functionbar_help
Font Size
functionbar_contact
Aboutus
Search Results for ""
Share
Top Stories

Published On December 23, 2020

TECHNOLOGY

The Drugs Machines Create

The idea of having computers design new therapies has slowly been gaining ground. In the COVID-19 crisis, it may have found its moment.

To pharmacologists, the world is a constant interplay of tiny, complex shapes. The modern task of making a successful drug is, at its most fundamental level, a hunt for one specific shape that can travel through the body and lock on to another specific shape—one molecular end link of a chain of microscopic events that end in illness. It is finding a piece in a biological jigsaw puzzle.

Solving this puzzle plays out through many iterations in the search for new disease treatments, including the task of developing effective therapies for COVID-19. The first step is to identify the right target. In the case of the SARS-CoV-2 virus, for instance, researchers identified 29 proteins, including the “spike” proteins that extend outward from the viral envelope, giving the virus its crown shape. They can identify which protein might make the best target but latching on is easier said than done. Proteins can be exquisitely complicated, and researchers must map out exactly what those target proteins look like.

The next challenge, then is to find the right molecule to bind to it. That search might start with sorting through more than 100 million known “small molecules”—drugs whose active ingredient contains fewer than 100 atoms—that have been identified during the past century. Researchers might also look at synthesizing a new one. Ten years ago, chemists estimated the total number of possible small molecules to be 1 novemdecillion, which is a million billion billion billion billion billion billion—or 1,000 times the number of atoms in the solar system.

If locating those matches of drugs and proteins seems daunting, the process could soon get much simpler. Over the past decade or so, researchers have developed increasingly sophisticated computational models to test interactions between small molecules and intended disease targets. These in silico experiments are cheaper and faster than laboratory research and show the early, tantalizing promise of dramatically accelerating drug discovery.

Adding artificial intelligence to this process—something that has only happened in a significant way over the past two years—could help further. Researchers are teaching algorithms to churn through existing drugs to find matches with target molecules. In cases where a match doesn’t exist, the programs could theoretically be able to design a small molecule from scratch that might succeed where existing drugs fail. Future AI models may predict not only which compounds will effectively bind to pathogens, but also how they’ll affect other systems in the body. 

AI suggests a tantalizing, low-cost solution to the twin problems of time and money. “At Exscientia, we are using AI to precision engineer new drugs for the clinic to treat targeted diseases,” says Andrew Hopkins, founder and CEO of Exscientia, in Oxford, United Kingdom. “AI design can shorten the optimisation path, thus simultaneously cutting discovery time and generating higher quality compounds not typically achieved by conventional methods.” In 2020 Exscientia reported that it was the first company to use AI to design a drug called DSP-1181 for treating obsessive compulsive disorder to enter into human clinical trials. And at least one COVID-19 therapy, identified using AI tools, has entered the development pipeline.

“AI is incredible in the way that it can find connections that no human could predict, doing things the brain isn’t capable of,” says oncologist Justin Stebbing at Imperial College London, who is taking part in an international collaboration that was among the first to use AI to identify a potential COVID-19 treatment.

COMPUTERS ENTERED THE WORLD of drug discovery in the 1980s as a faster and more efficient way to determine how well certain compounds might bind with therapeutic targets, and early forms of computer-aided drug design, or CADD, have played a major role in developing drug treatments. Notable early successes include the antihypertensive drug captopril, which debuted in 1981, and three therapeutics used to treat human immunodeficiency virus, all approved in the mid-’90s.

Twenty-first century advances in genetics pushed the field forward. The Human Genome Project, tasked with creating a complete map of human genes, eventually allowed researchers to identify new mutations that might be implicated in disease, as well as the proteins encoded by those gene variants. An improved ability to identify mutations in cancer has also led to an explosion in targeted therapies. Some of these can extend patients’ lives, and there have been a few game changers such as Gleevec, which has largely transformed chronic myeloid leukemia into a chronic, treatable disease.

Sorting through such vast stores of data to find matches for these mutant proteins turned out to be a task well-suited to computers, and AI tools may represent one of the best ways to find new treatments. “AI can screen billions of compounds in just one day while human experts, even using high-throughput screening tools, can do only a million compounds per day,” says MIT computer scientist Wengong Jin, one of a group of researchers using AI to not only design next-generation antibiotics, but look for vulnerable treatment targets in COVID-19 pathways.

Until COVID-19 vaccination is worldwide—a task that might lie years or decades in the future—finding ways to treat the infected will remain a critical task. Earlier this year, using an AI platform developed by the company BenevolentAI, Stebbing and his colleagues found that baricitinib, an FDA-approved rheumatoid arthritis drug, might be able to help coronavirus patients. The researchers knew from early investigations that SARS-CoV-2 attaches to cells in the lungs and other organs using a protein called ACE2. So they were looking for an existing drug that might interfere with that process.

To find one, they began by creating a “knowledge graph”—a sprawling database compiled using machine learning to search scientific literature about viruses, diseases, genes and pathogens. Next they used what was known about how the coronavirus infects human cells—it works its way in through a process called endocytosis, which utilizes enzymes within the cell—to train their algorithms.

