Computational Tool Could Improve Clinical Success Rate Of Drugs
By Deborah Borfitz
September 8, 2020 If physicians, computational scientists, and translational researchers more often collaborated to bring computation methods to the fore, it might help improve the notoriously low clinical development success rates for investigational drugs. Exhibit A is a newly developed translatable components regression (TransComp-R) model, developed by researchers from Purdue University and the Massachusetts Institute of Technology (MIT), which recently identified an overlooked biological mechanism possibly responsible for resistance to the drug infliximab commonly prescribed to patients with inflammatory bowel disease (IBD).
Doug Brubaker, assistant professor of biomedical engineering at Purdue who led development and testing of the TransComp-R computational model as a postdoctoral associate at MIT, says the framework would be potentially useful for anyone trying to understand if results of studies with animal models of disease can be replicated in humans. “It is well documented at this point that taking observations as-is from a mouse… more often than not don’t generalize well across species.”
A comprehensive survey of clinical success rates across the drug industry, published in Nature Biotechnology (DOI: 10.1038/nbt.2786), found that only one out of every 10 new compounds ever achieve approval by the U.S. Food and Drug Administration. Concerningly, the success rate for phase 3 development—which accounts for more than one-third of all R&D spending and over half of all clinical trial costs—was pegged at 60% for drugs across indications.
The idea behind TransComp-R, says Brubaker, is to allow researchers “to make an inference about which aspects of biological mechanisms in an animal study are most likely to translate to humans.” The results of those better-informed preclinical studies could improve the design of subsequent clinical trials, as well as confidence in go and no-go decisions on investigational drugs.
TransComp-R relies on common machine learning methods familiar to most computational biologists and statisticians, Brubaker says, combining principal component analysis and regression modeling based on those projections to identify the variables that matter most. “We found a way to take these well-established methods apart and use different pieces of them in new ways.”
The code for TransComp-R recently published in Science Signaling (DOI: 10.1126/scisignal.aay3258), along with findings of the proof-of-concept study and information on how to repurpose the framework to different kinds of animals, disease areas, and questions.
Proteomics from mice and transcriptomics from humans prior to treatment were used for the IBD study, says Brubaker, and all of the data have been in the public domain for several years. The mouse data came from two studies with fewer than 10 samples each and the human data from a single study involving more than 50 patients with IBD. Follow-up experiments were conducted to prove that the therapeutic resistance mechanism predicted by TransComp-R was in fact at work, Brubaker says.
The analysis on the mouse and human samples were done on biopsies that contained a broad mix of cell types, he explains, and researchers needed to know which cell types in a Crohn’s disease patient were expressing the genes in the resistance signature and then see if they could target this resistance mechanism to potentially make anti-TNF therapy more effective.
Collaborators at Vanderbilt University Medical Center provided single-cell RNA sequencing data on a patient not responding to anti-TNF therapy so researchers could figure out which cell types were expressing the genes in the resistance mechanism predicted by TransComp-R, he says. Observing that those genes were predominantly expressed in immune cells, they did a second experiment treating those cells in blood samples from healthy donors—with an inhibitor of the resistance mechanism, with anti-TNF therapy and then a combination of the two.
Inhibiting the resistance signature was found to enhance the ability of anti-TNF therapy to inhibit inflammatory cytokines, says Brubaker. The key takeaway is that TransComp-R can be used with available data to produce a testable biological hypothesis.
Other researchers might similarly look back at published studies to extract new learnings about disease biology, continues Brubaker. In principal, any kind of data could be analyzed with TransComp-R.
The tool essentially consolidates thousands of measurements from an animal model to just a few data coordinates for comparing with humans, he says. The dwindled-down data explain the most relevant sources of biological differences between the animal model and humans and can be used to train predictive models of a human’s response to therapy.
IBD was chosen as the starting point for the TransComp-R model because of the well-known clinical problem—about half of all people who take infliximab end up becoming resistant or unresponsive to it, Brubaker says. Trying to figure out why is also hampered by a gap in the literature when it comes to detailed, protein-level measurements from the site of disease in the human intestine during disease. Such measurements are only available from mouse models of the disease.
One potential follow-up line of investigation is to find ways to test for resistance markers in IBD patients to potentially inform which treatment physicians prescribe, says Brubaker. Another is to develop therapies around this resistance mechanism. The TransComp-R framework might also be applied to new types of questions and disease areas.
Douglas Lauffenburger, Ph.D., a professor of biological engineering, chemical engineering, and biology at MIT, as well as senior author on the Science Signaling paper, now has a new study underway using TransComp-R to translate data from COVID-19 vaccine trials in animals to humans, Brubaker notes. The goal is to identify the correlates of protection in the animals that might also be true about humans in the quest for an effective vaccine.