Fig. 6

MIOSTONE learns meaningful and discriminative representations. MIOSTONE’s internal neuron representations of samples are projected onto a two-dimensional Principal Component space and evaluated their efficacy in distinguishing between different disease subtypes. MIOSTONE’s family level representations are compared to the last-layer latent representations of other tree-aware methods, with the exception of Ph-CNN on the HMP2 dataset, as Ph-CNN is not scalable for that dataset. MIOSTONE’s representations show significantly improved separation between disease subtypes, indicating that the model’s internal representations effectively capture diverse disease-specific signatures. This separation is quantitatively assessed using the silhouette value