Artificial intelligence (AI) –  which learns from computer algorithms how to unravel complex genomic data, such as gene expression patterns of disease – is poised to transform every dimension of drug development, clinical research, and medical treatment. But it also may achieve another milestone for humans – lowering the price of medicines.

Many AI experts describe the value of this remarkable technological advance as reducing, if not eliminating, the inefficient, time-intensive trial-and-error process of innovation.  A Pharmaceutical Research and Manufacturers of America (PhRMA) study has shown that only about 12% of drugs entering clinical trials achieve regulatory approval.

AI, or machine learning – and its more complex offspring, deep learning – aim to change that ratio of success-to-failure. AI technology, the experts say, has the potential to turn drug candidates into bedside medicines much faster and much more efficiently than conventional, hands-on research.

In short, AI represents the beginning of the industrialization of biotechnological and pharmaceutical innovation.

This doesn’t mean AI will replace research scientists, says NuMedii CEO Gini Deshpande, who prefers the term “augmented intelligence.” It will take a combination of human and machine intelligence “to streamline the discovery of relevant disease biology, thereby reducing time for discovery, but more importantly, significantly improving the probability of success.”

And assuming it works, Andreesen Horowtiz general partner Vijay Pande observes, it could support reductions in the price of drugs in two ways.

First, drug companies won’t have to pass along to payers the costs of all their clinical trial failures; that 90% of drug candidates that never reach approval. And second, by getting to market faster, companies will have more patent-protected years to amortize their R&D costs.

So far, biotech and pharmaceutical companies have been slower than other industries to embrace AI. But they are beginning to take notice. “As with any new technology in the bio space,” says Pande, “it takes a while to get adopted. There is a point where it starts working; and that’s what we and others are proving now, that it’s working.”

WuXi AppTec – a leading global pharmaceutical and biopharmaceutical open access capability and technology platform – assists biotech and pharmaceutical companies from discovery to manufacturing and beyond. An important element of this support involves offering a communications platform to facilitate the exchange of ideas among the most innovative companies and the creative people behind them.

In this installment of WuXi’s new communications platform on the future of drug discovery and development, leading experts in applying artificial intelligence discuss the prospects for this exciting new technology.  They include NuMedii CEO Gini Deshpande, Exscientia CEO Andrew Hopkins, Numerate CEO Guido Lanza, Andreessen Horowitz General Partner Vijay Pande and WuXi NextCODE CEO Hannes Smarason. Their complete interviews also are available on this website.

The excitement surrounding the application of artificial intelligence (AI) to biomedical innovation is based on success in other fields where AI algorithms have trained machines in learning how to recognize faces, how to talk, how to drive cars, how to play games, and how to compose music.

As described in an article in Molecular Therapeutics, “The type of learning required in these tasks is representation learning; that is detecting, or classifying patterns, or representations from raw data.”

This makes AI and its subsets, machine learning and deep learning, a natural fit for mining and relating the mountainous genotypic and phenotypic data being collected worldwide in public and private databases, in hospitals and doctors’ offices, in academic research journals and in individual wearable health monitoring devices.

Pande of Andreessen Horowitz observes, “I think biologists have recognized for decades, maybe centuries, that biology is very complicated…so complicated, the human brain couldn’t understand all of it; and that the limitations we see in understanding biology translate into the complications we see in late-stage clinical trials.”

AI’s evolving proficiency in pattern recognition and representation learning leading to more accurate predictions of outcomes make it the most promising research tool yet in unraveling the genomic-level complexity of disease pathways, designing therapeutic interventions, and identifying patients who will benefit.

“Biology may be too complicated for the human brain,” Pande says, “but may not be too complicated for certain types of AI. AI can integrate the data in a way the human brain can’t, and then be able to guide researchers into interesting new places to look.”

The various descriptions of just what AI is and what it means underscore the promise it engenders for demystifying biology and improving health care for patients.

WuXi NextCODE CEO Hannes Smarason observes AI transforms research from a hypothesis-driven exercise to a data-driven process. For example, he explains, “A scientist is about to run an experiment and before he knows it, the knowledge system pings something and says, ‘Here are the three things you should do;’ or, why don’t you look at these four results? I already did the experiments for you.’”

Exscientia CEO Andrew Hopkins adds that AI “systematizes work that has to date been extremely manual and whose results have been highly dependent on the expertise of those put on the project.”

