“The body’s immune system is very good at identifying cells that are acting strangely. These include cells that could develop into tumors or cancer in the future,” explained Federica Eduati, Ph.D., from the department of biomedical engineering at Eindhoven University of Technology. “Once detected, the immune system strikes and kills the cells.” But that is not always with the case with cancer, which has sophisticated methods of cloaking itself from the immune system.
Now, researchers at the Eindhoven University of Technology report in the journal Patterns that they have combined machine learning with immunotherapy to find hidden tumors cells in the human body.
“Cancer cells can leverage several cell-intrinsic and -extrinsic mechanisms to escape immune system recognition,” wrote the researchers. “The inherent complexity of the tumor microenvironment, with its multicellular and dynamic nature, poses great challenges for the extraction of biomarkers of immune response and immunotherapy efficacy.”
Immune checkpoint blockers (ICB) are a type of immunotherapy that tell the immune cells to ignore the “shutdown orders” coming from cancer cells. Although ICB has been effective in treating many different types of cancers, only a third of patients respond to the treatment.
To predict whether a patient will respond to ICB, the researchers first sought to obtain particular biomarkers in tumor samples from patients.
“Tumors contain more than just tumor cells, they also contain several different types of immune cells and fibroblasts, which can have a pro- or anti-tumor role, and they communicate with each other,” explained Oscar Lapuente-Santana, a PhD researcher in the computational biology group. “We needed to find out how complex regulatory mechanisms in the tumor microenvironment affect response to ICB. We turned to RNA-sequencing datasets to provide a high-level representation of several aspects of the tumor microenvironment.”
The team searched the microenvironment of tumors using computational algorithms and datasets from previous clinical patient care.
“RNA-sequencing datasets are publicly available, but the information about which patients responded to ICB therapy is only available for a small subset of patients and cancer types,” said Eduati. “So, we used a trick to solve the data problem.”
The researchers picked out several substitute immune responses from the same datasets to be used as an indicator of the effectiveness of ICB.
“A significant challenge with this work was the proper training of the machine learning models. By looking at substitute immune responses during the training process, we were able to solve this,” said Lapuente-Santana.
“We used machine learning to look for associations between the derived system-based features and the immune response, estimated using 14 predictors (proxies) derived from recent publications,” wrote the researchers. “We considered these proxies as different tasks to be predicted by our machine learning models and used multi-task learning algorithms in order to learn all tasks jointly.”
The researchers observed that their machine learning model outperforms biomarkers currently used in clinical settings to assess ICB treatments.
“Mathematical models can provide a big picture of how individual molecules and cells are interconnected, while at the same time approximate the behavior of tumors in a particular patient. In clinical settings, this means that immunotherapy treatment with ICB can be personalized to a patient. It’s important to remember that the models can help doctors with their decisions on the best treatment, they won’t replace them,” said Eduati.
In addition, the model also helps in understanding which biological mechanisms are important for the biological response. “Another advantage of our approach is that it does not require a dataset where patients’ responses to immunotherapy are known for model training,” wrote the researchers.
More experiments are needed before these results turn to clinical settings.
