BioLM-NET: an interpretable deep learning model combining prior biological knowledge and contextual LLM gene...

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Deep neural networks (DNNs) are increasingly used in biological research, often incorporating gene pathway information.
• Existing DNNs frequently overlook crucial protein-protein interactions (PPI) and protein-DNA interactions (PDI).

Purpose of the Study:

• To introduce BioLM-NET, a deep learning framework that integrates multi-omics data with biological knowledge.
• To enhance the prediction accuracy of disease subtypes and patient outcomes by incorporating PPI and PDI.

Main Methods:

• BioLM-NET fuses gene expression and DNA methylation data with PPI and PDI knowledge.
• An attention-based pathway layer aggregates omics signals, utilizing a pretrained large language model (LLM) for gene embeddings.
• The framework was evaluated on single-cell colorectal cancer, TCGA cancer subtypes (BRCA, GBM, COAD), and Alzheimer's disease (ROSMAP) datasets.

Main Results:

• BioLM-NET significantly outperformed baseline and state-of-the-art methods (P-NET, PASNet) on scTrioseq2, TCGA-COAD, and ROSMAP datasets.
• It achieved comparable results to SVM and Dense neural networks on TCGA-BRCA data.
• Ablation studies confirmed the critical role of PPI, PDI, and the attention-based pathway layer.

Conclusions:

• BioLM-NET demonstrates superior performance in disease prediction by integrating multi-omics data and biological interactions.
• The model's key features are enriched in significant Gene Ontology (GO) terms and KEGG pathways.
• These features hold potential as biomarkers or therapeutic targets for various diseases.