TL;DR
This paper tackles the challenge of understanding how SARS-CoV-2 interacts with human RNA-binding proteins (RBPs) by creating the first comprehensive, nucleotide-resolution in silico map of these interactions. By leveraging deep learning models trained on experimental data, the authors provide a valuable resource for researchers to explore viral hijacking of host pathways and identify potential therapeutic targets.
What Problem Does This Paper Address?
RNA-binding proteins (RBPs) are crucial players in cellular processes and are known to be manipulated by viruses to facilitate infection. SARS-CoV-2, the virus responsible for COVID-19, is no exception. However, experimentally mapping the precise binding sites of numerous human RBPs on the viral RNA genome is a complex and time-consuming task, especially given the emergence of new viral variants. This lack of a detailed map hinders our understanding of how the virus hijacks host cellular machinery and limits the development of targeted antiviral strategies. This paper aims to bridge this gap by computationally predicting these RBP-RNA interactions at a fine-grained, nucleotide level.
Methods
The core of this study lies in the application of two deep learning methods: pysster and DeepRiPe. These convolutional neural network (CNN) based models were trained on existing experimental data (eCLIP and PAR-CLIP) to predict RBP binding preferences.
- Model Training: pysster models were trained individually for each RBP using eCLIP data. DeepRiPe models, pre-trained on PAR-CLIP data, were also utilized.
- Performance Evaluation: Rigorous evaluation was performed using metrics like Area Under the Precision-Recall Curve (auPRC) and Spearman correlation coefficient (SCC) to ensure model quality. A "Performance-In-Practice" (PIP) metric was introduced to assess model performance on full-length transcripts.
- Binding Site Prediction: The trained models were used to scan the SARS-CoV-2 genome, predicting single-nucleotide binding probabilities. Consecutive high-scoring and significant positions were then merged to define RBP binding sites.
- Variant Impact Analysis: The models were used to score the impact of mutations found in 11 SARS-CoV-2 variants of concern, predicting "gain-of-binding" (GOB) or "loss-of-binding" (LOB) events.
- Feature Attribution: Integrated Gradients were employed to understand which sequence features drive RBP binding, aiding in motif identification.
- Comparative Analysis: The approach was extended to predict RBP binding on the genomes of six other human coronaviruses to assess evolutionary conservation.
Datasets
- Training Data:
- eCLIP datasets for 150 RBPs from the ENCODE project (HepG2 and K562 cell lines).
- PAR-CLIP datasets for 59 RBPs (used for pre-trained DeepRiPe models).
- Viral Genomes:
- SARS-CoV-2 reference genome.
- Genomes of 6 other human coronaviruses (SARS-CoV-1, MERS, HCoV-OC43, HCoV-HKU1, HCoV-NL63, HCoV-229E).
- Sequences from 11 SARS-CoV-2 variants of concern.
- Functional and OMICs Data:
- Data from 22 studies related to COVID-19/coronavirus research, including interactomes, computational predictions, PPIs, multi-omics data (proteomics, transcriptomics, etc.), CRISPR screens, GWAS, and patient OMICS data.
Key Insights and Ideas
- Deep Learning for RBP-RNA Interaction Prediction: The study demonstrates the power of deep learning, specifically CNNs, to accurately predict RBP binding sites on viral RNA, even for RBPs lacking well-defined sequence motifs.
- Nucleotide-Resolution Mapping: The creation of a single-nucleotide resolution map is a significant advancement, providing a much finer view of RBP-RNA interactions than previous methods.
- Functional Clustering of RBPs: The predicted binding sites reveal that RBPs can be clustered based on their binding patterns across the viral genome, suggesting functional relationships and potential roles in viral regulation.
- SECRETE Motif Association: A notable finding is the association of certain RBPs with the "SECRETE" motif, a cis-acting RNA element implicated in viral replication and host cell processes.
- Variant Impact on Binding: The analysis of SARS-CoV-2 variants highlights how specific mutations can lead to significant gain or loss of RBP binding, potentially influencing viral fitness and host immune response.
- Cross-Coronavirus Conservation: The comparative analysis reveals conserved RBP binding patterns across different coronaviruses, particularly among highly pathogenic strains, suggesting fundamental host-pathogen interactions.
- Identification of Clinical Biomarkers: The integration of OMICs and functional data allowed the identification of RBPs like MBNL1, FTO, and FXR2 as potential clinical biomarkers for COVID-19 severity.
Key Results
- High-Quality Models: The study generated a set of 88 high-quality prediction models for human RBPs, with stringent performance evaluation ensuring reliability.
- Comprehensive Binding Map: An in silico binding map of human RBPs on the SARS-CoV-2 genome at single-nucleotide resolution was constructed.
- Variant Impact: Analysis of 11 viral strains identified numerous variants that cause gain or loss of RBP binding, some of which impact multiple RBPs simultaneously.
- Cross-Coronavirus Binding: RBP binding patterns were found to be largely conserved across seven coronaviruses, with highly pathogenic viruses showing similar profiles.
- Functional Annotation: A large fraction (72.4%) of the predicted RBPs were found to have direct experimental evidence of binding to SARS-CoV-2 RNA, and 86.2% were implicated in host-pathogen PPI networks or altered cellular pathways.
- Clinical Relevance: 44 RBPs showed nominal significance in GWAS studies related to COVID-19 severity, and a small set of RBPs were identified as potential clinical biomarkers.
Limitations and Future Work
- In Silico Predictions: While the models were rigorously validated, they are still predictions and experimental validation is crucial for confirming specific binding events and their functional consequences.
- Model Overlap: The study notes that different RBP prediction studies have limited overlap, highlighting the need for comprehensive and standardized approaches.
- Experimental Data Limitations: The availability and quality of experimental eCLIP and PAR-CLIP data can influence model performance.
- Future Work: The authors suggest that their comprehensive resource will aid researchers in assessing the impact of newly identified viral variants. Further experimental validation of predicted binding sites and their functional roles in viral replication and pathogenesis is warranted. The identified RBPs with clinical relevance also warrant further investigation.
Summary updated