In medical research and in healthcare, time is of the essence. It’s always best to diagnose a disease early, treat it before it progresses and cut years off drug development to speed life-saving therapies to patients. Yet even today, these processes can be slow and error-ridden. Artificial intelligence can help. The stories below highlight startup companies that use AI to spot patterns that humans would miss. In the process, they aim to discover novel drug candidates, make diagnosis more accurate, improve telehealth and deliver personalized care.
AI FOR WHAT AILS YOU
The news in October 2024 that Sir Demis Hassabis would receive a Nobel Prize in chemistry for his work on predicting protein structures put the recent rise of artificial intelligence for drug discovery in the spotlight.
Powerful new AI models can comb through datasets to identify promising molecules and predict their interactions with biological targets. Other models can predict how safe and effective an experimental drug is likely to be. Together, these models are showing great potential in accelerating the process of drug development.
Efforts to apply AI to drug discovery have had to overcome a particular challenge: a paucity of data upon which to train models. There is usually only a limited amount of data on a potential drug and how it might affect a disease. “Many drug targets are novel, and they have very little or no known chemical matter that modulates them,” says Evan Feinberg, founder and CEO of Genesis Therapeutics.
To cope with the paucity of data, scientists are having to dig deep into their programming toolkit. A machine-learning model for drug discovery might start with data based on what’s known about the molecular structure of a disease-related receptor and then generate theoretical molecules until some are found that would bind to that receptor. To develop such AI-based methods to search for new drugs, though, companies must build custom algorithms and models. Genesis, for instance, approaches drug discovery with a combination of proprietary AI-based models and laboratory studies of molecules.
Hassabis, who is CEO at Google DeepMind and Isomorphic Labs, both owned by Google’s parent company, Alphabet, was instrumental in developing AlphaFold, one of the highest-profile AI platforms. The latest version, AlphaFold 3, predicts the 3D structure of proteins and how they interact with other biomolecules, such as receptors and enzymes, in the progress of disease. Scientists at Isomorphic are collaborating with teams at Eli Lilly and Novartis to combine AlphaFold 3 with other custom AI-based tools to discover antibodies and other new treatments that inhibit disease-related targets.
What is reportedly the first AI-designed drug, to treat a rare lung disease, started phase 2 clinical trials in June 2023. Biotech firm Insilico Medicine developed it using its PHARMA.AI software suite. This AI-driven approach analyzes disease targets based on many features, including how likely they are to be safely and effectively inhibited by potential drugs. “This holistic evaluation is crucial for the identification of viable therapeutic targets,” says Thomas Leichner, the firm’s head of strategy.
Genesis’s AI software for drug discovery—its GEMS platform (for Genesis Exploration of Molecular Space)—uses large language models to create billions of druglike molecules. Then, it uses machine-learning algorithms to predict a protein’s potency and selectivity for a specific disease target. In 2024 pharmaceutical company Gilead Sciences began using GEMS in collaboration with Genesis.
Although most academic scientists don’t have access to the new tools, more companies offer cloud-based services. For instance, Ginkgo Bioworks, which develops methods of programming cells to make drugs and other products, now offers its AI models via Google Cloud.
As the Ginkgo-Google collaboration suggests, AI models are only a tool in the process of drug discovery. “We will find the best new drugs by pairing the smartest scientists with the most powerful AI platform,” says Feinberg. “And I think that synergy will continue for some time.” —Mike May
CATCHING CANCER
When trying figure out if a patient has cancer, pathologists still do what they’ve done since the early 20th century: peer into microscopes at tissue samples from a biopsy or surgical procedure and try to identify the presence of cancer cells. “The process hasn’t changed much in 100 years,” says Andy Beck, a pathologist and CEO of PathAI.
Pathology is a field that seems ripe for assistance from recent advances in artificial intelligence. Beck’s company is one of several that are developing AI models to make the process of diagnosis more efficient and accurate.
PathAI trained its AI models on digitized images of more than five million pathology slides containing 15 million annotations. It also supplemented those images with data on genomes and molecular biomarkers that pathologists don’t typically consider, with the goal of generating insight into how particular patients might respond to different treatment options. The models were also trained to identify the tumor’s microenvironment, including changes in the blood vessels and noncancerous cells surrounding the tumor, because these changes can provide additional insights into the tumor’s aggressiveness and its vulnerability to different treatments.
