Event details
Talk Title: Classifying mutations in cancer samples using the Darwin machine learning tool (runner up in the Centre for Machine Intelligence Hackathon)
Speaker: Clare Horscroft
Abstract: Identifying mutations in tumour tissue is a complex process due to the underlying biological processes these tissues undergo to become malignant. Normally the identification is made with the aid of matched normal-tissue samples; however, normal tissue is not always available. Lack of paired normal tissue may leave high number false positives that are difficult to filter out without validation or time consuming manual curation. Our hypothesis is that machine learning (ML) can make this process more efficient by classifying the somatic mutations, from germline mutations and noise. Darwin is an automated machine learning tool that allows the implementation and optimisation of multiple ML methods simultaneously and efficiently, to identify which would be the most appropriate method to fine-tune. We used Darwin to train models and implemented 5-fold cross-validation. Using Darwin’s neuroevolution method the model was trained for different amounts of time, at intervals of 5 minutes, to determine the optimal training time. The model that was the most successful was a random forest classifier trained for 25 minutes. This random forest model achieved an average precision of 90.30 and average recall of 92.94. This indicates that the model would perform well in a clinical setting, saving money, time and valuable resources.
In addition, Yang Hu will be leading the following paper: “Automated acquisition of explainable knowledge from unannotated histopathology images”. Please read the paper and join in our discussion (
https://www.nature.com/articles/s41467-019-13647-8 ).