Deep Learning-Based Localization for RNA Spatial Transcriptomics

Nanostring spatial molecular imaging platform (SMI) leverages single molecule hybridization-based sequencing chemistry to spatially profile expression of 1000+ genes at single cell and sub cellular resolution across tissue samples. The current RNA localization algorithm relies on commonly used methodologies where potential target locations are segmented from raw images and fit with a function that approximates a point spread function. While efficient and fast, this method will fail in the multi-emitter regime. Due to the desire to increase plexity and/or use reduced codesets, it is critical to evaluate other localization schemes that are designed for densely packed data. DECODE is a state-of-the-art deep learning-based approach that enables high precision localization microscopy in dense images. This student team worked to deliver modified DECODE modules that are tailored towards high density SMI samples. The student team also worked to develop models with adequate consideration towards compensating raw data that has highly variable background and large signal dynamic ranges and outperforms standard localization methods in SMI data by an order of magnitude with respect to density, defined as having a Jaccard index >~70%. The student team also worked to provide a Python module to enable 3D clustering of localizations to find the best fit Z-plane to allow for code matching via Nanostring’s matching module.

Latest news