At Relation, we are pioneering the application of graph machine learning (ML) and active learning - we call it ActiveGraph ML - to fundamental problems in drug discovery.

Machine learning has seen great strides in the past handful of years, with graph neural networks (GNNs) arriving as the "third wave" of modern ML after computer vision and natural language. The maturation of graph ML has been marked by advances including graph convolutional networks (Kipf and Welling, 2016), message passing schema (e.g. GraphSage (Hamilton and Ying, 2017)), graph attention (Veličković et al., 2017) and the close relationship between graph neural networks and another breakthrough technology, Transformers. There have been further advances in causal structure learning, including such algorithms as NOTEARS (Zheng et al., 2018).

Relation’s platform is built around these, and adds unique insights, into a causal understanding of complex biological graphs comprised of single cell and molecular data. This gives an ML-native understanding of complex biology whilst maintaining interpretability – the ability to integrate, parse and draw inferences in a scalable manner, capturing relationships with omics at single-cell resolution.

In conjunction with graph machine learning, we use active learning, a branch of machine learning that focuses on allowing the models to guide which data we acquire.

Together, these technologies allow for the full integration of model optimisation with a wet lab and human scientists - a “Lab-in-the-Loop”. This enables our ML platform to sit within a tight experimental feedback system. Using our wet-lab techniques we can accurately predict gene expression, and changes in gene expression, directly from DNA sequence and predict the effects of gene perturbation on cellular phenotype - accelerating and radically improving new drug target identification and validation.