BOSTON—Gilead Sciences' nonalcoholic steatohepatitis (NASH) prospect selonsertib may have failed two phase 3 studies, but the company is putting the data from those studies to work. Part of that involves bringing better, noninvasive diagnostic tests to clinical trials, but also working to improve the current gold standard: invasive liver biopsy.
“We’re very much focused on trying to develop novel endpoints to enable or improve the evaluation of new therapies in clinical trials. The entire field recognizes liver biopsy is suboptimal,” Rob Myers, vice president and clinical fibrosis research lead at Gilead, told FierceBiotech.
At the annual meeting of the American Association for the Study of Liver Diseases, the company presented a trio of abstracts from studies that worked with data from the nearly 1,700 patients who took part in the STELLAR-3 and STELLAR-4 studies. In both studies, selonsertib did worse than placebo at improving fibrosis, or scarring, in patients with advanced NASH.
One of the abstracts (poster 1734) found noninvasive tests reflected patients’ responses to treatment better than liver biopsy did. The study found that patients’ scores on tests—such as a blood test for liver scarring and an ultrasound-based test for liver stiffness—“are associated with each other and other clinical parameters.” If the scores worsened, so did the risk of a patient’s fibrosis getting worse and progressing to cirrhosis, Myers said. If they improved, so did the other parameters, including tests of liver function, liver stiffness, markers of metabolism and noninvasive tests of fibrosis.
That finding wasn’t true for liver biopsy, where “only histologic parameters improved in responders defined by liver histology.”
Translation? “In patients who had a histologic response, really all that changed were the features that are on the slide,” Myers said. A patient whose disease improved after treatment was found to have less scarring based on what a pathologist saw in a tiny sliver of liver tissue. But they did not improve based on all the other tests “that are important to clinicians and patients.”
The difference was so stark, Gilead chalked it up to “sampling error of liver biopsy,” meaning that tissue samples taken from different parts of the same person’s liver can turn up vastly different results.
“It’s long been known to be a challenge with liver biopsy in liver diseases including NASH and it creates problems when we’re trying to evaluate new therapies when we can’t trust a gold standard,” Myers said.
The findings support that the noninvasive tests Gilead looked at can be used to monitor patients and as endpoints in clinical trials. But to get there, the company—and the field—needs to figure out “what a relevant change in these noninvasive tests would be in terms of association with prognosis.” Namely, how much of a change in a particular blood test or imaging-based test score would translate into a clinical benefit for patients?
“We’re working towards that based on all the data sets that we have and what will be particularly important for us is to see the results of our ATLAS data at the end of the year to see what type of changes we’re observing with our therapies and then determine the potential benefits of those changes based on all of our data sets in NASH,” Myers said. The ATLAS study is testing selonsertib and two other NASH hopefuls, firsocostat and cilofexor.
In the meantime, Gilead and digital pathology player PathAI are hard at work on improving liver biopsy. One of the other abstracts (oral abstract 187) found that deep learning models were able to identify key features of NASH, such as fat accumulating in the liver and inflammation, as well as pathologists did. The second (poster 1718) found that the models could go beyond traditional staging of NASH to characterize patients’ disease.
“The advantage of the approach is we’ve been able to show that fibrosis is very heterogenous. If you take patients who’ve been told they have cirrhosis by a pathologist, they’re not all created equally; the deep learning models really reveal the heterogeneity of fibrosis in those patients,” Myers said.
Such models could automate the interpretation of liver biopsies and make it objective and reproducible.
“Liver biopsy currently has multiple limitations, one important one being that pathologists often disagree with each other on the grading and staging of NASH,” Myers said. “[Another issue] is when individual pathologists look at the same biopsy multiple times, in 20% to 40% of cases, they’ll give a different answer.”
They could also help the field look beyond the discrete grading and staging systems of NASH: “You grade things from 0, 1, 2 or 3, for example, and they lack sensitivity. The beauty of the PathAI system is it’s more quantitative and may improve sensitivity to detect changes in histology attributable to treatment,” he said.
Moving forward, Gilead is discussing the data it has on noninvasive tests with regulators and is collaborating with PathAI to figure out how to get their deep learning models approved with regulators for use in clinical trials.
“Right now, we live in a world where liver biopsy is necessary for clinical trials evaluate the effects of our therapies. We would like ultimately to move to a place where we can just use noninvasive tests as endpoints. But in the interim, while we’re generating data to achieve that goal, we’re trying to improve upon liver biopsy as best we can,” Myers said.