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

Protein-DNA recognition is essential for various biological processes such as transcriptional regulation, DNA replication, chromatin remodeling, and genome editing. Typically, proteins and DNA sequences interact with each other, but current language models that learn this cross-molecular dependence are limited. Researchers have developed Protein-DNA Language Models (PDLMs) to address this issue. PDLMs are interaction-aware language models that jointly represent protein and DNA sequences.

Building upon ESM2, PDLMs were pre-trained using masked language modeling on paired protein-DNA sequences. The models utilize a shared encoder to place amino acids and nucleotides in the same molecular context, enabling conditional sequence scoring and transferable pair representations. To enhance learning of interaction-relevant regions, the researchers introduced a domain-aware masking strategy that increased masking probability within DNA-binding domains.

When evaluated on downstream tasks, PDLMs outperformed independent-encoder baselines in protein-DNA pair classification. The models were also able to predict mutation-associated binding changes, recover known changes in DNA-recognition specificity through zero-shot scoring, and prioritize variants with higher editing activity. Furthermore, PDLMs learned representations that supported the prediction of base editor activity across different studies.

After fine-tuning on transcription factor binding data, PDLMs were able to predict DNA motifs directly from protein sequences. This framework for interaction-aware biomolecular language modeling has potential applications in protein engineering and genome editing.