Network-based drug repurposing for schizophrenia

Despite recent progress, the challenges in drug discovery for schizophrenia persist. However, computational drug repurposing has gained popularity as it leverages the wealth of expanding biomedical databases. Network analyses provide a comprehensive understanding of transcription factor (TF) regulatory effects through gene regulatory networks, which capture the interactions between TFs and target genes by integrating various lines of evidence. Using the PANDA algorithm, we examined the topological variances in TF-gene regulatory networks between individuals with schizophrenia and healthy controls. This algorithm incorporates binding motifs, protein interactions, and gene co-expression data. To identify these differences, we subtracted the edge weights of the healthy control network from those of the schizophrenia network. The resulting differential network was then analysed using the CLUEreg tool in the GRAND database. This tool employs differential network signatures to identify drugs that potentially target the gene signature associated with the disease. Our analysis utilised a large RNA-seq dataset comprising 532 post-mortem brain samples from the CommonMind project. We constructed co-expression gene regulatory networks for both schizophrenia cases and healthy control subjects, incorporating 15,831 genes and 413 overlapping TFs. Through drug repurposing, we identified 18 promising candidates for repurposing as potential treatments for schizophrenia. The analysis of TF-gene regulatory networks revealed that the TFs in schizophrenia predominantly regulate pathways associated with energy metabolism, immune response, cell adhesion, and thyroid hormone signalling. These pathways represent significant targets for therapeutic intervention. The identified drug repurposing candidates likely act through TF-targeted pathways. These promising candidates, particularly those with preclinical evidence such as rimonabant and kaempferol, warrant further investigation into their potential mechanisms of action and efficacy in alleviating the symptoms of schizophrenia.


Methods
Replication with CommonMind Consortium -HBCC Brain Bank cohort -RNA sequencing data: Dorsolateral prefrontal cortex (DLPFC) RNA sequencing data were accessed from the CommonMind Consortium [1].After quality control, a total of 237 post-mortem samples belonging to the HBCC Brain Bank were collected from 156 unaffected control subjects and 81 people with schizophrenia.Genes being expressed at more than 0.5 count per million (CPM) in at least 30% of samples were kept for downstream analyses.
To generate input for the co-expression network, the R package variancePartition (version 1.26.0) was used to create expression residuals [2].Covariates (i.e., diagnosis, sex, RNA integrity number, cell type composition, age of death, intronic rate, intragenic rate, intergenic rate, and ribosomal RNA rate) were regressed out from the full model (i.e., excluded effects).Then the main variable of diagnosis and the intercept were added back in to pertain residuals for the downstream comparison (schizophrenia versus controls).The expression residuals underwent pre-processing (removal of genes with no counts, taking the average of duplicated genes), before being calculated for coexpression using Pearson correlations for network construction.

-Gene co-expression regulatory networks
The R package PANDA was used to build the bipartite gene regulatory network that linked transcription factors (TFs) to their target genes [3].PANDA integrates three sources of information to infer the TFgene regulatory network: TF physical protein-protein interactions (TF -TF links), gene co-expression (gene -gene links) and TF motif binding sites (TF -gene links) [3].
TF protein-protein interactions (PPI) were obtained from the STRING database [4].A threshold of 0.7 (high confidence) was applied to the combined score to convert the score to binary (0 implies no interaction and 1 implies high likelihood of interaction).Binding motifs were acquired from previous studies [5,6], where TF binding domain sequences (i.e., motifs) were scanned for their presence in the promoter regions of genes where transcription initiates.
Expression residuals, TF PPI and binding motifs were inputted in PANDA with the following non-default parameters to make sure only mutual connections shared by PPI, co-expression and TF motifs were considered in the networks: mode = "legacy", remove.missing.motif= True, remove.missing.ppi= True, remove.missing.genes= True.Two separate regulatory networks were built for schizophrenia cases and unaffected control subjects.Edge weight of each network implied the strength of connection of TFs and genes, reflected via Pearson's correlation coefficient between the TF and the target gene.
-Differential schizophrenia network and finding drug repurposing candidates To find the differences in regulation in schizophrenia patients as compared to unaffected control subjects, the two corresponding regulatory networks were first aligned and filtered to keep intersections of genes and TFs only.Then the differential network was estimated by subtracting the edge weights of the unaffected control network from those of schizophrenia network.The 100 top positively differential TFs and 100 top negatively differential TFs based on the differential targeting score were submitted to the CLUEreg tool of the GRAND database [32] which utilises differential network signatures to find drugs that potentially target the disease's gene signature.

