Sequence Feature Representation Based on k-mer Tokenization: A Comparative Study of Methods and the Impact of Data Scale on Model Performance

This paper systematically investigates the impact of different k-mer tokenization strategies on sequence feature representation and model performance. Using the TCR-epitope affinity prediction task as an example, we integrated data from public databases (VDJdb, IEDB Receptor, PIRD, McPAS) to construct a large-scale dataset and randomly extracted <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{1 0, 0 0 0}$</tex> samples (small-scale), 200,000 samples (medium-scale), and used the full dataset (422,164 samples, large-scale). Under a unified simple convolutional neural network, four tokenization strategies (1-mer, 2-mer, 3-mer, and 4-mer) were compared. Experimental results show that smaller <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{k}$</tex> values converge faster on small-scale datasets, while the 3-mer strategy, which captures more contextual information, achieves the highest AUC on the large-scale dataset. Although 4-mer tokenization captures longer local features, its training time increases significantly with only limited performance gains. These findings provide data-driven guidance for selecting appropriate <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{k}$</tex>-mer tokenization strategies in bioinformatics tasks.