For Deshpande of NuMedii, AI enables researchers to find patterns “one might be trying to look for, or patterns one didn’t know or wasn’t expecting to find.”

Numerate CEO Guido Lanza says AI introduces a “true learning loop for the industry. The idea that all decisions can be guided by all preceding successes and failures is profound.”

So when will all this algorithmic prowess be unleashed? Most experts agree biotech and pharma companies are about 10 years away from fully integrating AI into their R&D operations. But they also agree the technology will become an essential tool for drug development.

“I don’t know if there is going to be a place for a biotech or pharmaceutical company that does not use AI,” Smarason says. “This is one of those technologies that is so comprehensive and ubiquitous that it’s going to be very hard to compete without having fully understood and embraced the technology.”

Where are we now?

Much of the initial appeal of AI for biopharma companies is in streamlining R&D operations and increasing the speed and success rate of clinical trials. It’s not surprising.

PhRMA estimates the average cost of developing a new drug is $2.6 billion, which includes the cost of failures, spent over about 10 years with most of that time consumed in testing the drug candidates in clinical trials.

A report from TechEmergence, which tracks AI applications across all industries, says increasing the success rate for new drug development just a few percentage points, from 12% to about 14%, could save the biopharma industry billions of dollars.

In PhRMA’s report on drug development costs, the association acknowledges that one element of successful R&D and clinical research has been “luck,” which is precisely what AI companies hope to reduce, if not eliminate.

“What we’re starting to see is that pharma is getting very interested in this,” says Pande, who joined Andreesen Horowitz to lead its venture capital investments in AI biopharma start-ups, following his tenure at Stanford University where his research laboratory still bears his name.

“What we’re seeing already,” he adds, “and something I’ve done with my research at Stanford, is that with machine learning and artificial intelligence you can go very quickly from one (drug candidate) hit to getting through preclinical trials, and by quickly, I mean that process could take a year or less.”

Genomics, Pande says, sets up well for AI. “The central challenge is how can you represent biopharma data in a way that’s amenable? Certain types of data sets work very well, like genomics. It is almost like a one dimensional picture and it carries over beautifully in areas such as diagnostic tests,” he says. “One of our portfolio companies, Freenome, is a great example. They can use genomics plus AI to identify early stage cancers in the blood, something you just couldn’t do before.”

WuXi NextCODE’s Smarason agrees that AI has been the missing piece in understanding genomics. “One of the things we’ve seen come out of our work in deep learning,” he says, “has been this ability to identify the true causality in relationships within the biology in question. We can find the causal genes or pathways that drive disease, and the hope is it will allow us to have a much more powerful starting point for the development of treatments.”

For example, Smarason says, his company worked with Yale University’s Department of Medicine in discovering a previously unknown mechanism in development of the human vascular system, which has implications for better understanding both the cardiovascular system and the vasculature of cancer tumors.

“Our deep learning algorithms made a prediction that a particular mechanism was a key driver in the pathway for development of the vascular system,” he says. “Yale biologists then validated that discovery in an animal model.” As a result, Smarason observes, researchers now have a “whole new druggable pathway” to explore.

NuMedii’s Deshpande says her company “has been pioneering the combined use of big data, AI and systems biology since 2010 to accelerate the discovery of precision therapies. AI approaches,” she adds, “are a natural fit to harness big data as they provide a framework to ‘train’ computers to recognize patterns and sift through vast amounts of new and existing genomic and other biomedical data to unravel diverse complex biological networks involved in disease processes.”

NuMedii, she says, is using “multiple AI methods, ranging from classical machine learning techniques to newer deep learning systems to rapidly discover connections between medicines and diseases at a systems level. Our AI approaches are also being used to identify subsets of patients and therapies that are likely to modulate these complex networks for each patient subgroup.”

Exscientia is using AI to automate the design of new drug molecules. “Our approach,” says Hopkins, “has already demonstrated the potential to improve the time required from project start to (drug) candidate to just one quarter of the industry average. The validation of our approach is the fact that we now have the first molecule heading to the clinic that was discovered within only 12 months of starting the project from just a target product profile, using our platform. For us, AI drug design can bring profound strategic advantages.”

