NEW YORK (GenomeWeb) – Mitra Biotech, an academic spinoff focused on personalized cancer care, is validating a test called CANScript that it hopes will predict the best treatment options for patients more accurately than genomic tests and more quickly than with avatar mice.
In a Nature Communications paper published in late February, researchers from Bangalore, India-based Mitra, Harvard University, Massachusetts Institute of Technology, the Indian Institute of Science, and elsewhere described the development and validation of CANScript.
The test involves first creating an ex vivo tumor ecosystem that closely reflects the unique heterogeneity and architecture of a patient's tumor as it exists inside the body; testing different factors indicating tumor response, such as cell proliferation and death, using up to 17 assays; and feeding the findings of these assays into a machine learning algorithm to calculate a so-called "M-Score" (on a scale of one to 100) that predicts the likelihood the patient will respond to a specific treatment strategy.
"The higher the M-Score, the greater the chance of a particular drug combination working for the patient under consideration," Mitra CEO and founder Mallik Sundaram told GenomeWeb. A score of greater than 25 suggests that a patient is likely to respond to a therapeutic combination, while a score lower than 25 signals non-response.
Privately held Mitra spun out of efforts at MIT and Harvard in 2009. According to Sundaram, CANScript can overcome the limitations of currently available genomic tests, which gauge specific mutations or interrogate certain drug pathways. Mitra is hoping to market CANScript as a predictive test that can guide treatment decisions with chemotherapeutics, biologics, monoclonal antibodies, and immunomodulators.
"We are collecting convincing data for each class of molecules," Sundaram said. "Further, CANScript has been tested in multiple solid cancers, as well as hematological cancers."
As described in the Nature Communications paper, CANScript analyzes tumor explants on ex vivo plates coated with tumor matrix proteins (TMPs) that are specific for the grade and type of the patient's tumor. "The composition of various classes of proteins varies with tumor type and grade," Sundaram explained. Mitra scientists have identified and optimized the composition of tumor matrix proteins for different indications, he added.
As part of this process of mimicking the tumor microenvironment in explants, the CANScript system utilizes autologous sera, which have critical growth factors that help conserve tumor characteristics in explants as if they were in the body.
"The TMPs in conjunction with autologous ligands provide structural support, as well as mimic cell signal mechanisms … to tumor cells better than other commercially available matrix proteins like collagen or gelatin," Sundaram said.
Commercial matrix proteins such as collagen "are good from a structural support point of view but are not sufficient to encompass the biological context of key complex matrix proteins in maintaining tumor viability, proliferation, and signaling," he added. Mitra scientists also use the patient's immune cells from blood to maintain the immune microenvironment.
Next, the CANScript test system assesses tumor response to different agents employing a range of assays, using for example, measures of the protein Ki-67; uptake of ATP and ADP to gauge cell proliferation; or assessments of proteins caspase-3 and caspase-8 to track apoptosis. The data from these assays is then put into a machine learning algorithm to calculate the M-Score and predict treatment responses.
In one of the many investigations described in the paper, researchers compared the ability of avatar mice (or human tumor-derived xenotransplants) and CANScript explants in predicting responses to a regimen of docetaxel, cisplatin, and 5-fluorouracil. There was "excellent correlation" in predicting responders between the two approaches, lead author Pradip Majumder from Mitra and colleagues wrote in the paper.
Mitra believes the CANScript system could offer certain advantages to xenotransplants, which are increasingly utilized in precision cancer treatment efforts. For example, the implantation and initial growth of the tumor in the mouse avatar can take as much as three months, and has a success rate of between 30 percent and 75 percent depending on the tumor type, Sundaram said. Then, the tumors have to be grown in a second set of mice, which can take another three to six months, with a success rate of between 50 percent and 90 percent.
So, arriving at molecularly informed treatment strategies using avatar mice may take between six months to a year. Meanwhile, "the patient's tumor does not remain static during this time period," Sundaram said. "So even if we assume that the tumor does grow in mice, the response one would see in mice would hardly match that of the human tumor because of the long time interval."
