The cell-type specific interaction based drug repurposing for psychiatric disorders

Abstract

Psychiatric disorders severely challenge the public health, but the development of new psychotropic medications cannot keep pace with clinical needs. The rapid accumulation of single-cell RNA data provides opportunities to develop new computational approaches of drug repurposing. However, there is a lack of dedicated computational tools for psychiatric disorders in this field. In this work, we developed a cell-type specific interaction based computational pipeline scPsyDrug to prioritize the drugs for psychiatric disorders. This tool integrated the single-cell transcriptomics, protein-protein interactions, drug-target interactions and psychiatric risk genes to construct cell type-specific interactions and seed genes for drug prioritization. We first evaluated the scPsyDrug performance using the publicly available single cell RNA datasets derived from the clinical patients, including Major Depressive Disorder, Schizophrenia and Parkinson’s disease, in which scPsyDrug covered lots of approved drugs and clinical-trial compounds in the top recommended candidates. We next applied scPsyDrug to Zbtb18^+/− mice, an anxiety-like mouse model, to screen the potential drugs. The scPsyDrug prioritized Forskolin and Chlorpropamide as the top candidates, while the following drug treatment demonstrated that these two compounds could respectively rescue the anxiety-like behaviors in Zbtb18^+/− mice. We also used scPsyDrug to screen herb ingredients, and validated the efficacy of Osthole in rescuing the anxiety-like behaviors in the mouse model. Collectively, the scPsyDrug can be applied to different kinds of psychiatric disorders for drug repurposing, and we also provided experimental evidence for the potential roles of Forskolin, Chlorpropamide and Osthole in anxiety intervention.

Introduction

The prevalence of psychiatric disorders substantially rises in recent years, greatly challenging public health. Despite the increased social impact, the global disease burden imposed by psychiatric disorders remains significantly underestimated [1, 2]. Meanwhile, the development of new psychotropic medications has progressed slowly and has not kept pace with clinical needs [3], and many patients experience limited efficacy or develop resistance to existing treatments, leading to an urgent demand for innovative therapeutic strategies. Drug repurposing has emerged as a particularly promising approach to leverage the existing pharmacological and clinical knowledge to accelerate the therapeutic rediscoveries for approved or previously characterized compounds. The growing interest in drug repurposing for psychiatric disorders reflects the substantial challenges that hinder traditional drug development, including long development pipelines, high financial investment, and a high failure rate. These challenges are further complicated by the complex and heterogeneous pathophysiology of psychiatric disorders and the inherent difficulties associated with target identification and validation, which collectively contribute to the low success rate of psychotropic drug development [4].

Currently, the rapid growth of high-throughput omics data, especially the single-cell RNA sequencing (scRNA-seq), has greatly accelerated the data-driven drug repurposing approaches. By utilizing the large-scale multi-omics datasets, researchers are now able to link gene expression signatures with disease-associated molecular mechanisms, thereby identifying compounds with potential antipsychotic or neuro-modulatory effects [5]. For example, Gupta et al. integrated single-cell transcriptomic profiles from the prefrontal cortex of patients with bipolar disorder, autism spectrum disorder, and schizophrenia to construct transcriptional regulatory network modules for graph neural network-based prioritization of candidate risk genes, ultimately identifying drugs associated with these risk modules and genes [6]. Cheng et al. analyzed RNA-seq profiles from seven brain regions of patients with Alzheimer’s disease to construct region-specific co-expression networks for identifying potentially driver genes, and then used the transcriptional reversal strategy to screen compounds capable of reversing disease-related gene expressions [7]. Similarly, Truong et al. constructed transcription factor-target regulatory networks using postmortem brain RNA-seq data from individuals with schizophrenia, and employed differential network signatures to prioritize potential therapeutic candidates [8].

Despite the success of utilizing scRNA-seq to construct gene regulatory networks or co-expression modules for drug repurposing, these works did not provide general software, making it difficult to directly use and evaluate these methods. In addition, the protein-protein interaction (PPI) networks were widely used to elucidate the propagation of drug effects across molecular pathways since the approved or investigational drugs mainly target the receptors and/or kinases involved in protein-protein interactions and signal transductions. However, there remains a lack of software that integrates the scRNA-seq data and PPI networks for drug repurposing oriented toward psychiatric disorders. These factors collectively inspired us to develop the computational pipeline scPsyDrug to integrate scRNA-seq, PPI network, drug-target interactions and psychiatric risk genes to prioritize approved drugs, investigational compounds and herb ingredients for psychiatric disorders at cell-type specific resolution. We first evaluated the applicability of scPsyDrug in different kinds of psychiatric disorders by using publicly available scRNA-seq data derived from patients, and clinical knowledge on approved and clinical-trial drugs. We also established an anxiety-like mouse model to screen potentially repurposed drugs by using scPsyDrug, and we found and validated that the Forskolin, Chlorpropamide and Osthole had the potentials for anxiety intervention.

