A mechanism that detects cancers that cannot be diagnosed by the available means

Researchers in the United States have come up with a new methodology that may help detect unknown cancers even before a patient develops symptoms.
The methodology, developed by researchers at the Koch Institute for Integrative Cancer Research at MIT and Massachusetts General Hospital, is based on deep learning by taking a close look at gene expression programs related to early cell development and differentiation.
“Sometimes you can apply all the tools that pathologists offer, and it remains unanswered,” says Salil Garg, a researcher at the Charles and Jennifer C. Johnson Institute for Cancer Research, a clinical investigator at the Koch Institute and a pathologist at Massachusetts General Hospital. An AI tool like this is capable of identifying types of cancer with a high degree of sensitivity and accuracy.”
Garg is the lead author of the new study, which was published Aug. 30 in the journal Cancer Discovery.
The first step in choosing the right cancer treatment is determining the type of cancer the patient has, including identifying the organ or part of the body from which the cancer began.
In some cases, the origin of the cancer cannot be determined, even with extensive testing. Although these primary unknown cancers tend to be aggressive, oncologists must treat them with untargeted therapies that often have severe toxicity and lead to low survival rates.
Hence, the importance of early diagnosis of the disease.
Analyzing differences in gene expression between different types of unknown primary tumors is a problem that machine learning can solve. Cancer cells look and act very differently than normal cells. This is partly due to extensive modifications in how their genes are expressed. Thanks to advances in characterizing individual cell traits and efforts to catalog different cellular expression patterns, abundant data exists containing clues about how different cancers arise and where they originate.
The researchers compared two large cell atlases and identified associations between cancer cells and embryonic cells: The Cancer Genome Atlas, TCGA, which contains gene expression data for 33 types of tumors, and the Mouse Organogenesis Cell Atlas, MOCA), which identifies 56 separate pathways of embryonic cells during their development and differentiation.
The map of associations between developmental gene expression patterns in tumor cells and embryonic cells was then turned into a machine learning model. The researchers divided the gene expression of tumor samples from the Cancer Genome Atlas into individual components corresponding to a specific time point in the growth pathway, and assigned a mathematical value to each of these components. The researchers then built a machine-learning model, called a Developmental Multilayer Perceptron (D-MLP), that scores a tumor from its developmental components and then predicts its source.
After training, the D-MLP machine learning model was applied to 52 new samples of particularly challenging cancers of unknown primary type that could not be diagnosed with the available tools. These represented the most challenging cases at Massachusetts General Hospital over a four-year period starting in 2017. Interestingly, the model categorized tumors into four categories, yielding predictions and other information that could guide diagnosis and treatment of patients.
For example, one sample came from a patient with a history of breast cancer who showed signs of aggressive cancer around the abdomen. Oncologists, at first, were unable to find the mass of the tumor, and they could not classify the cancer cells using the tools they had at the time. However, the new D-MLP model predicted ovarian cancer in the woman. Six months after the patient was first seen by doctors, an ovarian mass was finally found that proved to be the origin of the tumor.
In the future, the researchers plan to increase the predictive power of their model by incorporating other types of data, particularly information from radiology, microscopy and other types of tumor imaging.
“Evolutionary gene expression represents just one small slice of all the factors that can be used to diagnose and treat cancers,” Garg asserts, and goes on to say that “integrating radiology, pathology and gene expression information together is the real next step in cancer medicine.”