Artificial Intelligence Is Helping Biotech Get Real – Genetic Engineering & Biotechnology News
AI may be used to extract insights from millions of experimental affinity measurements and thousands of protein structures to predict the binding of small molecules to proteins. This approach is being realized by Atomwise, the developer of AtomNet, a structure-based, deep convolutional neural network designed to predict the bioactivity of small molecules for drug discovery applications. AtomNet helped Stanford University researchers screen 6.8 million small molecules for their ability to target Miro1, a protein implicated in Parkinson’s disease. In the space filling structure for Miro1 shown here, the area in white represents the screening site. The most promising compound, the Miro1 Reducer, appears in the inset.
Artificial intelligence (AI) may sound futuristic, but it already exists in many everyday technologies. For example, it gives our handheld devices voice and facial recognition capabilities. AI is also making its presence felt in biotechnology, where it has become integral to many aspects of drug discovery and development.
AI applications in biotech include drug target identification, drug screening, image screening, and predictive modeling. AI is also being used to comb through the scientific literature and manage clinical trial data.
By leveraging machine learning, AI can manage disparate clinical trial datasets, enable virtual screening, and analyze vast amounts of data. Besides reducing clinical trial costs, AI can gain otherwise unobtainable insights and feed them back into the drug development process.
AI technologies to serve the biotech industry are being developed by several companies. Their services are rapidly becoming indispensable as older methods like classical statistical analysis or manual image scanning reach their practical limits.
A new world of abundance
Atomwise was the first company to apply a common type of machine learning, convolutional neural networking (CNN), to drug design and discovery. CNN is used in familiar everyday applications like Alexa’s speech recognition technology or Facebook’s image tagging feature. Atomwise has 550 ongoing machine learning projects focusing on problems like hit discovery, potency optimization, selectivity optimization, and off-target toxicity testing.
According to Atomwise CEO Abraham Heifets, PhD, there is virtually no limit to the number of small-molecule compounds that can be virtually screened using Atomwise’s algorithm. “We did the largest screen in human history recently—12 billion molecules,” says Heifets. Most of those 12 billion molecules don’t exist in nature and have never existed. Any of them, however, can be synthesized by Atomwise partners. These vendors can quickly synthesize interesting candidate molecules from a set of basic building blocks and deliver them within four to six weeks. Heifets asserts that the availability of those theoretical compounds is increasing rapidly, and soon it may be possible to screen 100 billion molecules.
The problems of drug screening shift with a large abundance of molecules. “If you have 100 billion molecules and you have a 99% accurate model, it sounds pretty good.” Heifets explains. “But that means you have a 1% false-positive rate, and your right answer will be swamped in a billion false positives. In fact, what you need to fruitfully, productively work in this new world of abundance are computational techniques that are far in excess of 99% accuracy—99.999%.”
Like many other disruptive technologies in pharmaceutical research—technologies such as CRISPR gene editing, proteolysis targeting chimera–induced protein degradation, and RNA interference—AI is generating a lot of excitement. “AI,” Heifets maintains, “promises to [help drug developers] go after previously intractable targets.” He points out that early interest in disruptive technologies such as AI tends to precede incredible clinical results. “It is,” he emphasizes, “the promise of opening up greenfield opportunities.”
Autonomous AI in real-world use
Machine learning also offers exciting opportunities in the realm of clinical diagnostics. For example, AI technologies for medical applications are being developed by Eyenuk. The company’s first product to market, EyeArt®, uses AI to detect disease from retinal images. In a clinical trial that included 942 patients and spanned 15 medical centers around the United States, sensitivity for detection of diabetic retinopathy was over 95%.
Eyenuk, an artificial intelligence medical technology and services company, has developed the EyeArt platform, which uses artificial intelligence to detect diabetic retinopathy. EyeArt can autonomously analyze a patient’s retinal images to robustly detect signs of disease and return an easy-to-read report in under 60 seconds.
EyeArt was developed using machine learning to train its algorithms on nearly 2 million images. “Now imagine training your resident with so many images,” says Eyenuk CEO Kaushal Solanki, PhD. “That’s just not possible.”
Experts recommend annual screening for diabetic retinopathy for everyone with diabetes. Currently, the United Kingdom is the only country in the world screening more than 80% of its diabetes population. That’s nearly 2.5 million patients whose retinal images would need to be individually reviewed by experts every year.
The U.K. National Health Service (NHS) carried out a health technology assessment in which EyeArt was compared with several competing technologies. EyeArt was found to be superior by a broad margin. The assessment’s findings, which were published in 2016, showed that EyeArt has 99.6% sensitivity for detecting proliferative disease, and 93.8 % for referable disease. The NHS is now shifting its workflow to adopt EyeArt for its screening programs. Meanwhile, pilot assessments at six centers in the United Kingdom have been completed.
