By Teresa Meek
A
mong the business processes most in need of reform, drug discovery likely is near the top of the list. The average cost to bring a new drug to market is $2.6 billion—and the average time it takes is 10 years. Researchers have been trying to accelerate drug discovery with computers for decades, with limited success.
Artificial intelligence (AI) is exhibiting real promise. By finding the best drug candidates, shedding new light on disease and crunching an unprecedented amount of patient data, AI is revolutionizing drug discovery and development from start to finish.
Big data strategies deploying AI capabilities, such as predictive modeling and analysis of sensor data, could generate up to $100 billion in value annually across the U.S. healthcare system, according to McKinsey. Most pharma companies already have an AI program in place. A study by Carnegie Mellon and a German university suggests AI could lower their discovery costs by 70%.
Here are some of the ways AI is making a big difference in finding and testing new drugs.
Zeroing In On Winners
With its mind-boggling number-crunching ability, AI can evaluate potential drug candidates exponentially faster and better than humans. It does so with artificial neural networks, which mimic the human brain in recognizing patterns and adapting to change, but take in and process far more information than our brains could ever handle.
Judging a potential drug candidate involves examining billions of data points in a thousand or more data types, says Andrew A. Radin, the CEO of AI-driven drug discovery company twoXAR.
Computational neural networks are ideal for unearthing those possibilities that stand the best chance of working, including some that scientists may not have considered.
A number of the tests that researchers carry out during the drug discovery process still use biochemical assays. More recently, researchers developed a method that exposes cells representative of a specific disease to many drug candidates simultaneously. A high-quality camera observes these assays, collecting thousands of images of the cells and documenting their reaction to the drug candidates. This method is called high-content screening (HCS).
The analysis of these images was initially a more manual process. Eventually researchers began using computers to extract features from images of the cells in the HCS assay to develop a classification system for assessing drug candidate performance. But that process remained time-consuming.
“When you’re looking at thousands and thousands of cancer cells, it’s daunting,” says Michael McManus, principal engineer and senior health and life sciences solution architect at Intel. “Allowing neural networks to do the work is far faster and improves accuracy.”
Of course, neural networks need to be trained in what to look for before they begin their jobs. But faster data processing is speeding up that task, too. A neural network that Novartis and Intel developed reduces the time required to train an AI system on the images from the HCS process from 11 hours to just 31 minutes.
Speed and accuracy are critical factors in drug discovery and development, where a mere 14% of candidates emerge from the exhaustive testing process to gain FDA approval.
Just getting from an initial idea to the start of the animal-testing phase typically takes four to six years, Radin noted. Using AI to zero in on the right compounds has allowed twoXAR to reduce that time to about three months.
“Traditionally, you’d test a hundred compounds, and after taking on a sizable team and spend, you might get one or two that show efficacy,” he says. “We test 10 drugs and three will show signs of efficacy. We do it all computationally.”
AI also frees scientists from choosing which kinds of information to study, thereby reducing bias.
“Software has no opinions, and it’s not emotional,” Radin says.
Defining The Enemy
In parallel with identifying strong drug candidates, scientists create a 3D model of the disease protein the drug candidates are designed to fight. The 3D model is used to perform molecular dynamics and “docking” studies to further evaluate the drug candidates. Achieving a 3D model of the diseased protein has, however, traditionally taken place via crystallography, in which scientists grow crystals of the disease protein and elucidate its structure using X-rays. Unfortunately, not all proteins can be crystallized—and even when they can be, the process can take years, McManus says.
Over the past several years, drug discovery researchers have started using a new Nobel Prize-winning process called cryogenic electron microscopy (cryo-EM), which can shorten the process of obtaining a 3D model of the disease protein from years to weeks.
Rather than growing crystals, researchers chill a solution of the disease protein to produce a nearly frozen solution containing lots of protein particles. These particles represent the disease protein in random orientations in the solution. A software algorithm picks the best particles for targeting with the electron beam of the electron microscope. Once the protein particles are selected and bombarded, the electron microscope collects terabytes of images. Other software algorithms, some using AI, then analyze these images to create a detailed 3D model of the disease protein. With the 3D model in hand, researchers can use computers to further study the interaction of the drug candidates with the disease protein.
During the recent Zika virus outbreak in Brazil, scientists used cryo-EM to create a high-resolution 3D image of the virus within months, giving them a powerful tool in the search for a treatment.
Learning More From Patients
Once researchers have a strong drug candidate that fits a disease, they move on to clinical trials, where AI helps in several ways.
First, by analyzing complex genomic data from patients who have the disease, AI allows drug discovery researchers to select participants whose genetic profile suggests they will benefit from the drug candidate undergoing testing in the trial.
“There’s a lot of genetic diversity on Earth. Before genetic screening, some people admitted to a clinical trial may not have benefited from a drug because they may not have had the exact same protein for which the drug was created. In addition, there are other potential issues, such as toxicity. A person may have the correct protein, but their liver cannot metabolize the drug, leading to a toxic reaction,” McManus explains. “For others, the drug may have no effect. Researchers are selecting people for clinical trials with high confidence that they’ll benefit from the drugs.”
Then, once trials start, AI helps scientists collect more and better information from patients.
Traditionally, drug trial participants are required to keep a daily journal documenting drug use and symptoms, and they must periodically visit a regional medical center.
But journaling is subjective. Patients often forget to note symptoms or describe them in varying ways. In addition, trial participants don’t always show up for appointments—and may forget to mention important information when they do.
AI-based systems, such as Intel’s Pharma Analytics Platform, relieve clinical trial participants of many of their burdens while giving researchers better information.
Instead of keeping a journal and visiting a doctor far from home, patients don a wearable connected device that can measure heart rate, breathing, gait and a host of other biometric measures. Data collects in a continuous stream and securely uploads to the cloud, where it undergoes analysis to reveal disease progress and measure the drug’s effectiveness in combating the disease.
By providing objective data and analyzing it in vast quantities, AI can reduce the cost of drug discovery trials and speed time to market for new drugs. It also helps researchers learn more about how drugs are metabolized. In addition, these wearable devices let trial physicians learn about unintended adverse reactions to the drug far quicker than they would if they waited for the patients to visit the doctor again.
Israeli pharma company Teva used the Intel Pharma Analytics platform to develop a method of analyzing data from patients with Huntington’s disease, a neurological disorder that causes motor problems. In a phase 2 clinical trial, accelerometers and gyroscopes recorded the patients’ body movements. Algorithms were developed to measure the severity of chorea of the upper limbs—a condition involving involuntary jerky motions, especially in the shoulders, hips and face. This taught researchers more about the progress of the disease and the trial drug’s effect on it.
“A clinical trial is not just about the drug candidate,” McManus says. “It’s about both the positive and negative impacts of that drug candidate on the body. Today, we can get a better sense of those impacts through wearables.”
From finding promising drug candidates to creating disease-fighting models and running better clinical trials, AI is transforming the entire drug discovery process. As data accumulates, this young technology is bound to develop still more ways to improve drugs and get them—more quickly—to the people who need them.
Teresa Meek lives and works in Seattle. With over 15 years of experience in communications, she has also written for the Miami Herald and Newsday.
Credit: SolStock/GettyImages
