GINI: from ISH images to gene interaction networks

Area of Science:

• Developmental Biology
• Computational Biology
• Genomics

Background:

• Accurate inference of gene interactions is crucial for understanding complex biological systems, particularly in multicellular organisms like Drosophila.
• In-situ hybridization (ISH) provides spatial-temporal profiling of gene expression but lacks robust analytical tools for network inference.
• Existing methods often rely on scalar gene intensity, missing the rich information in spatial patterns.

Purpose of the Study:

• To present GINI, a novel machine learning system for inferring gene interaction networks from Drosophila embryonic ISH images.
• To develop a method that represents gene activity using spatial patterns rather than scalar intensity.
• To create statistically sound and biologically meaningful gene networks from image data.

Main Methods:

• GINI utilizes a computer-vision-inspired vector-space representation of gene expression spatial patterns from ISH images.
• A multi-instance-kernel algorithm is employed to learn a sparse Markov network model.
• The system represents each gene as a vector-valued spatial pattern, capturing spatial similarity.

Main Results:

• GINI effectively infers gene interaction networks from both synthetic and curated datasets.
• Application to a large collection of Drosophila ISH images yielded novel and interesting gene interaction predictions.
• The system demonstrates the capability to handle multiple images per gene through multi-instance kernels.

Conclusions:

• GINI offers a powerful approach for inferring gene interaction networks directly from ISH image data.
• By leveraging spatial patterns, GINI provides more biologically meaningful insights than traditional methods.
• The developed system advances the analysis of large-scale gene expression imaging data for biological discovery.