SpliceTransformer predicts tissue-specific splicing linked to human diseases

Ningyuan You, Chang Liu, Yuxin Gu, Rong Wang, Hanying Jia, Tianyun Zhang, Song Jiang, Jingsong Shi, Ming Chen, Min‐Xin Guan, Siqi Sun, Shanshan Pei, Zhihong Liu, Ning ShenNature Communications 2024 · DOI

Abstract

We present SpliceTransformer (SpTransformer), a deep-learning framework that predicts tissue-specific RNA splicing alterations linked to human diseases based on genomic sequence. SpTransformer outperforms all previous methods on splicing prediction. Application to approximately 1.3 million genetic variants in the ClinVar database reveals that splicing alterations account for 60% of intronic and synonymous pathogenic mutations, and occur at different frequencies across tissue types. Importantly, tissue-specific splicing alterations match their clinical manifestations independent of gene expression variation. We validate the enrichment in three brain disease datasets involving over 164,000 individuals. Additionally, we identify single nucleotide variations that cause brain-specific splicing alterations, and find disease-associated genes harboring these single nucleotide variations with distinct expression patterns involved in diverse biological processes. Finally, SpTransformer analysis of whole exon sequencing data from blood samples of patients with diabetic nephropathy predicts kidney-specific RNA splicing alterations with 83% accuracy, demonstrating the potential to infer disease-causing tissue-specific splicing events. SpTransformer provides a powerful tool to guide biological and clinical interpretations of human diseases.

Introduction

RNA splicing is a critical regulatory layer in the human genome, where non-coding introns are removed and coding exons are joined. While many genetic mutations disrupt this process, their effects are often tissue-specific—a mutation might impair splicing in the brain while leaving it intact in the liver. Existing tools, such as SpliceAI, excel at general splicing prediction but frequently overlook these tissue-specific nuances. Furthermore, many pathogenic mutations are "synonymous" or "intronic," often leading them to be ignored in standard clinical pipelines. To address this, we developed SpliceTransformer (SpTransformer), a model designed to predict these "hidden" effects across the human body using only genomic sequences.

Methods and Architecture

SpTransformer utilizes a Transformer-based architecture, specifically a variant of the Sinkhorn Transformer, to handle long-range sequence dependencies up to 10,000 base pairs. This is essential for capturing the complex genomic context required for splicing.

Overview of the SpTransformer architecture and its tissue-specific prediction workflow.

The model employs a dual-encoder strategy: one encoder is trained on human data from GTEx V8, while the other leverages cross-species mammalian data (including rhesus macaques, mice, and rats) to capture evolutionarily conserved splicing signals. Multi-head attention allows the model to simultaneously monitor proximal splice sites and distal cis-regulatory elements. To ensure balanced performance across 15 representative tissues, a weighted geometric mean loss function was implemented, preventing bias toward tissues with larger data availability.

Key Results

SpTransformer achieves state-of-the-art performance, reaching approximately 85% Top-k accuracy and 91% PR-AUC, consistently outperforming previous models such as SpliceAI and Pangolin across various tissues.

SpTransformer outperforms existing models in both general accuracy and tissue-wise PR-AUC across 15 human tissues.

Our analysis of the ClinVar database revealed that splicing alterations account for 60% of intronic and synonymous pathogenic mutations. Crucially, these tissue-specific splicing alterations often align with clinical symptoms independent of gene expression levels. For instance, in brain disease cohorts involving over 164,000 individuals (including ASD, SCZ, and BD), we identified a significant enrichment of brain-specific splicing alterations.

In a practical clinical application, SpTransformer predicted kidney-specific splicing issues in patients with diabetic nephropathy using DNA from blood samples with 83% accuracy. This demonstrates the model's potential as a "liquid biopsy" tool, allowing clinicians to infer molecular events in hard-to-access organs without direct invasive procedures.

Discussion

SpTransformer represents a significant advancement in functional genomics by providing a high-resolution map of tissue-specific splicing. While the current model covers 15 tissues, future work will aim to expand this resolution to more niche tissues and individual cell types. Additionally, while the model is highly effective at identifying common splicing events, predicting rare, individual-specific variants remains a challenge. Nonetheless, SpTransformer offers a powerful framework for the biological and clinical interpretation of human diseases, moving us closer to personalized genomic medicine.

Summary updated

Abstract

We present SpliceTransformer (SpTransformer), a deep-learning framework that predicts tissue-specific RNA splicing alterations linked to human diseases based on genomic sequence. SpTransformer outperforms all previous methods on splicing prediction. Application to approximately 1.3 million genetic variants in the ClinVar database reveals that splicing alterations account for 60% of intronic and synonymous pathogenic mutations, and occur at different frequencies across tissue types. Importantly, tissue-specific splicing alterations match their clinical manifestations independent of gene expression variation. We validate the enrichment in three brain disease datasets involving over 164,000 individuals. Additionally, we identify single nucleotide variations that cause brain-specific splicing alterations, and find disease-associated genes harboring these single nucleotide variations with distinct expression patterns involved in diverse biological processes. Finally, SpTransformer analysis of whole exon sequencing data from blood samples of patients with diabetic nephropathy predicts kidney-specific RNA splicing alterations with 83% accuracy, demonstrating the potential to infer disease-causing tissue-specific splicing events. SpTransformer provides a powerful tool to guide biological and clinical interpretations of human diseases.
BibTeX
@inproceedings{you2024splicetransformer,
  author = {Ningyuan You and Chang Liu and Yuxin Gu and Rong Wang and Hanying Jia and Tianyun Zhang and Song Jiang and Jingsong Shi and Ming Chen and Min‐Xin Guan and Siqi Sun and Shanshan Pei and Zhihong Liu and Ning Shen},
  title = {{SpliceTransformer predicts tissue-specific splicing linked to human diseases}},
  booktitle = {Nature Communications},
  year = {2024},
  doi = {10.1038/s41467-024-53088-6},
}
Figure