Single-Cell Multimodal Prediction via Transformers.

Area of Science:

• Single-cell multi-omics data analysis
• Computational biology
• Machine learning for genomics

Background:

• Multimodal single-cell technologies enable deeper cellular state and dynamics understanding.
• Modeling complex interactions within multimodal single-cell data presents significant challenges.
• Existing graph neural network (GNN) methods construct static graphs, limiting downstream task integration and facing limitations with deep layer stacking.

Approach:

• Propose scMoFormer, a transformer-based framework for end-to-end multimodal single-cell data analysis.
• scMoFormer integrates external domain knowledge and models intra- and cross-modality interactions.
• The framework addresses limitations of static graphs and deep GNN architectures.

Key Points:

• scMoFormer demonstrates superior performance across various benchmark datasets.
• The framework effectively models complex interactions within and between different data modalities.
• Achieved a Kaggle silver medal (Top 2%) in a NeurIPS 2022 competition without ensembling.

Conclusions:

• scMoFormer offers a powerful and flexible approach for multimodal single-cell data analysis.
• The transformer-based architecture provides significant advantages over traditional GNN methods.
• Publicly available implementation facilitates further research and application in single-cell biology.