Materials and methods

Public datasets for scPsyDrug

We downloaded the human PPI network constructed by Cheng et al. [7] as the reference PPI network to build cell-type specific interactions for single cell transcriptomics (Table S1). We extracted all drug-target interaction pairs from the DrugBank database (released on 2025/01/02) [9], and retained only those annotated for human or mouse. The herb ingredients and their targets were obtained from the HERB resource [10]. KEGG Drug IDs were retrieved for each drug to obtain corresponding drug classification information from the KEGG MEDICUS repository (https://www.genome.jp/ftp/kegg/medicus/drug/). Blood-brain barrier (BBB) permeability of drugs was evaluated using admetSAR 3.0 [11], and only compounds with BBB ≥ 0.7 were retained. Approved psychiatric medications were exempted from BBB evaluation and were directly included in the curated dataset. The psychiatric disorder risk (PSD) genes were collected from Gene2Phenotype (G2P) [12], OMIM [13], SFARI Gene [14] as well as literature search [15], yielding a total of 4148 PSD genes (Table S2). For drugs explicitly documented as being approved for the major depressive disorder (MDD), schizophrenia (SCZ), or Parkinson’s disease (PD) in the DrugBank database, we annotated them as approved drugs for MDD, SCZ, and PD, respectively. By contrast, compounds documented as ongoing or completed clinical trials for MDD, SCZ, or PD were classified as investigational compounds for the treatment of MDD, SCZ, and PD, respectively.

As for the input single cell data used in model evaluation, the single nucleus RNA sequencing (snRNA-seq) derived from the prefrontal cortex of MDD patients [16] and PD patients carrying LRRK2 variants [17] were obtained from the GEO database (GSE144136 and GSE272760, respectively), and scRNA-seq data derived from SCZ patients were downloaded from the Zenodo repository (10.5281/zenodo.6583332) [18]. In addition, we collected bulk RNA datasets for MDD from GSE102556 [19, 20], SCZ from GSE87194 [21], and PD from GSE136666 [22]. All datasets were downloaded in original matrix format with annotated cell types. The count matrices from MDD, PD and SCZ were respectively normalized using NormalizedData function followed by log scale conversion, and the 3000 highly variable genes were selected using FindVariableFeatures function and the features were scaled and centered using ScaleData function with default parameters.

The computational pipeline of scPsyDrug

In overall, the scPsyDrug utilized scRNA-seq data derived from control and disease groups as input, which were automatically integrated to reference PPI networks, drug-target interactions and PSD genes for drug prioritization (Fig. 1). For the input cell types and gene expressions derived from users, the differentially expressed genes (DEGs) and PPIs were subjected to cell-type specific filtering. For cells from disease group, the cell-type specific PPI networks were constructed in a curated reference PPI network, in which the differentially expressed PSD genes were designated as seed genes. Network modules significantly enriched for these seed genes were identified via a hypergeometric test. Then the network proximity algorithm was used to quantify the shortest path distances between drug targets and module seed genes in each cell-type specific network. For a given compound, the mean network proximity score between its all targets and one network module was computed, and then the proximity scores across all modules were averaged. After this calculation, the proximity scores among all compounds were normalized by using Z-score. Finally, each drug was assigned a drug repurposing score to rank the compounds for drug repurposing potentials.

Fig. 1. Computational pipeline of scPsyDrug.

Single cell transcriptomics were integrated with reference protein-protein interactions to construct cell-type specific networks. The cell-type specific differentially expressed genes (DEGs) and psychiatric disease risk (PSD) genes were used to define seed genes in cell-type specific network. Network proximity between drug targets and the seed genes were calculated in each cell-type specific module, which was used to rank the drugs.

Specifically, for each cell type derived from scRNA-seq data, genes expressed in fewer than 10% of cells were removed. The reference PPI network was subsequently filtered by excluding gene pairs in which the expression proportions of both genes were below 5%. For the remaining genes, co-expression patterns were estimated using SCINET [15], and cell-type specific PPI networks were constructed by mapping these co-expression relationships onto the reference interactions. Community detection was performed using the Louvain algorithm [23] to identify network modules. Differentially expressed PSD genes for each cell type were designated as seed genes. To mitigate noise arising from highly dispersed seed genes, we applied a hypergeometric enrichment test to identify modules significantly enriched for seed genes:

$$P\left(X=k\right)=\frac{\left(\begin{array}{c}\hfill M\hfill \\ \hfill k\hfill \end{array}\right)\left(\begin{array}{c}\hfill N-M\hfill \\ \hfill n-k\hfill \end{array}\right)}{\left(\begin{array}{c}\hfill N\hfill \\ \hfill n\hfill \end{array}\right)}$$ 1

Modules with a P value < 0.05 were considered significantly enriched for the seed genes.

