Pattern Recognition for Protein Crystallisation Strategies

Protein crystallography is of fundamental importance to structural biology and drug design, but producing crystals is a major bottleneck in the pipeline. Crystallisation is a trial and error process and scientists do not currently have the tools to correlate the hidden relationships between the properties of a protein that are known or can be measured and the experimental conditions in which they will crystallise. In collaboration with AstraZeneca, this project aims to use machine learning and statistical pattern recognition to reveal such relationships, thereby improving future experiments and accelerating the drug discovery process.
The results from many experiments, both successful and unsuccessful, will be required but information on failed experiments is rarely recorded by crystallographers and, once conditions producing crystals have been identified, alternative successful conditions are often not considered. However, the use of robotics to perform crystallisation trials not only increases throughput and reduces costs, but allows the results to be recorded by imaging systems. The MARCO (MAchine Recognition of Crystallization Outcomes) project showed that automated image analysis can be used to classify experimental results, having a 94% correct classification rate [1]. Work on a custom image classifier using convolutional neural networks is already underway at AstraZeneca. In this project, the student will incorporate Active Learning [2] into the training process to improve performance and use additional information from UV images to achieve higher accuracy. As the main aim of the project, the student will then use the annotated crystallisation images to connect compounds or proteins with experimental conditions using novel graph-based networks [3, 4].