Predicting RNA secondary structure by a neural network: what features may be learned?

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in genomics

Background:

• Deep learning models can learn data representations without explicit programming.
• Understanding RNA structure is crucial for biological function.
• Predicting RNA secondary structure from sequence is a fundamental bioinformatics challenge.

Purpose of the Study:

• To develop a deep learning model (PredPair) for predicting RNA base pairs and structure solely from sequence.
• To investigate the internal representations learned by the model regarding RNA structural features.
• To evaluate PredPair's performance against existing methods and its ability to predict complex structures like pseudoknots.

Main Methods:

• A deep learning neural network (PredPair) was trained on RNA sequences to predict base pairing.
• No prior knowledge of thermodynamics or spatial structures was incorporated during training.
• Model performance was assessed using experimental accessibility data (DMS-Seq) and compared to thermodynamic models.
• t-distributed Stochastic Neighbor Embedding (t-SNE) was used to visualize sequence embeddings.

Main Results:

• PredPair successfully learned Watson-Crick and wobble base-pairing rules.
• The model developed internal representations of stacking energies and helices.
• Predictions correlated with experimental data on nucleotide accessibility.
• PredPair achieved performance comparable to state-of-the-art thermodynamic methods, with success in predicting pseudoknots.
• Sequence embeddings clustered according to Rfam families, aligning with biological classification.

Conclusions:

• Deep learning models can effectively predict RNA secondary structure and learn biologically relevant features from sequence data alone.
• PredPair offers a novel deep learning approach for RNA structure prediction, demonstrating capabilities beyond traditional methods, including pseudoknot prediction.
• The learned representations and clustering by Rfam families suggest PredPair captures fundamental aspects of RNA sequence-structure relationships.