Replication with PsychENCODE BrainGVEX cohort
-RNA sequencing data: Dorsolateral prefrontal cortex (DLPFC) RNA sequencing data were accessed from the PsychENCODE [7].After quality control, a total of 364 post-mortem samples belonging to the BrainGVEX cohort were collected from 259 unaffected control subjects and 95 people with schizophrenia.Genes being expressed at more than 0.1 transcripts per million (TPM) in at least 25% of samples were kept for downstream analyses.
To generate input for the co-expression network, the R package variancePartition (version 1.26.0) was used to create expression residuals [2].Covariates (i.e., diagnosis, sex, RNA integrity number) were regressed out from the full model (i.e., excluded effects).Then the main variable of diagnosis and the intercept were added back in to pertain residuals for the downstream comparison (schizophrenia versus controls).The expression residuals underwent pre-processing (removal of genes with no counts, taking the average of duplicated genes), before being calculated for coexpression using Pearson correlations for network construction.

-Gene co-expression regulatory networks
The R package PANDA was used to build the bipartite gene regulatory network that linked transcription factors (TFs) to their target genes [3].PANDA integrates three sources of information to infer the TFgene regulatory network: TF physical protein-protein interactions (TF -TF links), gene co-expression (gene -gene links) and TF motif binding sites (TF -gene links) [3].
TF protein-protein interactions (PPI) were obtained from the STRING database [4].A threshold of 0.7 (high confidence) was applied to the combined score to convert the score to binary (0 implies no interaction and 1 implies high likelihood of interaction).Binding motifs were acquired from previous studies [5,6], where TF binding domain sequences (i.e., motifs) were scanned for their presence in the promoter regions of genes where transcription initiates.
Expression residuals, TF PPI and binding motifs were inputted in PANDA with the following non-default parameters to make sure only mutual connections shared by PPI, co-expression and TF motifs were considered in the networks: mode = "legacy", remove.missing.motif= True, remove.missing.ppi= True, remove.missing.genes= True.Two separate regulatory networks were built for schizophrenia cases and unaffected control subjects.Edge weight of each network implied the strength of connection of TFs and genes, reflected via Pearson's correlation coefficient between the TF and the target gene.
-Differential schizophrenia network and finding drug repurposing candidates To find the differences in regulation in schizophrenia patients as compared to unaffected control subjects, the two corresponding regulatory networks were first aligned and filtered to keep intersections of genes and TFs only.Then the differential network was estimated by subtracting the edge weights of the unaffected control network from those of schizophrenia network.The 100 top positively differential TFs and 100 top negatively differential TFs based on the differential targeting score were submitted to the CLUEreg tool of the GRAND database [32] which utilises differential network signatures to find drugs that potentially target the disease's gene signature.

Overlapping significant repurposing candidates (q-value < 0.05) from the top 100 repurposing candidates of each dataset between three datasets: CommonMind Consortium -MSSM -Pitt -Penn Brain Bank (CMC_MPP_current), CommonMind Consortium -HBCC Brain Bank (CMC_HBCC), PsychENCODE -BrainGVEX Repurposing candidates from analyses on CommonMind Consortium -HBCC Brain Bank cohort
t it u t io n I n s t it u t io n : c e ll F r a c _ il r _ 1 s c a le ( I n t r o n ic R a t e ) s c a le ( I n t r a g e n ic R a t e ) I n s t it u t io n : a g e O f D e a t h I n s t it u t io n : c e ll F r a c _ il r _ 2 R I N s c a le ( I n t e r g e n ic R a t e ) c e ll F r a c _ il r _ 3 c e ll F r a c _ il r _ 2 D x a g e O f D e a t h I n s t it u t io n : c e ll F r a c _ il r _ 3 s c a le ( r R N A R a t e ) E t h n ic it y c e ll F r a c _ il r _ 1 D x : R e p o r t e d _ G e n d e r P M I R e p o r t e d _ G e n d e r R e s id u a ls Variance explained (%)