[Paper Review] Computational neuroanatomy and gene expression: optimal sets of marker genes for brain regions

Pascal Grange, Partha P. Mitra|Cold Spring Harbor Laboratory Institutional Repository (Cold Spring Harbor Laboratory)|May 11, 2012

Gene expression and cancer classification12 references5 citations

TL;DR

This paper proposes quantitative, optimization-based criteria to identify optimal sets of marker genes for brain regions using the Allen Mouse Brain Atlas. By combining localization and fitting scores—formulated as generalized eigenvalue and L1-L2 regularized optimization problems—it identifies gene sets with near-perfect spatial localization across major brain regions, except the pallidum, using weighted combinations of genes with both positive and negative coefficients.

ABSTRACT

The three-dimensional data-driven Allen Gene Expression Atlas of the adult mouse brain consists of numerized in-situ hybridization data for thousands of genes, co-registered to the Allen Reference Atlas. We propose quantitative criteria to rank genes as markers of a brain region, based on the localization of the gene expression and on its functional fitting to the shape of the region. These criteria lead to natural generalizations to sets of genes. We find sets of genes weighted with coefficients of both signs with almost perfect localization in all major regions of the left hemisphere of the brain, except the pallidum. Generalization of the fitting criterion with positivity constraint provides a lesser improvement of the markers, but requires sparser sets of genes.

Motivation & Objective

• To develop quantitative, data-driven criteria for identifying marker genes that accurately reflect the spatial boundaries of brain regions in the adult mouse brain.
• To address the limitation of single-gene markers by generalizing the concept to optimal sets of genes with weighted coefficients.
• To improve spatial localization and fitting of gene expression patterns to brain region shapes using mathematical optimization techniques.
• To evaluate the performance of these methods on the Allen Reference Atlas, focusing on the left hemisphere's Big12 regions.
• To explore the trade-off between fitting accuracy and sparsity in gene sets, particularly under positivity constraints.

Proposed method

• Defines a localization score as the fraction of a gene’s expression energy contained within a target brain region, normalized over the whole brain.
• Uses a generalized eigenvalue problem to find optimal gene sets that maximize alignment with the region’s shape, minimizing error in fitting the region’s expression profile.
• Applies L1-L2 regularization (elastic net-like) to the fitting problem, balancing fitting error and sparsity of the gene set coefficients.
• Employs a parameterized optimization framework where the regularization parameter Λ controls the trade-off between fitting accuracy and sparsity.
• Uses the Allen Reference Atlas to define region-specific indicator vectors χω and co-registered gene expression data E from in situ hybridization at 200 μm voxel resolution.
• Computes and compares fitting scores and localization scores across single genes and optimized gene sets across all Big12 brain regions.

Experimental results

Research questions

• RQ1 What quantitative criteria can reliably rank genes as markers of specific brain regions based on spatial expression patterns?
• RQ2 Can sets of genes with optimized coefficients (positive and negative) achieve near-perfect localization of brain regions compared to single genes?
• RQ3 How does the inclusion of both positive and negative coefficients in gene sets affect the accuracy and sparsity of regional marker identification?
• RQ4 What is the trade-off between fitting accuracy and sparsity when imposing positivity constraints on gene set coefficients?
• RQ5 To what extent can the proposed optimization framework detect biologically meaningful gene combinations that define brain region boundaries?

Key findings

• Optimal gene sets with coefficients of both signs achieve near-perfect localization (localization score ≈ 1) for all major brain regions in the left hemisphere except the pallidum.
• The best-fitting gene sets, derived via L1-L2 regularized optimization, significantly outperform single genes in fitting accuracy, with the largest improvement observed in the midbrain.
• For the midbrain, the best-fitting set of 8 genes (e.g., Slc17a6, Ephb1, Sema3f) achieved a fitting score substantially higher than the best single gene (Slc17a6).
• Under positivity constraints, the improvement in fitting score is less dramatic, but sparser gene sets are obtained, with the number of genes in the best-fitted sets ranging from 7 to 14 across regions.
• The method identifies region-specific optimal gene sets, with the cerebral cortex and hippocampus requiring up to 14 genes in their best-fitted sets.
• The framework reveals that complex, non-trivial combinations of genes—rather than single markers—best capture the spatial expression patterns of brain regions, especially in complex or poorly localized regions.