In this paper, we present StructTokenBench, an integrated framework for evaluating protein structure tokenization methods that segment protein 3D structures into discrete or continuous representations. Unlike existing benchmarks, we comprehensively evaluate the quality and efficiency of tokenizers by focusing on fine-grained local substructures. The evaluation results show that no single model has an advantage in all benchmarking aspects, which leads to the discovery of low codebook utilization. In response, we develop a strategy called AminoAseed to improve tokenizer utilization and quality by improving codebook gradient updates and optimally balancing codebook size and dimensionality. AminoAseed achieves an average performance improvement of 6.31% on 24 supervised learning tasks compared to the ESM3 model, with sensitivity and utilization increased by 12.83% and 124.03%, respectively. The source code and model weights are available on Github.