Channeling the data deluge
Traditional methods of data analysis in drug discovery work best with straightforward, homogenous data. However, those methods fall short when the data becomes complex, for example, when patient records chronicle multiple diagnoses, comorbidities, complex treatment plans, and many encounters with clinics and clinicians. AI can integrate that information, analyze it, and produce stratified patient groups. That ability to handle complex, multivariate data is revolutionizing the design and execution of clinical trials.
Sensyne Health is at the forefront of this clinical data movement. Sensyne’s chief of translational medicine, Rabia T. Khan, PhD, says that the traditional model of drug discovery, which burns through billions of dollars and still produces high failure rates, is unsustainable. She adds, however, that AI promises to reduce costs and failures. Sensyneis partnering with the NHS to capture patient data and enable the stratification of patients for clinical trials.
“Because this data is so noisy and sparse and heterogeneous, you really do need AI,” Khan argues. “If you use classical approaches, you can’t identify subpopulations in heart failure. But when you go on to use more complex machine learning–based methods, you’re able to identify subpopulations of heart failure and show that there are more than just the two well-known subgroups of heart failure.
“There are actually multiple subgroups. We’re currently looking into how drugs are effective in many of those different subgroups.”
She predicts that eventually the industry will move away from classical randomized controlled trials and toward virtual trials. Enabled by AI, virtual trials will do the heavy lifting, providing much of the information that used to require expensive human trials. In fact, this information will be available for a prospective drug before the the drug is ever tested in human beings.
“Instead of taking something from an abstract idea in a dish all the way through to clinical work,” Khan says, “we will start with real-world data, link that to patient samples, and use that for drug discovery, and then we will feed the same information back into the clinical trial.”
Another company focused on managing clinical trial data is Precision Medicine Group. The company’s Precision for Medicine business recently acquired QuartzBio, an AI platform that analyzes biological and clinical data streams to extract knowledge and insights to accelerate drug development.
In clinical trials, there have been many attempts to funnel data from diverse sources to clinical trial investigators, who would then be able to quickly and flexibly interpret the data holistically. These attempts seldom succeed, suggests Cliff Culver, senior vice president of Precision Medicine Group.
“All of that data is generated independently and lives in disconnected formats all over the place,” he says. “For a drug company, multiple people would work for multiple weeks or months to pull all of that together, particularly when the focus involves linking quantifiable ‘reportables’ back to source information—like images or sequencing data—to enable ongoing quality control.
“The result is that analysis is meaningfully delayed, typically until after a trial is complete. And there’s rarely bandwidth to undertake deep data integration across trials within an organization. We do it as a trial is unfolding, so the company has regular insight into what’s happening, and then we do it at an enterprise level to maximize the utility of the data.”
Some of the approaches enabled by the QuartzBio platform bear a superficial resemblance to the unguided AI analyses used by Google, Netflix, or other large technology companies. Still, finding biological insights is different from choosing a movie you might like.
“In drug development, we typically have smaller and less dynamic data sets, and we have a material need to get beyond correlation to understand what’s happening biologically,” says Culver. “Our secret sauce is our ability to both pull all of that data together so you have the richest possible data set to explore and then apply the right data-driven or computational biology analysis, the AI, to derive actionable meaning.”
Concerto HealthAI is a precision medicine company with a major focus in oncology. It uses AI and machine learning to reveal how patients respond to treatments in real-world settings. The company’s work can guide pharmaceutical research, inform outcomes research and value-based studies, and accelerate drug development.
Concerto HealthAI CEO Jeff Elton, PhD, offers as an example of a typical cancer patient being diagnosed for the first time. That patient would initially be staged based on tumor location and size, metastatic status, and other features of the disease. Then this information would be used to guide treatment. Then, additional information—“on the fly” information about the patient’s progress or changes in disease status of the disease—may become available. It will not, however, automatically become part of the patient’s electronic medical record.
This sort of on-the-fly information is critical for pharmaceutical companies. For example, accurate information on staging at different timepoints in the patient history helps pharmaceutical companies understand why a treatment is or isn’t effective.
“We develop AI models that read the record, compute a stage, and give an accuracy score based on everything that’s in the record—imaging reports, molecular reports, and things of that nature,” says Elton. “We even arrange for analyses to accept data only if it is of a given accuracy.”
Concerto HealthAI’s technology can also be used to make predictions, such as whether a patient is going to be a responder or whether the response is durabile. These predictions can be valuable in the design of a clinical study. And they can offer clinical researchers a direct line of sight into the standard of care, which is important because often physicians will not place patients in a trial if it is too much of a burden compared to the standard of care. Concerto HealthAI’s models can predict patient burden for trial design purposes.
What all these capabilities allow researchers to do, Elton says, is run on-the-fly analyses, accomplishing in real time what would otherwise have required weeks of data preparation.