Science & Technology

AI for drug development: How trustworthy is the tech?

Pharma companies today deploy artificial intelligence with a potential to reduce the drug development cycle by half 

By Rohini Krishnamurthy
Published: Monday 10 April 2023

Illustration: Yogendra Anand

Illustration: Yogendra Anand

In 2022, Hong Kong-based pharmaceuticals company Insilico Medicine, along with researchers from Stanford University in the United States and the University of Toronto, Canada, began a unique experiment, powered by three artificial intelligence (AI) tools.

The experiment was to find a cure for Hepatocellular carcinoma, the most common form of liver cancer.

In the first stage, the company’s biocomputational platform called PandaOmics scanned through endless research papers online to shortlist 20 potential targets — either proteins or enzymes — for drugs to act on. Of these, the researchers selected cyclin-dependent kinase 20 (CDK20) — an enzyme that had topped the list.

In the second stage, the researchers turned to AlphaFold, a Google-owned open access AI system that predicts a protein’s (all enzymes are proteins) three-dimensional (3D) structure. Of the 20,000 proteins present in the human body, structures for only 10 per cent are known, and CDK20 was one of the unknowns. Knowing the protein’s shape is necessary to find chemicals capable of binding and inactivating it.

Finally, the researchers fed the 3D structure into Chemistry42, an AI algorithm owned by Insilco Medicine, which designed 8,918 new chemicals with a high probability of attaching to CDK20. The researchers used computational modelling to virtually fit the chemicals with CDK20 and shortlisted seven candidates showing the best affinity for the target. They then constructed physical tests to assess how strongly the seven candidates attached to the enzyme and picked the chemical showing the best affinity.

The entire process took just 30 days. It would have taken at least six months (even a year) if all 8,918 chemicals were tested individually. The results were published in January 2023 in the journal Chemical Science.

James Collins, professor of biological engineering at Massachusetts Institute of Technology, US, found potential in the study. “They looked at a single target and then used very focused simulations around that target,” he said.

In the last decade, a wave of pharma companies that use AI at every step of the drug discovery process have emerged on the scene and created drugs that are under clinical trials. The AI drug discovery companies are, in fact, catching up with Big Pharma, which uses AI to a much less degree.

In 2021, some 20 AI drug discovery companies had 160 candidates in the discovery and preclinical stages and 15 in clinical development, according to a March 2022 analysis in Nature Reviews Drug Discovery. Compared to this, the top 20 pharma companies in terms of revenue have roughly 330 candidates in the discovery and preclinical stage and approximately 430 in Phase I of three clinical development phases.

Exscientia, a United kingdom-based AI-drug discovery start-up, is positive that AI will discover and develop all new drugs by 2030.

The ability of AI companies to complete the entire discovery process and preclinical stages in less than four years gives them an edge over their Big Pharma competitors, which take five to six years to reach that stage.

Developing a new drug costs about $2.6 billion and an average of 12 years, according to a 2019 report by UK-based consultancy Deloitte. AI can make predictions significantly faster and cheaper. As the technology develops, it could further accelerate the drug discovery process, the Nature analysis predicts.

A branch of AI called machine learning (ML) is widely used in drug discovery. ML algorithms make decisions or predictions using data. They first undergo a learning phase, where researchers train them with data. For example, clinicians can feed in chest X-rays labelled by humans as cancerous or non-cancerous and the machine is then exposed to a new set of data to make predictions.

AI models can analyse vast quantities of data from different sources far beyond the capacity of humans. “Humans are good at strategy and creative thinking but not great at making decisions from large datasets,” Garry Pairaudeau, chief technology officer at Exscientia, told Down To Earth (DTE).

Traditional methods

The drug discovery phase covers target identification, hit (molecules displaying the desired action or potential drug) generation, hits-to-lead (where hits are optimised to produce leading compounds), lead optimisation (leads are fine-tuned further) and pre- clinical trials.

Traditionally, scientists comb through past studies to identify targets, which can be an enzyme or a protein linked to the disease. This is the first and the most critical step in drug discovery. The time required in this step is undefined, and can be from weeks to months, even years.

Next, researchers try to decipher the target’s 3D structure. To decode the structure, scientists freeze a solution containing the protein and then expose it to electrons. This produces microscope images of individual molecules, which are reconstructed to create the 3D shape or structure. This process can take a month.

Bryan L Roth, professor of pharmacology at the University of North Carolina School of Medicine, US, told DTE that this experimental method is more accurate in finding out shapes than a predicted one.

Source: “AI in small-molecule drug discovery: a coming wave”, Nature Reviews Drug Discovery, 2022; 
US National Institutes of Health’s; company websites The next step is to design and screen chemicals with the potential to bind to the target.

Researchers use automated robots to assess hundreds of thousands of chemicals a day to identify hits. This process can last from a week to three months, according to a 2011 paper in Nature Reviews Drug Discovery.

