Deciphering enhancer sequence using thermodynamics-based models and convolutional neural networks.

Area of Science:

• Genomics
• Computational Biology
• Developmental Biology

Background:

• Deciphering the sequence-function relationship in enhancers is vital for understanding gene regulation and non-coding genetic variants.
• Existing quantitative models for predicting enhancer function lack systematic comparison, hindering mechanistic insights.

Purpose of the Study:

• To systematically compare various biophysical and machine learning models for predicting enhancer function.
• To develop and validate a novel deep learning model, CoNSEPT, for unbiased prediction of enhancer grammar and function.
• To infer underlying regulatory mechanisms of enhancer activity from model performance.

Main Methods:

• Utilized a dataset of Drosophila neuroectodermal enhancers to test multiple quantitative models.
• Performed rigorous comparisons of thermodynamics-based models with varying activation, repression, and cooperativity mechanisms.
• Developed and applied a convolutional neural network (CNN) model, CoNSEPT, for predicting enhancer function across conditions.

Main Results:

• Identified evidence supporting established mechanisms like cooperative activation and short-range repression.
• Found that different models, while fitting data similarly, offer varying degrees of mechanistic interpretability.
• CoNSEPT demonstrated effectiveness as a general-purpose tool for predicting enhancer function and suggesting regulatory mechanisms.

Conclusions:

• Systematic model comparison is essential for advancing the understanding of enhancer function and regulatory mechanisms.
• The developed CoNSEPT model provides a powerful, interpretable tool for predicting enhancer activity and uncovering sequence-based regulatory logic.
• This work highlights the utility of sophisticated computational models in deciphering complex biological systems and guiding future research.