SCI: AI in pharma: Helping science to find drugs faster

BY JASMIN FOX-SKELLY

Pharma companies are increasingly harnessing AI to accelerate drug discovery. But will AI transform the hunt for new medicines, or are the benefits overhyped? Jasmin Fox-Skelly reports

Discovering a new drug used to take more than a decade of research and billions of pounds of investment. Many candidates fail along the way, with the pharma industry spending an estimated $50 to $60bn/year on failed cancer drug trials alone[1]. However, AI algorithms can cut the time needed to identify a lead compound to a matter of weeks, if not days.

‘The first step in drug discovery is identifying a ‘hit’ or lead compound, and that has been entirely solved by AI,’ says Michele Vendruscolo, Professor of Biophysics at the University of Cambridge, UK.

Traditionally, drug discovery starts with ‘high throughput’ screening of thousands of chemical compounds to find which ones, if any, bind to and interact with the target protein, which can take up to two years. AI can dramatically speed up this process. Chemical libraries contain information about hundreds of thousands of molecules, including their chemical structures and physical characteristics.

Based on this information, AI models can predict how a molecule will bind to and interact with its target protein, by rapidly analysing vast quantities of data. Millions of potential candidates can be screened, allowing researchers to focus on the most promising leads.

‘We now have decades of results from X ray crystallography and micro spectroscopy studies,’ Vendruscolo explains. ‘We know the structure of hundreds of thousands of proteins. AI is very good at learning from that huge amount of structural data.’

Higher throughput

According to Vendruscolo, AI is now as accurate as experimental screening methods at predicting protein structures and how small molecules interact with them. However, so far this has only proved true for proteins that undergo a folding process to form 3D stable structures with binding pockets into which small molecules can fit.

‘The textbook definition of proteins is that they fold into their native state and then they function,’ says Vendruscolo. ‘But it turns out that about one third of human proteins don’t do that.’

These other ‘intrinsically disordered’ proteins don’t acquire a single conformation, so their structures cannot be determined by standard experimental and computational methods. Many are involved in human diseases, including neurodegenerative conditions such as Alzheimer’s and Parkinson’s. Until recently these proteins were considered ‘undruggable’.

‘The traditional concept of binding in drug discovery is called lock and key, where you have pockets in the surface of the protein which are like locks, and then the small molecule can fit into the crevice and that is like a key,’ says Vendruscolo. ‘But if the protein doesn’t [have a defined structure] there is no lock and no key. There are no approved drugs for clinical use that target dissolved proteins because there are no pockets.’

Vendruscolo and his team have used AI to identify a new way of binding, in which the small molecules don’t need a pocket. The team focused on amyloid beta, an intrinsically disordered protein implicated in Alzheimer’s disease. Clumps of amyloid beta form structures called plaques, which accumulate around neurons, causing them to die.

In soon-to-be published work, the team used AI to quickly screen a chemical library containing millions of small molecules and identified five compounds for further investigation. These compounds don’t bind to a pocket but rather ‘dance around’ the disordered amyloid beta protein before binding to the protein molecules, stabilising them and stopping them from clumping together. 

‘The rules for disordered bindings are likely to be far, far more complex than the rules for lock and key,’ says Vendruscolo. ‘However, AI programs based on deep neural networks with trillions of parameters could learn to understand them.’

The genetic makeup of lung cancer tumours in people who have never smoked is completely different to that of smokers. In never-smokers, tumours typically have large numbers of mutations in kinase receptor proteins on the surface of tumour cells.

Predicting whether a drug will be toxic to humans is one area in which AI has made little ground, due to the lack of data available to train models.

Antimicrobial search

AI’s ability to predict how molecules bind to proteins has also been used by researchers developing new classes of antibiotics. Most antibiotics in use today were discovered over 50 years ago, with pharma companies mostly abandoning the quest to develop new ones, partly because of the rapidity of bacterial resistance.

The hope is that AI can solve this problem by looking for antimicrobials in unusual places, or by designing entirely new molecules that don’t exist in nature. Such drugs may be harder for bacteria to evolve resistance to.

César de la Fuente, Presidential Associate Professor at the University of Pennsylvania, US, has spent over a decade using AI to trawl genetic databases looking for antimicrobial molecules. In 2023, he and his team searched the DNA of extinct Neanderthals and Denisovans looking for genes that might code for antimicrobial peptides[2].

They found one peptide, neanderthalin-1, which – when recreated in a lab – was effective at treating bacterial infections in mice. Then in 2024 they used a new AI deep learning model called APEX to trawl the ‘extinctome’ – the 208 extinct organisms with sequenced genomes available to science[3].

Most recently, de la Fuente used AI to scan 233 species of Archaea – ancient single celled microbes – which yielded over 12,000 antibiotic candidates. Some 80 of these were synthesised, with tests showing 93% were effective at neutralising six common pathogenic bacteria in the lab[4].

‘We’re doing experiments to see how different bacteria evolve resistance to different molecules,’ says de la Fuente. ‘We’re doing this both from the bacterial perspective, by mining the genome of ancient bacteria to see what mutations arise, but also from the molecule perspective to see what particular sequences of amino acids are less likely to lead to resistance.’

The hope is that AI could then be used to predict how and when bacteria might evolve resistance to any new drugs.

‘By compiling all of this information it should be possible to train AI models that can give us probabilistic outputs of potential mutation trajectories into the future, and then that can inform how we try to develop new drugs,’ says de la Fuente.

