iDNA-ABT

iDNA-ABT predicts DNA methylation sites across multiple species to identify epigenetic modifications relevant to gene regulation and disease.

Key Features:

• Adaptive embedding (BERT): Uses adaptive embedding based on Bidirectional Encoder Representations from Transformers (BERT) to learn sequence representations.
• Transductive information maximization (TIM) loss: Integrates transductive information maximization (TIM) loss during training to enhance discriminative feature learning.
• Automated feature learning: Automatically and adaptively learns distinguishing features from biological sequences without manual feature extraction.
• High-latent information capture: Captures high-latent information from sequences to improve predictive accuracy.
• Multi-type methylation prediction: Predicts three distinct types of DNA methylation as reported in benchmarking results.
• Benchmark performance: Demonstrates improved performance compared to existing state-of-the-art methods in comparative experiments.
• Robustness across species: Exhibits strong adaptability and robustness across various species relative to six handcrafted feature encodings.

Scientific Applications:

• Epigenetic site identification: Identification of DNA methylation sites for studies of gene regulation and disease mechanisms.
• Cross-species epigenomics: Comparative analysis of DNA methylation patterns and cross-species differences in epigenetic regulation.
• Dichotomous classification tasks: Application to dichotomous methylation detection tasks where TIM loss is shown to be effective.

Methodology:

Adaptive embedding using Bidirectional Encoder Representations from Transformers (BERT) combined with transductive information maximization (TIM) loss, with comparative benchmarking against six handcrafted feature encodings and existing state-of-the-art methods.