1. How is artificial intelligence changing drug discovery?

Artificial intelligence (AI) can process vast amounts of information and absorb old and new scientific research, sifting information for possible breakthroughs much faster than human researchers alone. In the past few years, tools like Google-backed AlphaFold, a network that can determine a protein’s structure from its building blocks, are revolutionising the pace and practice of drug discovery, and bringing AI medicines closer to market.

In its report ‘Unlocking the potential of AI in drug discovery’, the Wellcome Trust identified five major areas where AI is impacting drug discovery:

  • identifying and validating new targets for drug discovery efforts
  • identifying small molecule compounds and optimising them for favourable properties
  • designing vaccines, with a primary focus on mRNA-based vaccines
  • identifying and optimising antibody structures, binding and other properties
  • evaluating the safety of a potential treatment or vaccine

As well as developing medicines for diseases that do not currently have effective treatments, novel medicines can be developed for conditions that are currently treatable, but where treatments have bad side effects.

In an article for ‘Drug Discovery Today’, Debleena Paul et al explain that as well as indicating which molecules could be therapeutic, AI can create models which predict absorption, and so recommend how a medication might be administered, to the extent of designing “optimal tablet geometry”.

Paul et al confirm that AI predictions still need to be tested in labs and then in humans. From target identification, through testing and approval, drugs typically take 10 to 12 years to become available, though AI can accelerate the initial identification process. The exception is where potential novel uses of existing medicines are discovered. If a medication is already approved for use in humans it may be able to fast-track to phase 2 clinical trials and could then be deployed more quickly.

2. What has been discovered?

The Wellcome Trust lists 73 AI-derived clinical assets that it has catalogued, most of which are in clinical trial phases, including Covid-19, HIV and flu vaccines.

In June 2023, Insilico Medicine announced it had created “the first fully generative AI drug to reach human clinical trials”. The drug, INS018_055, aims to treat the chronic lung disease idiopathic pulmonary fibrosis. AI was used both to discover the target and to design the drug. Insilico Medicine’s CEO Alex Zhavoronkov spoke to Vox: “Many of those drugs that are true AI drugs, they were created just a couple years ago, so they didn’t have the time to get into the human clinical trials. We are I think the first one with a true generative AI drug”.

AI has also been used to make discoveries in the field of senolytics or ‘anti-ageing’, prompting headlines like ‘The end of ageing?’ in BBC Science Focus. This is based on a study published in ‘Nature Communications’ from Vanessa Smer-Baretto et al at the University of Edinburgh. Senolytics are a form of drug that can slow ageing and prevent age-related diseases by killing ‘senescent cells’. Left in situ, these damaged cells can cause inflammation and damage neighbouring cells.

Smer-Baretto et al searched the scientific literature for 58 senolytics, and fed these and non-senolytics into the AI so that it could learn the characteristics. They then asked the AI to identify possible senolytics from a list of 4,340 molecules. It returned a list of 21 likely molecules within five minutes. The potential drug candidates were then tested in the lab on healthy and ageing cells. Of the 21 identified molecules, three could eliminate the ageing cells, while keeping healthy cells alive. They are currently undergoing further testing, including testing on tissue models.

In February 2020 a team of researchers at the Massachusetts Institute of Technology (MIT) used artificial intelligence to discover an antibiotic capable of killing E. coli, as well as an antibiotic-resistant strain of the common bacterial infection, acinetobacter baumannii. They named it ‘Halicin’ after Hal, the artificial intelligence in ‘2001: A Space Odyssey’. Demonstrating the potential of AI to make novel associations, Halicin is “structurally distinct from prior antibiotics”. It is at the stage of positive results treating infections in mice studies.

