanced beyond Phase 2 trials. In itself, this should not be surprising: drug development takes years and requires passing through several stages of testing.

But this is precisely where the gap between reality and the industry’s original promises emerges. Supporters of AI have repeatedly claimed that the technology would significantly shorten drug development timelines. So far, there is not enough convincing evidence of such acceleration.

AI Is Already Helping Make Treatment Decisions

At the same time, it would be wrong to claim that AI provides no practical value at all in pharmaceuticals.

For example, Moderna and Merck have begun a Phase 3 trial in which AI is used to select tumor targets for specific cancer patients. After that, patients are prescribed an already existing drug.

However, this is a somewhat different scenario. Here, AI helps make more precise treatment decisions for a specific person rather than completely transforming the process of creating a new drug from scratch.

Such applications may become an important part of medicine even without the emergence of fully AI-created drugs.

The Most Expensive Problems Begin After Drug Discovery

Some investors believe that AI’s next major impact should appear not only in molecule discovery.

Robert Nelsen of Arch Venture Partners points to the need for a much larger volume of biology data, including cellular and population-level data. In his view, together with regulatory changes and advances in computing technologies, this could significantly transform the pharmaceutical industry.

Bijan Salehizadeh of NaviMed Capital, by contrast, warns that AI may for now be used primarily where it is easiest to implement: in biostatistics, in processing data after patient enrollment is completed, and in preparing documents for regulators.

At the same time, the most expensive and complex tasks remain unresolved. These include conducting large studies with hundreds or thousands of clinical sites around the world, launching studies, negotiating with sites, recruiting patients, and continuously monitoring data.

The result is that AI is already capable of helping at certain stages of development, but it has not yet demonstrated an ability to remove the main obstacles that make creating a new drug a long and expensive process.

A Gap Remains Between the Promises and the Proven Impact

Today, AI is confidently used in certain tasks in pharmaceuticals and biomedicine, but that is still not the same as a revolution in drug development.

The main question now is not so much whether an algorithm can come up with a promising molecule, but whether the entire system — from laboratory data to large-scale clinical trials — can turn that prediction into an effective drug.

For now, it is precisely at this stage that AI’s advantages remain far less convincing than the promises surrounding the technology.