Fig. 2
Construction of the interpretable deep-learning age-predictor (DL-XAI clock). (a) Detailed diagram of the MGRL framework. A meta-cell’s input graph with a meta-cell’s expression values overlayed on this network defines the input to a graph convolutional neural network called DeeperGCN, to capture joint expression and topological features through two graph convolution layers that enable nodes to aggregate and update information from their neighbors. A mean pooling strategy is then applied to merge individual node representations into a unified graph-level feature vector as final topological information. Meanwhile, the MLP extracts node features from the expression data without consideration of network topology. The framework then fuses the joint topological-expression embeddings (topology view) with those derived from only expression (feature view) This fused embedding is finally fed into a fully-connected neural network to predict chronological age of the meta-cell. (b) Illustration of how PGExplainer is applied to our proposed MGRL method to extract the predictive subnetworks. The process begins by computing latent variables Ω from the edges of the original graph, which inform the edge distributions that underpin the explanatory framework. A random sampled graph is constructed by selecting the top-ranked edges based on Ω, and this graph is subsequently input into a trained MGRL model to generate the prediction (\(\:{y}_{s}\)). Finally, MLP parameters are optimized by minimizing the Mean Squared Error (MSE) between the original prediction (\(\:{y}_{o}\)) and the updated prediction (\(\:{y}_{s}\)).