3. How can AI help personalise medicine?

AI also provides opportunities for personalised medicine. UK-based company Exscientia has developed a system that pairs individual patients with the precise drugs they need, considering the biological differences between individuals. ‘MIT Technology Review’ describes how this worked in the case study of ‘Paul’, an 85 year old whose blood cancer had proven resistant to six courses of chemotherapy. The researchers took a small sample of tissue from Paul, including both normal cells and cancer cells. The sample was split into more than a hundred pieces and each exposed to different possible drugs. Machine-learning models trained to identify small changes in cells then ‘watched’ the samples. The top candidate was not a viable option due to Paul’s state of health, but the second most likely option led to full remission. It was a Johnson & Johnson cancer drug that in general trials had not demonstrated efficacy for Paul’s type of cancer.

AI also has potential uses in implantation as part of personalised drug treatment. The journal ‘Science Robotics’ has published research from scientists at MIT and the University of Galway who have created an AI-enabled soft robotic implant. Implantable drug delivery devices can be used to regularly dose patients when they need it, for example releasing insulin to treat diabetes. However, patients’ bodies often react to a ‘foreign body’. The AI device would be able to sense when the body is beginning to reject it and change its shape to avoid scar tissue build up, as well as maintain treatment dosage and delivery despite tissue changes.

4. What concerns are there about AI drug discovery?

In ‘Nature Machine Intelligence’, Fabio Urbina et al have warned that the same AI that can help researchers learn which drugs are safe for humans, can be used to determine which are most harmful. A discussion at an international security conference led to Urbina’s team from Collaborations Pharmaceuticals, publishers of a toxicity database, testing whether AI toxicity models could be used to generate new potential biochemical weapons. Discovering thousands of potentially toxic substances in their experiment, they said: “By inverting the use of our machine learning models, we had transformed our innocuous generative model from a helpful tool of medicine to a generator of likely deadly molecules”.

Urbina et al call for strong ethical training of researchers, and screening before access to databases is made available. However, they also indicate concerns that “a global array of hundreds of commercial companies offering chemical synthesis” in a poorly regulated environment make it possible that bad actors could design and acquire toxic agents. They also warn against autonomous synthesis—allowing AIs to generate compounds they design without intervention.

The Wellcome Trust also points to many AI algorithms, tools and databases being patented or otherwise protected, making them unavailable to the broader research community. It raises concerns about the availability and completeness of databases, but highlights work to address this: the US’s National Institutes of Health is funding projects to clean and standardise datasets so they can be more easily used for machine learning; the Wellcome-Sanger African Genome Variation Project is “laying the foundations for generating high-quality genomic datasets in Africa”.

Concerns have also been raised that funding and development will be limited to certain areas of medicine. The Wellcome Trust found that private funding of AI drug discovery is skewed towards the most commercially profitable therapeutic areas, with around 70% of AI-related investments in the last five years being made in oncology, neurology, and Covid-19.

Indeed, ‘Drug Discovery Today‘ explains that some pharmaceutical companies are using AI to determine which avenues will be profitable for drug generation. For example, E-VAI is an analytical and decision-making AI platform developed by AI-first pharma company Eularis, which uses machine learning algorithms:

to create analytical roadmaps based on competitors, key stakeholders, and currently held market share to predict key drivers in sales of pharmaceuticals, thus helping marketing executives to allocate resources for maximum market share gain, reversing poor sales and enabled them to anticipate where to make investments.

The Wellcome Trust argues that focus on already “well-served” commercially viable therapeutic areas can amplify disparities in health inequalities globally. For example, there is less research into infectious diseases which are leading causes of ill-health and death in much of the world.

5. What is the government’s approach to regulation?

The government policy paper, ‘A pro-innovation approach to AI regulation’ sets out five principles for the responsible development and use of AI across the economy, and intends those principles to be interpreted and acted upon by existing regulators. In the UK, the pharmaceutical industry is regulated by the Medicines and Healthcare products Regulatory Agency. The Ada Lovelace Institute, in its report ‘Regulating AI in the UK’, describes the pharmaceutical sector as “comprehensively regulated, with well-resourced regulatory bodies that are able to shape organisational practices through effective enforcement and the setting of incentives”, providing an opportunity to help “mitigate AI harms”.

6. Read more

Cover image by danilo.alvesd on Unsplash.