NEW YORK (GenomeWeb) – Researchers at the University of California, San Diego School of Medicine and the Moores Cancer Center have developed a method of generating virtual cancer cells — colloquially called cancer avatars — based on patients' genomic profiles that can be used to identify drugs to potentially treat their cancers more effectively.
A Journal of Translational Medicine paper published this week describes the development and application of the model to find treatments for glioblastoma (GBM). The model is a version of an existing one used by Cellworks, a therapeutics design company that was founded in 2005 and uses simulation techniques to screen data from studies in order to predict clinical outcomes.
Cellworks, which has offices in the US and India, has used its modeling techniques in rare disease studies, but the new study marks the first time that the approach has been used for cancer, Sandeep Pingle, a project scientist at UCSD and the lead author on the study, told BioInform. Researchers from Cellworks are co-authors on the study.
The team believes that its model-based approach could help researchers use genomic information to group similar patients for clinical trials and only test therapies on patients who the models predict will respond to them. "The advantage of computational modeling is the ability to incorporate the wealth of genomic and proteomic information on cancer cells and to screen drugs and combinations of drugs much faster and cost effectively," Santosh Kesari, the study's senior author, said in a statement. Kesari is the director of the Moores Cancer Center's neuro-oncology division and a professor in UCSD's neurosciences department. Pingle is one of the scientists in his lab.
According to the JTM paper, the researchers built the basic model by "curating data from the literature and aggregating functional relationships between proteins." Full details of the protocols used to build the models based on this information are available in the supplementary documents provided along with the study, but essentially the result of the process is a virtual cell model that includes representations of signaling pathways and metabolic networks that are found in normal, healthy cells and known to play roles in cancer's development and survival.
Their model includes representations of growth factors such as EGFR, PDGFR, and FGFR; cytokines and chemokines such as IL1, IL4, and IL6; GPCR-mediated signaling pathways; mTOR signaling; and more. The version of the model that was used in the glioblastoma study included "more than 4,700 intracellular biological entities and ~6,500 reactions representing their interactions, regulated by ~25,000 kinetic parameters" providing "comprehensive and extensive coverage of the kinome, transcriptome, proteome, and metabolome," the researchers wrote.
To make the cells cancerous, the team introduced distortions such as mutations and copy number variants that are known to be present in the specific tumor being studied. "We also created in silico variants for cancer cell lines, to test the effect of various mutations on drug responsiveness … by adding or removing specific mutations from the cell line definition," the paper explains.
Although the method can be applied to cancer in general, the UCSD researchers and their collaborators chose to focus on GBM which is one of the more common types of adult brain cancer with a median survival of 15 months, although some estimates have that figure somewhere between 12 and 14 months. Earlier this year, the New York Genome Center launched a study focused on finding better treatments for GBM using a prototype of IBM's Watson designed specifically to handle genomic data for clinical research studies.
They tested the in silico GBM cell lines that they created for the study using the methods described above on data from a separate study run by Garnett et al that analyzed the responses of 639 cancer cell lines to 130 therapies. The researchers also performed a prospective study where they compared the predictions of their GBM cell lines to in vitro experiments conducted with cell lines derived from fresh glioblastoma samples taken from patients at UCSD.
In the retrospective analysis that only looked at a subset of the drugs and genes analyzed in Garnett et al — the UCSD models accounted for about 45 cell lines and 70 drugs from the larger study — the GBM models were able to accurately predict 22 of 25, or approximately 85 percent, testable associations, according to the paper.
For the prospective analysis, the researchers tested the drug sensitivities of virtual models of eight of the cell lines that they'd generated from the UCSD patients and compared them to the results of drug sensitivity testing experiments using the real cell lines. Here they tested ten anti-cancer drugs on the virtual models and cell lines. Of the 80 in silico predictions, just over 76 percent "showed agreement with in vitro experimental results," the researchers reported.
For their next steps in the glioblastoma study, the team intends to incorporate additional data into its models, for example things like mutation data, Pingle told BioInform. For the prospective study that used cells from the UCSD GBM patients, the researchers only used CNV data and suspect that might be why they achieved only 75 percent agreement between the simulations and actual results. "A more comprehensive and accurate profile would require additional data (gene mutations, methylation status, etc., along with copy number variation)," they wrote. "This would help us develop a more representative avatar and would likely improve the accuracy of our drug response predictions and provide higher correlation with experimental data."
In addition, Pingle and his colleagues will use the models to explore responses to therapy combinations including looking at responses to specific drug dosages. Other plans for future versions of the model include incorporating information on the tumor microenvironment, "including aspects of angiogenesis, hypoxia, and tumor-associated inflammation," the paper states.