This article has been corrected to reflect that Cellworks was mistaken when it provided information about grant-seeking activities. The company is not pursuing a grant from DARPA.
NEW YORK – Cellworks is advancing multiple clinical decision support tools based on a computational biology model through validation studies, betting that they will more accurately predict cancer patients' responses to treatment than marketed genomic tests and help oncologists practice precision oncology amid rapidly advancing science.
The South San Francisco, California-based company markets two products, Singula and Ventura, that analyze patients' next-generation sequencing data based on a computational model that simulates the interactions of more than 6,000 genes, 30,000 molecular species, and 600,000 molecular interactions within a tumor. Singula predicts patients' responses to standard-of-care, front-line therapies, and Ventura predicts and ranks combinations of US Food and Drug Administration-approved drugs, including off-label indications that refractory cancer patients are likely to respond best to. Both products are based on a therapy response index generated by Cellworks' computational biology model.
Unlike many other platforms in development that inform treatment selection for cancer patients, Cellworks' model is not primarily driven by artificial intelligence. Instead, the developers have constructed a simulation of cell biology from the ground up, using the fundamental mathematical equations that govern molecular interactions, including mass action equations, diffusion equations, and Michaelis-Menten enzyme kinetics. Some of the equations come from the scientific literature, and some Cellworks generated in-house using public and private databases. Then, Cellworks layers on top of that framework multiomics data from each patient, including whole-exome sequencing, gene expression, and methylation patterns.
The result, said Cellworks Chief Medical Officer Michael Castro, is a "personalized avatar" of each patient's tumor that provides a "network view" of the disease rather than limited snapshots of one or more cancer-related genomic alterations. The simulation runs in the cloud, Castro said, because it's "too vast" for a single computer to run.
"There are a great many paradigms targeting [mutations in genes such as] EGFR, ALK, NTRK, and BRAF, but invariably the patients who have these [alterations] also have many other aberrations in their tumors, and those other aberrations determine whether a drug is going to work for nine months, five months, or not at all," Castro said. "When you get that information and see the signaling pathway consequences, it's a revelation in terms of explaining the hallmark behaviors of cancer." He added that reports generated from the model often contain hundreds of aberrations that never appear in commercial next-generation sequencing reports.
Castro highlighted mismatch repair deficiency as an example of a cancer driver defect simulated in the model. Typically, mismatch repair deficiency is understood as a deletion or mutation in one of a small number of genes. However, Castro said, Cellworks' biosimulation model takes into consideration interactions along the entire dysregulated pathway such as the opening of chromatin during transcription or acetylation of mismatch repair proteins. "If your cancer has acetylases or deacetylases that are missing, you're going to have some degree of mismatch repair deficiency," Castro said. "There's about 27 genes or so involved in this biosimulation, some of which are histone-modifying enzymes or microRNA."
Using this personalized avatar of the patient's tumor, the model can then simulate responses to "the entire therapeutic armamentarium," Castro said, including targeted cancer drugs, chemotherapy, radiation, non-oncology drugs, and various combinations.
"This is not very different from running a whole slew of experiments in a traditional wet lab, except we are doing this in silico," said Cellworks CEO Yatin Mundkur. "And a luxury of doing in silico is that you can get almost infinite measurements."
Although there are some machine learning elements in the model, such as for weighting individual phenotypes, Castro does not consider the computational biology model to be a form of machine learning or AI. "Machine learning for biology is going to be relatively limited because of the tremendous heterogeneity in individuals with disease," Castro said. That means that databases for statistical machine learning models would have to contain billions of cases to cover the range of genomic diversity for a single disease. "We need a mechanistic model which tells us how multiple aberrations are going to impact the transcriptional organization of the tumor," he added.
Cellworks is in the process of validating the model by studying survival outcomes in clinical datasets. In a real-world study published in ESMO Gastrointestinal Oncology in January, researchers retrospectively evaluated a biosimulation therapeutic-response index in 270 patients with operable gastroesophageal adenocarcinoma enrolled in the UK Esophageal Cancer Clinical and Molecular Stratification International Cancer Genome Consortium study. All patients had whole-genome sequencing from biopsied or surgically resected tissue within 12 months of receiving standard neoadjuvant chemotherapy treatments according to clinical guidelines in the UK.
The National Comprehensive Cancer Network and European Society for Medical Oncology recommend preoperative chemotherapy plus radiation and perioperative chemotherapy alone for patients with gastroesophageal cancer. However, evidence is unclear on whether any particular neoadjuvant therapy regimen is better than another, and thus far, robust biomarkers of therapy response have not been identified and validated. Cellworks researchers tested the association of the therapeutic-response index with disease-free survival, overall survival, and tumor regression grade.
They found that patients with unfavorable therapeutic-response index scores had worse outcomes, and the index was significantly associated with all three endpoints. The therapeutic-response index functioned better than clinical parameters in the study, Castro said, adding that he hopes validation studies will continue to show that the index helps clinicians make much better decisions with the index than without it.
Cellworks has also carried out validation studies in non-small cell lung cancer, acute myeloid leukemia, and glioblastoma, and Castro said the company will present results from a lung cancer study at ESMO's upcoming annual meeting in September.
It has been a challenge obtaining randomized data sets to validate the model, however, and Castro said that Cellworks is continuing its validation efforts through collaborations with organizations such as Avera Cancer Institute and Miami Cancer Institute. The company is also interested in partnering with drug developers to rescue failed drugs and inform development strategies for new agents.
In the meantime, in Castro's own California practice at the Beverly Hills Cancer Center, he is using Cellworks reports to inform therapy choices for patients. "Sometimes, what we find is that the standard of care is perfect," Castro said. "But there are other times we find out that the standard of care does not apply for the individual, and there's a not only plausible but mechanistically compelling reason. So, I use this information to help patients."
Cellworks is also exploring whether its computational biology model can help a molecular tumor board within the myCare-101 study. "The study is a joint learning [opportunity] between the molecular tumor board, the physicians, and Cellworks to best present this information in a digestible form for the clinician," said Mundkur.
In the study, Cellworks is analyzing patients' next-generation sequencing data using its computational biology model and returning patient-specific reports to molecular tumor board members, who will then complete a survey evaluating how the reports helped them recommend treatments for patients. The survey queries molecular tumor board members about the usefulness of the information in the report, whether the report supported their recommendations or provided a recommendation they wouldn't have otherwise made, and whether the reports informed their treatment recommendation or changed their mind. One of the participating cancer centers, Avera Cancer Institute, has ordered 400 reports from Cellworks, representing one-third of the study's goal of about 1,200 reports.
"Even for well-resourced and highly skilled molecular tumor boards, the rapid pace of knowledge advancement and the increasing number of patients undergoing molecular profiling present significant challenges," said Tobias Meissner, manager of cancer genomics at Avera. "By participating in this study, we can explore how technology can support interpretation of molecular data and selection of treatments, which is a crucial step in evaluating the practical application of this technology for patient care."