Genetic-linkage mapping of complex hereditary disorders to a whole-genome molecular-interaction network

Abstract

Common hereditary neurodevelopmental disorders such as autism, bipolar disorder, and schizophrenia are most likely both genetically multifactorial and heterogeneous. Because of these characteristics traditional methods for genetic analysis fail when applied to such diseases. To address the problem we propose a novel probabilistic framework that combines the standard genetic linkage formalism with whole-genome molecular-interaction data to predict pathways or networks of interacting genes that contribute to common heritable disorders. We apply the model to three large genotype-phenotype data sets, identify a small number of significant candidate genes for autism (24), bipolar disorder (21), and schizophrenia (25), and predict a number of gene targets likely to be shared among the disorders.

Figure 1.

An example of genetic-linkage data used as the input to our analysis and the resulting network of top-scoring genes for the three disorders. (A) Standard multipoint-linkage analysis of human chromosome 15 for 94 schizophrenia families (schizophrenia-no-x analysis). Each line above the chromosome map represents the linkage signal for one family. Also shown are the positions of genetic markers on the chromosome map and the set of top-scoring candidate genes. In this case, four genes (CYFIP1, UBE3A, OCA2, and TJP1) have significant linkage statistics. (B) The molecular network obtained by superimposition of the 70 best 10-gene clusters for each of the three disorders analyzed in this study (autism-no-x, bipolar-no-x, and schizophrenia-no-x analyses). Arrows indicate two genes (UBE3A and TJP1) discussed in the main text. Note that the COMA and TAM genes are not yet approved by HUGO (see the genes_not_in_HUGO.xls supplemental file).

Figure 2.

Analysis of the 14 top-scoring 10-gene clusters for the schizophrenia data (schizophrenia-no-x). (A) Each cluster is shown separately, where the vertex size represents the cluster probability estimated for the corresponding gene. We used the color of the cluster to encode cluster LOD scores. (B) Position of all genes represented in the 14 clusters on human autosomes. (C) Molecular network combining the 14 clusters in one graph. In this case, the color and size of nodes indicate gene-specific P-values associated with each gene.

Figure 3.

Analysis of the 14 top-scoring 10-gene clusters for the autism data (autism-x-rec); see Figure 2 for explanation of panels A–C. Note that the PKC and TAM genes are not yet approved by HUGO (see the genes_not_in_HUGO.xls supplemental file).

Figure 4.

Analysis of the 14 top-scoring 10-gene clusters for the bipolar disorder data (bipolar-no-x); see Figure 2 for explanation of panels A–C.

Figure 5.

Molecular networks combining the 100 best 10-gene clusters for autism (A) and bipolar (B) disorder and the 50 best 10-gene clusters for schizophrenia (C). The color and size of nodes in all three networks indicate gene-specific P-values (autism-no-x, bipolar-no-x, and schizophrenia-no-x analyses). Note that the LOC347422, PKC, TAM genes are not yet approved by HUGO (see the genes_not_in_ HUGO.xls supplemental file).