When cancer immunotherapy succeeds, the effects can be transformative. For some patients, it even extends life expectancy. For many patients, however, it ends up being an expensive disappointment. The reasons for this disparity remain poorly understood, but a growing number of researchers and startups are betting that the analytical capabilities of artificial intelligence (AI) might be just what the field needs. AI, they hope, will release immunotherapy’s latent power.
Tumors often protect themselves by figuratively dumping cold water on the immune response. For example, some cancers stimulate checkpoint proteins, signaling molecules that stud immune cells and induce immunosuppression, sparing tumor tissue from attack. This immunosuppression can be lifted with checkpoint inhibitor drugs, which unleash dormant immune cells, instigating attacks that can beat back even advanced disease.
One recent study of a combination of checkpoint inhibitors, nivolumab and ipilimumab, showed that 22% of patients with metastatic melanoma had no detectable cancer at five years post-treatment. Fifty-two percent of these patients were still alive at five years, in contrast to a historic survival rate of 22.5%. But there are currently no good means for doctors to reliably predict who will benefit or for how long, and the response rate for checkpoint inhibitors is generally even lower in cancers other than melanoma.
Checkpoint inhibitors could be far more effective if a technology were available that could predict their effects. In fact, such a game-changing technology exists in the form of AI platforms that incorporate machine learning techniques and statistical algorithms. Using such platforms, scientists and clinicians can identify potentially meaningful patterns in extremely complex data—patterns that would be all but impossible for a human to discriminate.
Making the most of cancer DNA
AI is helping researchers seek out biological signatures that distinguish treatment-responsive or -unresponsive tumors. For example, at the Institute of Cancer Research in London, a research team led by Anguraj Sadanandam, PhD, has been using machine learning to classify various cancers based on gene expression profiles collected from biopsies. These profiles can identify active biological pathways and cell types present within the tumor.
For example, in a study from last summer, his team used this approach to bin breast cancers into five distinct subtypes, two of which exhibited characteristics that suggest they may respond well to immunotherapy. “Quite a few companies are excited to collaborate with us,” asserts Sadanandam. “We can see if patients might respond to a checkpoint inhibitor or another therapy.”
Sadanandam’s team is hardly alone in working to realize the promise of AI in cancer immunotherapy. There is, in fact, a cancer research community eagerly developing AI applications ranging from patient selection to clinical trial execution to the discovery of new immunotherapies.
A notable member of this community is Cambridge Cancer Genomics (CCG), a University of Cambridge spinoff focused on computational tools for identifying genomic features that offer clinically actionable insights to oncologists. The company’s flagship platform, OncOS, is designed to work with data obtained from either conventional biopsies or less invasive liquid biopsy assays that characterize tumor DNA circulating in the bloodstream.
John W. Cassidy, PhD, co-founder and CEO of CCG notes that although biologists have been tinkering with machine learning for some time, clinical researchers have started exploiting its power only recently. “That field has exploded,” says Cassidy, “and now we’ve got super-fast graphics processing units, cloud computing infrastructure, and lots of mathematicians and statisticians that want to work on interesting problems.
“We have super-fast, super-scalable bioinformatic processes that can take the raw sequencing data all the way through to very high confidence machine learning–based cancer variants. These data may reveal specific mutations that can guide clinicians to treatments that specifically target abnormal genes. In addition, these data may identify more nebulous characteristics that might inform success or failure of immunotherapy.
Studies have shown that tumors with widespread genomic mutations—known as a high tumor mutational burden—are more likely to produce clearly abnormal proteins that set off immunological alarms. Cassidy notes that CCG has had great success in training OncOS to detect such profiles, alongside genomic features that reveal the heterogeneity of the tumor and other biologically important characteristics.
“We’ve looped a lot of that tumor mutational burden work into creating complex biomarkers for tumor response in lung and colorectal cancer immunotherapies,” he asserts. “It’s a genome-driven picture of whether the tumor is immunologically active or silent.”
Considering other biomarkers
DNA sequences and gene expression give only part of the patient stratification picture. Filling out the picture—and building more sophisticated tumor classifications—will require multiple sources of biological data. Sadanandam’s team is currently working with clinicians’ reports to better understand how treatment history and outcomes correlate with molecular tumor features, as well as imaging data from pathology slides and radiological scans.
Such visual data have guided oncologists for decades, and they can give a surprisingly rich view of a tumor’s growth and invasive behavior as well as its interactions with the immune system. For example, a new approach that relies on CT scans has been developed by Hugo Aerts, PhD, director of the Computational Imaging and Bioinformatics Laboratory at Dana-Farber Cancer Institute/Harvard Institutes of Medicine.
