The cost of drug discovery and subsequent development is a massive challenge in the pharmaceutical industry. A typical drug can cost upwards of $2.5 billion and a decade or more to identify and test a new drug candidate .
In this article, we will be taking a look at the spiraling costs of drug development and how artificial intelligence may be the solution to reducing costs and speeding up drug discovery.
Increasing costs in drug development drive the need for change
These costs have been increasing steadily over the years, and pharmaceutical manufacturers are constantly seeking ways to improve efficiency, save time and money, and speed up research progress.
Automation in the lab is one example; tasks that were traditionally carried out by technicians can now be done by machines. Increasingly sophisticated assays to detect new drug candidates have also helped to slash development time. Now, a new ally has arrived to aid drug development – artificial intelligence – and a powerful ally it is.
Artificial intelligence (AI) has already made great progress in aiding drug discovery, and a number of pharmaceutical companies, such as GlaxoSmithKline, are seeking aid from this new technology. AI can learn and discover new things rapidly, and these discoveries can be turned into actionable knowledge that can be applied to the drug development process.
It’s all about the data
One of the primary reasons that AI has such great potential in drug development is that there is a huge amount of health data available right now in the public health system. Clinical trials, health records, genetic profiles, preclinical studies and a wealth of other information is readily available to use for drug development.
AI is very well suited to dealing with laborious tasks that use a great deal of data and are traditionally very time-consuming for people to perform. Pharmaceutical companies have vast drug libraries containing the details of many compounds and test results. AI is able to rapidly assess this information and, in conjunction with researchers, can help to facilitate drug discovery and new insights faster.
AI uses algorithms that can discern patterns and trends in the vast pools of data it is given, and this allows it to sift through that information, identify useful data and even develop hypotheses like a researcher would, only far faster. Machine learning allows the AI to be trained so that it is able to work out the solution to a problem instead of relying on the answers being given to it by programmers.
A tide of information
Even in the field of aging research alone, there are dozens of new research publications arriving every week, and it can be a struggle for researchers to keep up with current progress. The amount of new information arriving in the wider health sector is truly mind-boggling, and discoveries made in one field of research may have implications in other fields. However, these discoveries can be missed by researchers focused on one field, such as biomedical gerontology and the study of aging.
This is where the potential of AI excels; a research team has little chance of being able to process the huge amount of scientific data arriving on a daily basis, but an AI is ideally suited to handling this task. AI is able to analyze billions of research papers, allowing it to correlate and connect new information with existing data and to link direct relationships in those data sources to create a list of established facts.
Guided by researchers, the AI can then use these facts to make connections that lead to the creation of many potential hypotheses. These hypotheses can then be tested by researchers, thus speeding up the pace of progress.
Screening for drugs
The majority of pharmaceutical companies screen large numbers of compounds and molecules searching for potential drug candidates. These candidates are then further tested in the hope that they will pan out as the bases of new drugs. The problem with traditional screening is that this takes a great deal of time and, as a result, is very expensive. AI could do this laborious screening with less time and cost.
Insilico Medicine is working on a new deep learning technique: the generative adversarial network (GAN). GAN uses two competing neural network models to create new data that is indistinguishable from real data. In other words, it is able to “imagine” new types of molecules with the potential to combat cancer.
The generative model tries to create output that “looks like” real data, and the discriminative model takes input from both the generative model and real data and tries to distinguish between the two sources. Generative models have been used to create things like images, text and even speech, but its use in drug discovery is a first.
Insilico reports that it has given its neural network historical biological and chemical data and that it has been able to “imagine” 69 new molecules that may help find solutions to cancer.
The same company also used AI to help predict the therapeutic use of drugs. The company gave the system experimental data for 678 drugs and their effect on gene expression profiles in three human cell lines. The AI system was able to classify the drugs into therapeutic use categories with 54.6% accuracy in identifying one out of twelve of the drugs’ therapeutic uses.
Normally, researchers would have had to make the same predictions with hours of testing and experimentation, so this represents a huge time-saving. Even the wrong answers the AI gave were helpful, as they showed secondary uses for drugs that had not been previously considered by researchers.
The field of AI is still in its early days, but the initial results are already starting to give us an idea of the great potential that AI has. Reducing screening times, aiding new drug candidates and finding the most effective drugs for specific diseases at a speed that humans cannot achieve is compelling, and we believe that AI will increasingly become part of the medical landscape.
 DiMasi, J. A., Grabowski, H. G., & Hansen, R. W. (2016). Innovation in the pharmaceutical industry: new estimates of R&D costs. Journal of health economics, 47, 20-33.