GraSP Gene Targets to Hierarchically Infer Sub-Classes with CuttleNet

This paper introduces a machine learning approach, GraSP, for retinal cell classification that addresses key challenges in spatial biology, alongside a novel neural network architecture, CuttleNet, tailored for class and subclass inference with incomplete datasets. We propose an innovative, unbiased gene selection method that utilizes simple neural networks for each target cell subclass, such that Gradient Selected Predictors (GraSP) corresponding to gene importance are found for each. This approach significantly outperforms traditional machine learning techniques and expert-selected gene targets, reducing the necessary genes for classification from over 18k to 300 within the murine retina. Such reduction is crucial for advancing spatial biology, particularly in mapping retinal cell subclasses. Furthermore, our hierarchical architecture inspired by the organization of the cephalopod nervous system, CuttleNet, adeptly handles the pervasive issue of missing data in disjointed single-cell RNA sequencing datasets. CuttleNet operates by first classifying cell classes using consistently measured genes, then dynamically routing to subclass-specific subnetworks that leverage all available data for subclass classification. CuttleNet establishes a new standard in handling systematically missing data, offering substantial improvements over existing models in our targeted application.