A Comparison of Tokenization Impact in Attention Based and State Space Genomic Language Models

Genomic language models have recently emerged as powerful tools to decode and interpret genetic sequences. Existing genomic language models have utilized various tokenization methods including character tokenization, overlapping and non-overlapping k-mer tokenization, and byte-pair encoding, a method widely used in natural language models. Genomic models have significant differences from natural language and protein language models because of their low character variability, complex and overlapping features, and inconsistent directionality. These differences make sub-word tokenization in genomic language models significantly different from traditional language models. This study explores the impact of tokenization in attention-based and state-space genomic language models by evaluating their downstream performance on various fine-tuning tasks. We propose new definitions for fertility, the token per word ratio, in the context of genomic language models, and introduce tokenization parity, which measures how consistently a tokenizer parses homologous sequences. We also perform an ablation study on the state-space model, Mamba, to evaluate the impact of character-based tokenization compared to byte-pair encoding. Our results indicate that the choice of tokenizer significantly impacts model performance and that when experiments control for input sequence length, character tokenization is the best choice in state-space models for all evaluated task categories except epigenetic mark prediction.