2018.06.13 Release

ExaWizards, an AI start-up, has successfully developed AI core technology and is implementing full-fledged measures for practical application to significantly improve the productivity of drug discovery

〜 Developing AI for the activity prediction, visualization and formation of compounds using a Graph Convolutional Network ~

ExaWizards Inc. (Minato-ku, Tokyo, Representative Director & President: Ko Ishiyama, hereafter “ExaWizards”) has jointly developed AI technology for learning and predicting the properties of small-molecular compounds, with Okuno Lab., Graduate School of Medicine, Kyoto University (hereafter “Kyoto University”) and Drug Development Data Intelligence Platform Group, Medical Sciences Innovation Hub Program, Institute of Physical and Chemical Research (hereafter “Institute of Physical and Chemical Research”). ExaWizards will work on developing an AI model that can be used in each of the medical, drug discovery and sales processes.

1.Background and purpose of the development

Pharmaceutical companies have been facing problems such as the increasing cost of new drug development, shrinkage of new drug pipelines, and reductions in drug prices. Accordingly, as they seek to significantly improve their R&D efficiency, pharmaceutical companies are looking to use AI and have high expectations. In response, ExaWizards has established AI technology covering compound activity prediction, visualization and compound formation, and has developed AI to support the search for and optimization of lead compounds, which conventionally takes a long time and high cost. Also, this AI can be used not only for analyzing compounds but also can be easily applied to a gene network and paper citation network. By developing AI technology for drug discovery with high applicability in Japan, we are helping to shorten the R&D period and reduce costs.

2.Overview of the joint development

 Based on a learning model using Deep Learning: “Graph Convolutional Network (GCN),” ExaWizards, Kyoto University and the Institute of Physical and Chemical Research have jointly developed this AI as described below**

(1)Model for predicting mutual action between compounds and proteins

This model learns protein activity based on the compound structure and enables prediction. This increases the efficiency of screening candidate compounds in drug discovery. In benchmark testing, the model has achieved an accuracy equivalent to that of DeepChem***, an existing drug discovery library, while maintaining high applicability.

(2)Visualization of compound structures supporting compound design

Although it has been difficult to present the reasons for the predicted activity of compounds on proteins using conventional Deep Learning technology, this model enables effective parts of activity expression in a compound and ineffective parts to be “visualized”.

(3) Compound formation model for proposing new compounds

A wide variety and number of suitable compounds for drug discovery among existing compounds can be proposed. This model replaces the compound design proposal process conventionally implemented by researchers based on their own experience and makes the process more efficient.

3.Future prospects

In the future, we will expand and enhance the functions for assisting drug discovery by using GCN with high applicability and through collaboration with pharmaceutical companies, and will develop services that contribute to improved productivity of all business processes of pharmaceutical companies, including a service to assist sales activities in drug discovery. Through these activities, we aim to contribute to progress and innovation across the medical product industry, and thus extend healthy life and improve QOL as part of our companywide mission of solving social issues.

*     GCN is a new technology for learning the “linkages and relations” represented by compounds and for extracting information. Thanks to this characteristic, it has high applicability regardless of the joint authorship relation in gene network analysis and papers, analysis of the relevance of researchers, etc.

**   This is developed as part of the activities by LINC (Life Intelligence Consortium).

***  DeepChem is a Python library developed by Stanford University researchers. As it has an activity prediction function with GCN and other appropriate functions, it has been attracting global attention.

 

[Profile of ExaWizards Inc.]

Company name: ExaWizards Inc.

Address: 5F, Sumitomo Hamamatsucho Bldg., Hamamatsu-cho 1-18-16, Minato-ku, Tokyo

Founded: February 2016

Representative: Representative Director & President: Ko Ishiyama

Business: Development of AI-enabled services for industrial innovation and solving social problems

URL: https://exawizards.com/

<Contact> 
ExaWizards Inc. Public Relation Division
E-mail:publicrelations@exwzd.com