We quantified the network proximity of each drug target to each module by iteratively calculating the sum of the shortest path lengths from the target to all seed genes within the module:

$${Prox}_{tm}=\frac{{\sum }_{s\in S_m}d\left(s,t\right)+{\sum }_{t\in T}{min}_{s\in S_m}d\left(s,t\right)}{\mathrm{\mid \mid }{S}_m\mathrm{\mid \mid }+\mathrm{\mid \mid }T\mathrm{\mid \mid }}$$ 2

where T denotes the drug target set, S_m denotes the seed gene set in module m, d(s,t) denotes the shortest path length between drug target t and seed gene s, and ||S_m|| and ||T|| denote the seed gene number in module m and drug target number in target set T. Then the adjusted proximity score ZProx_tm was computed as:

$$Z{Prox}_{tm}=\frac{w_t{Prox}_{tm}-mean\left(w_t{Prox}_{tm}\right)}{\mathrm{s}d\left(w_t{Prox}_{tm}\right)}$$ 3

$$w_t=\frac{1-s*{priority}_t}{\mathrm{m}ean\left(1-s*{priority}_t\right)}$$ 4

where mean and sd denote the mean value and standard deviation operations respectively. In Eq. 4, each drug target was assigned a different weight w_t by considering its properties. Specifically, if the drug target is both the PSD gene and differentially expressed gene, the priority was set to 1. If the drug target is a PSD gene but not differentially expressed gene, the priority was set to 0.5. If the target is neither a PSD gene nor differentially expressed gene, the priority was set to 0. The shrinkage coefficient s was set to 0.2 in Eq. 4. We used ZProx_tm < = −1.5 to select drugs in this work. For cell types with the distributions shifted toward high values, the ZProx_tm < = −1.0 can also be used.

Since one drug can target multiple proteins, the repurposing score DRScore of one drug was calculated by averaging the network proximity of all drug targets:

$$DRScore=\frac{{\sum }_{m\in M}{\sum }_{t\in T_d}{ZProx}_{tm}}{\mathrm{\mid \mid }{T}_d\mathrm{\mid \mid }\cdot \mathrm{\mid \mid }M\mathrm{\mid \mid }}$$ 5

where T_d denotes the target set for drug d, M denotes the seed gene modules, and ||T_d|| and ||M|| denote the number of drug targets in T_d and the number of seed gene modules respectively.

Method comparison

We compared scPsyDrug to three signature-reversal drug repurposing tools, WebCMap [24], metaLINCS [25], and ConnectivityMap [26], by using snRNA-seq datasets from MDD, SCZ, and PD. The default parameters recommended by each tool were used, and we employed the same cell-type specific DEGs as input for all four methods. The MeanRank, which is defined as the average rank of identified FDA-approved drugs and clinical-trial compounds, was used as metric to measure the performance. Thus, lower MeanRank value indicates higher top ranks in drug prioritization. We also performed comparison between scRNA-seq and bulk RNA-seq data in our model. To mimic the network construction in scPsyDrug, we used bulk RNA-seq derived from patients to calculate the Pearson correlation coefficients between two genes for each interaction edge in the reference PPI, and the interactions with absolute Pearson correlation coefficient greater than 0.7 were retained for network construction. The other computational steps were the same as scRNA-seq.

Mice experiments and behavioral tests

The Zbtb18^+/– mice (GemPharmatech, Nanjing, China; strain no. T033873) were housed in the SPF barrier facility of the Laboratory Animal Center of Yunnan University. Mice were maintained under standard environmental conditions, including a temperature of 20–26 °C, relative humidity of 40–70%, and a 12 h light/12 h dark cycle. Mice were euthanized by carbon dioxide (CO₂) inhalation in accordance with institutional animal care guidelines, and the prefrontal cortex and the hippocampi were dissected carefully for snRNA-seq and RT-qPCR.

The Zbtb18⁺/⁻ male mice were randomly assigned to two groups to receive drug treatment and vehicle control, respectively. In all three drugs, the treatments were conducted around from the start of week 11 to the end of week 12, once daily for exactly 14 consecutive days. Forskolin (Selleck, S2449) was dissolved in 5% Dimethyl sulfoxide (DMSO) to a concentration of 0.25 mg/mL and administered via oral gavage at a dose of 1.05 mg/kg once daily, while 5% DMSO was used as the vehicle control. Chlorpropamide (Selleck, S4166) was suspended directly in corn oil at a concentration of 4 mg/mL and administered via oral gavage at a dose of 32.5 mg/kg once daily, with corn oil serving as the vehicle control. Osthole (Selleck, S2337) was suspended directly in corn oil at 15 mg/mL and administered via oral gavage at a dose of 100 mg/kg once daily, while corn oil was used as the vehicle control.

In the light-dark box test, the male Zbtb18^+/+ and Zbtb18^+/– mice (day 13 in drug administration) were acclimated to the testing room for at least 3 h. Each mouse was then placed in the white compartment of the light-dark box. After removing the partition between the light and black compartments, mice were allowed to freely explore both compartments for 5 min. Mouse activity was recorded to analyze the total distance traveled, number of entries into the light compartment, time spent in the light compartment, distance traveled within the light compartment, and number of transitions between compartments. In the open field test, the male Zbtb18^+/+ and Zbtb18^+/– mice (day 14 in drug administration) were acclimated to the testing room for at least 3 h. The open field area was divided into 16 equal squares, with the central four squares designated as the central zone and the remaining squares as the peripheral zone. Each mouse was placed in the central zone and allowed to freely explore all zones for 10 min. Mouse activity was recorded to analyze the total distance traveled, number of entries into the central zone, time spent in the central zone, and distance traveled within the central zone. The 6 Zbtb18^+/+ mice treated with DMSO were used as vehicle control for the Forskolin treatment, while 7 mice treated with corn oil were used as vehicle control for both the Chlorpropamide and Osthole treatments since these two drugs were dissolved in the same corn oil in drug treatments. In Forskolin test, 7 Zbtb18^+/- mice were treated with DMSO as vehicle control, and 8 Zbtb18^+/- mice were treated with Forskolin which was dissolved in DMSO. In Chlorpropamide test, 6 Zbtb18^+/- mice were treated with corn oil as vehicle control, while 6 Zbtb18^+/- mice were treated with corn-oil-dissolved Chlorpropamide. As for Osthole, 6 Zbtb18^+/- mice were treated with vehicle control, and 8 Zbtb18^+/- mice were treated with corn-oil-dissolved Osthole.

