Trainable subnetworks to disentangle structure knowledge in pretrained protein language models - View it on GitHub