FURI | Summer 2021

Inference of Brain Morphology from Spatial Gene Expression

The Allen Brain Atlas offers an atlas of gene expression across the whole mouse brain, but spatial transcriptomic datasets are beyond direct interpretation. Reducing dimensionality with principal component analysis (PCA) and partitioning with k-means clustering yields labels that reflect classical anatomy, albeit imperfectly and at the cost of all gene information. Factorization machines (FM) potentially offer an improvement by using supervised learning to retain input data relationships while producing a classified output. The researchers compare the accuracy of this technique with both the previous unsupervised PCA strategy, and a logistic regression for a supervised baseline, and show favorable results.