Why biotech needs knowledge graph convolutional networks for discovery

Description

Robustly making sense of complexity in domains such as finance, marketing, biotech, and cybersecurity remains an unsolved problem. In order to combat complexity, you need a method to maintain data integrity of highly interconnected data at scale. You need to build AIs that can ingest and augment the insights of that data.

James Fletcher dives into the Grakn knowledge graph, which provides the means to keep interrelated data organized at scale and answer questions of that data. However, learners are still needed to interrogate this data.

KGCNs have a clear goal: provide a tool (under the Apache license) that can embed an understanding of a user’s knowledge base. Grakn’s approach leverages the Grakn knowledge graph, automated deductive reasoning, and deep learning with TensorFlow to build a multifaceted solution. A KGCN is a learner that interfaces to a knowledge graph. This network concept can be specialized for applications such as drug discovery, drug repositioning, ontology merging, recommender systems, structural classification, and link prediction. They’re agnostic to the structure and type of the data.

Statistical approaches alone are not sufficient to tackle the complexity of AI challenges today. Being smarter with the data you already have is critical to achieving machine understanding of any complex domain. To build smarter, you must merge and combine different factions of AI, including those less fashionable than deep learning. As such, KGCNs demonstrate the usefulness of combining a connectionist deep learning approach with a symbolic approach.

Grakn implements symbolic reasoning at the database level and is a true mathematical knowledge base—a typed hypergraph as a totally new system for knowledge representation. Logical deduction over the stored knowledge is fully automated via the definition of logical rules. All of this is available with one-click production deployment on the Google Cloud Platform (GCP) and AWS.

New research that has been fully implemented in Python is available from GitHub under the Apache license. It’s ready to apply reasoning and deep learning over any dataset out of the box. James’s demo looks at a specific use case of link prediction in biotechnology. See the GitHub for more information.