Protein function prediction is an essential problem in the field of computational biology and has a significant influence on many applications like drug discovery and antibody design. A long-lasting challenge for this problem is the data imbalance issue in existing protein function data, known as the long-tail problem: a small set of function classes are associated with most of the annotated proteins, while the majority of classes only have very few labeled proteins. This leads to the biased performance of current machine learning-based protein function prediction models, which focus more on optimizing prediction accuracy for well-represented function classes while performing poorly for underrepresented classes. In this work, we developed MSRep, a deep learning-based framework for protein function annotation aimed to mitigate the data imbalance issue. The inspiration for MSRep comes from an interesting phenomenon, called neural collapse (NC), observed in high-accuracy deep classification models. MSRep leveraged an NC-inspired loss to refine a pretrained protein language model embedding into a new space with NC properties, where both the well- and under-represented classes are equally represented in the embedding space. Evaluations across four protein function annotation tasks suggested that MSRep provides accurate predictions for both well- and underrepresented classes, outperforming several state-of-the-art annotation tools.