Michael A. White

Hardcoding biological priors directly into the tokenizer via statistically calibrated transcription-factor identities beats standard genomic tokenization on 54 of 55 downstream tasks while unlocking direct mechanistic interpretability.