New AI model improves prediction of cancer immunotherapy success

Cancer immunotherapy drugs known as immune checkpoint inhibitors (ICIs) can be miracle drugs for cancer patients, curing some patients and turning a deadly disease into a manageable chronic disease in others. However, these drugs are only effective for some patients, and little is known about why. Knowledge gaps negatively impact patient prognosis, clinical trial recruitment, and research that could lead to new treatments.

A new artificial intelligence model called COMPASS, developed by researchers at Harvard Medical School and their colleagues, improves predictions of which patients are most likely to respond to ICIs. The model, which used data from previously treated patients, outperformed the best existing approach by 8.5%. It makes predictions based on patient tumor gene activity and provides a rationale for its output.

If these results are validated in future clinical trials, COMPASS could lead to better personalized medicine for cancer patients, more efficient trial enrollment for new treatments, and new drug targets for researchers to explore.

Details of the results will be posted on July 3rd. natural medicine.

ICIs are an exciting treatment that has transformed cancer treatment over the past decade by engaging the immune system to fight and destroy cancer cells. By leveraging cutting-edge AI capabilities, we can identify patients who are most likely to respond to a particular ICI before they receive the drug. ”

Marinka Zitnik, lead study author, Associate Professor of Biomedical Informatics, HMS Blavatnik Institute

Potentially powerful cancer treatment

The first ICI was approved by the U.S. Food and Drug Administration in 2011. These drugs, made possible in part by the work of HMS scientists, target proteins on the surface of tumor cells or T cells, such as PD-L1, PD-1, and CTLA-4. These proteins act as an invisible cloak, protecting cancer cells from immune attack. ICIs disrupt this interaction, allowing cancer cells to be recognized and destroyed by the immune system again.

For some patients with certain types of cancer, ICIs are a literal lifeline, extending survival far beyond what was previously thought possible. For example, former U.S. President Jimmy Carter survived for nine years after being diagnosed with stage IV melanoma that had spread to his liver and brain, a feat largely attributed to taking a PD-1 inhibitor called pembrolizumab.

But those who are responding to President Carter and other ICIs represent only a small portion of patients receiving these drugs. Clinical trials have shown that ICIs are successful in only 10 percent to 40 percent of patients, depending on the type of cancer. People who don’t respond risk having potentially serious side effects, as well as wasting time on ineffective treatments while their cancer progresses.

Several machine learning approaches and biomarkers have been used to predict which patients are most likely to respond to ICIs. For example, responses are associated with an immunoinflammatory tumor microenvironment (characterized by tumor infiltration of immune cells), whereas unresponsive tumors are often so-called immune deserts.

However, a significant number of patients respond to these drugs in unexpected ways, negatively impacting the reliability of these predictions.

“Understanding who addresses ICI is not a small knowledge gap,” says Zitnik, who is also an associate professor at Harvard’s Kempner Institute for Natural and Artificial Intelligence. “This is one of the central unresolved questions in oncology.”

A compass that points the way for responders

Zitnik and her colleagues developed COMPASS to solve this problem. This model predicts ICI responses by analyzing immune cell status, tumor-microenvironment interactions, and the activity of approximately 16,000 genes with known roles in signaling pathways.

COMPASS is designed with what is called a conceptual bottleneck transformer architecture. Rather than spewing out unexplained black box predictions, it provides human-interpretable results and provides a rationale for the output.

The researchers trained COMPASS using data from 10,184 tumors across 33 cancer types from the Cancer Genome Atlas, a public database containing genetic sequence and molecular data from primary cancers and matched normal samples. Using this data, the AI ​​program “learned” which gene activity correlated with responders and non-responders to different types of ICIs.

The team then fine-tuned this training using results from 16 clinical trials that tested the effectiveness of different ICI regimens in seven types of cancer. To assess the model’s success, they removed individual clinical trials from this tweak one by one and asked COMPASS to predict ICI responders and non-responders in the missing trials.

Their results showed that COMPASS outperformed the best existing approach for predicting ICI response by nearly 10% on average. This increased accuracy holds true under different conditions, including different cancer types, ICI drugs, gene transcription sequencing platforms, and biopsy sites.

The results were interpretable, allowing the team to explain the unexpected results among the ICI response outliers. For example, gene expression in some non-responders of immunoinflamed tumors correlated with processes that interfere with the immune response. Conversely, the gene expression signatures of immune desert tumor responders often suggested biological processes promoting other types of immune activation.

Future direction

If these results hold true in prospective clinical trials, COMPASS could be used in cancer clinics as a decision aid to help physicians determine which patients would benefit most from ICIs, Zitnik explained.

This tool could also be a boon for ICI clinical trials by helping trial participants enroll optimal participants and giving those participants the best chance of a meaningful response.

And because COMPASS results are interpretable, they could generate new hypotheses about how the immune system fights cancer, which in turn could lead to new drug targets, Zitnik added.

She and her colleagues plan to test whether incorporating additional data into COMPASS can further improve accuracy. This could include details from the patient’s electronic medical record, such as medical history, disease comorbidities, and previous responses to other drugs or treatments, as well as single-cell sequencing data that may reveal the role of different cell populations in the ICI response.