Joint protein-DNA language modeling enables prediction and engineering of DNA recognition specificity

Protein-DNA Language Models (PDLMs) improved protein–DNA pair classification and enabled zero-shot prediction of mutation-linked binding specificity changes, including variants with higher editing activity.

• Why it matters: Protein–DNA recognition depends on coupled protein and DNA sequences, but existing language models that learn this cross-molecular dependence during pretraining remain limited.
• What they did: PDLMs were built by continued pretraining from ESM2 using masked language modeling on paired protein–DNA sequences with a shared encoder and domain-aware masking that increases masking probability within DNA-binding domains.
• The result: PDLMs outperformed independent-encoder baselines on pair classification, predicted mutation-associated binding changes, recovered known specificity shifts via zero-shot scoring, and supported cross-study base editor activity prediction and motif prediction from protein sequences after fine-tuning.

AI summary of the abstract · may contain errors · not medical advice · read the paper.