Automated Image Analysis with Deep Learning to drive Cardiac Safety Profiling

Cardiotoxicity is a crucial consideration during early stages of drug development to de-risk new therapies and maximise patient safety. Current in vitro models are restrictive as they are either highly target-centric (e.g. hERG) or have limited translatability for structural cardiotoxicity due to simplified cell systems or analysis methods. In answer to this, a deep learning based image analysis approach (Genedata Imagence) was deployed on a 3D cardiac microtissue (CMT) high content imaging assay, with the aim of deriving increased mechanistic insight and increased sensitivity.

One of the main challenges here was to maintain tissue-level phenotypic information while achieving a high enough object count for robust classification at scale. A solution was found in analysing 50-100 tissue subregions from widefield or 2D projections of the CMT images. This enabled phenotypic classification on a tissue level while maintaining high enough throughput for production-level profiling. Applying this approach, a neural network was trained using a collection of known cardiotoxins in live CMTs to annotate phenotypes that represent a potential cardiac safety risk. Around a dozen distinct phenotypes were identified and trained to a high degree of confidence, based on a triple stain for mitochondria, ER and nuclei. To guide in the phenotype interpretation for a range of end-users, the most similar phenotypes were grouped to fall into 6 categories, but drill-down options were kept available for detailed analysis. In addition, the neural network was also able to identify and flag a broad spectrum of interference, including (cellular) debris and fluorescence artefacts.

From a test set of roughly 80 in-house compounds, phenotypic classification achieved excellent correlation with the legacy segmentation-based analysis approach. In addition, we observe several benefits of the deep learning approach: 1) Direct correlation of test compounds to known cardiotoxins that were used to train the network enables increased mechanistic insight; 2) Identification of novel phenotypes even while the assay is in production, allows for the continuous improvement of the neural network and phenotypic clustering of test compounds; 3) Increased assay robustness; 4) Time-savings through automated analysis of production-level screens and easy of development/transfer. Finally, early data suggests we observe higher assay sensitivity, which could enable detection of cardiotoxins that were previously missed.