Gene expression inference based on graph neural networks using L1000 data.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Gene expression profiles reflect cellular states and aid in discovering functional gene connections.
• The L1000 technology provides cost-effective gene expression data for numerous conditions.
• Existing gene expression inference methods, including linear and deep learning models, treat data as vectors.

Purpose of the Study:

• To investigate the effectiveness of nonlinear models based on graph structures for gene expression inference.
• To compare the performance of graph neural network (GNN) models against traditional linear and nonlinear non-GNN models.
• To assess the impact of feature selection and organ information on GNN performance.

Main Methods:

• Development and application of a graph neural network (GNN) model where genes are represented as nodes.
• Comparison of GNN model performance with linear regression and other nonlinear models.
• Evaluation of input feature selection strategies and the incorporation of organ-specific data.
• Assessment of cross-platform generality for gene expression inference.

Main Results:

• The GNN model significantly outperforms linear and nonlinear non-GNN models in predicting gene expression values and rankings.
• The GNN model achieves comparable performance using approximately 10-fold less input information.
• Strategic input feature selection and the inclusion of organ features further enhance GNN inference accuracy.
• The GNN model demonstrates robust cross-platform generality in gene expression inference.

Conclusions:

• Representing RNA expression data as a graph structure effectively captures complex, nonlinear gene correlations.
• GNNs offer a more accurate and efficient approach to gene expression profile prediction.
• This graph-based method advances the understanding of gene interactions and cellular states.