High Content Screening (HCS) has revolutionized the field of drug discovery by enabling the analysis of multiple parameters of cells in a single assay. HCS generates a vast amount of data that requires sophisticated analytical tools to extract meaningful information.
Using a standardized assay, such as cell painting, we generate datasets that:
1) can be compared to other cell painting assays and
2) are extremely dense with uninterpreted information, which makes it a great dataset for machine learning.
Machine learning and Artificial Intelligence (AI) is a crucial aspect of HCS as it enables us to extract more meaningful information from images, such as toxicity or bioactivity. In this blog post, we will discuss how we use AI in HCS and cell painting.
Cell segmentation is the process of defining the boundaries of individual cells within an image. Accurate cell segmentation is crucial for obtaining reliable data in HCS. Currently, the cells are segmented by simple methods, such as thresholds and watersheds. While this works mostly well, AI approaches consistently generate better results, leading to more accurate HCS assay results. At TK Analytics we use CellPose for better segmentation, which is an AI model specifically built for HCS data. It uses the U-Net convolutional neural network that has shown remarkable performance in segmenting cells from fluorescent images.
Phenotypic profiling is a powerful approach for studying cellular responses to perturbations. Phenotypic profiling involves measuring multiple parameters of cells, such as morphology, texture, and intensity. The original cell painting protocol uses CellProfiler for phenotypic profiling. Lately, several AI alternatives have been used and mostly show an improved performance. While CellProfiler extracts human-engineered features, AI tools learn on their own the best features to extract from the images. Such an approach consistently gives better results. At TK Analytics we use DeepProfiler to extract better features from cell painting.
Bioactivity prediction is a critical aspect of drug discovery. Cell-based assays are used to evaluate the biological activity of compounds. Typically, we would perform many different cell based assays to evaluate compounds activity, toxicity and other aspects. However, with AI we can significantly reduce the number of cell based assays by predicting the assay result from cell painting images. We can predict between 40 assays with AUROC>0.9 and up to 150 assays with AUROC>0.7.
Cell painting images can be also used to predict other compound bioactivity parameters, such as mechanism of action, toxicity and environmental effects. For toxicity prediction we also rely on compound structure predictions, which is thoroughly described in the article Toxicity prediction.
Data presentation and visualization are crucial for interpreting the results of HCS experiments. AI techniques are used to visualize large datasets, enabling researchers to extract meaningful information. Often we use t-SNE or UMAP dimensionality reduction techniques to visualize the phenotypic profiles of different cells. Both methods are effective at visualizing the similarities and differences between cells. Other AI tools for data visualization include Hierarchical Stochastic Neighbour Embedding (HSNE), which can visualize high-dimensional data, and X-Vis, an interactive visualization tool for exploring large datasets.
To illustrate the application of AI in HCS, consider a study that used cell painting to screen the effects of different compounds on cell morphology. The study used CellPose to segment cells accurately and efficiently. The phenotypic features were extracted with DeepProfiler. The study then used machine learning to classify the different cell types based on their phenotypic features. The study also used deep learning to predict the toxicity and bioactivity of different compounds based on their phenotypic features.
AI is an indispensable tool in HCS, enabling the rapid and accurate analysis of large amounts of data. AI techniques such as deep learning, machine learning, and dimensionality reduction have been used effectively in HCS. Cell painting is a powerful approach for studying cellular responses to perturbations, and our AI techniques have been used to analyze the vast amounts of data generated from cell painting experiments. The future of drug discovery lies in the integration of AI and HCS, enabling the development of new therapies for a wide range of diseases.