Machine Learning Methods for Drug Evaluation and Treatment Assessment

Abstract

Drug preclinical test is a key step in evaluating the profile of drug treatment. Many drug tests have been designed for different diseases. For instance, researchers manually count the number of peristaltic waves of drosophila larvae to conduct the severity of amyotrophic lateral sclerosis (ALS). In other cases, pharmacologists have to count dead cells by visual scoring to assess the performance of chemotherapy treatment. Labeling the mitosis events is a time-consuming task, and thus are prohibitive for large scale drug screenings. Machine learning algorithms have allowed researchers to dramatically increase the throughput of analyzing a large amount of data. However, the current methods require massive ground truth annotations which is labor intensive in biomedical experiments. Approaches with few human interventions remain unexplored. This dissertation focuses on three tasks for drug evaluation and treatment assessment. First, we propose a machine learning method to evaluate the effectiveness of drug for ALS. This method leverages t-Distributed Stochastic Neighbor Embedding (tSNE) and statistical analysis to assess the locomotion behavior of drosophila larvae and compare the difference between groups with and without the testing drug. Second, we designed a first-of-the-kind weakly supervised deep neural network for dead cell detection and counting. Compared with many existing fully supervised approaches, our approach only requires image-level ground truth. We show classification performance compared to general purpose and cell classification networks, and report results for the image-level supervised counting task. Last but not least, we propose a sequence-level supervised neural networks model using convolutional long short-term memory (ConvLSTM) and convolutional layers to detect mitosis events at pixel-and-frame level. By using binary labels, the proposed network is able to localize the cell division spatially and temporally. We have evaluated our method with stem cell time-lapse images. With significantly less amount of ground truth in the training data, our method achieved competitive performance compared with the state-of-art fully supervised mitosis detection methods.