Scientists at Harvard Medical School have created a new artificial intelligence system that can accurately determine which cancer patients will respond to the class of cancer treatments known as immune checkpoint inhibitors. The new system called COMPASS is explained in a recent study published in the journal Nature Medicine and represents a huge step forward towards a more personalized approach to treating cancer patients.
What is significant about the discovery is that the current generation of immune checkpoint inhibitors, although very promising, works on a limited number of patients. Determining beforehand which patients are responsive to such treatments will mean that a lot of patients do not have to suffer any side effects.
In order to grasp the importance of this AI software in more detail, one must first be aware of what the drugs in question are.
Also referred to as ICIs (immunotherapy checkpoint inhibitors), these drugs represent a form of cancer therapy that operates on the same principles but is fundamentally different from conventional chemotherapy. While the latter targets cancerous cells, ICIs operate based on activation of a patient's natural immune system, allowing them to target cancer cells independently of any external influence.
The use of immune checkpoint inhibitors revolutionized cancer treatment and was successfully used even in extreme cases, such as that of former US President Jimmy Carter, who survived nine years after his stage IV melanoma diagnosis thanks to the use of pembrolizumab, an ICI drug.
On the other hand, things do not always go as smoothly for cancer patients. In fact, depending on the type of cancer, clinical trials show that only 10 to 40 percent of patients respond to the use of ICIs. For the remaining majority of patients, taking such a drug poses a risk of severe complications due to its ineffectiveness.
Physicians have long sought to identify, in advance, those patients that are most likely to be candidates for immunotherapy. Existing predictive tools have been based on a variety of factors, including whether a high percentage of immune cells has infiltrated the tumor and genetic markers of the tumor itself.
The challenge has been that existing predictive approaches are simply unreliable. As mentioned, a tumor that exhibits high levels of immune cells, referred to as an "immune-inflamed" tumor, is predicted to respond, whereas so-called "immune desert" tumors, which have low levels of immune cells, are not predicted to respond. But in reality, there are many patients who exhibit unexpected responses to such drugs.
Marinka Zitnik, an associate professor of biomedical informatics at Harvard Medical School who was involved in the development of the research project, called this one of the key unsolved problems in cancer therapy. "Predicting response to ICIs is a major unsolved knowledge question," she said. "This is a crucial problem of oncology."
Unlike other previous methods, the new algorithm uses an entirely different technique. The model considers the activity of almost 16,000 genes that have been previously shown to be involved in the following processes:
- State and behavior of immune cells
- Tumor-microenvironment interactions – interaction of the tumor with its surrounding environment within the body
- Multiple cell signaling pathways that regulate communication among cells
The ability of the model to consider so many different genes at the same time allows for a better understanding of how the tumor and immune system of a particular patient will respond to the drug in question.
According to researchers, COMPASS belongs to the category of "foundation models." Foundation models is a concept used in the field of artificial intelligence for a system that is initially trained on broad data and then applied to many different situations.
In this case, that implies that COMPASS is not restricted by any specific kind of cancer or drugs.
An aspect of COMPASS worth highlighting here is that it is not your run-of-the-mill "black box" artificial intelligence system, in which some model produces a prediction without any indication of how it came up with said prediction. Rather, COMPASS provides an explanation for its rationale when it makes a prediction about a certain patient.
That point is crucial when it comes to the application of artificial intelligence in healthcare, since physicians must rely on the reasoning process of a tool in addition to the end result of its operation.
The researchers tested their COMPASS tool using data from 16 different clinical groups of patients or cohorts, which included a variety of cancers and immunotherapies. In each test, COMPASS was found to outperform the best of the currently available prediction methods by 8.5 percent, which is a significant achievement for the field of prediction, as improvements in accuracy have proven to be quite challenging.
Since the model is designed to be broadly applicable rather than specific to any particular case, scientists consider that it will eventually prove to be useful in a broad range of clinical contexts rather than one particular type of cancer only.
If the results of future trials continue to support these findings, then COMPASS may have some significance in several ways in the field of cancer therapy:
- Personalized treatment approaches. There will be an evidence-based decision about whether a certain patient can benefit from a certain immunotherapy agent without having to go through the trial-and-error process of therapy.
- Patient safety. In the case when the patient will not respond to the treatment, the unnecessary risks can be prevented as there will be no need to experience the side effects of the drug that was not going to treat them and another approach will be considered by doctors.
- Designing effective clinical trials. Pharmaceutical companies and other researchers could rely on the COMPASS tool in order to select the right type of patients to participate in immunotherapy trials and make drug trials easier and faster.
- Identification of new scientific research topics. Being able to explain the reason for its predictions, COMPASS will help to identify the new biological connections that were unknown before.
The COMPASS model from Harvard scientists belongs to an array of recent initiatives in bringing artificial intelligence into more parts of the process of diagnosing cancers and planning their treatment. In addition to the COMPASS, Harvard scientists created yet another AI technology known as CHIEF that analyzes images of digital scans of tissue slides containing cancer tumors and predicts a number of things related to a certain type of cancer and its patient: whether a cancer is detected, what its genetic signature is, and how long a person may survive.
At the same time, CHIEF can detect traits in the tissue around a tumor that may be important for predicting a patient's reaction to chemotherapy, radiation, and immunotherapy.