Aerts and colleagues have developed a CT-guided AI strategy for classifying lung and skin cancers. The researchers have used the strategy to identify a patient cohort that achieved 25% greater one-year survival from immunotherapy relative to an unselected group.
Informing clinical trials
AI-assisted tumor profiles are also valuable in the clinical trial process. For example, the blood-based liquid biopsies used by CCG can be repeatedly collected with minimal inconvenience to patients, which means that clinical researchers can monitor the dynamic response to treatment. This is particularly useful for immunotherapy, notes Cassidy, where the signatures of response are not as clear cut as with other treatments.
“When you look at scans of the tumor, it actually expands if the drug is working and shrinks afterward,” he says. “So, it’s very difficult to track whether treatment is working with traditional metrics.”
Molecular profiles can be much more informative in terms of tracking the ongoing impact of such therapies, and CCG is now involved in clinical trials in Singapore and California that use multi-timepoint sampling to profile patients who benefit or fail to respond to immunotherapy.
The potential that AI holds to improve clinical trials is also being explored by Mateon Therapeutics. The company is using machine learning and other AI approaches to evaluate its pipeline of immunotherapies for late-stage cancers such as gliomas, pancreatic cancer, and melanoma.
One such therapy is the company’s lead immunotherapy product, OT-101. This is an antisense RNA drug that alleviates tumor-mediated immune suppression by inhibiting a signaling protein called tumor growth factor-beta (TGF-β).
Mateon recently acquired PointR Data, a company specializing in AI- and blockchain-based technologies. According to Mateon CEO Vuong Trieu, PhD, the merger positions the company to leverage these tools to conduct trials that collect richer information about study participants and enable study designers to respond more quickly to signals of efficacy or failure to respond.
“In clinical trials, we are typically forced to wait until the end of the trial and after data lock to look at the data and make decisions,” says Trieu. “We need to be able to access the data in real time without being penalized on the statistical significance—this will allow us to run adaptive trials that are smaller and faster.”
To achieve this, Mateon will be using AI to actively collect and analyze data from clinicians throughout the trial, with blockchain used to ensure the integrity of the data being harvested. Trieu says that his company will test this approach in upcoming trials of OT-101 in the United States and China.
Expanding the immunotherapeutic arsenal
Trieu believes that some promising drug candidates might be hiding in plain sight, with intriguing leads from the literature buried under a nonstop avalanche of new articles. “Approximately 2.5 million new scientific papers are published each year,” says Trieu. “We can use deep neural networks to understand the abstracts of these millions of papers.”
When these computational insights are coupled with with other data from public databases and Mateon’s proprietary resources, it should be possible, Trieu believes, to expose promising therapeutic candidates that were overlooked after their initial publication, or even to rehabilitate failed drugs that might prove effective when translated to other indications.
The immunotherapies now being used in the clinic are mostly a one-size-fits-all proposition, with checkpoint inhibitors that target a handful of immune signaling pathways. But researchers also see a lot of potential in more personalized strategies that use custom-made vaccines to elicit an anticancer immune response based on specific abnormal proteins found on a given patient’s tumor cells.
Last April, Evaxion Biotech began dosing patients in its first clinical trial of EVX-01, a customized vaccine formulated based on insights from the company’s Pioneer platform, which uses machine learning to comb through DNA and RNA data to find clear signatures of neoantigens that arise exclusively from tumor-associated mutations. NEC OncoImmunity is employing machine learning to a similar end with its Immune Profiler software, and it has announced partnerships with companies focused on the development of personalized cancer vaccines and cell therapies in the United States, Europe, and China.
Although more accessible than in the past, the discovery-phase work associated with AI analytics can be expensive and technically demanding. “This is highly computationally intensive,” explains Sadanandam. “We are trying to use statistical methods that can shrink the data to lower dimensions and remove the noise, so that the amount of computing power required is also reduced.”
Cassidy also notes that there are many challenges associated with the sharing and analysis of patient data, particularly across international borders, and that the field is still working to establish best practices. “There’s a lot of work that could go into standardizing the analytics,” he points out. “We all have different filters and ground truth rules and so on.”
But it is also clear that such heavy computational artillery will be essential for deconvoluting how a patient’s immune system is interacting with their cancer and identifying the best tactics to stir it to action. “As an industry, we have to move to individualized therapy that leverages all of the omics data available from patients,” insists Trieu. This point is reinforced by Cassidy, who envisions a “hyper-personalized” medicine approach, where not only the drugs themselves, but also the dose, timing, and sequence in which they are administered must be carefully managed to achieve the best benefit.
“When you’re talking about side-effect profiles, different drugs stacked together, and other things like that,” Cassidy declares, “you’re going to have to use some sort of machine learning–based technology to understand how to do that in a clinical setting.”