RNA extraction and RT-qPCR

RT-qPCR Total RNA was extracted with the TransZol Up Plus RNA Kit (TransGen Biotech, ER501-01-V2) from these regions. Total RNA was converted to cDNA with the HiScript II Q RT SuperMix for qPCR Kit (Vazyme, R223-01). cDNAs were used for quantitative real-time PCR, which was performed using the ChamQ Universal SYBR qPCR Master Mix (Vazyme, Q711-02) in the QuantStudio^TM 5 Real-Time PCR System. All samples were assayed in triplicate. The condition of PCR was the following: 95 °C for 30 s, followed by 40 cycles of 95 °C for 10 s, and 60 °C for 30 s. The Actb gene served as the internal reference, and relative gene expression levels between groups were calculated using the ΔΔCt method (Table S3). For Adcy1 and Abcc8, 6 Zbtb18^+/+ mice and 5 Zbtb18^+/- mice were treated with vehicle control, and 6 Zbtb18^+/- mice were treated with Forskolin or Chlorpropamide treatment. As for Parp1, the sample sizes were 5 Zbtb18^+/+ mice with vehicle control, 5 Zbtb18^+/- mice with vehicle control, and 6 Zbtb18^+/- mice with Osthole treatment.

The snRNA sequencing on Zbtb18 mice

Following euthanasia, brains from 12 weeks old male Zbtb18^+/– and Zbtb18^+/+ mice were rapidly dissected on ice, and the hippocampus and prefrontal cortex were isolated. Tissues were immediately snap-frozen in liquid nitrogen and stored at −80 °C until use. Nuclei extraction and cDNA library construction were performed by the company following the manufacturer’s protocol of the MGI Single-Nucleus RNA Library Preparation Kit V3.0. The paired-end sequencing was performed on the MGISEQ-2000 platform. Raw reads were processed using dnbc4tools (v2.1.0), aligned to the mouse reference genome GRCm38, and subjected to rna run processing for Unique Molecular Identifier (UMI) recognition and quantification. Low-quality reads were removed using default parameters, and the gene expression matrices were then generated. In snRNA-seq, the hippocampal and cortical tissues derived from three Zbtb18^+/+ mice were pooled respectively, and the same tissues derived from three Zbtb18^+/- mice were also pooled for mutant group. Three technical replicates were performed for both Zbtb18^+/+ group and Zbtb18^+/- groups.

Clustering and cell-type annotation

Data preprocessing was performed using the Seurat R package (v5.0.3) [27]. Cells were first filtered using the criteria 100 < nFeatures < 7500 and 200 < nCount < 20,000. The mitochondrial-encoded genes in the expression matrix were removed. Data from the hippocampus and prefrontal cortex were processed independently. NormalizedData function was used to normalize UMI counts to 10,000 counts per cell, followed by log scale conversion. The genotype annotation was added to metadata of the Seurat object. The 3000 highly variable genes were selected by FindVariableFeatures function, and the features were scaled and centered by ScaleData function with default parameters. RunPCA function was used for dimensionality reduction with the first 50 principal components (PCs).

To align shared cell types and states between samples and enable downstream cross-genotype comparisons, nuclei from Zbtb18^+/– and Zbtb18^+/+ mice were integrated using the IntegrateLayers function with the HarmonyIntegration method. Neighborhood relationships were computed using FindNeighbors (default parameters), and clusters were identified using FindClusters with a resolution of 1.7 for hippocampal nuclei and 1.4 for cortical nuclei by using the Shared Nearest Neighbor (SNN) modularity optimization algorithm. Following unsupervised clustering, cell-type identities were assigned for each cluster. Marker genes were computed using FindMarkers (min.pct = 0.1, logfc.threshold = 1). Cell-type annotation was guided by transcriptional profiles reported by Yao et al. for the mouse hippocampus and cortex [28], together with reference marker sets curated in the Allen Brain Atlas [29]. Marker genes satisfying log2FC > 1 and adjusted p-value < 0.05 were compared against reference marker gene panels to assign each cluster to its corresponding cell type.

