One language model reads protein and DNA together: untrained scoring recovers MutH substrate shifts and flags 6 of 82 variants for better editing

Protein–DNA recognition specificity can now be learned directly from paired sequences: one shared encoder places amino acids and nucleotides in the same context, and a DNA sequence's likelihood depends on the protein paired with it. Earlier protein language models such as ESM-DBP, CLAPE and PDNAPred strengthened only the protein side, and cross-molecular foundation models LucaOne and OmniBioTE did not explicitly model cognate molecular interactions during pretraining. On the 2,558-pair held-out test set, PDLM-150M-domain reached an AUROC of 0.940 and an AUPRC of 0.903 for pair classification, versus 0.823 and 0.683 for the strongest independent-encoder baseline, ESM2-150M plus DNABERT-2; the advantage appeared on mismatched, shuffled and random DNA negatives, while models performed similarly on non-DNA-binding-protein negatives.

The method writes protein and DNA as one sequence into a single Transformer, gives each modality its own vocabulary and output head, and masks residues inside DNA-binding domains more often. Pretraining continues from ESM2, inheriting the protein vocabulary and embeddings, while DNA vocabulary and embeddings are randomly initialized and masked language modeling acts on both modalities. The pretraining corpus is 51,329 protein–DNA pairs curated and deduplicated from BioLiP2, AptaDB, REBASE and CIS-BP, split by protein sequence into training, validation and test sets; domain-aware masking uses HMMER and Pfam annotations, with masking probability 0.30 inside domains, 0.10 outside and 0.15 for unannotated proteins, while DNA masking stays at 20%.

The same conditional scoring gives the direction of mutation effects on recognition specificity without task-specific training, and narrows the candidate set before experiments. The mitoBE system lacks a practical high-throughput framework for screening MutH variants, since each variant must be built separately and evaluated with site-specific cellular editing assays. PDLM-35M-domain assigned positive pseudo-log-likelihood changes for E91A/F94A relative to wild-type MutH on GATA, GATG and GATT while GATC stayed nearly unchanged; scoring 4,123 single substitutions in the E91A/F94A background across 16 GANN contexts then prioritized 82 variants, and 6 of them showed significantly higher mean GACA editing than E91A/F94A.

The demonstrated scope is computational evaluation plus cellular editing validation for MutH: pair classification, ΔΔG regression, cross-dataset TadA activity prediction and off-target scoring, transcription-factor motif prediction, and variant screening at the GACA site in the mitoBE system. The intended readers are researchers working on protein–DNA recognition modeling and base-editor protein engineering.

The authors note that PDLMs are limited by the scale and diversity of available paired protein–DNA data, which are far fewer than protein-only sequences and unevenly distributed across protein families and assay types, so generalization to distant homologues and unseen protein families remains to be established. The benefits of model scale and domain-aware masking vary with task and model size, with no single universal configuration, and the model did not recover the purine-preference shift observed for the FOXA1 R219S mutation.