OmicTransformer

Multi-omics integration is crucial for comprehensive understanding of complex diseases, but it poses challenges due to high dimensionality, noise, sparsity, and incomplete data. Here we present OmicsTransformer, a novel deep learning framework that addresses these challenges by constructing relation-aware feature embeddings and leveraging a transformer architecture for global feature extraction. We introduce a specialized masked-loss function that uses random feature masking to encourage the model to capture intrinsic correlations between biomolecular features and to filter out noise. Each omic data modality is processed by a transformer encoder, and an uncertainty-based weighting mechanism then integrates the modalities, dynamically balancing their contributions. In evaluations on three benchmark datasets (including classification tasks in Alzheimer’s disease, glioma, and breast cancer subtyping), OmicTransformer significantly outperforms state-of-the-art multi-omics methods, with 2-4% higher accuracy. Ablation studies confirm that the masked-loss guided embedding and the transformer encoder are critical to performance. Moreover, applying modality dropout during training further improves robustness to missing omics, simulating real-world scenarios with incomplete data. Using a gradient-based approach, OmicTransformer can also identify important biomarkers associated with disease subtypes, many of which align with known biological findings. Our work demonstrates a powerful and interpretable transformer- based strategy for multi-omics data integration, setting a new state-of-the-art in multi-omics classification and offering insights for precision medicine applications.