Enhancer recognition and prediction during spermatogenesis based on deep convolutional neural networks.

Area of Science:

• Genetics
• Molecular Biology
• Bioinformatics

Background:

• Enhancers are critical for regulating gene expression, particularly during spermatogenesis.
• While ChIP-Seq technologies illuminate enhancer-DNA and histone modification relationships, predicting enhancers solely on sequence and modification patterns remains challenging.

Purpose of the Study:

• To develop and validate a convolutional neural network (CNN) model for predicting enhancers involved in spermatogenesis.
• To assess the model's performance, reliability, and generalizability across different cell types within spermatogenesis.

Main Methods:

• A CNN model was trained using experimentally verified enhancers (P300 locus) and non-enhancers (promoters).
• Transfer learning was employed for rapid model adaptation to various spermatogenesis cell types.
• Model interpretability was achieved through convolution layer visualization and alignment with the JASPAR database.

Main Results:

• The CNN model demonstrated high accuracy in predicting spermatogenesis enhancers, aligning with key transcription factors.
• The model exhibited superior generalization ability compared to the gkmSVM algorithm.
• The developed CNN model is computationally efficient with a simple structure, avoiding overfitting.

Conclusions:

• The proposed CNN model offers a reliable and efficient method for enhancer prediction in spermatogenesis.
• The model's ability to generalize and its alignment with known transcription factors underscore its biological relevance.
• An enhancer recognition website was created to facilitate further research and collaboration.