Researchers at the University of California (UC) San Diego School of Medicine have developed a drug discovery approach that uses machine learning to search for disease targets and then predict whether a potential drug is likely to succeed in Phase III trials and achieve FDA approval. Demonstrating their approach using inflammatory bowel disease (IBD) as a test case, the researchers replaced two steps in traditional preclinical drug discovery workflows with the new approaches developed within the UC San Diego Institute for Network Medicine (iNetMed).
They say the study findings could measurably change how researchers sift through big data to derive meaningful insight, which will then be of significant benefit to patients, the pharmaceutical industry, and health care systems. “Our approach could provide the predictive horsepower that will help us understand how diseases progress, assess a drug’s potential benefits, and strategize how to use a combination of therapies when current treatment is failing,” said Debashis Sahoo, PhD, who leads the Center for Precision Computational System Network (PreCSN), the computational arm of iNetMed, and is an associate professor in the departments of pediatrics and computer science at UC San Diego School of Medicine and UC San Diego. Sahoo is also co-senior author of the researchers’ study, which is published in Nature Communications, and titled, “Artificial intelligence guided discovery of a barrier-protective therapy in inflammatory bowel disease.”
Drug discovery in its current state, is “… wasteful and fraught with increasing trends of failures, implying the existence of uncertainty and impreciseness in the process,” the authors stated. Researchers are now increasingly looking to use computational algorithms to sort through “big data”—such as transcriptomics—they suggest, and build networks, as a way to visualize complexity, increase understanding of complex human diseases, and potentially prioritize targets. “Modeling human diseases as networks simplify complex multi-cellular processes, helps understand patterns in noisy data that humans cannot find, and thereby improves precision in prediction.”
Pradipta Ghosh, MD, senior author of the study and a professor in the departments of medicine and cellular and molecular medicine at UC San Diego School of Medicine, explained further: “Academic labs and pharmaceutical and biotech companies have access to unlimited amounts of ‘big data’ and better tools than ever to analyze such data. However, despite these incredible advances in technology, the success rates in drug discovery are lower today than in the 1970s. This is mostly because drugs that work perfectly in preclinical inbred models, such as laboratory mice, that are genetically or otherwise identical to each other, don’t translate to patients in the clinic, where each individual and their disease is unique. It is this variability in the clinic that is believed to be the Achilles heel for any drug discovery program.”
The new strategy for drug discovery, developed by Ghosh, Sahoo, and colleagues, employs what they call a network-based approach that uses artificial intelligence to prioritize target identification, and then guides validation in “network-rationalized preclinical mouse and patient-derived organoids.”
For their newly published paper, the researchers demonstrated their approach using a disease model for inflammatory bowel disease (IBD), as a test case. IBD is a complex, multifaceted, relapsing autoimmune disorder characterized by inflammation of the gut lining. And because it impacts all ages and reduces the quality of life in patients, IBD is a priority disease area for drug discovery. However, it is a challenging condition to treat because it doesn’t behave in the same way in different patients. And despite being at the forefront of biomedical research, the team noted, therapies that can restore and/or protect the integrity of the gut barrier in IBD haven’t yet been developed.
The researchers’ new approach is carried out in four stages. The first, target identification step applies an AI methodology developed by the PreCSN, which helps to model a disease using a map of successive changes in gene expression at the onset and during the progression of the disease. “A database containing 1,497 human gene expression data from both 1,263 human samples and 234 mouse samples was mined to build a validated Boolean implication network-based computational model of disease continuum states in IBD,” the team explained. Paths, clusters, and a list of genes in the network-based model are prioritized to discover clinically actionable drug targets.” The search yielded PRKAB1 as a potentially barrier-protective therapeutic target.
What sets this mapping apart from other existing models is the use of mathematical precision to recognize and extract all possible fundamental rules of gene expression patterns, many of which are overlooked by current methodologies. The underlying algorithms ensure that the identified gene expression patterns are “invariant” regardless of different disease cohorts. In other words, PreCSN builds a map that extracts information that applies to all IBD patients. “In head-to-head comparisons, we demonstrated the superiority of this approach over existing methodologies to accurately predict ‘winners’ and ‘losers’ in clinical trials,” said Ghosh.
As outlined in their paper, step 2 of their approach is preclinical validation in a human-like mouse model. For their IBD test case, two PRKAB1-specific agonists were successfully tried in mice, the authors reported. Tests showed that the identified PRKAB1-agonists ameliorated colitis in a mouse model of colitis, which the authors claimed “… validate the use of PRKAB1-agonists as barrier-protective therapy and provide preclinical proof of concept and mechanism ….” The final stage of target validation in preclinical models (this is step 3), was conducted as a first-of-its-kind Phase “0” clinical trial using a living biobank of organoids created from IBD patients at the HUMANOID Center of Research Excellence (CoRE), which is the translational arm of iNetMed.
This Phase “0” approach involves testing the efficacy of the drugs identified using the AI model on the human disease organoid models. These are human cells cultured in a 3D environment that mimic diseases, but outside of the body. In this case, they used an IBD-afflicted gut-in-a-dish. Biopsy tissues for the study were taken during colonoscopy procedures involving IBD patients. Those biopsies were used as the source of stem cells to grow organoids.
“The ‘Phase 0’ trial concept was developed because most drugs fail somewhere between Phases I and III,” said Soumita Das, PhD, co-senior author of the study, director of the HUMANOID center, and an associate professor in the department of pathology at UC San Diego School of Medicine. “Before proceeding to patients in the clinic, ‘Phase 0’ tests efficacy in the human disease models, where ineffective compounds can be rejected early in the process, saving millions of dollars.” In their reported study, one of the PRKAB1-agonists was successfully tested in the Phase 0 trial in patient-derived orgnaoid models. The experiments showed that the PRKAB1-agonists protected the epithelial barrier in organoid models, and restored the leaky barrier in patient-derived organoids.
Step 4 of the process is then predicting success in Phase III trials, an approach that involves identifying all past and current clinical trials of drug targets, including FDA and failed targets.
Through their study, the researchers found that the computational approach had a surprisingly high level of accuracy across diverse cohorts of IBD patients. Das said their results threw up two major surprises. “First, we saw that despite being away from the immune cells in the gut wall and the trillions of microbes that are in the gut lining, these organoids from IBD patients showed the tell-tale features of a leaky gut with broken cell borders. Second, the drug identified by the AI model not only repaired the broken barriers, but also protected them against the onslaught of pathogenic bacteria that we added to the gut model,” Das added. “These findings imply that the drug could work in both acute flares as well as for maintenance therapy for preventing such flares.”
“We demonstrate how these four steps synergize and aid in the modeling of fundamental progressive time series events underlying complex human diseases and exploit such insights to improve precision when developing disease-modifying drugs,” the team stated. “This AI-assisted approach identifies a first-in-class gut barrier-protective agent in IBD and predicted Phase-III success of candidate agents.”
“Our study shows how the likelihood of success in Phase III clinical trials, for any target, can be determined with mathematical precision,” noted Sahoo. The authors said that next steps in their development will include testing whether the drug that passed the human Phase “0” trial in a dish can pass Phase III trials in clinic; and whether the same methodologies can be used with other diseases, ranging from diverse types of cancers and Alzheimer’s disease to non-alcoholic fatty liver disease. “Our blueprint has the potential to shatter status quo and deliver better drugs for chronic diseases that have yet to have good therapeutic solutions,” claimed Ghosh.