OR WAIT null SECS
Existing software tools cannot take into account the complexity of disease.
The pharmaceutical industry is currently challenged to accelerate the development of new drugs while reducing the cost of these efforts. Drug repositioning is an effective approach for achieving these goals. Evaluating existing compounds that have been proven safe-whether approved, still in clinical trials, or previously found to be ineffective for the diseases they were originally targeting-takes less time and money. The challenge is to identify the right existing compounds to investigate for a given disease.
Drug repositioning today is currently pursued using one of five key approaches, according to Aris Persidis, co-founder and president of Biovista, which is developing a pipeline of repositioned drug candidates in neurodegenerative diseases, epilepsy, oncology, and orphan diseases. Phenotypic screening is the one method that involves laboratory testing. In this approach, a drug substance is tested in parallel against a small set (10–50) of animal models or cell lines. “This approach is neither comprehensive nor systematic by any means, but if one is lucky and gets a hit, then a potential new use of a drug in a specific new disease has been identified with some initial experimental proof,” Persidis says.
To achieve systematic or comprehensive repositioning, computer-based approaches must be used. There is no single preferred digital tool or technology for accomplishing drug repositioning because different tools are optimally suited for different strategies, according to Hermann Mucke, CEO of H.M. Pharma Consultancy, a firm that provides drug development services, including support for drug repositioning activities. “Many pharmaceutical companies use their own in-house algorithmic solutions or leverage technology platforms developed by specialists in repositioning, such as Biovista, Excelra, etc.,” he notes.
In addition to in-silico modeling and target docking using algorithmical solutions beyond medicinal chemistry, Mucke points to data mining of side effects and drug interactions as a way to find valuable clues for APIs formulated into marketed drugs. “Crucial information can also be extracted from the peer-reviewed and patent literature by creative pharmacologists who can think out of the box and ‘connect the dots’,” he adds.
The other four approaches identified by Persidis overlap somewhat with these strategies and include literature mining, pathway mining, adverse-event matching, and gene-regulation mining. “Pathway mining involves the use of pathway keywords to search the literature and other gene/protein pathway databases where a specific pathway is linked to different diseases, and then match drugs known to hit that pathway across the different diseases,” he explains.
Similarly, in adverse-event matching, lists of adverse events are used to mine the labels of drugs for the same adverse events. “If different drugs that work in different diseases have the same adverse events, then maybe they can work in the same diseases,” Persidis observes. Gene regulation mining uses data from microarrays to find genes that may be up/down-regulated in a disease microarray, and then the literature is searched to find drugs known to have the opposite effect.
Unfortunately, there are few data sources that are specifically tuned to the needs of drug repurposers, according to Mucke. In addition, analysis tools are not yet sufficiently advanced to assist without extensive training in pathway and systems medicine.
The existing approaches have all yielded some successes and failures, adds Persidis. They fall into the “quick/easy/cheap category,” however, and are limited in their applicability. “To do repositioning is easy, to do it well is hard. The key limitation of these existing computational approaches is that they are focused on reducing repositioning to one parameter (gene up/down, common side effects, etc). As a result, they do not take into account the complexity of diseases and their mechanisms of action, or the complexity of patient populations,” he said.
Going forward, both Persidis and Mucke believe that the application of artificial intelligence (AI) to drug repositioning will take the field much further than it can go today. “AI is emerging as the best tool to do repositioning well. New AI-driven technology shows the best promise, because AI is able to integrate and use many different types of data much more effectively than before,” Persidis explains.
Adds Mucke: “Advanced AI technologies that can look for connections they have not been specifically trained to identify will revolutionize drug repositioning within the next few years, taking us beyond mere expert systems.”
It is not just existing compounds that have potential for drug repositioning; their derivatives can also be explored for therapeutic activity. “Once you have a repositioning candidate compound identified for your disease target (or vice versa), you can let loose all the medicinal chemistry and pharmacology tools that are out there,” says Mucke. Derivatives can be generated using known quantitative structure-activity relationship approaches that have been widely employed in drug discovery for many years, according to Persidis.
Of course, the parent compound must be identified first, which is a repositioning exercise. “The real challenge is not the derivatives, but the starting points. Once you have a drug, then deriving different versions can be done, but you can’t derive a different version unless you know where to start from. The difficulty is finding a good starting point,” says Persidis.
H.M. Pharma Consultancy maintains the Discontinued Drug and Candidate Database (DDCD), a comprehensive machine-readable database of information on compounds that might be suitable for drug repositioning. The DDCD is regularly updated with information on approved drugs and drug candidates that might be available for repositioning efforts taken from peer-reviewed papers, patent documents, and chemistry, pharmacology, and development information in various public sources, according to Mucke.
The company uses the DDCD as an internal resource for drug development projects and for repositioning projects that it is completing as part of the European Union Research and Innovation program, Horizon 2020. “The DDCD has been constructed over the past 20 years and goes beyond any single public source in defining the pool of compounds that are potentially available for drug repositioning,” Mucke comments.
Over the past 10 years, Biovista has been developing an AI solution called Project Prodigy. “A major challenge in drug repositioning is to balance the insights into a new use with equivalent insights into potential side effects,” Persidis says. “Biovista’s AI program is one of very few equally able to identify both potential new uses and the specific subpopulations to include or avoid, based on a full understanding of the benefit/risk balance,” he notes.
The Project Prodigy AI is also one of few systems that does not limit itself to machine learning (ML), says Persidis. That is important because ML only affords ‘hits’ that are basically already known from training sets, even if hard to find. “Biovista’s Project Prodigy AI is capable of building entirely new clinical scenarios that have never been seen before and can be evaluated as truly novel. No training sets are needed or used. This ability provides a key advantage over typical ML-based approaches,” he states.
Project Prodigy, according to Persidis, has already been used in a number of projects. Internal efforts at Biovista have led to repositioned drugs with animal model/cell line validation and issued or granted patents in multiple sclerosis, epilepsy, anti-glomerular basement membrane disease, and some rare diseases such as Friedreich’s ataxia and Leber’s hereditary optic neuropathy. The company has also used the AI system in collaborations with major pharmaceutical companies, patient advocacy groups, and FDA.
Developing a drug is a long, expensive process. Repositioning is the most efficient way to get to a valid starting point. “Drug repositioning should be a standard operating procedure at any stage of development, especially early on. Its ‘right way’ systematic use also reduces development and investor risk across different phases of the portfolio,” said Persidis.
While none of the current digital tools available for repositioning are capable of providing anything beyond assistance for pharmacologists who know how to use them and who can interpret their results properly, according to Mucke, this situation will change soon with the application of continually improved AI technology.
“AI has already shown its promise, with actual results achieved much more efficiently than other approaches. Now, it is beginning to be understood as one of the most powerful ways forward for drug development and healthcare cost containment,” Persidis concludes.
Vol. 42, No. 9
When referring to this article, please cite it as C. Challener, “Can Artificial Intelligence Take the Next Step for Drug Repositioning?" Pharmaceutical Technology 42 (9) 2018.