A Pretraining-Retraining Strategy of Deep Learning Improves Cell-Specific Enhancer Predictions.

Area of Science:

• Genomics
• Computational Biology
• Molecular Biology

Background:

• Cis-regulatory elements (CREs), particularly enhancers, are crucial for gene regulation.
• Predicting enhancer locations genome-wide (discriminative enhancer predictions, DEP) is important, but predicting their cell- or tissue-specific activity (tissue-specific enhancer predictions, TSEP) is more critical.
• Existing deep learning models excel at DEP but struggle with TSEP due to limited tissue-specific training data.

Purpose of the Study:

• To develop a novel training strategy for accurate tissue-specific enhancer predictions (TSEP).
• To address the challenge of limited training data for specific cell or tissue types in enhancer prediction.
• To improve upon existing methods for identifying functional enhancers in a context-dependent manner.

Main Methods:

• Developed a "pretraining-retraining strategy" (PRS) for TSEP.
• Stage 1 (Pretraining): Trained a deep learning model on comprehensive enhancer data for general enhancer prediction (DEP).
• Stage 2 (Retraining): Fine-tuned the pretrained model using limited tissue-specific enhancer samples for TSEP.

Main Results:

• PRS demonstrated strong performance in DEP, achieving an AUC of 0.922 and GM of 0.696 on the FANTOM5 dataset.
• The retraining stage required only ~20 additional epochs to achieve high performance for TSEP across 23 tissues/cell lines.
• PRS significantly outperformed gkm-SVM (mean GM 0.806 vs. 0.528) and other state-of-the-art methods (DEEP, BiRen) for TSEP.

Conclusions:

• The pretraining-retraining strategy (PRS) is a reliable and effective method for tissue-specific enhancer predictions (TSEP).
• PRS successfully applies transfer learning principles to overcome data limitations in predicting cell- and tissue-specific enhancer function.
• This approach offers a significant advancement for understanding gene regulation in specific biological contexts.