Differential expression and enrichment analysis

Differentially expressed genes (DEGs) for each cell cluster from MDD, SCZ, and PD patients, as well as Zbtb18^+/– mice, were identified using the FindAllMarkers function (min.pct = 0.1, test.use = “wilcox”, logfc.threshold = 0.2 for Zbtb18^+/– datasets and logfc.threshold = 0.1 for public MDD, SCZ, and PD datasets). We used a relatively lower cutoff (|log2FC| ≥ 0.1) to retain modest but potentially meaningful signals for public MDD, SCZ and PD since the datasets derived from the postmortem of human patients may vary substantially across individuals and cell types. The more stringent cutoff (|log2FC| ≥ 0.2) was used in Zbtb18^+/– mice since the snRNA-seq data derived from fresh mouse sample have higher sequencing quality. The fold change log2FC > 0 indicated upregulation in MDD, SCZ, PD, or Zbtb18^+/– samples, whereas log2FC < 0 indicated downregulation. Functional enrichment analyses of DEGs were performed using clusterProfiler (v4.10.1) [30], including KEGG pathway enrichment [31] and Gene Ontology (GO) enrichment analyses [32]. To reduce redundancy among GO terms with high semantic similarity, the simplify function was applied with default parameters.

Statistical analysis and visualization

All statistical analyses were performed using R. We found that the behavioral and qPCR data did not always follow the normal distribution and there was heterogeneity of variance across groups. Therefore, the Wilcoxon rank-sum test was used for statistical comparisons. The permutation test was used for the network proximity comparison between cell type A and cell type B. Since the targets identified in cell type A and cell type B may not completely overlap, the missing proximity values in each cell type were imputed using the maximum proximity in cell type A and cell type B. Subsequently, the proximity values of shared targets between cell type A and cell type B were randomly permuted 10,000 times to calculate the null distribution, and then the one-sided empirical p-value was computed for the observed mean network proximity difference between cell type A and cell type B. The center line of the boxplot represents the median. Network visualization was performed using Cytoscape (v3.10.2) [33], and other data visualizations were generated using R.

Results

scPsyDrug well prioritized therapeutic compounds for psychiatric disorders

By systematically integrating the curated resources, we collected a compendium of 4148 PSD genes (Fig. 2A, Table S2) and human PPI interactions comprising 15,931 proteins and 217,016 interactions (Fig. 2A, Table S1) to serve as the reference seed genes and PPI network throughout our analyses. Drug-target pairs were obtained from DrugBank (3614 compounds and 17,624 interaction pairs) and HERB (191 traditional Chinese medicines and 6205 interaction pairs) databases. In the DrugBank derived repertoire, we further identified 43 FDA-approved drugs and 109 investigational compounds in clinical development for MDD (Fig. 2A, Table S4), 32 FDA-approved antipsychotic drugs and 80 clinical-stage candidates for SCZ (Fig. 2A, Table S5), as well as 25 FDA-approved drugs and 55 investigational drugs for PD (Fig. 2A, Table S6).

Fig. 2. The scPsyDrug evaluations on public datasets.

A Summary of data resources. B–D Evaluations on Major Depressive Disorder (MDD), Schizophrenia (SCZ), Parkinson’s disease (PD) by using scRNA-seq datasets derived from patients, respectively. In each subfigure, the left and right panels present the number and percentage of drugs identified by scPsyDrug.

We then applied scPsyDrug to the snRNA-seq dataset derived from the prefrontal cortex of MDD patients, and observed that, among the top 100 ranked drugs, scPsyDrug successfully captured 21 FDA-approved antidepressant drugs and 34 investigational compounds in clinical trials for MDD (Fig. 2B), with layer 4 excitatory neurons (ExN L4) contributing the largest number of candidates: 14 approved and 5 clinical-stage compounds, respectively (Figure S1). Similarly, scPsyDrug retrieved 16 FDA-approved antipsychotics and 9 investigational compounds for SCZ within the top 60 candidates (Fig. 2C, Figure S3). In addition, we applied scPsyDrug to the snRNA-seq dataset generated from the prefrontal cortex of PD patients. In the PD cohort, among the top 40 ranked drugs, the algorithm retrieved 14 FDA-approved drugs and 16 investigational compounds for PD (Fig. 2D, Figure S5). In addition, we compared scPsyDrug to the single-cell signature-reversal methods, and the results showed that the FDA-approved drugs and investigational compounds identified by scPsyDrug exhibited lower MeanRank than those identified by other methods (Figure S1A, Figure S2C and Figure S4C), suggesting that scPsyDrug achieves good performance in single-cell based drug repurposing.

scPsyDrug enabled cell-type resolved prioritization for psychiatric disorders

To investigate whether the prioritization scores produced by scPsyDrug exhibited cell-type dependent variation, we first performed drug repurposing by pooling all cells together, but only 6 drugs indicated for MDD were retrieved within the top 80 ranked drugs. By contrast, when the analysis was resolved at the cell-type specific level, scPsyDrug identified 19 and 12 MDD drugs within the top 80 candidates in ExN L4 and ExN L2, respectively (Fig. 3A and B). In addition to pooling all cells, we also compared the scRNA-seq and bulk RNA-seq data for drug prioritization, and the results indicated better performance of single-cell data in drug repurposing (Figure S4E). Notably, we observed that a substantial fraction of approved and investigational MDD drugs were selectively prioritized in ExN L4 via the muscarinic acetylcholine receptor CHRM1 (Fig. 3C, Table S7), a muscarinic acetylcholine receptor that is predominantly expressed in the prefrontal regions to mediate excitatory cholinergic neurotransmission [34]. Among these, the FDA-approved antidepressant Amitriptyline and Ketamine, a rapid antidepressant agent in treatment-resistant MDD patients [35], were also prioritized.

