How AI Is Driving Innovation And De-Risking R&D

By Grant Wishart and Guido Lanza

Novel drugs are not the proverbial needle in a haystack, waiting to be discovered. They are invented by scientists trained to identify which molecules are most likely to be potent and efficacious using medicinal chemistry and a range of in vitro and in vivo assays.

So why is it so hard to get it right? It’s no secret that developing a drug is an arduous and expensive process. Though it is certainly true that failure is an essential and inescapable part of scientific research, the extremely low probability of success begs the question: Are we really learning from our mistakes?

A big part of this struggle is that drug developers are often seeking targets that are becoming harder and harder to crack, and, once cracked, the likelihood of target modulation resulting in the desired clinical efficacy is small. If developers are lucky, their promising compound will hit the laboratory equivalent of a bull’s-eye. But fundamentally, finding that compound can take far more time and resources than most drug developers have, leaving patients waiting in vain for better treatments and cures.

The good news is that AI can provide a better path forward — and perhaps a paradigm shift in how we select the novel drugs that will go on to make a difference. With AI there is real progress being made in determining, early on, with a greater degree of accuracy, which molecule might be a hit, and which one won’t.

AI is not a new tool in drug development. Look around and you can find examples of how AI is shaking up the pharmaceutical industry. It is being used to support diagnostic decision-making in the medical imaging space and driving sustainable histology in preclinical pathology. The complexity of drug development and the huge amounts of data being generated in the process make drug discovery an ideal arena for the application of AI tools.

For example, the data-generating biology and chemistry conducted in the earliest stages of drug discovery are being combined with next-generation AI platforms to make drug discovery and development more efficient and less costly. Predictive models driven by AI are helping to alleviate the uncertainty of discovery by providing developers with a better understanding of which candidates are advanceable — saving precious time and money while lowering risk.

Additionally, integrating molecular design capabilities and biological assay data in an active learning loop can drive the design process in partnership with medicinal chemistry expertise. These predictive models also can provide insight into in vitro and in vivo models that can shape which experiments are necessary to validate safety and efficacy.

A SEARCH ENGINE FOR SMALL MOLECULES
Here’s how this relationship might work. Suppose you want to redesign the features of a successful class of cancer drugs to enable them to pass the blood-brain barrier and target brain metastases. We know few oncology drugs can do this, and there are very few known examples of compounds that can. What does exist, however, are a lot of existing data about anticancer activity of those drugs. So, if you create a model that identifies all compounds with similar effects and targets of the original drug and build a separate model to identify any compound that has ever gotten into the brain, you can hopefully begin to find and optimize promising compounds.

AI systems have been aided by tremendous strides in bioinformatics. The amazing and early success of DeepMind’s AlphaFold program, whose database of predicted protein structures now tops more than 200 million, freely shared, is having a profound impact on scientific research. At the same time, “on demand” databases are cropping up that contain structures of chemical compounds that have probably never been made but that chemists believe could be readily synthesized using well-validated chemical reactions and available reagents. These databases are dramatically expanding the chemical space — a kind of final frontier for proteins.

AI and machine learning applications are also helping to upend the darkness and uncertainty of drug discovery by narrowing down the vastness of chemical space into more manageable sections for faster searchability. In this vast chemical space, AI can be used by scientists to probe the biology of potential compounds and identify their best features. From there they can create a more qualified series of compounds and further optimize them to deliver preclinical candidates for IND-enabling studies.

AI CAN’T DO IT ALL
However, AI is only part of the equation in our efforts to drive better drug discovery. Over the last two decades we have seen hundreds of millions invested in the AI drug discovery industry with mostly limited returns on investment. It has become clear that drug discovery is a complex problem for AI alone. You also need the pragmatic approach taken by laboratories through assay development, medicinal chemistry, and high throughput screening in combination with sophisticated analytical technologies, state-of-the-art equipment, and the input and experience of hundreds of drug discovery scientists — from medicinal chemists to structural biologists — to provide empirical data and make these platforms truly deliver what they have been developed for. Furthermore, animal models continue to be needed for things such as toxicology and safety testing and are required by the regulatory agencies, such as the FDA, to move investigational new drugs (INDs) to first-in-patient studies. Although the predictive capabilities of AI and associated models continue to improve, they are not yet able to replace many of these experiments that take place in the lab, and as such — at least for the moment — we need a combination of traditional and modern approaches to achieve the best outcomes. In the future, AI may be able to limit the number of experiments needed in the lab or replace them altogether, but for now both are required for the best result.

AVOID THE HYPE
As AI becomes better at predicting outcomes — information that we can now only gain through experiments in the lab — it may be possible to reduce experimental testing further, thereby decreasing costs and speeding up the process. For many, it remains an ongoing challenge separating the hype from the actual hits. AI’s influence is measurable but still incremental in drug R&D, as the impact is limited by AI being deployed within the existing silos to improve existing workflows Most off-the-shelf AI approaches, such as the ones that have made high-profile impacts in image recognition, are seldom relevant in spaces where the data are often multifaceted, not well understood, expensive to collect, and highly biased.

Everyone loves the Hollywood blockbuster, but the notion that machines alone are creating cures for diseases is exaggerated. In a traditional drug discovery framework, it can be challenging for drug developers to foresee obstacles eventually encountered in preclinical and clinical trials. If we had the data and AI modeling technology to predict the journey of a candidate earlier in the process, plus human expertise, we would finally increase our success rate and get better molecules into the hands of patients faster. We are aiming for this as our Hollywood ending.

GRANT WISHART, PH.D., is senior director, small molecule drug discovery and Logica lead, at Charles River.

GUIDO LANZA is VP of integrated research and Logica general manager at Valo Health.