Machine Learning at GSK with Kim Branson - #536

Literature Mining

GSK leverages AI to mine scientific literature, extracting valuable insights using advanced NLP techniques. Kim Branson explains that their team employs BERT-based models to perform entity and relationship extraction from vast sources like PubMed and bioRxiv. This process involves creating semantic triples, which are stored in a massive 500 billion node graph, enabling efficient data retrieval and analysis.

We mine all that out. We run all this sort of thing over the literature, we pull out all those semantic triples, and we stick that into a really big store. So we have a graph that's like 500 billion nodes, right? It's huge.

--- Kim Branson

This approach allows GSK to utilize existing scientific knowledge efficiently, avoiding the need to start research from scratch 1.

   

Knowledge Graphs

Knowledge graphs play a crucial role in GSK's research, enabling the management and utilization of scientific knowledge. Kim Branson highlights how these graphs allow researchers to extract subgraphs focused on specific interests like genes and diseases, facilitating link prediction and hypothesis testing. By integrating historical data and experimental results, these graphs become a dynamic reference tool for biologists and researchers.

You can import that knowledge as well into the knowledge graph. So it becomes a growing reference space to use.

--- Kim Branson

This system enhances sample efficiency and supports the development of more accurate machine learning models by focusing on unknowns rather than relearning known information 2 3.