These hits are further fine-tuned by making small changes to the structure until they become leads or leading molecules. Researchers keep tweaking the design and testing it in cells grown in the laboratory in a cycle, until they land a “near-perfect” candidate, which binds selectively to its target and nothing else to avoid side effects.

Following this is the preclinical step, where it is tested on animals to see if it causes toxicity.

Based on the results, more structural improvements are made to the lead. The final candidate enters human clinical trials.

Malaria, E Coli drugs?

AI can speed up the early discovery process by scouring information from the public and gathering data from the drug developing company’s internal experiments. “By pooling all the information, AI connects the dots effectively, allowing us to make inferences that were probably unimaginable earlier,” Pairaudeau said.

For instance, Debopam Chakrabarti, a drug discovery scientist from the University of Central Florida, US, used AI to develop a different approach for identifying hits (potential drugs) against malaria. Instead of identifying a target and uncovering its structure, he decided to use AI to screen chemicals that could be detrimental to plasmodium, the parasite that causes malaria.

His team trained an ML algorithm with 13,446 publicly available hit compounds against malaria. Later, they fed the tool with 2,400 macrocyclic compounds chosen at random for use as potential malaria drug since they have been explored for treating bacterial infections.

The researchers ran the ML algorithm and physical experiments to screen all 2,400 compounds. ML’s accuracy in predicting potential drugs ranged 70-80 per cent, comparable to experimental results. The findings were published in a 2020 Frontiers study.

Collins and his team also used AI to predict potential drugs against Escherichia coli (E Coli), bacteria known to cause severe stomach cramps, bloody diarrhoea and vomiting.

The tool was first trained with 2,500 molecules, including about 1,700 approved drugs and 800 natural products. It then analysed 6,000 compounds from a library, which hosts a repository of drugs, many of which are US’ Food and Drug Administration-approved, to identify compounds that could inhibit E coli growth.

It predicted 99 molecules as hits, of which 51 were accurate, as established by experiments. One compound, halicin (HAL), stood out. It is structurally different from conventional antibiotics, a feature essential to counter antimicrobial resistance, which was responsible for 1.27 million deaths worldwide in 2019 alone. HAL also kills Mycobacterium tuberculosis and carbapenem-resistant Enterobacteriaceae, acting as a potential broad-spectrum agent. The findings were published in a 2020 Cell study.

AI hyped?

While AI-led drug companies have tasted initial success, there is no guarantee that they will reach the market. For example, Exscientia had designed a compound against obsessive-compulsive disorder for Japanese company Sumitomo Dainippon Pharma, which started the drug’s clinical trials in 2020. But in 2021, the company dropped the potential drug from clinical trials without citing any reason.

Most potential drugs fail clinical trials because the wrong target was chosen. “The first big problem in drug discovery is that we do not understand how drugs work to a great extent. The other big problem is that we do not clearly understand which targets to hit with our drugs to treat a particular disease,” Roth explained.

Take the case of Alzheimer’s. Billions of dollars have been pumped into developing drugs that target beta-amyloid, a small protein. It aggregates to form plaques, which may disrupt cell function. “Nothing has come of that. The hypothesis that beta-amyloid was causing the disease was probably wrong,” he added. An AI cannot do much if our understanding of the disease is incomplete or poor.

AI suffers from another issue. The technology does not develop an understanding of the subject, creating a “black box”. It does not reveal the logic behind its prediction or decision, leading to misinterpretation and bias.

For example, when researchers used AI to diagnose skin tumours from medical images, they found it more likely to flag metre rulers as malignant, not the tumour. The algorithm was trained on medical images containing a ruler to indicate scale. So it “inadvertently ‘learned’ that rulers are malignant”, as per a 2018 study in the Journal of Investigative Dermatology.

This has also created trust issues. According to a 2019 Global CEO survey by international services firm PricewaterhouseCoopers, 84 per cent of CEOs agree that AI-based decisions must be explained to be trusted. Biologists have the same concerns. “People are increasingly developing explainable AIs to understand what led the model to make a certain prediction,” Collins said.

The current verdict is that AI is far from perfect. According to Roth, the AI-predicted and confirmed hit rates are never higher than 50 per cent. But Pairaudeau argued that technology will get better over time as it learns. The algorithm learns from data fed into it from each cycle of optimising the design, synthesising the chemical and testing it.

Researchers also said that AI cannot take over the drug discovery process. “Humans will need to validate predictions experimentally in a laboratory. There’s no substitute for that,” Chakrabarti said.

Collins agrees. He expects the technology will be more widely used in the future. “AI is not displacing old efforts but complementing them.”

As of now, there are no AI designed and developed drugs in the market. Since only 10 per cent of the drugs pass clinical trials to reach markets, AI proponents hope the technology would improve this figure. Until then, “I would not hype it up,” Collins noted.

Subscribe to Daily Newsletter :

Comments are moderated and will be published only after the site moderator’s approval. Please use a genuine email ID and provide your name. Selected comments may also be used in the ‘Letters’ section of the Down To Earth print edition.