"
Team:Berkeley/Project/Automation
From 2012.igem.org
We can accommodate libraries of up to one million members through our MiCodes design scheme. With a data set this large, we needed to develop methods to make MiCodes a high-throughput screening technique. We realized that two aspects of microscopy needed to be automatable: image acquisition and image processing.
How the cell segmentation was done in Matlab.
How we created pipelines in Cell Profiler to detect organelles.
We applied machine learning concepts in order to tackle automation of MiCode identification. The following illustrates general workflow principles for machine learning:
1. Design of features associated with organelles through image analysis software.
2. Iterative Machine learning -> computer presents small sample of cells to biologist, who batch scores and returns categorized cells.
3. Computer scores phenotypes among the sample and refines rules based on biologist scores.
4. When accuracy of identification through the rules is sufficient, classification is automatically run on all cells in experiment.
Because we had to begin by constructing our MiCodes phenotypic library, we did not have sufficient time to build a comprehensive training set to process and refine the rules and algorithms for scoring phenotypes. As such, we developed a substantial part of the workflow, the feature design, to demonstrate the feasibility of automated MiCode identification.
We imaged a small sample size of about 100 cells with varying phenotypes, and determined feature sets associated with each of the four organelles. We used CellProfiler, an open-source software designed for quantitative analysis of cellular phenotypes. The software is available at their website. We used this software to build pipelines that demonstrate the feasibility of applying machine learning for identify cellular phenotypes.
