Researchers from the University of California San Diego School of Medicine have described a new approach that uses artificial intelligence to hunt for disease targets and then predicts whether a drug is likely to receive FDA approval.
Hurdles in drug discovery
In its current state, drug discovery is a wasteful process with a high rate of failures. Researchers have access to vast amounts of big data and advanced analytic tools to process this data. However, despite advances in technology, the success rates in drug discovery are lower today than they were in the 1970s. One of the main reasons for this is that drugs that work perfectly in preclinical models, which are genetically or otherwise identical to each other, don’t translate to patients in the clinic, where each individual and their disease is unique.
The use of computational algorithms to sort through big data and build networks to visualise this complexity has become a popular approach to understand complex diseases and prioritise targets. First, relationships are identified between pairs of genes using symmetric computational frameworks such as linear regression. Following that, gene co-expression networks (GCNs) are built by focusing on pairwise gene similarity scores. GCN-based analyses have helped formalise network medicine as a field and deliver many successes in drug repositioning, drug-target discovery and drug-drug interactions. Identification of drugs that can predictably re-set the network in complex multi-component diseases has been a topic of interest for decades, in the search for novel drug targets.
Using artificial intelligence to predict drugs likely to get FDA approval
In this study, the researchers presented a different network-based approach for drug discovery, which uses AI to prioritise target identification and then guides its subsequent validation in network-rationalised preclinical mouse and patient-derived organoid models in a total of 4 steps. An overview of these 4 steps can be seen in figure 1, which has been taken from the article by Sahoo et al. (see the full citation at the end of this article).
The researchers used the disease model for inflammatory bowel disease, which is a complex multifaceted, autoimmune condition characterised by inflammation of the gut lining. Due to the effect the condition has on quality of life and the vast age groups affected by the condition, IBD is a priority area for drug discovery. Moreover, the condition is notoriously challenging to treat since no two patients behave similarly.
Target identification using artificial intelligence
The first step in the researcher’s method involves target identification. The team used an AI method previously developed by The Centre for Precision Computational System Network (PreCSN). This AI approach helps model a disease using a map of successive changes in gene expression at the onset and during the progression of a disease. This mapping approach goes beyond other methods as it uses mathematical precision to recognise and extract all possible fundamental rules of gene expression patterns, many of which are overlooked by current methodologies.
Moreover, the underlying algorithms ensure that the identified gene expression patterns are ‘invariant’, or never changing, regardless of the different disease cohorts.
Target validation to predict drugs likely to receive FDA approval
The final step involves target validation in preclinical models. This step was conducted in a first-of-its-kind Phase ‘0’ clinical trial, which uses a living biobank of organoids created from IBD patients at The HUMANOID Centre of Research Excellence (CoRE).
This phase ‘0’ trial concept was developed by the researchers because most drugs fail somewhere between phase I and III in clinical trials. Before proceeding to testing in patients in the clinic, these phase ‘0’ tests measure efficacy in human disease models, where ineffective compounds can be rejected early in the process, thereby having the potential to save millions of dollars.
Evaluation of the AI approach for predicting drugs likely to receive FDA approval
During evaluation of their approach, the researchers discovered two major surprises. The first of which is that despite the cells used to produce the organoids being away from the immune cells in the gut wall, these organoids from IBD patients showed the tell-tale features of a leaky gut with broken cell borders. Secondly, the drugs identified by the AI model were not only able to repair the broken barriers, but also protect them against the onslaught of pathogenic bacteria that the team added to the gut model. These findings suggest that the identified drugs could work in both acute flares as well as for maintenance therapy for preventing flares.
The researchers also found that the computational approach had a high level of accuracy across diverse cohorts of IBD patients, and were able to use their Phase ‘0’ approach to develop a first-in-class therapy to restore and protect the leaky gut barrier in IBD.
Therefore, based on the success of their approach, the lead researcher believes that their method could be used to help clinicians understand how diseases progress, assess a drug’s potential benefits and strategize how to use a combination of therapies when current treatment plans are failing.
Next steps
Using AI, the researchers created an approach that is able to accurately identify disease targets and predict drugs that are likely to receive FDA approval, therefore overcoming current hurdles in the drug discovery pipeline. The authors of this study hope to go on to test whether the drugs that pass the human phase ‘0’ trials can pass phase III trials in the clinic. The team also want to assess 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.
Sahoo, D., Swanson, L., Sayed, I.M. et al. Artificial intelligence guided discovery of a barrier-protective therapy in inflammatory bowel disease. Nat Commun 12, 4246 (2021). https://doi.org/10.1038/s41467-021-24470-5
Image credit: makyzz – FreePik
