ThermoNet

ThermoNet predicts RNA-binding protein (RBP) specificity by integrating sequence contexts and thermodynamic ensembles of RNA secondary structure using a sequence-embedding convolutional neural network.

Key Features:

• Sequence-Embedding Convolutional Neural Network: Employs a sequence-embedding CNN that generalizes k-mer based methods via joint learning of convolutional filters and k-mer embeddings to capture RNA sequence contexts.
• Thermodynamic Ensemble Integration: Integrates thermodynamic ensembles of RNA secondary structures by averaging deep-learning predictions over ensemble members to account for structural variability.
• Handling Structural Variability: Leverages thermodynamic averages to address multi-modal or variable RNA structures and improve prediction accuracy for structured RNAs.
• Structural Probability Input: Incorporates predicted or experimental RNA structural probabilities as input features alongside sequence data.
• Training and Performance: Trained on large-scale datasets and reported to outperform methods including RCK and DeepBind on benchmark experiments.

Scientific Applications:

• Gene Expression Regulation: Predicts RBP specificity to inform studies of gene expression regulation.
• RNA-Mediated Processes: Supports analysis of RNA-mediated enzymatic processes influenced by RBP binding.
• Experimental Interpretation: Aids interpretation of in vitro and in vivo RBP–RNA interaction experiments, including contexts where RNAcompete-derived datasets are used.
• Structured RNA Analysis: Facilitates investigation of RBP binding on structured RNAs where secondary structure affects specificity.

Methodology:

Train a sequence-embedding convolutional neural network on large-scale datasets using sequence data and predicted or experimental RNA structural probabilities, and average model predictions over thermodynamic ensembles of RNA secondary structures.