Is AI-driven drug discovery the next big thing? Not quite yet

For decades, time and money have been two of the greatest limiting factors in new-drug development. For the period from 2009 to 2018, the mean research and development cost per new drug was about $985.3 million.^[] The median duration of non-oncological trials ranges between 5.9 and 7.2 years, and oncology trials are about 13.1 years long.^[]

In recent years, AI has been heralded as a much-anticipated solution to lowering drug-development costs and shortening the timeline. Has it lived up to the expectations? Recent AI advancements offer some insights.

Alphabet/Google break the fold

In November 2020, Google/Alphabet subsidiary DeepMind introduced AlphaFold, an AI system that predicts protein structures accurately.^[] DeepMind billed this as a “solution to a 50-year-old grand challenge in biology”: the protein-folding problem, or, determining the shapes of folded proteins.

With proteins, form often follows function, and knowing protein shapes may prove beneficial for developing new therapeutics, according to DeepMind. As of July 2022, DeepMind stated that AlphaFold “reveals the structure of the protein universe.”^[] Before AlphaFold, we knew about 1 million protein structures. 

But has AlphaFold protein mapping led to any clinical breakthroughs? So far, the results appear to be incremental.

For example, researchers publishing in February 2023 reported that AlphaFold had accelerated the discovery of a novel CDK20 small molecule inhibitor, which could prove useful in combating hepatocellular carcinoma.^[] They discovered the molecule in 30 days after synthesizing only seven compounds. Whether this compound will make it through the clinical trial process is another matter, but the discovery process marks a first.

“This work represents the first example of successfully utilizing AlphaFold predicted protein structures for hit identification for a novel target,” the researchers wrote, adding that they are investigating additional applications.

But other than this study, AlphaFold’s clinical relevance remains a work in progress—albeit a fascinating work in progress, given the current competitive, evolving AI landscape.

ChatGPT has entered the chat

It also performed “at or near the passing threshold” for the Step 1, Step 2CK, and Step 3 exams for the USMLE.^[]

The USMLE researchers commented on the significance of their findings: “We believe that LLMs such as ChatGPT are reaching a maturity level that will soon impact clinical medicine at large, enhancing the delivery of individualized, compassionate, and scalable healthcare,” they wrote.

ChatGPT and AlphaFold are drastically different. But both may have roles to play in the new-drug discovery process, in addition to the day-to-day life of clinicians.

Publishing in Clinical and Translational Medicine, researchers theorized about how ChatGPT might prove useful in drug discovery.^[] Some of the assistance may be incremental, such as generating summaries of previous research or assisting with clinical trial recruitment, data management, or patient education. They also speculated that AI models and ChatGPT may one day rely on image recognition to pinpoint and categorize chemical formulas or molecular structures, which could in turn facilitate the design of new structures.

Researchers writing a preprint article in ChemRxiv posited that ChatGPT may potentially help identify new targets, capture the current state of research, assist in new drug design, and enhance the pharmacodynamics and pharmacokinetics of new drugs.^[] The new knowledge it could generate may support early decision-making in the drug discovery process. 

Nevertheless, it has some important limitations. 

The authors also noted that ChatGPT couldn’t handle certain complex computational calculations.

To put ChatGPT to the test, MDLinx asked it how it might be of assistance, and it cited:

• Literature review and analysis

• Data organization and interpretation

• Collaboration and communication with colleagues

• Writing grants and proposals 

• General brainstorming

It also provided a caveat:

“While I can be a valuable tool for researchers in the drug discovery process, I should be used as a supportive resource to complement, not replace, their expertise and judgment,” ChatGPT wrote. “Researchers should always verify the information I provide, consider alternative perspectives, and maintain a critical mindset when utilizing my assistance.”

Challenges ahead

As ChatGPT intimated, there are some challenges ahead for AI in drug discovery. Researchers explored these in a preprint article published on arXiv, an open-access archive of non–peer-reviewed scholarly articles hosted by Cornell University.^[]

One of those challenges is a lack of suitable data. The researchers explained that if the accessible data is limited, of low quality, or inconsistent, this can translate to questionable results.

There are also important ethical considerations on this new frontier. 

Bias and equity concerns exist, too, wrote the authors. Could AI-influenced drug development exclude specific groups of people?

Finally, with more data come more privacy and security issues. Breaches could prove costly for individual people and companies. As this field develops, the researchers admonished, fellow researchers and policymakers need to proceed with caution and to implement “technical solutions, regulatory frameworks, and public education.”

“Until AI can be trusted to produce reliable and accurate information, it should be used with caution in the world of science,” the researchers wrote. “It is essential to carefully evaluate the information provided by AI tools, and to validate it using reliable sources.”