From Trainable Negative Depth to Edge Heterophily in Graphs