Searching the knowledge graph, the algorithms zeroed in on proteins called AAK1 and GAK that play a key role in connecting the virus to ACE2 receptors. The AI immediately produced a list of 47 known AAK1 and GAK inhibitors, six of which have proven in past experiments to be especially effective. But five of the six, used to treat cancer, have severe side effects. The sixth, baricitinib, has been safer, and the AI predicted it could block viral infection of lung cells. Because the algorithm had access to dosing and safety information—based on baricitinib’s use in arthritis treatment—it was able to predict that a dose of 2 or 4 mg of the drug, given once daily, would effectively block AAK1 and GAK.

Each of these steps—compiling a big database, searching for information about pathways, making inferences about what data from past studies says about COVID-19—could have been done by a person, but the complexity of that entire task meant it likely would have taken years—compared with several hours using AI—to arrive at confident predictions about baricitinib, says Stebbing.

Once baricitinib’s promise had been identified, the drug was tested on liver organoids—biological models engineered to model organ processes—then in a small human clinical trial. In May, the U.S. National Institute of Allergy and Infectious Diseases joined with Eli Lilly to launch a phase 3 trial testing baricitinib in combination with remdesivir, an antiviral compound originally developed to treat Ebola. Results published in The New England Journal of Medicine on December 11 suggest they reduced time to recovery for those hospitalized with COVID-19. (In November, the FDA issued an Emergency Use Authorization for the distribution and emergency use of baricitinib to treat hospitalized patients with COVID-19).

Other possible COVID treatments came earlier this year from Insilico Medicine in Hong Kong. Previously, in August 2019, the company had announced that its researchers used AI to produce, in less than two months, a compound that targets a protein connected to fibrosis and other diseases. Using the same methods, the company identified the potential COVID-19 therapy—but in this case, AI was used to design a molecule rather than to find one that was already known. That becomes complicated, because no existing information about dosing or safety yet exists. Insilico used tools called generative adversarial networks (GANs), which are designed to synthesize predictions based on training data. When their GANs looked at data on coronavirus proteins, it came up with a new molecule whose structure should, in theory, inhibit them. Since then, the company has reported developing other newly synthesized compounds, including some that target a COVID-19 enzyme.

THE REACH OF AI technology may also help with another drug-development crisis—finding new antibiotics for treatment-resistant pathogens that already kill tens of thousands of people and infect millions more in the United States. As these superbugs become more prevalent, existing antibiotics become less effective, and researchers are scrambling to find ways to treat infections that don’t respond to available therapies. 

In February 2020, researchers at MIT reported using deep learning to identify a new and effective antibiotic. The researchers first taught their algorithm to recognize what an antibiotic looks like by training it on 2,500 molecules that (at least in lab tests) can kill stubborn, pathogenic strains of E. coli. Then they used the software to evaluate 6,000 other known compounds. It chose one that appeared likely to work as an antibiotic, didn’t look like any other antibiotics available and had a low risk of toxicity.

Tests in the lab showed that the new compound would work against resistant strains of E. coli, C. diff and other pathogens. Then, in mice, the researchers tested it against a strain of Acinetobacter baumannii that’s resistant to all known antibiotics. The experimental drug, which the researchers named halicin after Hal, the AI system in the movie “2001: A Space Odyssey,” cleared the infection is less than a day.

Bolstered by those results, the researchers then set their algorithm loose on a much larger database containing billions of compounds. It identified six more potential antibiotics that, like halicin, were structurally unlike any existing ones. “Once we trained this model, it learned to recognize what compounds will likely kill bacteria,” says MIT’s Wengong Jin, who says the group hopes to test those compounds soon. The group is also investigating ways to use the same approach to test drug combinations for treating COVID-19.

Still, as promising as AI drug development may be, it’s hardly a mainstream approach. A survey of 330 scientists involved in drug discovery, conducted in December 2017, found that 40% weren’t familiar with AI approaches—and those who did know of it tended to overestimate what it could do. Moreover, some early successes—in silico’s newly designed drug, as well as DSP-1181, the new drug for obsessive compulsive disorder—target the same proteins as existing FDA-approved drugs, and it’s not yet known how the new drugs compare to the older treatments.

There’s also the matter of getting support from pharmaceutical companies that could underwrite clinical trials and scale up production of AI-discovered drugs. So far, except in the case of baricitinib, drug companies have been slow to embrace this approach. Yet most experts say these obstacles are surmountable, and typical of emerging technologies. The case of COVID-19 shows that a speedup is possible and has the potential to be beneficial. Other ongoing research suggests that an AI approch could be a formidable tool against other diseases, too. Justin Stebbing’s group at Imperial is now partnering with BenevolentAI to identify new treatments for glioblastomas, and Insilico has been applying its approach to triple-negative breast cancer.

In all of this, the primary potential benefit is speed. “Normally, bench to bedside takes years,” says Stebbing. “We can do better. This is computer to laboratory to bedside to an FDA approval at breakneck speed.”