GraSP Gene Targets to Hierarchically Infer Sub-Classes with CuttleNet

Abstract

This paper presents a machine learning approach for retinal cell classification, overcoming key constraints in spatial sequencing. We introduce a novel neural network training strategy that effectively classifies cells using just 225 genes, enabling cutting-edge techniques to provide spatial insights into retinal function. Inspired by biological perception, we also develop CuttleNet, a specialized architecture mirroring coarse-to-fine processing. Through hierarchical routing and subclass-specific subnetworks, CuttleNet sets a new benchmark in our limited application space. This is an important advance in spatial biology enabled by machine learning.