NEW YORK – The use of artificial intelligence in precision medicine research and development surged in 2024 as researchers in industry and academia explored ways to leverage the technology to improve cancer diagnoses, predict outcomes, and personalize therapies.
Once a novelty, AI tools have become commonplace in healthcare software used to analyze patient data in diagnostic testing and radiological imaging. In 2024, researchers continued to advance machine learning-based systems that assist doctors in interpreting those images and test results, attracting interest from investors and industry partners.
For example, Zephyr raised $111 million in a Series A round in 2024 to develop AI tools to stratify patients with cancer and cardiometabolic diseases and predict their responses to treatment. VieCure raised $45 million to advance its AI-enabled clinical decision support platform. And Tempus, which bills itself as an "AI-enabled precision medicine company" and uses AI to analyze data, provide clinical decision support, match patients to clinical trials, and develop better tests and drugs, raised $410.7 million through an initial public offering.
Biopharma companies are also increasingly integrating AI tools to run clinical trials more efficiently and make better decisions about which drugs to advance in their pipelines. Pharmaceutical giants like AbbVie, Roche, and AstraZeneca all invested in AI in 2024.
In January, AbbVie inked a deal with ConcertAI and Caris Life Sciences to advance its precision oncology pipeline and optimize clinical trials using Caris' real-world, multimodal clinical and genomic database and ConcertAI's research-grade clinical data. Roche teamed up with PathAI in February to develop AI solutions that can interpret results from digital pathology-based companion diagnostic tests, and in September, Roche agreed to distribute Stratipath's digital pathology solution for breast cancer risk stratification. In October, AstraZeneca struck a deal with Owkin to use Owkin's data network and AI tools to optimize prescreening for germline BRCA1/2 mutations directly from hematoxylin and eosin (H&E)-stained pathology slides. The UK-based pharma giant also partnered with Seoul, South Korea-based firm Lunit in November to develop AI-driven biomarker prediction tools.
Cancer screening and detection
Some of the most advanced AI technologies for precision medicine focus on cancer screening and detection. This year, Qure.ai received US Food and Drug Administration clearance for its AI-driven chest CT tool, which helps radiologists and pulmonologists locate incidental lung nodules in routine chest X-rays. Qure.ai Chief Business Officer Bhargava Reddy said this approval was the first of its kind and was followed by an additional FDA approval of tools that estimate nodule volume and predict nodule volume doubling time.
The Mumbai, India-based company is already marketing the technology, dubbed qXR, globally in low- and-middle income countries, where many patients don't have access to CT screening for cancer or other lung conditions like tuberculosis. However, the firm has now set its sights on the US market, where it sees an opportunity to use its AI technology to screen patients for cancer when they have chest X-rays for other complaints.
Reddy pointed out that in the US, only smokers and others at elevated risk of lung cancer receive routine lung cancer screening via CT scan, even though nonsmokers account for 10 percent to 20 percent of lung cancer diagnoses. Qure.ai's message has resonated with US hospitals, according to Reddy. "The number of hospital consultations, which we are able to have, has led us to believe that at least 10 to 15 hospital systems in the US will be adopting [qXR] in 2025," he predicted.
Imidex is another company aiming to enter the lung cancer screening space with an AI-driven tool for early lung cancer detection from chest X-rays, called VisiRad XR. The FDA cleared the tool in 2023, and Imidex inked deals with Spesana and Orbit Genomics this year around development of the technology.
Earlier cancer detection through the use of AI has applications beyond lung cancer, as well. Helio Genomics aims to expand adoption of its AI-driven multi-analyte liver cancer test, HelioLiver, and is studying potential development of additional tests for colon, stomach, and lung cancer. The firm submitted a premarket approval application to the FDA in the second quarter of 2024, seeking market authorization of HelioLiver as a Class III medical device.
AI-driven biomarkers
Many AI applications in precision medicine leverage its potential to predict treatment responses, outcomes, and disease recurrence, or even to guide treatment choices for patients. In 2024, researchers made strides toward development of biomarker-based tests that can be used alongside conventional genomic tests to answer questions about patient care. These up-and-coming tests use a variety of inputs, including H&E slide images, genomic data, routinely collected clinical data, and liquid biopsy information.
