Abstract: Deep learning-based methods to analyze imaging data can revolutionize high-content imaging (HCI). These methods enable the training of models that can detect subtle phenotypic changes caused by chemical or genetic perturbations. This allows researchers to explore gene and drug similarities cost-effectively and at scale, aiding in gene function analysis, drug mechanism identification, and other screening endpoints. However, these AI models can be sensitive to unexpected factors, leading to inaccurate predictions unrelated to biology. This challenge is further exacerbated by the “black-box” nature of these models which makes them difficult to interrogate and shrouds the underlying signals that underpin the model predictions.
In this talk, we will share best practices and techniques developed by Spring Science over the past 5 years to optimize AI-based analytics for HCI data. We'll discuss current AI approaches for classification and similarity scoring, recommend ideal experiment and plate designs, provide image quality control tips, and explore data normalization and alignment options.