Gene-MOE: A Sparsely Gated Cancer Diagnosis and Prognosis Framework Exploiting Pan-Cancer Genomic Information

Area of Science:

• Genomics
• Computational Biology
• Artificial Intelligence

Background:

• Accurate cancer genomic diagnosis and prognosis are crucial for effective medical treatment.
• Deep learning, particularly Transformer models, has advanced cancer analysis precision.
• The potential of novel architectures like sparsely gated Mixture of Experts (MOE) for cancer prognosis and classification remains underexplored.

Purpose of the Study:

• To introduce Gene-MOE, a novel framework leveraging MOE and Mixture of Attention Expert (MOAE) layers for enhanced cancer diagnosis and prognosis.
• To address overfitting by integrating pan-cancer information from 33 cancer types via pre-training.
• To evaluate Gene-MOE's performance against state-of-the-art models in cancer survival analysis, classification, and subtyping.

Main Methods:

• Development of the Gene-MOE framework incorporating sparsely gated MOE and MOAE layers.
• Pre-training strategy utilizing pan-cancer genomic data from 33 distinct cancer types to mitigate overfitting.
• Comparative analysis of Gene-MOE against established models using Concordance Index for survival analysis, accuracy for classification, and log10 P-values/significant clinical markers for subtyping.

Main Results:

• Gene-MOE achieved superior Concordance Index in survival analysis for 12 out of 14 cancer types.
• The model reached 95.8% accuracy in classifying 33 cancer types, outperforming existing methods.
• Gene-MOE demonstrated top performance in cancer subtyping metrics for seven of nine cancers.

Conclusions:

• Gene-MOE significantly enhances accuracy in cancer genomic diagnosis, prognosis, and classification.
• The framework's innovative use of MOE and MOAE layers, combined with pan-cancer pre-training, effectively addresses analytical challenges.
• Gene-MOE shows substantial potential for improving downstream genomic analysis tasks in oncology.