Inferring RNA-binding protein target preferences using adversarial domain adaptation.

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Accurate identification of RNA-binding protein (RBP) target sites is crucial for understanding their biological roles.
• Experimental data from in vivo and in vitro methods reveal similar but distinct RBP binding preferences, posing challenges for computational models.
• Existing computational methods struggle to generalize across different datasets due to domain shift.

Purpose of the Study:

• To develop a computational framework, RBP-ADDA, that integrates in vivo and in vitro RBP binding data.
• To address the domain shift problem between different experimental datasets using adversarial domain adaptation (ADDA).
• To improve the prediction accuracy of RBP binding sites, particularly for in vivo data.

Main Methods:

• Implemented an adversarial domain adaptation (ADDA) framework (RBP-ADDA) to integrate disparate RBP binding datasets.
• Pre-trained a source network on in vitro data and adapted it to in vivo data using ADDA.
• Fine-tuned the predictive model using fused in vitro and in vivo data.
• Applied data augmentation for RBPs with limited in vivo data.

Main Results:

• RBP-ADDA demonstrated superior performance in modeling in vivo RBP binding data compared to existing methods, as validated by Pearson correlations.
• The framework also enhanced predictive accuracy on in vitro datasets.
• Data augmentation strategies further improved prediction performance for datasets with sparse in vivo information.
• Integrated Gradients analysis revealed key nucleotide positions critical for RBP recognition.

Conclusions:

• RBP-ADDA effectively leverages complementary information from in vivo and in vitro datasets to improve RBP binding site prediction.
• The ADDA technique provides a robust solution for domain shift challenges in computational RBP binding analysis.
• The framework offers insights into RBP recognition mechanisms through interpretable feature analysis.