Applying Deep Learning to Biomarker Development & Drug Discovery

While there is still skepticism with regard to the use of deep learning in biomarker development and drug discovery using blood biochemistry and transcriptomic data, there are multiple applications that are already effective.

Earlier this year, Alex Zhavoronkov, CEO of Insilico Medicine, spoke at the RE•WORK Deep Learning in Healthcare Summit, presenting a range of deep learned biomarkers of ageing and deep learned predictors of biological age. The bioinformatics firm revealed that they have figured out how to teach artificial intelligence (AI) to predict the therapeutic use of new drugs before they’re even tested.

The predictors are trained on hundreds of thousands of human biochemistry and cell count samples linked to chronological age, gender and health status. Transcriptomic and signalomic ageing markers and predictors of chronological and biological age and cross-species comparison were also discussed. Insilico believe AI will transform biomarker development and drug discovery much sooner than most pharmaceutical companies and regulators expect. I asked Alex a few questions to learn more.

What areas of healthcare have the biggest potential for disruption by AI?
In my opinion, the frontiers with most potential for disruption by AI are drug discovery and biomarker development. There is a massive amount of data linked to a massive number of chemical entities, including transcriptional response, CRISPR/CAS9, toxicity, and adverse effects, coupled with other in vitro, in vivo and clinical data in public repositories and available through research collaborations. While highly heterogeneous, this data can already be used to develop accurate deep-learned tissue-specific and liquid biopsy biomarkers of multiple diseases and to predict the therapeutic use of thousands of chemical entities. When combined with human domain expertise in drug discovery, rapid and cost-effective validation, and intellectual property management, deep-learned markers and in silico drug efficacy predictors can uberize the pharmaceutical industry. AI is already being applied in personalized and experimental medicine to predict the most effective treatments for rare diseases. Several countries previously considered to be developing countries are preparing to launch programs to use AI to streamline drug discovery and development, so soon we may see new players emerging in some of the most unexpected places and geographies. AI will level the ground in drug discovery and allow for new entrants and new geographies to enter the market.

What are the main risks associated with applying deep learning methods into healthcare and medicine?
Since millions of people are dying of aging and age-related diseases every year and there are virtually no cures for chronic conditions, the greatest risk is in delaying the advancement of deep learning and other AI techniques in healthcare. It is time for some of the thought leaders in artificial intelligence including social networks, IT companies and banks to prioritize increasing healthy productive longevity of their customer base over more immediate profit-generating alternatives like entertainment. For many of these companies, a focus on longevity research could provide a major competitive advantage. That being said, considering the high level of regulation and bureaucracy in medical practice, there may be risks associated with AI projects lacking domain expertise and resulting in negative publicity leading to increased regulation. Investors betting on AI projects in healthcare must ensure that their portfolio companies have a trail of research publications in the field and have experience working with the pharmaceutical companies and avoid excessive capitalizations and Theranos-like experiences. On the other hand, regulation may be beneficial in areas like insurance and human resources, where AI with domain expertise may lead to significant inequality and exclusion.

What developments can we expect to see in deep learning in the next 5 years?
Deep learning is likely to be commoditized within two years to the extent that a high school student will be able to replicate the work of some of the leading teams, and we will see more integration of transfer learning, reinforcement learning, memory, attention and variable action-value management. But when it comes to practical applications, in my opinion, deep learning itself has a limited number of applications and it needs to be combined with other machine learning methods and clear domain expertise. Systems utilizing deep reinforcement learning with domain expertise in drug discovery may yield unprecedented breakthroughs. Many pharmaceutical companies are starting to make baby steps in utilizing deep reinforcement learning for drug discovery. At Insilico we are working on a system to discover drug candidates and repurposing candidates that are most efficacious and least toxic for specific diseases and looking for new approaches to treat individual patients with rare diseases. Some companies will go beyond drugs and identify nutritional and lifestyle interventions that are likely to benefit individual patients. And insurance companies are likely to use deep learning to accurately predict various events, including mortality. 

© RD MEDIA. Alex Zhavoronkov on stage at the Deep Learning in Healthcare Summit, 7-8 April 2016. 

How will hospital systems change with the application of AI?
Nowadays it is difficult to forecast more than five years into the future in many areas of science and technology, but considering how regulated the healthcare industry is, many of the advances in AI will not significantly transform healthcare in that time frame. Most likely we are going to see mobile consumer devices being integrated into medical practice with AI providing early diagnosis, lifestyle recommendations, nutritional adjustments, etc to consumers and patient monitoring and warning notices to physicians. One of the first areas to go through a transformation will be medical imaging, where vast amount of data is available to build intelligent systems that may go beyond decision support and diagnostics and into treatment. We are likely to see fully-automated Cyberknife-like systems that will only require doctor’s supervision and approval for treatment.

What impact will AI have on aging research?
Aging is one of the most complex and multifactorial processes killing millions every year and causing more pain and suffering than any other disease. And while our understanding of many of these processes has improved dramatically over the past two decades, there are many blank spots that need to be filled with human experimental data to build a framework for system-wide interventions. We are likely to see deep-learned multi-factorial biomarkers of aging trained on large time-series data sets. We are presenting one of these at RE-WORK. Another area full of low-hanging fruits is drug repurposing, where deep learning is already helping find existing drugs that are reasonably safe to use with alternative therapeutic uses. These drugs may be tested on humans using deep-learned biomarkers. We are likely to see some of these studies within a year or two. The area of regenerative medicine will also benefit from AI. We already have deep-learned systems that can accurately classify various cell and tissue types, perform quality control and “understand” embryonic development. These systems will help identify new factors to direct horizontal and vertical cellular reprogramming both in vitro and in vivo, provide new tools for tissue and organ engineering, 3D bioprinting, in vivo regeneration, and cell therapy, and help discover new drugs against fibrosis and other diseases. And finally, AI will help us link many of the aging processes on a cross-species level. We are already training some of the markers on human data and testing them on mouse and even fly data and vice-versa. Or training the DNNs on data from various species to understand the common features and their importance. In my opinion, many of the animal studies performed today are rather useless and countless numbers of mice are being sacrificed every day without generating meaningful data that can be extrapolated to human-level research. AI-driven cross-species analysis will help us tailor these experiments to be more relevant to humans and reduce animal experimentation to a bare minimum.

What area of deep learning advancements excites you most?
I am mostly excited by transfer learning techniques applied to genomic and transcriptomic data in the context of biomarker development, drug discovery and repurposing. We can really minimize animal experimentation and make research more relevant to humans in early pre-clinical R&D stages.

The Deep Learning in Healthcare Summit will be next held in London on 23-24 February 2017, and in Boston on 11-12 May 2017. We will also be holding a Women in Machine Intelligence & Healthcare Dinner in London on 12 October 2016 to celebrate the women advancing this field.

Future AI & Machine Learning focused events of 2016 include:
  • Machine Intelligence Summit, Berlin, 29-30 June 2016
  • Deep Learning in Finance Summit, London, 23 September 2016
  • Deep Learning Summit, London, 22-23 September 2016
  • Deep Learning Summit, Singapore, 20-21 October 2016
  • View all upcoming RE•WORK summits here.

    Deep Learning Machine Learning Clinical Patient Care Genomics Deep Learning Summit Healthcare Deep Learning in Healthcare Summit MedTech


      As Featured In


      Partners & Attendees

      Rbc research.001
      Mit tech review.001
      Kd nuggets.001
      Maluuba 2017.001
      This website uses cookies to ensure you get the best experience. Learn more