Medicine, Meet AI: How Machine Learning is Transforming Drug Discovery

Integrating laboratory capabilities with an AI drug discovery to improve chances of candidate success

This is the start of Eureka's ongoing series, Accelerating Innovation, about new technologies that are impacting the pace of drug discovery and development. Todays is about how the data and technology in AI drug discovery platforms are catching up.

It is a long road to bring a medicine to patients. Before a drug can be made, sold, or tested, the hit compound must first be found before it can be optimized and developed into a usable drug. The selection process is every bit as important as its optimization. Drug discovery and development is a notoriously expensive and inherently uncertain process; seemingly promising candidates fail routinely to the gauntlet of required tests.

However, advances in artificial intelligence in the pharmaceutical industry and machine learning are posed to dramatically increase the rate at which promising compounds are identified and optimized. It is increasingly common to see pharmaceutical companies partnering with nimble AI-powered biotech companies and Contract Research Organizations to develop AI-drug discovery platforms in the development of candidate drugs. But the proof is in the pudding; the best outcomes for AI-drug discovery platforms arise where this state-of-the-art technology is combined with high-quality data.

Choosing the right AI drug discovery platform is critical

The drug discovery process begins as a research hypothesis: that a drug with the specific effect — such as modulating a specified protein channel — will be beneficial in the treatment of some disease. And so, the search begins. Promising available compounds are checked, and if you’re lucky, you find one with the right effect. But fundamentally, finding that compound can take far more time and resources than most of us have, so choosing the right platform to bridge the gap is critical. Guido Lanza, Vice President of Integrated Research at Valo Health, says “We have a plethora of rich data on past discovery and development successes and failures. Through our fully integrated AI-driven platform, we can use this data to capitalize on successes and avoid future mistakes — optimizing the entire process. It compounds too, so we get better and smarter at every turn.”

The growing use of machine learning and data have revolutionized entire fields. To truly maximize the impact of algorithms, domain specific approaches are needed to adequately address the complexity of patterns observed in the lab, in animal models, and, hopefully, in humans. Most off-the shelf approaches, such as the ones that have made high-profile impacts in image recognition, are seldom relevant in spaces where the data is often multifaceted, not well understood, expensive to collect, and highly biased. In other words, recognizing a cat in a photo, winning at alpha go, and designing a drug are all fundamentally different problems.

While machine learning has been used in drug discovery before, it is only recently that the technology and data have caught up. “Models are better, algorithms are better, better datasets, better computational resources…We’re really just starting to see the potential of that on drug discovery projects now, and it's going to play a big part of what we do in the future,” says Grant Wishart, Director of Computer Aided Drug Discovery at Charles River.

AI-Enabled Drug Discovery Solutions

The partnership between Charles River Laboratories and Valo Health has led to the development of Logica^TM, a recently launched joint product offering using artificial intelligence to rapidly deliver optimized preclinical assets to clients. Because Charles River’s platform delivers two different biochemical products — leads and candidates — there is tremendous potential for clients. Extremely promising compounds can be delivered for further testing, and development candidates can be readied for Investigational New Drug (IND) applications.

This is a natural extension of the work Contract Research Organizations (CROs) do already by providing trained laboratory technicians and scientists for experimenting, testing, and developmental services. The difference here is that by using AI drug discovery platforms, the product can be the valuable compound itself. By integrating their laboratory capabilities with Valo’s AI platform, they have demonstrated the ability to produce advanceable lead series with a success rate greater than 90%. This has enormous implications for the future.

When it comes to actual implementation, the question is one of data generation and quality. “It is all about how fast I can close the learning loop,” says Lanza. “How fast can I get the results for an experiment… and feed it back in?” For both quantity and quality of experimental data, there are few peers to rival Charles River Laboratories. Wishart is optimistic about the possibilities: “There's a great potential [with AI drug discovery platforms] to influence what we do with our clients and really redefine how we do drug discovery. We were looking for a partner that could enable that, and Valo was the perfect match. Being able to take Valo's established machine learning platform and combine it with Charles River's drug discovery and experimental success allows us to create a product that will be extremely valuable to our clients.”

Lanza is quite clear that AI is not a “borg” and not some magic bullet, it simply helps maximize the value of the information they already have and collect. The distinction, and the improvements in the drug discovery process, become obvious and tremendously exciting when considered in a real-world example.

The problem was this: Valo wanted to re-design the features of a successful class of cancer drugs, and they wanted it to be able to have it pass the blood-brain barrier and target metastases in the brain. The current drugs cannot do this, and there are very few known examples of compounds that can. But while there were no known examples, Valo does have a lot of existing data about the anticancer activity of those drugs. A model was made to identify all compounds with similar effects and targets of the original drug, and a separate model was built to identify any compound that has ever gotten into the brain. “And if I bring these two models together,” says Lanza, “I’ll can start with a small number of much more promising compounds.” Fundamentally, it’s about improving the probability of success of the process to make better decisions on which compounds to make and test.

A search engine for small molecules

AI upends the darkness and uncertainty of drug discovery: “At any given point, I can tell you how likely I am to succeed, depending on which models I'm bringing to bear and deploying against enormous spaces of chemistry,” says Lanza. “It's not mathematically that hard if I can take a million compounds and get point 0.1 percent hits.” Essentially, it’s making a search engine for small molecules.

But on a practical, commercial level, removing the uncertainty around testing makes it possible to change the way business is done. “The product offering we've put together is a highly value generation-tied business model based on delivering advanceable leads, advanceable chemistries, and candidates ready to enter the clinic. It’s focused on the deliverable and the value of the deliverable, and not an effort and cost-based relationship,” says Lanza. It opens the door for the customers to pay for results, not efforts, aligning Charles River Laboratory’s incentives with the customer internal value drivers. “That’s that kind of alignment that drives faster innovation,” says Lanza.

In April, Charles River and strategic partner, Valo Health, launched Logica, a product that leverages Valo’s AI-Powered Opal Platform^TM Logica’s Accelerated Lead (Logica-AL) and Candidate (Logica-C) programs are designed to deliver preclinical assets, namely leads and candidates, with characteristics that are optimized to reflect a client’s desired target product profile. By seamlessly integrating world-leading laboratory capabilities with AI-driven molecular design and optimization and a trusted track record of clinical candidate delivery, this approach has a demonstrated ability to produce advanceable lead series with a >90% success rate.