Two Neurons Is All You Need — A Case Study on Interpretability in Protein Models