A Multimodal Biomedical Transformer Fusion Network for Disease-Level Rare-Disease-Inheritance Classification Using...

Area of Science:

• Computational biology and bioinformatics
• Medical informatics
• Genetics and genomics

Background:

• Rare disease inheritance classification is challenging due to incomplete, heterogeneous, and imbalanced knowledge.
• Disease-level inheritance modeling can aid rare-disease knowledge organization and hypothesis generation.
• Existing methods struggle with the complexity and data scarcity inherent in rare diseases.

Purpose of the Study:

• Introduce RareFusion-Net, a multimodal benchmark framework for disease-level inheritance classification.
• Evaluate the impact of integrating ontology-enriched text, epidemiological metadata, and gene associations on prediction accuracy.
• Provide a tool for knowledge modeling in rare-disease resources, not for individual patient diagnosis.

Main Methods:

• Developed RareFusionBalanced, a gated multimodal fusion model combining disease text, metadata, and gene information.
• Utilized ontology-enriched text as the primary semantic modality, supplemented by tabular and gene data.
• Employed robust training techniques including balanced regularization, selective transformer fine-tuning, and prediction aggregation.

Main Results:

• RareFusionBalanced achieved 0.7382 test accuracy, 0.6284 macro-F1, and 0.9183 macro-AUC.
• Ontology-enriched text was the strongest modality; gene associations provided complementary value.
• The multimodal model showed slight improvement over text-only models, but gains over simpler baselines were not statistically significant.

Conclusions:

• RareFusion-Net offers a practical benchmark for ontology-aware rare-disease inheritance modeling.
• Results indicate selective multimodal benefits and highlight challenges with minority classes and statistical significance.
• Further validation and interpretability studies are needed to fully leverage multimodal approaches in rare disease research.