The tumor microenvironment isn't maintained once implanted in mice, he added. Comparatively, the focus with CANScript is to retain as much of the patient's inherent tumor characteristics in the model system. Furthermore, Mitra claims that it takes only a week from the time the tumor is removed from the patient to when the firm reports results from CANScript to the doctors.
In the paper, researchers also described how they trained the machine learning algorithm so that CANScript could predict treatment benefit using explants from biopsies of more than 100 colorectal cancer and head and neck squamous cell cancer patients. The head and neck cancer patients had received docetaxel, cisplatin, and 5-fluorouracil and the colorectal cancer patients had received Erbitux (cetuximab) and FOLFIRI (folinic acid/fluorouracil/irinotecan). The researchers aimed to design the test to have a high sensitivity rate for accurately identifying true positives (i.e. potential responders), and achieved a sensitivity of 96.77 percent on the training set.
Then, Majumder and colleagues used CANScript on a test set of 55 patients – 42 with head and neck squamous cell cancer and 13 with colorectal cancer. The test demonstrated 91.67 percent specificity and 100 percent sensitivity. The researchers moved on to train the machine learning algorithm to differentiate between partial, complete and non-responders, with 87.27 percent accuracy.
All 13 colorectal cancer patients had the wild-type form of the KRAS gene according to mutation testing and received Erbitux as indicated in the drug label. However, only three of them actually responded to the drug – one had a complete response, two had partial responses, and 10 experienced disease progression. Meanwhile, CANScript predicted two patients would have a complete response, two would have partial responses, and nine would be non-responders. As such, CANScript in one case predicted a non-responder to be a complete responder.
The 42 head and neck squamous cell cancer patients in the test set received docetaxel, cisplatin, and 5-fluorouracil according to the standard of care, but 14 did not respond. CANScript identified 13 as non-responders. The study authors characterized the mistakes made by CANScript as "benign" errors, since there were no cases where the test identified responders as non-responders. While such errors could result in a patient who is unlikely to respond to receive treatment and experience side effects, it also means that no patient who would respond to chemotherapy is denied a drug based on a false prediction, the researchers noted in the paper. They added that the platform needs to be further validated on larger sample sizes.
According to Mitra, between 100 mm3 to 200 mm3 of fresh tumor tissue is required for CANScript to test whether a patient will respond to four drug combinations; 300 mm3 of tissue is necessary to test six drug combos. Since "the tumor tissue is cultured in its native state," Sundaram explained that "the tumor behavior, irrespective of the genetics, drug mechanism, signal pathways involved, is observed on introduction of different drugs." Genetic testing requires the same amount of tissue from the patient, he noted. "In both cases additional sample, if available, would be preferred to account for the necrotic tissues that may form part of the biopsy sample."
Because CANScript is a highly personalized test, the company believes it will be useful globally for patients of different races and ethnic groups. Mitra, which is ventured backed by Accel Partners and Tata Capital Innovation Fund, has introduced the test initially in a number of Indian cities, but is planning to launch it in other markets in the future with local partners. The company currently has operations in Chicago and Boston.
"For high-end, first-in-class diagnostics, the R&D and validation phases are the longest," Sundaram said. "We are actively building sizable data comparing CANScript outcomes and corresponding clinical outcomes for many solid and hematological cancers." Mitra has also partnered with a number of Indian and American research groups, including Tata Memorial Hospital in Mumbai, Mazumdar-Shaw Cancer Center in Bangalore, HCG Cancer Care Network headquartered in Bangalore, and the Cancer Treatment Centers of America.
Mitra operates a repository housing more than 600 annotated refractory tumors. The company also has a wet lab and employs approximately 30 scientists who are conducting research studies and are further validating CANScript.
CANScript is Mitra's first test. The company is also eyeing collaborations with drugmakers to use CANScript and xenograft mouse models to inform investigational drug programs. "At our core, we are a research organization and have ongoing studies to validate CANScript for additional cancer types," Sundaram said.