You can’t afford clinical trials? Emulate them with AI!
AI is providing a breakthrough in drug repurposing by emulating randomized clinical trials on the real-world patient records
Creation of new drugs is slow and expensive. Nowadays the early stages of the drug discovery pipeline, from identifying molecular target to predictions of toxicity and the ease of chemical syntheses, could be accelerated with computational techniques to such extent that they are no longer considered a bottleneck. However, the clinical trials remain very slow, very expensive and very tedious.
The stage of clinical trials is what hampers not only quick development of new medicines, but also repurposing of existing drugs for off-label usages. The later becomes more and more important because existing drugs have well-known safety profile, their adverse effects are already studied, the dosage is known and possible interference is mostly established. Thus using them to treat other diseases is often more promising than developing new drugs from scratch.
Currently identification of the off-label usage of existing drugs is pretty much random and is based on either anecdotal clinical evidence or theoretical analysis of molecular mechanisms of action. The clinicians may notice some positive or negative off-label effects in patients with co-morbidities directly or by the retrospective analysis of the patient records. The cases of unexpected drug interference may also give a clue to their off-label action. Careful analysis of molecular mechanisms of diseases may reveal similarities between two unrelated conditions and suggest the off-label usage of the drugs approved to only one of them. No need to say, that all these approaches are extremely ineffective and do not allow for any systematic search for drug repurposing.
Moreover, when promising off-label usage is identified for some drug, additional clinical trials are usually necessary to confirm the hypothesis and to get reliable information on efficacy and preferred dosage.
Ironically, modern medical databases often contain enough information not only to identify prospective off-label usages, but also to avoid or shorten additional clinical studies. Indeed, we now have diverse real-world data: patient electronic health records, patient surveys, insurance claims, bills and prescriptions for buying medicines, etc. This records are linked to the patient profiles, including basic demographics and the whole health history. All this enormous body of information could be analyzed to detect hidden correlations between on-label drug usage and their off-label influence on various co-morbidities. Clearly, this job is an ideal task for AI, which excels in finding hidden correlations and patterns.
In the recent work published in the Nature Machine Intelligence journal, researchers from USA attempted drug repurposing for a coronary artery disease (CAD)— one of the most common diseases, with millions of real-world patient records available.
An approach is to emulate randomized clinical trials for the repurposing of available on-market drugs by means of AI-assisted analysis of the real-world patient records.
The authors used impressive 107.5 million distinct depersonalized patients records containing healthcare claims information from employers, health plans and hospitals. Among them ~1.1 million CAD patients were identified.
Than, demographic characteristics (age and gender), diagnosis codes (~57000 distinct codes) and prescribed medicines were extracted. The drugs in such huge databases could be designated by brand or generic name or the national registration code. Additional round of pre-processing was needed to regularize this data and to get the dataset of 1,353 distinct drugs, which could be screened for repurposing.
The next crucial step was to identify clinical events, which are relevant for the coronary artery disease. In order to do this the standardized codes from International Classification of Diseases (ICD-9/10) were use by the experts to provide a set of definitions, which were than searched in the database. The same approach was used to compile the list of the most probable confounded variables, such as sex, age, co-morbidities, co-prescribed drugs, etc.
The deep learning model was based on recurrent LSTM neural network. The output was the treatment effect estimation — the measure of the treatment efficacy by particular drug on the level of population between treatment and control groups. The groups themselves where formed automatically from the general dataset, which is one of the most important innovations of this approach.
As a result the AI model have identified nine drugs, which could be used to treat a coronary artery disease. For them the estimate of the treatment effect is negative, which means that using them in patients with CAD is beneficial in comparison to not using on the population level. Three of these drugs (amlodipine, diltiazem and rosuvastatin) are already well known as approved CAD treatments, but the other six are repurposed drug candidates.
This work shows an unexpected power of AI-based drug repurposing methods, which allow to skip the clinical trials stage providing that enough real-world data is available. The analysis of the patients records and prescriptions on the very large datasets allows to emulate the randomized clinical trials on existing population. Definitely, this method would only work for common diseases, where the necessary amount of data is recorder, but even with this limitation it may revolutionize the field in the nearest future.
Receptor.AI is not standing away from the emerging trend of drug repurposing. We are developing our own AI-based module, which allows to search for off-label usages of existing drugs. This module would be integrated with our drug-target interaction and clinical trials prediction modules for identifying the best possible candidates for repurposing and off-label usage for a given target.
