Deep Learning Techniques Speed Drug Discovery

(Gerd Altmann, Pixabay)

3 Sept. 2019. Researchers developed and demonstrated techniques with deep learning to discover new small molecule drugs ready for validation and lab animal tests in 21 days. A team from Insilico Medicine in Hong Kong and WuXi AppTec in Shanghai, with colleagues from University of Toronto, described their techniques in yesterday’s issue of the journal Nature Biotechnology (paid subscription required).

The team of computer scientists and computational chemists led by Alex Zhavoronkov, CEO of Insilico Medicine, are seeking faster and more reliable methods to discover new treatments for disease with a good chance of success. In a paper published in October 2018, Zhavoronkov highlighted a role for artificial intelligence to reverse the declining productivity of the pharmaceutical industry, with high failure rates of new drugs, and costs for developing new therapies costing in the billions of dollars.

Zhavoronkov and co-author Alán Aspuru-Guzik, a computational chemistry professor at University of Toronto, and colleagues earlier developed a deep learning application called a Reinforced Adversarial Neural Computer model for drug discovery. Deep learning is a form of machine learning and artificial intelligence that makes it possible for systems to discern underlying patterns in relationships, and build those relationships into knowledge bases applied to a number of disciplines.

The Reinforced Adversarial Neural Computer adapts a form of deep learning called a generative adversarial network using two sets of algorithms that test each other while learning the underlying patterns. Those repeated rounds test the authenticity of new data generated by the model, eventually producing optimized data points. Generative adversarial networks are being applied to a range of intellectual endeavors, including art and creative writing, with Aspuru-Guzik studying this form of deep learning in chemistry.

In their new paper Zhavoronkov, Aspuru-Guzik, and associates refine these deep-learning algorithms to develop a generative tensorial reinforcement learning, or Gentrl, model for drug discovery. Gentrl learns the molecular structures of chemicals, and encodes the relationships between those structures and their properties. Real-world data have missing values, which are recorded by the model and form a baseline for generating new molecules. Gentrl then freezes each instance of missing data, and uses generative algorithms to test chemical properties to find those with the highest reward.

The researchers tested Gentrl to find new small-molecule drugs that block discoidin domain receptor 1, an enzyme associated with cirrhosis, or scar tissue in the liver, and other diseases. The team reports Gentrl identified potent compounds against the enzyme target in 21 days. From that first round, the field was narrowed to two compounds that succeed on collagen cells in lab cultures. The researchers then tested the top candidate in lab mice, which shows favorable chemical activity against discoidin domain receptor 1.

Insilico Medicine makes code for the Gentrl model available through Github. The company develops artificial intelligence applications for extending human life and drug discovery.

More from Science & Enterprise:

• Aging Therapy Network Adds $100M to Venture Round
• Hispanic Neuro Disease Landscape Mapped in New Study
• Darpa Funds Early Disease Detection System
• A.I. Finding Cancer Treatments in Generic Drugs
• Sanofi, Alphabet Partner on A.I., Analytics

*     *     *