Text this: Trainable subnetworks reveal insights into structure knowledge organization in protein language models.