OmicsTransformer: self-supervised masked consistency and uncertainty-aware fusion for robust multi-omics prediction

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Current multi-omics integration models face challenges with high dimensionality, data redundancy, missing assays, and incomplete pathway information.
• Existing methods often rely on heuristic graph construction or fixed knowledge bases, limiting their flexibility and biological interpretability.

Purpose of the Study:

• To develop a novel framework, OmicsTransformer, for direct learning of biologically meaningful patient manifolds from high-dimensional multi-omics data.
• To overcome limitations of existing models by avoiding heuristic graph construction and fixed knowledge-base constraints.

Main Methods:

• OmicsTransformer projects omics modalities into latent patches and uses a Transformer encoder to model dependencies.
• It enforces masked semantic consistency via an Exponential Cosine Consistency Loss and fuses modalities using sample-specific uncertainty.
• The framework is implemented in PyTorch and its source code is publicly available.

Main Results:

• OmicsTransformer demonstrated strong performance across eight diagnostic and prognostic cohorts.
• Achieved 89.4% accuracy for TCGA-BRCA subtyping and 90.6% AUC for TCGA-LGG grading.
• Significantly improved recurrence prediction accuracy over DeepKEGG, outperforming it by up to 21.5 percentage points.

Conclusions:

• OmicsTransformer offers a powerful, flexible approach to multi-omics integration for cancer research.
• The framework successfully learns patient-specific biological features directly from complex omics data.
• Identified reproducible cross-modal biomarker cores and novel progression drivers through variance-weighted attribution.