Evaluation of Coding Schemes for Transformer-based Gene Sequence Modeling

Abstract

Currently, many studies view DNA sequences as a special type of language and utilize Transformers to model them. These studies use fixed-length k-mer segmentation and BPE subword tokenization but lack a systematic evaluation to determine which is superior. We compare k-mer segmentation with k=1,3,4,5,6, a 4,096-token BPE vocabulary, and three positional encoding methods-sinusoidal, AliBi, and RoPE. Each configuration is trained from scratch in 3, 6, 12, and 24-layer Transformer encoders and evaluated on GUE benchmark dataset. In general, BPE delivers higher and more stable performance across tasks by compressing frequent motifs into variable-length tokens, reducing sequence length, and improving model generalization. RoPE excels at capturing periodic motifs and extrapolating to long sequences, while AliBi also performs well on tasks driven by local dependencies. In terms of depth, we observe significant gains when increasing layers from 3 to 12, with only marginal improvements or slight overfitting at 24 layers. This study provides practical guidance for designing tokenization and positional encoding in DNA Transformer models.