The new AI tools aren’t designed to replace pathologists, but to guide doctors toward faster, more accurate and detailed diagnoses by highlighting salient elements in the sample, counting cell types and quantifying various other abnormal features. The results, says Beck, can in some cases disambiguate tricky diagnostic calls. Getting an early, definitive diagnosis of some diseases can be tricky, even for specialists. New AI tools are lending a hand, helping doctors more quickly arrive at clearer, more detailed diagnoses of cancer, inflammatory disease and other illnesses. “There can be disagreements among pathologists over what the same slide is showing,” he says. “The algorithms can provide much more granularity and reproducibility.”
By improving diagnostics, AI tools will hopefully make it possible for doctors to catch diseases in their earliest stages, when they’re more likely to respond to interventions. Catching cancer at the early stages would provide a boost to average healthspan because risks go up dramatically with age—90 percent of cancer cases are diagnosed in people older than 50, according to data from the American Cancer Society.
It’s not just the difficulty of tissue diagnosis that opens the door to improvements from AI. It’s also a shortage of pathologists—the 15,000 working in the U.S. are far fewer than needed to ensure fast, thorough turnaround of all the tissue biopsies being produced. At the same time, demand is expanding as the population’s average age increases, people live longer, and more and better cancer screening turns up more cases.
Exacerbating the problem is the fact that a full cancer diagnosis now calls for greater scrutiny of samples, as slight differences in tumor cells of a given type of cancer can impact the choice of treatment. “As more therapeutic options become available, pathologists need to be as predictive as possible to give patients personalized treatments,” says Beck. —David H. Freedman
AI is helping doctors deliver accurate, actionable diagnoses.
Phil Wheeler
DR. BOT WILL SEE YOU NOW
At age 22, Susan Conover developed a mole that looked worrying and decided to have it checked out. Her primary-care doctor told her he’d typically send her to a dermatologist, in a process that would have delayed its removal and biopsy by three months, she recalls. “But we don’t have time to waste,” he told her.
Instead, her primary-care doctor did a biopsy. The results: melanoma. “If I had waited to go to a dermatologist,” she says, “I might not be here.”
Conover’s story had a happy ending—the cancer hadn’t spread. But not everyone is so lucky. Skin disease affects one in four people in the U.S., and trained specialists are in short supply. A 2018 survey found that in Boston, which boasts state-of-the-art hospitals, patients waited an average of 52 days for an initial consultation and four to nine months to see a specialist.
Her terrifying experience galvanized Conover to do something about the situation. In 2017 she founded a company, Piction Health, to build a mobile app to help primary-care physicians recognize melanoma from images.
Then came the pandemic and the rise of telehealth. More than 96 percent of dermatologists offered their services via telehealth during the pandemic. At the same time, artificial-intelligence models were getting better at diagnosing moles and rashes from images. In a 2017 study, researchers showed that AI models could identify skin cancer from images as well as board-certified dermatologists. In a 2020 study, led by a team at Google Health, AI performed on par with specialists and better than general practitioners and nurses in recognizing conditions commonly seen in primary care.
Piction changed course in part to capitalize on these trends. It developed an AI model and trained it on 750,000 images from more than 200 dermatologists across 20 countries—including Bolivia, South Africa, India, Tunisia and some from Europe—to ensure a representative mix of skin tones. In preliminary analyses of the same 26 common skin conditions used in the Google Health study, “our AI models generate the top-5 conditions list with the same accuracy level as a dermatologist,” says Pranav Kuber, a co-founder and chief technology officer. The company claims its AI model makes accurate identification and treatment faster and easier for doctors. Its early data, which has yet to be published, suggests that it can reduce their evaluation time from 15 minutes to three.
In December 2022 Piction opened its own online-first clinic. Patients submit photos to Piction dermatologists, who use the firm’s AI model to produce a short list of potential diagnoses and identify patients who, like the young Conover, should be fast-tracked for in-person examination in partner clinics, where specialists make the diagnoses and fashion a treatment plan. The goal, says Kuber, is that patients can be seen “within a two- to three-week period instead of waiting months.”
As of October 2024, 3,000 patients had used Piction’s clinic. So far, it is available in Connecticut, Florida, Massachusetts, New Hampshire and Washington. The service is covered by several major insurance companies, or patients can pay $119 out-of-pocket for each consultation.
Eleni Linos, a professor of dermatology and epidemiology who directs the Stanford Center for Digital Health, and who has no connection with Piction, says: “I’m really optimistic about how this technology can help patients get the best care they can get, while at the same time helping doctors.” —Esther Landhuis
Explore the emerging science of healthspan in other stories in this special report.
