Deciphering Cell Type Abundance in Proteomics Data Through Graph Neural Networks

Area of Science:

• Proteomics
• Bioinformatics
• Computational Biology

Background:

• Proteomics sequencing advances cell-type signature exploration in tissues for disease insights.
• Current proteomic technologies lack resolution, mixing cell types and hindering accurate profiling.
• Existing cell-type deconvolution methods, primarily for transcriptomics, face challenges with proteomic data's weak correlation and divergent quantification.

Purpose of the Study:

• To introduce GraphDEC, a novel graph neural network (GNN)-based method for precise cell type deconvolution in proteomic profiling data.
• To address the limitations of existing methods that ignore higher-order relationships within proteomic datasets.
• To enhance the inference of cellular composition from complex proteomic samples.

Main Methods:

• GraphDEC simulates bulk samples from single-cell proteomic data to generate reference datasets.
• An autoencoder extracts low-dimensional representations for constructing sample similarity relationships.
• A GNN with a multi-channel mechanism and hybrid neighborhood-aware approach processes integrated proteomic and similarity data.
• Multiple loss functions (triplet, domain adaptation, MSE) optimize the model and mitigate batch effects.

Main Results:

• GraphDEC achieves state-of-the-art performance on diverse synthetic and real-world spatial proteomic datasets.
• The method demonstrates strong generalization capabilities across different sequencing technologies and species.
• GraphDEC shows high efficiency when applied to transcriptomics data, indicating broad applicability.

Conclusions:

• GraphDEC represents a significant advancement in cell-type deconvolution for proteomic data.
• The GNN-based approach effectively leverages higher-order sample relationships for improved accuracy.
• GraphDEC offers a robust and versatile tool for analyzing cellular composition in complex biological samples.