In addition to designing molecules for biopharma companies, Exscientia also plans to develop some of its own discoveries. “Our company values working with collaborators,” Hopkins says, “but we are also developing our own portfolio of compounds. Our systems are highly scalable so we do not want to be constrained by only working on targets that are selected by partners.”

Numerate, says Lanza, was a pioneer “at a time when nobody was looking for AI. We did this by starting with a team of both computer scientists and drug hunters – people with compounds in the clinic and on the market.”

The company is applying AI to small molecule drug discovery. “From a scientific point,” Lanza says, “our differentiators are around our translational capabilities. First, we are able to work on emerging biology with extremely small datasets, the kind not suitable for deep learning approaches. Second, our modeling is based on 3D ligand formation. The other translational axis is around our ADME (absorption, distribution, metabolism and excretion) and toxicity prediction capabilities.”

Where are we headed?

NuMedii’s Deshpande sums up for much of the AI biopharma industry where these companies are headed.

“Our goal,” she says, “is to speed up drug discovery, cut R&D costs and decrease failure rates in clinical trials, all of which can eventually lead to better, more precise medicines.”

In addition, such a significant transformation in R&D, whose costs have been steadily increasing, should logically get passed onto consumers, or payers, in the form of lower prices.

A 2014 study by the Tufts Center for the Study of Drug Development revealed the average pharma R&D costs for approval of a new drug increased 145% over 10 years.

Pande predicts AI could bring about lower drug prices by achieving two significant industry milestones. “One is greatly shortening the amount of time for development,” he says, “compressing the time for development will be huge.”

The reasoning here is that getting to market faster soothes one of the major headaches for biopharmaceutical companies – lack of patent protection for market exclusivity to recoup R&D expenses.

In a May 23, 2016 article in the scientific journal SpringerPlus, researchers observed that from initial patent filing to regulatory approval takes an average of 12 to 13 years, leaving 7 to 8 years of patent-protected market exclusivity. They concluded that was “insufficient time for most new drugs to recoup the up-front R&D costs and earn a positive return on this investment.”

“If you can get these drugs out faster,” Pande notes, “you can have more time to amortize the cost over more years. That’s particularly appealing.”

The second way AI could affect drug prices is increasing the success rate of clinical trials. According to an October 3, 2016 article in Clinical Leader, the cost of clinical trial failures is “estimated between $800 million and $1.4 billion,” equaling from about one-third to more than half of the average $2.6 billion cost of new drug development.

Clinical trial failures, Pande says, have “to be paid for somehow, and so if you can have fewer failed trials, of course that would decrease the cost.”

Beyond these “two easy wins” Pande says, AI could have an impact on drug prices in unexpected ways. “One area we are seeing is the rise of digital therapeutics,” he says. “These are software programs for behavioral therapies, typically some version of cognitive behavioral therapy.”

For example, he notes, a diabetes prevention program conducted by the U.S. Centers for Disease Control and Prevention showed a behavioral therapy approach to Type 2 diabetes was more efficacious than the drug, metformin.

Digital therapies also could be developed for anxiety, depression, even Alzheimer’s disease, Pande suggests, providing alternatives to medicines.

“Another area is an interesting twist on this,” he says. “You could combine a digital therapeutic and an existing small molecule (drug). You can take even a generic and make it best in class with low cost.”

Most of these anticipated R&D saving measures, however, are long-term, and there are still challenges to overcome in the application of AI.

No AI-driven medicine has yet received regulatory approval, Smarason says. “But I would say we are definitely in the sub-10 year time frame before we see something important come out where people can directly tie it back to AI.”

Among the challenges, according to Hopkins, is the requirement of “domain expertise to know what a reasonable question to ask is. There are cases where questions being poised are too ambitious, the associated data insufficient and the AI question poorly constructed.”

Deshpande says another challenge is access to “high quality and consistent data to train algorithms for accurate pattern recognition. Often data are kept in silos and spread across organizations.”

Lanza adds that overcoming traditional R&D culture is also a challenge. “AI is, by its nature, not meant to be interpretable, but used more as a black box,” he says. “I often hear that in order to believe the predictions, the scientists want to know how those were arrived at by the AI. This is the wrong way to think about AI generally. The point is these algorithms can see signals in data that are either too narrow or too broad for a human to see. If we put a requirement on AI that it generates human-interpretable results, we are likely limiting the AI to the least interesting problems.”