In an example of how AI can pair with liquid biopsy technologies to monitor a patient's response to therapy, early this year, Novigenix introduced an AI-driven liquid biopsy platform, LITOSeek, as a tool to track the immune system's response to cancer immunotherapy. The technology analyzes patients' immune responses using RNA analytics, a task Novigenix asserts would not be possible without AI. The firm aims to deploy the technology initially in clinical trials, and eventually, to develop it as a real-world test to inform treatment for cancer patients.
Digital pathology approaches have benefited from the availability of large sets of consistently formatted data, such as whole slides or radiology images, and in 2024, R&D activity around these technologies intensified. Algorithms that ingest these images can extract features and classify images for tasks, such as prescreening patients for genomic mutations, tumor subtyping, and risk stratification.
For example, this year, as part of its mission to develop a portfolio of AI-based companion diagnostics, BioAI published an analysis showing that an AI model could detect RET alterations from H&E slides from lung cancer patients with 100 percent sensitivity and 63.3 percent specificity, meeting the performance requirements for a reliable method of prescreening patients prior to genomic testing and potentially fast-tracking them onto targeted therapy.
And in November, Ataraxis AI introduced an AI-driven digital pathology test in November that it believes may offer faster, more accurate, and more cost-effective results than conventional genomic testing. Ataraxis Breast was developed on the company's pan-cancer foundation model, Kestrel. Unlike many other AI models trained on digital pathology images to make histology-specific predictions, Ataraxis trained Kestrel on images across solid tumor types based on the theory that it would learn not just about one type of cancer but also more broadly about tumor biology and histopathology.
In a validation study, Ataraxis compared the test's risk score predictions for breast cancer patients to the Exact Sciences' Oncotype DX Breast Recurrence Score and found that Ataraxis Breast had a 30 percent greater accuracy for predicting recurrence than Oncotype DX and was able to reclassify patients deemed to be at intermediate risk by Oncotype DX into low- or high-risk groups. However, like most other AI-based precision oncology tools, Ataraxis Breast has not yet been tested in large, randomized, prospective clinical trials. Until that happens, it is unclear whether the results seen in validation studies will translate to patient care.
Looking ahead to 2025
Ataraxis' approach incorporates a couple of features that represent emerging trends that will become increasingly significant in 2025. The larger training dataset put it in the category of a foundation model, which is a model trained on vast, widely variable datasets that can be applied to a range of use cases. Ataraxis also uses a self-supervised algorithm called DINOv2, borrowed from Meta, that allows it to handle millions of unlabeled image segments. This circumvents a common limitation of typical digital pathology models, since labeling by pathologists is time-consuming and not feasible on such a large scale.
"It's still early days" for AI models that predict patient outcomes and treatment response, said Carlo Bifulco, chief medical officer of Providence Genomics, the genomics research arm of Oregon's Providence health system. Yet, he sees "room for growth" in terms of the size of training datasets and the number of parameters the models can manipulate. "What you've seen so far in this space is a beginning," Bifulco said.
Looking ahead to 2025, Bifulco, who contributed to the development of Ataraxis Breast and another digital pathology foundation model based on DINOv2 called GigaPath, expects to see the larger foundation models make breakthroughs in basic biology. He predicts there will be more multimodal models incorporating different types of data, particularly multimodal digital pathology models that combine radiology, whole slides, and other image types into a single model.
Tinglong Dai, a professor at Johns Hopkins University's Carey Business School and co-chair of the Johns Hopkins Workgroup on AI and Healthcare, expects to see more divergence between models for diagnosis, prognosis, and treatment navigation in 2025, as development of those models are sharpened for their different purposes.
In particular, when choosing interventions for a patient, Dai said a different type of algorithm called causal inference, which uses AI to determine cause-and-effect relationships instead of correlations, will be necessary. "Causal inference is a very powerful tool, but [currently] it's just too slow," Dai said. "Imagine if we could do … millions of causal inference tasks in one day. That will be a game changer. It's hard to imagine that doesn't happen in 2025."