James Collins, Professor of medical engineering and science at the Massachusetts Institute of Technology (MIT), US, is also using AI to discover new antibiotics. In 2020, his team identified a compound, halicin, that can kill E. coli and other bacteria resistant to antibiotics[5]. Halicin is an entirely new class of antibiotic, which attacks bacteria using a method never seen before – by disrupting their ability to maintain an electrochemical gradient across their cell membranes. More recently in 2025, his team asked AI to design novel hypothetical antibiotic molecules that could prove effective against superbugs[6]. The idea was to generate new structures to reduce the chance of bacteria evolving resistance. The two antibiotics with the most potential were then synthesised, with tests showing both were effective at killing MRSA and antibiotic-resistant gonorrhoea in the lab. They were also able to cure mice infected with these bacterial strains.

‘AI has the potential to transform the way we look at antibiotic drug discovery,’ says Hazel Barton, Professor of geological sciences at the University of Alabama, US. ‘For example, by discovering new and novel molecules that there’s no way nature could synthesise, because the chemistries are impossible for an organism to replicate.’

Intrinsically disordered proteins don’t acquire a single conformation. Many are involved in human diseases, including Alzheimer’s and Parkinson’s. Until recently these proteins were considered ‘undruggable’.

AI could be used to predict: whether a molecule will be toxic before it enters human trials; whether a molecule can cross the blood brain barrier; and how drugs given in combination could interact.

Later stages

But, while AI is excellent at uncovering new ‘hit’ compounds, this is just one part of the drug discovery process.

‘Most drugs don’t fail at the hit discovery step,’ says Vendruscolo. ‘In particular, you need to find out if the compound is toxic, if it’s effective enough, and if there is a dosage that can be translated from mouse studies to people in clinical trials. These parameters are largely discovered by trial and error with huge amounts of uncertainty, which in turn translates to high failure rates.’

However, some companies believe AI could solve at least some of these problems too. In the US, Lantern Pharma, a Texas-based oncology biotech company, has used AI to try to understand why certain subsets of cancers respond better to treatment than others. To do this, the team fed machine learning models huge amounts of clinical trial data, to discover if the genes or proteins produced by a particular tumour can be used to predict whether a particular drug will be successful at treating it.

‘The initial intent was to rescue drugs that failed in late-stage cancer trials,’ says Panna Sharma, CEO of Lantern Pharma.

Lantern Pharma initially focused on LP-300, a drug that failed five Phase 3 trials for lung and breast cancers between 2006 and 2013. Subsequent analysis by company researchers showed that LP-300 was effective in treating non-smokers with certain types of lung cancer.

‘It was aimed at the wrong patient population when it failed,’ says Sharma. ‘So we used AI to really determine the mechanism of action at play in never-smokers.’

The results were astonishing. It turns out the genetic makeup of lung cancer tumours in people who have never smoked is completely different to that of smokers. In never-smokers, tumours typically have large numbers of mutations in kinase receptor proteins on the surface of tumour cells. Kinase receptors primarily bind to growth factors, triggering a cascade of events inside the cell which ultimately drives cell growth, survival and migration. Mutations in kinase receptors can lead to uncontrolled cell growth and division, contributing to the development and progression of many aggressive cancers.

‘Our drug works by binding to the kinase receptor, which are usually over-expressed in these never-smokers, and denaturing it so it slows the cancer’s ability to drive cancer growth,’ says Sharma.

LP300 is currently being tested in a Phase 2 clinical trial for non-smokers with advanced primary adenocarcinoma of the lung, a type of non-small cell lung cancer (NSCLC).

By helping to find the right patients for a particular drug, Sharma believes that AI could significantly speed up the drug discovery process.

‘Once you think you have a molecule, selecting the right indication is really important,’ says Sharma. ‘There are drugs that are going to be much more potent and much more effective in certain cancers versus others, and if you can understand that and understand the rationale for that early on, it’s going to help you position this molecule clinically.’

Clinical positioning can often take between one to three years, but Sharma believes AI can compress that timeline from 50 to 75%. He contends that AI could be used to predict: whether a molecule will be toxic before it enters human trials; whether a molecule can cross the blood brain barrier; and how drugs given in combination could interact. Even the process of human clinical trials – the longest part of the drug discovery process where most failures occur – could be streamlined by AI. For example, by targeting patients whose genomic profile suggests the highest probability of benefiting from a drug.

‘You will get higher [trial] success rates because you’re going to go after more targeted populations, and you’re going to know more about the mechanisms,’ says Sharma. ‘The cost of those trials is also going to come down by maybe 20-30% because you’ll better understand which patients you really want in your trial, versus those you don’t.’

However other scientists are more sceptical about the extent to which AI can help with drug discovery. Despite the fact pharma companies started to use AI to find new drugs in the mid-2010s, no AI drug has yet to pass Phase 3 trials.

‘AI is very good when there is lots of data, but not very helpful when there isn’t,’ says Vendruscolo. ‘There are some problems that cannot be solved at the moment, and perhaps won’t be in the foreseeable future, because we don’t have enough data, or the data is sparse and is not organised well.’

He points to predicting whether a drug will be toxic to humans as one area where AI has made little ground, due to the lack of data available to train models.

‘A compound can be toxic in a million different ways, so this is where I think the hype of AI is meeting reality. My view is that people should value AI for the specific problem it can solve and be aware of the problems that cannot be.’