AI Exposes Patterns For Faster Drug Discovery
Contributed Commentary by Enakshi Singh
November 29, 2017 | It now costs an estimated $2.6 billion, and takes more than 10 years to develop and test a new drug, and bring it to market. Facing these steep odds, the pharmaceutical industry is constantly looking for new efficiencies that will save time, money, research, and resources. One promising approach is to use artificial intelligence (AI) and machine learning to do much of the mundane, data-driven work.
Though the results are just starting to emerge, AI is showing positive traction in drug development. AI platforms are helping pharma researchers identify new drug targets. AI is also showing the potential to decrease the time required to screen novel molecules. Moreover, it can help determine which drugs are most effective in treating specific diseases. Machine learning in particular can help pharma recover the value of repurposed drugs by finding new therapeutic applications for them.
Today, drug discovery is a trial-and-error process that eats up enormous amounts of research time. AI can significantly narrow the focus of researchers by rapidly assimilating and analyzing the information in public and proprietary databases. Publicly available sources of de-identified patient data include clinical trials, electronic health records, medical images, and genomic profiles, all of which can be useful for drug development. AI learning systems can also quickly scan the millions of new research papers, as well as the massive libraries of compounds and test results maintained by pharmaceutical companies.
Using algorithms, AI systems can discover patterns in these data pools. AI rapidly learns how to hone in on key information and develop hypotheses; it can refine its answers over time. With the help of machine learning, including deep learning, the computer can figure out an unforeseen problem, rather than having the answers programmed into it.
Researchers can use AI tools to sift through the ongoing avalanche of scientific papers and find correlations in the trial data. These applications can compile a list of known facts about particular compounds, including how they may affect patient health and medical conditions. Based on these facts, AI systems can make connections that generate a range of hypotheses, using criteria set by the research team.
Testing Drug Candidates
After researchers have narrowed the focus of their discovery efforts, they must screen large numbers of molecules to find promising drug candidates. They then have to carry out a multitude of tests in the hope of finding a winner. All of this is an expensive exercise that takes a huge amount of time. But what if AI could do the tedious molecular screening?
Insilico Medicine, a Baltimore biotech company that develops innovative solutions for cancer research, has begun to test this concept. The company is using a relatively new deep learning technique known as a generative adversarial network (GAN). A GAN includes a pair of competing neural networks, including one that generates hypothetical data and another that tries to distinguish the fake data from real data. Insilico is the first company to use a GAN in cancer drug discovery.
Researchers in Insilico’s Pharma AI division have reported that its GAN has used historical biological and chemical data to “imagine” 69 new molecules with the potential to fight cancer. Insilico also has used AI to help predict the therapeutic use of drugs. Its AI platform was fed experimental data on 678 drugs and the effects they had on gene expression in three types of human cells. The AI platform developed an ability to classify drugs into therapeutic use categories, achieving 54.6% accuracy in identifying one out of 12 of a particular drug’s therapeutic applications.
While this may seem like a modest achievement, it actually represents a major step forward for researchers who otherwise would have to make these predictions through countless hours of experimentation. Moreover, some of the AI platform’s “wrong” answers were helpful, pointing to secondary uses for drugs that researchers had not considered.
These secondary uses can be vital to pharmaceutical companies that wish to recover some of their investment in the 90% of potential medicines that never make it to market. Some of these manufacturers are turning to AI to find new ways to repurpose drugs that have already gone through the early stages of testing. If they can find new indications for these medicines, the process of getting them approved will be faster and less expensive than if they had to start testing a new compound from scratch.
With the capability to look at 14 trillion data points in a single tissue sample, AI can help eliminate many of the years lost in trial-and-error drug development. By tapping into untapped data pools, AI has the potential to shave millions of dollars—or trillions in total-- off the cost of drug development.
Enakshi Singh is senior product specialist at SAP Health. She can be reached at Enakshi.email@example.com.