Motif-Vocab: StatisticallyCalibrated Transcription-Factor-Identity Tokenization forGenomic Language Models

Abstract

Tokenization is a central design choice in genomic language models, yet most deoxyribonucleic acid (DNA) tokenizers use characters, fixed-length k-mers, or frequency-derived subwords without explicitly using prior information about the specificity of DNA-binding regulatory factors. We introduce Motif-Vocab, a biologically informed tokenizer that scans both DNA strands for statistically calibrated motif matches, emits transcription-factor (TF) identity tokens, and applies nucleotide, $k$-mer, or byte-pair encoding (BPE) to unmatched sequence. Motif-specific null distributions put position-weight matrices (PWMs) of different lengths and degeneracy on a common significance scale; deterministic overlap rules make the representation reproducible. In controlled Bidirectional Encoder Representations from Transformers (BERT) pretraining on two billion base pairs, real motif libraries outperform randomized-motif controls on 54 of 55 in-scope downstream tasks. On a motif-disjoint recognition task derived from DART-Eval Task 2, TF-specific tokens improve macro-F1 by 0.040 over a position-matched generic motif token and by 0.033 over a matched no-motif tokenizer (95\% bootstrap confidence interval: 0.027--0.038). Motif tokens also receive stronger attribution and produce larger occlusion effects than shuffled controls. Dense no-motif tokenizers remain strong general-purpose baselines, including a near-tie on the five-task BERT-base panel. Thus, Motif-Vocab is not a universal accuracy replacement; it is a targeted, interpretable inductive bias for motif-sensitive genomic modeling.