The aim of this project is to use interpretable neural network architectures to segment single cell RNA-seq data.
Our way to design interpretable neural networks lies in the use of genomic ontology terms (see the Gene Ontology project) to expertly assign biological meaning to the neural network activations on its first hidden layer. This type of architectures has already been proposed in [1].
Whether we want to segment our data in a supervised or unsupervised way, two types of architectures can be considered: a MLP classifier with a final softmax activation or an autoencoder to do dimension reduction. In both tasks, the way to make these neural networks interpretable is the same: define expertly the relations between the input layer (genomic features) and the first hidden layer (GO terms). The architectures of these neural networks are then no longer dense.
[1] Peng, J., Wang, X. Shang, X. Combining gene ontology with deep neural networks to enhance the clustering of single cell RNA-Seq data. BMC Bioinformatics 20, 284 (2019). https://doi.org/10.1186/s12859- 019-2769-6