BashTheBug.net: Citizen Scientists consistently measure the growth of M.tuberculosis on a 96-well plate
2018-04-18T13:31:56Z (GMT) by
The Comprehensive Resistance Prediction for Tuberculosis International Consortium (CRyPTIC) was launched in March 2016 and is collecting >30,000 clinical M. tuberculosis samples worldwide. Each sample will have its genome sequenced, and its susceptibility to a panel of 14 anti-TB drugs determined using a 96-well microtitre plate. Since the goal of the project is to identify new genes that are involved in resistance and infer the effect of individual genetic mutations, reducing errors and biases in the phenotypic DST dataset is crucial. Each plate will, however, be primarily read by a scientist in each laboratory, potentially introducing biases. To both test this and to generate an unbiased dataset, we launched a Zooniverse citizen science website in April 2017 inviting members of the public to help us to analyse how well each sample of M. tuberculosis (MTB) grows in the presence of drug.
The photograph of each plate is first segmented and images containing the two control wells and a single strip of wells of one of the 14 anti-TB drugs are created by our AMyGDA software. These images are then shown to 15 different volunteers on the Zooniverse website, allowing a consensus to be obtained. Each lab in the CRyPTIC project was first required to grow and classify 31 known MTB strains, including replicates: the volunteers have classified all ~1,500 images from this initial dataset.
As of Apr 2018, 10,332 volunteers have done 743,668 classifications. The minimum inhibitory concentrations (MICs) returned by the BashTheBug volunteers have an essential agreement ~90% when compared to our reference dataset, comprising measurements made by experts using a Thermo Fischer Vizion. We shall also describe cases where the volunteers have identified anomalies missed by the experts, the features and types of growth that the volunteers struggle with and discuss the potential for training a machine learning algorithm on this dataset.
Consensus-based measurements of MICs, as done here by the BashTheBug volunteers, can be less biased and as accurate as those produced by individual experts.