CLADES - Contrastive Learning Augmented DifferEntial Splicing with Orthologous Positive Pairs.

Area of Science:

• Genomics
• Computational Biology
• Molecular Biology

Background:

• Alternative splicing (AS) generates diverse transcripts and proteins, but predicting context-specific splicing remains challenging due to limited data and experimental variability.
• Understanding sequence-based regulatory mechanisms governing AS is crucial for deciphering cellular complexity.

Purpose of the Study:

• To develop a robust method for learning sequence-to-splicing mappings using evolutionary conservation.
• To improve the prediction of differential splicing events (Δψ) across various biological contexts.

Main Methods:

• Proposed a contrastive representation learning framework leveraging evolutionary orthologous sequences as positive pairs.
• Utilized a discriminative objective to align embeddings of regulatory equivalents and emphasize conserved signals.
• Fine-tuned a supervised head for predicting Δψ and developed an interpretable splice-motif-aware classification framework.

Main Results:

• The contrastive pre-training significantly enhanced the prediction of exon inclusion changes (Δψ).
• Achieved strong performance in classifying splicing shifts (AUPRC/AUROC) and competitive regression results (RMSE, Spearman) on diverse benchmarks.
• Demonstrated the biological interpretability of the learned representations through splice-motif awareness.

Conclusions:

• Evolution-guided contrastive learning provides an effective and biologically principled approach for context-resolved splicing prediction.
• The developed method enhances understanding of sequence-based regulation in alternative splicing.
• This framework offers a powerful tool for analyzing differential splicing across tissues and cell types.