A sequence knowledge-guided deep learning method for single-cell multi-omics translation.

Area of Science:

• Computational Biology
• Genomics
• Proteomics

Background:

• Protein analysis is crucial for understanding biological processes and diseases.
• Proteome profiling is limited compared to RNA sequencing due to technical challenges and cost.
• Multi-omics technologies link transcriptome and proteome, enabling computational prediction of protein abundance.

Purpose of the Study:

• To introduce scProTrans, a deep learning framework for cross-omics translation at single-cell resolution.
• To predict protein abundance from transcriptome data, overcoming proteomic data acquisition limitations.
• To enhance multi-omics integration and downstream single-cell analyses.

Main Methods:

• Developed scProTrans, a deep learning framework utilizing sequence knowledge and multi-omics integration.
• Employed a hierarchical attention mechanism for gene/protein sequence alignment with cellular contexts.
• Utilized a bidirectional encoder for sequence-to-embedding-to-profile learning and modality translation.
• Incorporated cell-specific associations to capture dynamic gene-protein interplay.

Main Results:

• scProTrans surpasses state-of-the-art methods in single-cell protein abundance translation across 17 datasets.
• The framework improves protein prediction accuracy and preserves low-abundance protein signals.
• Extended scProTrans to tri-omics (ATAC-RNA-protein) scenarios with high cross-modal prediction concordance.
• Enhanced downstream analyses including cell clustering, subtype identification, and biomarker discovery.

Conclusions:

• Advanced multi-omics integration with a sequence-aware paradigm for cross-modal translation.
• Overcame key limitations in proteome data acquisition using computational prediction.
• Established a versatile platform for emerging multi-modal single-cell technologies with modular architecture and zero-shot capability.