Fig. 3. The scPsyDrug facilitated the cell-type specific drug discoveries in MDD.

A The prioritized drugs for MDD treatment (approved and investigational) in the top 80 candidates. ExN L2, ExN L4, ExN L5 excitatory neurons of cortical layers 2, 4, and 5, InN inhibitory neurons, ODCs oligodendrocytes, OPCs oligodendrocyte precursor cells, Astro astrocytes. B Drug numbers prioritized by all cells and different cell types in the top 80 candidates. C, D Drug-target networks derived from ExN L4 and Microglia, respectively. E Target proximity values between ExN L4 and other cells types. The top 10 drugs in ExN L4 were used as comparison. *0.01 ≤ p < 0.05; **0.001 ≤ p < 0.01; ***p < 0.001; ns p ≥ 0.05.

Beyond excitatory neurons, we also investigated the microglia populations, and some MDD-relevant compounds were prioritized (Fig. 3D, Table S7). For example, scPsyDrug discovered Imipramine (Table S7), a prototypical tricyclic antidepressant with KCND2 (Kv4.2) proposed as putative target [36], and Curcumin, which is annotated to target APP, was also prioritized. In a randomized, double-blind, and placebo-controlled clinical trial involving 56 MDD patients, Lopresti et al. demonstrated that Curcumin produced greater improvements in mood symptoms than placebo, thereby underscoring its potential as a candidate therapeutic for MDD [37].

To further demonstrate the cell-type specificity inherent in scPsyDrug, we evaluated the top 10 targets (ranked by DRScore) identified in ExN L4 across all cell populations. The result showed that these targets exhibited significantly higher absolute network proximity in ExN L4 than other cell types except ExN L2 (Fig. 3E), consistent with the higher numbers of detected drugs in ExN L4 and ExN L2. In addition, Clozapine was identified by CHRM1, while the clinical-trial compound DAS-431 IV was detected through CALY in excitatory neuron subpopulations (ExN L2, and ExN L4) of SCZ (Figure S2A-B, Figure S3, Table S8). Likewise, in PD neuronal subtypes (ExN 1, ExN 2), the CHRM1-centric network context facilitated the selective prioritization of the FDA-approved PD drugs Benzatropine, Biperiden and Profenamine, as well as the clinical-stage agents Caramiphen, Cycrimine, Doxylamine and Metixene (Figure S4A-B, Figure S5, Table S9).

Screening potentially repurposed drugs in Zbtb18^+/- mouse model

We next used the scPsyDrug to computationally screen the repurposed drugs in Zbtb18 haploinsufficient (Zbtb18^+/-) mice, an anxiety-like mouse model, followed by experimental validations. In our work, we found that Zbtb18^+/− mice exhibited anatomical deficits, including reduced brain length and width as well as decreased body and brain weight, compared to Zbtb18^+/+ controls (Fig. 4A, Figure S6A). Behavioral tests further revealed anxiety-like phenotypes: in the light-dark box test, Zbtb18^+/− mice showed a tendency toward reduced total locomotor distance, and exhibited significantly decreased distance traveled in the light compartment, fewer entries into the light compartment, and fewer transitions between the light and dark compartments (Fig. 4B, C). Consistently, in the open field test, Zbtb18^+/− mice showed reduced percentage of distance traveled in the central zone, fewer entries into the center, and reduced time spent in the center (Fig. 4D, E). Collectively, these tests indicated that the Zbtb18^+/− mice exhibited anxiety-like behaviors.

Fig. 4. Zbtb18^+/- mice exhibited abnormal brain morphology and anxiety-like behaviors.

A Zbtb18^+/- mice showed significantly reduced brain weight, body weight, brain length, and brain width compared to Zbtb18^+/+ mice. B Representative locomotor traces in the light-dark box test. C Compared to Zbtb18^+/+ mice, Zbtb18^+/- mice showed reduced total distance, distance in light box, entry number into the light box and crossing number compared in the light-dark box test. D Representative locomotor traces in the open field test. Shuttle times were calculated as light entries plus dark entries, minus one. E Compared to Zbtb18^+/+ mice, Zbtb18^+/- mice showed reduced percentage of distance travelled in the central zone, entry number into the central zone and time spent in the central zone in the open field test. A–E n = 13 for both Zbtb18^+/+ and Zbtb18^+/- groups. F UMAP visualization of hippocampus (left) and prefrontal cortex (right) cells. DG dentate gyrus granule neurons, CA cornu ammonis pyramidal neurons, L2/3/4/5/6 layer 2/3/4/5/6 respectively, CTX cerebral cortex.

