Learning on Structure-Activity Relationship from Heterogenous Chemical Compound Databases

2020 Fiscal Year Final Research Report

• Project/Area Number: 17K00320
• Research Category: Grant-in-Aid for Scientific Research (C)
• Allocation Type: Multi-year Fund
• Section: 一般
• Research Field: Intelligent informatics
• Research Institution: Kwansei Gakuin University
• Principal Investigator: Inokuchi Akihiro 関西学院大学, 理工学部, 教授 (70452456)
• Co-Investigator(Kenkyū-buntansha): 田中 大輔 関西学院大学, 理工学部, 准教授 (60589399)
• Project Period (FY): 2017-04-01 – 2021-03-31
• Keywords: 機械学習 / データベース / グラフ / 構造活性相関

Outline of Final Research Achievements

In this work, we investigated various approaches for learning on structure-activity relationship from heterogenous chemical compound databases. The first outcome is establishing a methodology for efficiently searching graphs contained in a query graph from a database consisting of a huge amount of graphs. Our approach is based on the prefix tree of graph codes that represent the graphs in the database. By using the prefix tree as an index, we simultaneously compute the subgraph isomorphism problem (which is known to be NP-complete) between the query graph and multiple graphs in the database. The second outcome is reducing the over smoothing phenomenon in Graph Convolution Networks in the deep learning research domain. In our approach, we combined Graph Convolution networks with the dense connection, and increased a certain percentage of prediction accuracies for various benchmark datasets compared with some conventional methods.

Free Research Field

人工知能

Academic Significance and Societal Importance of the Research Achievements

深層学習法は，現在，盛んに研究が実施されている研究分野である．その中で，様々なタイプのデータが解析されているが，グラフは非常に高い表現力を有していることが知られ，それに対するデータ解析手法や学習手法は非常に重要である．我々の成果は，データベースと機械学習の基礎研究分野への貢献であり，それらを発展させていくことで，実用化を目指せるものである．実用化の具体例の1つは，創薬化学分野である．