# Masked language modeling

PDLM concatenates a protein sequence and its DNA partner into one input and applies masked language modeling with a shared Transformer, so a DNA sequence is scored in the context of the protein paired with it. On a held-out test set of 2,558 protein–DNA pairs, PDLM-150M-domain reached an AUROC of 0.940 and an AUPRC of 0.903 for pair classification, against 0.823 and 0.683 for the strongest independent dual-encoder baseline, ESM2-150M plus DNABERT-2. Zero-shot conditional scoring recovered the known direction in which the MutH E91A/F94A variant extends preference to GATA, GATG and GATT, and 6 of the 82 variants prioritized on that basis showed significantly higher GACA editing than the parent.

PDLM concatenates a protein sequence and its DNA partner into one input and applies masked language modeling with a shared Transformer, so a DNA sequence is scored in the context of the protein paired with it. On a held-out test set of 2,558 protein–DNA pairs, PDLM-150M-domain reached an AUROC of 0.940 and an AUPRC of 0.903 for pair classification, against 0.823 and 0.683 for the strongest independent dual-encoder baseline, ESM2-150M plus DNABERT-2. Zero-shot conditional scoring recovered the known direction in which the MutH E91A/F94A variant extends preference to GATA, GATG and GATT, and 6 of the 82 variants prioritized on that basis showed significantly higher GACA editing than the parent.