(1021-B) AutoHCS: AI-based scoring of high-content screens results in morphological clustering that predicts mechanisms of action.
Monday, February 5, 2024
2:00 PM – 3:00 PM EST
Location: Exhibit Halls AB
Abstract: Modern drug development increasingly depends on high-content compound screens where automation is the key to rapid, impactful discoveries. AutoHCSTM is a multifaceted AI-based analysis tool developed by ViQi Inc. that automatically detects and scores phenotypic responses to compounds in high-content screens. Because the system does not depend on segmentation, it works non-parametrically with multichannel fluorescence, a combination of fluorescence and brightfield, or brightfield alone. The only inputs to the analysis are images from any automated plate imager and a plate map specifying concentrations, replicates, and controls. A few core AutoHCS analytical tools are: 1) comparing compounds of interest against negative and positive controls or target phenotypes 2) evaluating the dose response of compounds of interest and 3) computing morphological clusters across many different compounds of interest. Importantly, AutoHCS AIs can conduct each of these analyses independently or in combination. For example, comparing the dose response of a compound of interest against positive controls will determine which dose, if any, is most similar to a known target phenotype. Whereas, investigating dose-dependent responses independently of controls permits the discovery of novel phenotypes. AutoHCS can also be used to evaluate the ability of compounds to reverse a background phenotype, such as in screens for neuroprotectant drugs, anti-inflammatory drugs, and antivirals. AutoHCS entirely determines its training parameters using the experimental controls rather than user input, which eliminates subjective criteria selection that may bias phenotype scoring. It also allows multiplex scoring of screens both for a positive target phenotype and against negative phenotypes such as cellular toxicity. AutoHCS is cloud-based, so there is no software or specialized computing hardware to install locally. Accordingly, AutoHCS is scalable to millions of images and works regardless of contrast method, cell type, or cellular responses generated.
A key function of AutoHCS is its ability to morphologically cluster compounds according to their induced phenotype. To further investigate the significance of these clusters, we used AutoHCS to analyze a subset of the JUMP dataset, a public HCS dataset with many compounds and replicates. We converted our morphological clusters into gene lists using standard databases and conducted pathway analysis on these gene lists. We found that gene lists from our morphological clusters resulted in a high degree of overlap in database-queried mechanisms of action when compared with randomly generated clusters. As we build our knowledge base using tools like the JUMP dataset, this bioinformatics-based approach to cluster validation will not only allow us to make predictions about novel compounds, but may provide deeper insight into other AutoHCS analyses. With all its capabilities, AutoHCS harnesses the pattern recognition abilities of modern AIs to precisely score and phenotypically profile high-content screens in an entirely automated, objective manner.