A Novel Repetition Frequency-Based DNA Encoding Scheme to Predict Human and Mouse DNA Enhancers with Deep Learning

Area of Science:

• Genomics and Bioinformatics
• Computational Biology
• Molecular Biology

Background:

• DNA enhancers play a critical role in regulating gene expression, impacting development, homeostasis, and embryogenesis.
• Experimental identification of DNA enhancers is labor-intensive and costly, necessitating computational approaches.
• Existing deep learning methods show performance inconsistencies across different cell lines for DNA enhancer prediction.

Purpose of the Study:

• To develop a novel DNA encoding scheme for improved DNA enhancer prediction.
• To evaluate the efficacy of the proposed scheme using BiLSTM (Bidirectional Long Short-Term Memory) deep learning model.
• To compare the proposed scheme against established DNA encoding methods (EIIP, integer, atomic number).

Main Methods:

• A four-stage process involving data acquisition, DNA sequence encoding, BiLSTM model design, and performance evaluation.
• DNA sequences were numerically represented using the proposed scheme and compared with EIIP, integer, and atomic number encodings.
• Performance metrics included accuracy, precision, recall, F1-score, CSI, MCC, G-mean, Kappa coefficient, and AUC scores.

Main Results:

• Scenario 1 (Human vs. Mouse enhancer identification): The proposed scheme achieved 92.16% accuracy and 0.85 AUC, outperforming EIIP (89.14% accuracy, 0.87 AUC), atomic number (86.61% accuracy, 0.84 AUC), and integer (76.96% accuracy, 0.82 AUC).
• Scenario 2 (Enhancer identification and species origin): The proposed scheme yielded the highest accuracy (84.59%) and AUC (0.92), surpassing EIIP (77.80% accuracy, ~0.90 AUC), integer (73.68% accuracy, ~0.90 AUC), and atomic number (68.27% accuracy, 0.81 AUC).
• The proposed DNA encoding scheme demonstrated superior performance in both prediction scenarios.

Conclusions:

• The novel DNA encoding scheme is effective and successful for predicting DNA enhancers.
• This approach offers a more robust and accurate computational alternative to experimental methods.
• The findings highlight the potential of tailored encoding schemes in enhancing deep learning applications for genomic analysis.