There is a growing need for fast and accurate methods for testing developmental neurotoxicity across industrial, pharmaceutical, and environmental chemical exposures. Current approaches, such as in vivo animal studies and assays of primary cell cultures, suffer from challenges related to time, cost, and applicability to human physiology. Thus, we previously developed an alternative using machine learning to predict developmental neurotoxicity with gene expression data collected from complex human 3D bio-engineered tissue models. The 3D model has advantages in its similarity to developing human neural structures, but the complexity of the model requires extensive expertise and effort to employ. By focusing only on the goal of constructing an assay of developmental neurotoxicity, we proposed that a simpler 2D tissue model may prove sufficient. We thus compared the accuracy of predictive models trained on data from a 2D tissue model with those trained on data from our more complex 3D tissue model, and found the accuracy of the 2D model to be better than the 3D model.
In order to develop these and new tissue models, however, we must first understand how to produce cells of many different types. This requires determining sets of transcription factors that allow us to reprogram cells from one type to another, which typically necessitates both extensive breadth and depth of biological knowledge. Furthermore, due to exponential growth in scientific literature, it is becoming continually more challenging for wet-lab scientists to keep up with current knowledge and to prioritize their experiments appropriately. To address this challenge, we developed a simple text mining approach to rank transcription factors by their association with a keyphrase representing a biological concept of interest. We evaluated our text mining approach on multiple historical discoveries in cell reprogramming and found that we could have identified the important factors years in advance of the landmark publications for each.
Brief Bio (adapted from http://pages.cs.wisc.edu/~finn/ )
Finn Kuusisto earned his BA in Computer Science and Classics from Gustavus Adolphus College in 2007 and his PhD in Computer Science in 2015 from the University of Wisconsin-Madison. He conducted his graduate research in machine learning in the Computer Sciences department with Dr. Shavlik, Dr. Page, and Dr. Burnside. His work focused primarily on breast cancer prediction applications related to overdiagnosis, estimation of individualized treatment effects, and precision medicine in general.. His current work is with the Regenerative Biology lab at the Morgridge Institute to build models from genetic expression data that can predict when compounds are toxic to developing neurological tissues.