To utilize the scPsyDrug for the drug repurposing in this mouse model, we performed the snRNA-seq on the hippocampus and prefrontal cortex derived from Zbtb18^+/− and Zbtb18^+/+ mice. In the hippocampus, we obtained a total of 60,448 cells, including 28,684 cells from Zbtb18^+/+ mice and 31,764 cells from Zbtb18^+/- controls, which were subsequently clustered and annotated into 9 cell populations. In the prefrontal cortex, snRNA-seq yielded 93,957 cells in total, of which 46,968 and 46,989 cells originated from Zbtb18^+/+ and Zbtb18^+/- mice respectively. These cells were then classified into 9 cell populations (Fig. 4F). In the hippocampus, we identified 2242 DEGs in cornu ammonis pyramidal neurons (CA), 1473 DEGs in GABAergic neurons (GABA), 1087 DEGs in dentate gyrus (DG) granule neurons, 701 DEGs in astrocytes (Astro), 183 DEGs in oligodendrocytes (Oligo), 169 DEGs in subicular neurons (SUB) and 18 DEGs in microglia (Micro), respectively (Figure S6B). In the prefrontal cortex, we identified 1463 DEGs in L2/3/4 CTX neurons, 966 DEGs in L5/6 CTX neurons, 309 DEGs in astrocytes (Astro), 280 DEGs in GABAergic neurons (GABA), 192 DEGs in oligodendrocytes (Oligo) and 15 DEGs in microglia (Micro), respectively (Figure S6B). These cell-type specific DEGs were subsequently used in scPsyDrug to screen potential compounds for the Zbtb18^+/- mice.

Forskolin rescued the aberrant behaviors of Zbtb18^+/− mice

In the hippocampal DG population, scPsyDrug prioritized Forskolin as top 3 among all compounds derived from DrugBank (Fig. 5A–C), with Adcy1 and Kcnn2 as targets. Forskolin is a diterpenoid extracted from the roots of the Lamiaceae plant Coleus forskohlii and has long been recognized as one of the direct agonists of the adenylyl/adenylate cyclase (AC) family. By directly acting on the catalytic subunit of AC, Forskolin can enhance the conversion of ATP to cAMP, thereby rapidly and persistently elevating intracellular cAMP levels [38–40]. Consistent with this pharmacological profile, the interaction network centered on the seed genes Adcy1 and Kcnn2 (Fig. 5D) was enriched for cAMP signaling (Figure S8A). Moreover, Adcy1, a member of the AC family, was significantly downregulated in the DG of Zbtb18^+/− mice (Figure S10A-B). Then we treated Zbtb18^+/− mice with Forskolin by daily oral gavage for fourteen consecutive days, using 5% DMSO as vehicle control. Following the Forskolin intervention, Zbtb18^+/− mice exhibited the behavioral restoration in the light-dark box test, including the recoveries of total distance traveled, distance traveled in the light compartment, number of entries into the light compartment, and shuttling events between the two compartments (Fig. 5E). Similarly, in the open-field test, Forskolin rescued total locomotor activity, the percentage of distance traveled in the central zone, the number of center entries, and the time spent in the center (Fig. 5F).

Fig. 5. Forskolin rescued the anxiety-like behaviors in Zbtb18^+/- mice.

A Forskolin, targeting Adcy1 and Kcnn2, ranks top 3 in hippocampal DG. B The high rank of Forskolin is cell-type specific. C Proximity values of top 10 drug targets in DG. D PPI network of Adcy1 and its associated seed genes. E Forskolin treatment rescued the abnormal behaviors in Zbtb18^+/- mice in the light-dark test. From left to right: total distance traveled, distance traveled in the light compartment, number of light compartment entries, and shuttle frequency. F Forskolin rescued the abnormal behaviors in Zbtb18^+/- mice in the open field test. From left to right: the total distance traveled, percentage of distance traveled in the center, number of center region entries, and time spent in the center region. E, F Zbtb18^+/+ DMSO: n = 6, Zbtb18^+/- DMSO: n = 7, and Zbtb18^+/- Forskolin: n = 8.

Chlorpropamide partially rescued the aberrant behaviors of Zbtb18^+/− mice

In the L2/3/4 CTX of prefrontal cortex, scPsyDrug ranked the FDA-approved Chlorpropamide as top 5 among all candidates derived from DrugBank (Fig. 6A, Figure S7A-B), with Abcc8 as its target. Chlorpropamide is a first-generation sulfonylurea hypoglycaemic agent to treat diabetes via binding to the regulatory subunit of ATP-sensitive potassium (K_ATP) channels, sulfonylurea receptor 1 (SUR1) encoded by the Abcc8/ABCC8 gene. In our data, the interaction network constructed around Abcc8 and its seed genes (Figure S7C) was significantly enriched for pathways such as synaptic transmission, glutamatergic, regulation of neurogenesis, and neuron apoptotic process (Figure S8B), implying that the aberrant upregulation of Abcc8 observed in L2/3/4 CTX (Figure S10C-D) might be mechanistically related to these processes. Then we treated Zbtb18^+/− mice with Chlorpropamide by daily oral gavage for fourteen consecutive days, using corn oil as the vehicle control. Following the Chlorpropamide intervention, Zbtb18^+/− mice showed the behavioral recovery in the light-dark box test, including total distance traveled, distance traveled in the light compartment, the number of entries into the light compartment, and the number of shuttling events between the two compartments (Fig. 6C, Figure S9A).

