H2Opred

H2Opred predicts 2'-O-methylation (2OM) sites in human RNA to identify modification loci that affect RNA splicing, stability, and innate immunity.

Key Features:

• Hybrid deep learning architecture: Combines stacked one-dimensional convolutional neural network (1D-CNN) blocks with attention-based bidirectional gated recurrent unit (Bi-GRU-Att) blocks.
• Nucleotide-specific and generic models: Implements four nucleotide-specific models (A2OM, C2OM, G2OM, U2OM) and a generic model (N2OM).
• 1D-CNN feature extraction: 1D-CNN components extract feature representations from 14 conventional descriptors.
• Sequence embedding: Bi-GRU-Att blocks utilize natural language processing-based embeddings derived from RNA sequences to capture sequence dependencies.
• Integrated feature representation: Combines descriptor-based features and NLP-derived embeddings for joint prediction.
• Performance evaluation: Demonstrates superior predictive performance versus traditional single-feature machine learning models using cross-validation across five datasets.
• Generic-model robustness: The generic N2OM model exhibits consistently high accuracy in training and testing evaluations.

Scientific Applications:

• Mapping 2'-O-methylation sites: Identification of 2OM loci in human mRNA for studies of modification distribution and site-specific analyses.
• RNA biology research: Investigation of how 2'-O-methylation influences RNA splicing, stability, and innate immune recognition.
• Method comparison and benchmarking: Comparative evaluation of predictive methods and feature representations for post-transcriptional modification prediction.

Methodology:

H2Opred integrates stacked 1D-CNN blocks extracting features from 14 conventional descriptors with attention-based bidirectional GRU (Bi-GRU-Att) blocks using NLP-based RNA sequence embeddings, implements nucleotide-specific (A2OM, C2OM, G2OM, U2OM) and generic (N2OM) models, and evaluates performance via cross-validation on five datasets.