Decoding promoter activity from DNA sequence using pre-trained language models.

Area of Science:

• Genomics
• Computational Biology
• Molecular Biology

Background:

• Transcriptional regulation relies heavily on promoter architecture.
• Predicting promoter activity solely from DNA sequence is a significant challenge in genomics.

Purpose of the Study:

• To evaluate the capability of transformer-based DNA language models in deciphering regulatory logic within Drosophila core promoters.
• To assess the model's predictive performance and interpretability using known promoter elements.

Main Methods:

• Fine-tuning the DNABERT-2 DNA language model on a synthetic Drosophila core promoter dataset.
• Utilizing luciferase reporter assays in S2 cells for promoter activity measurement.
• Applying SHapley Additive exPlanations (SHAP) for model interpretation.

Main Results:

• The DNABERT-2 model achieved high accuracy (R² ≈ 0.91) when predicting promoter activity with distinct training and test datasets.
• The model demonstrated robust performance (R² ≈ 0.64) even when test sequences were entirely new.
• SHAP analysis identified known promoter elements (INR, TATA, DPE, etc.) as key predictive features, with position-dependent effects.
• Integrating hormonal and nucleosomal context improved the unified modeling of sequence and biological information.
• The model showed promoter-specific generalization capabilities and captured partial in vivo activity trends in independent embryo data.

Conclusions:

• DNA language models can learn interpretable sequence rules for promoter activity from controlled experimental data.
• Accurate in vivo promoter activity prediction necessitates the inclusion of broader regulatory contexts beyond DNA sequence alone.