ProtSAE: Disentangling and Interpreting Protein Language Models via Semantically-Guided Sparse Autoencoders | Synapse