A Contrastive Learning Pre-Training Method for Motif Occupancy Identification.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Motif occupancy identification predicts DNA motif binding to transcription factors.
• Current sequence-based methods lack biological interpretability in their DNA sequence representations.
• There is a need for interpretable and robust DNA encoding methods.

Purpose of the Study:

• To develop a contrastive learning method for pre-training interpretable and robust DNA encoding.
• To improve motif occupancy identification by enhancing sequence representations.
• To evaluate the performance and robustness of the proposed models.

Main Methods:

• Proposed two contrastive learning models: a self-supervised and a supervised model.
• Augmented DNA sequences using edit operations for contrastive learning.
• Utilized the Needleman-Wunsch algorithm for a sequence similarity criterion in self-supervised learning.
• Fine-tuned a deep neural network (DNN) classifier with the pre-trained encoder.

Main Results:

• Both contrastive learning models outperformed baseline end-to-end CNN and SimCLR methods.
• Achieved high Area Under the Curve (AUC) scores of 0.811 and 0.823.
• Demonstrated superior robustness compared to the baseline method, particularly with small sample sizes.
• The self-supervised model showed practical utility in transfer learning scenarios.

Conclusions:

• Contrastive learning effectively pre-trains interpretable and robust DNA encoders for motif occupancy identification.
• The proposed method offers significant improvements over existing approaches, especially in data-limited situations.
• The self-supervised model is a viable option for transfer learning in genomics.