Fig. 6. Chlorpropamide and Osthole partially rescued the anxiety-like behaviors in Zbtb18^+/- mice.

A Chlorpropamide, targeting Abcc8, ranked top 3 in cortical L2/3/4 CTX. B Osthole, targeting Parp1, exhibited relatively high rank in cortical L2/3/4 CTX. C Chlorpropamide and Osthole rescued the abnormal behaviors in Zbtb18^+/- mice in the light-dark test. From left to right: total distance traveled, distance traveled in the light compartment, number of light compartment entries, and shuttle frequency. Zbtb18^+/++ Corn Oil: n = 7, Zbtb18^+/-+ Corn Oil for CHL: n = 6, Zbtb18^+/-+CHL: n = 6, Zbtb18^+/-+ Corn Oil for OST: n = 6 and Zbtb18^+/-+ OST: n = 8. The same 7 Zbtb18^+/+ mice treated with corn oil were used as vehicle control for both Chlorpropamide and Osthole experiments since these two drugs were dissolved in the corn oil for treatment. CHL chlorpropamide, OST osthole.

Osthole partially rescued the aberrant behaviors of Zbtb18^+/− mice

We next screened the active compounds of traditional Chinese medicine to rescue the abnormal behaviors of Zbtb18^+/− mice, demonstrating the application of scPsyDrug in repurposing the herb ingredients. By using the HERB (v2.0) database, scPsyDrug prioritized the Osthole in L2/3/4 CTX (Fig. 6B, Figure S7D-E), with Parp1 identified as its target. Poly(ADP-ribose) polymerase 1 (Parp1) catalyzes post-translational modification of itself and a broad spectrum of protein substrates involved in diverse cellular processes [41]. The interaction network constructed around Parp1 and its seed genes (Figure S7F) was significantly enriched for pathways related to neurodegenerative disorders, excitatory synapses, calcium signaling, and the FoxO signaling pathway (Figure S8C), implying that the aberrant upregulation of Parp1 in L2/3/4 CTX (Figure S10E-F) might be coupled to these signaling pathways. Inspired by these findings, we treated Zbtb18⁺/⁻ mice with Osthole by daily oral gavage for fourteen consecutive days, with corn oil serving as the vehicle control. The results showed that the Osthole treatment rescued the behaviors of Zbtb18⁺/⁻ mice in the light-dark box test, including the total distance traveled, distance traveled in the light compartment, the number of entries into the light compartment, and the number of shuttling events between the two compartments (Fig. 6C). The abnormal behaviors in open field were also partially rescued by Osthole (Figure S9B).

Discussion

Inspired by the recent single-cell studies that the human patients with psychiatric disorders showed highly stratified pattern of transcriptional pathology across different cell populations and subpopulations, we developed computational framework scPsyDrug for psychiatric drug repurposing by integrating multiple layers of molecular information in a cell-type specific manner, rather than collapsing cellular information within highly heterogeneous tissues. These cell-type specific networks and seed genes can prioritize compounds whose targets are predicted to be more proximal to the dysregulated disease-associated genes, thus naturally yielding the divergent drug-repurposing landscapes across different cell populations. The evaluations on MDD, SCZ and PD validate the improved sensitivity of drug repurposing by using the cell-type specific networks against the pooled and averaged networks. Even in the subpopulation level, the excitatory neuron subtypes can exhibit better performance than the overall excitatory neurons. For example, the ExN L4 in MDD and SCZ and ExN2 in PD represent the excitatory subtypes harboring the greatest number of therapeutic compounds in their respective disorders. These results highlight the power of cell-type specific networks in capturing sensitive changes in different cell populations and subpopulations. However, the bulk RNA-seq data are still useful in measuring the overall gene expression changes for drug repurposing though they may dilute cell-type specific signals by averaging across heterogeneous tissues.

The scPsyDrug was optimized for psychiatric disorders in two key aspects. First, we used the predicted blood-brain barrier (BBB) permeability to keep the compounds relevant to central nervous system. Second, we used the psychiatric disorder risk genes as seed genes to define network modules and compute target-to-seed proximity. The disease specific modules will prioritize drugs for psychiatric disorders. We also compared the enriched module strategy to other network-based approaches, such as nearest neighbor and random walk, and the enriched modules exhibited better performance in current datasets (Figure S11). Thus, we implemented enriched modules in scPsyDrug. It should be noted that though the integration strategy of scRNA-seq and network-based framework in this work can be in principle extended to other fields (e.g. cancer), the detailed computational steps together with validations should be carefully optimized to improve the performance in relevant fields.

The Zbtb18^⁺/⁻ mouse model provides a direct experimental model to validate and apply scPsyDrug in psychiatric disease mouse model. ZBTB18 (also known as RP58/ZNF238) is a BTB-zinc finger transcription factor that is highly expressed in cortex and hippocampus [42, 43], and involved in the generation and migration of cortical projection neurons and pyramidal neurons during brain development [44–46]. Human genetic studies have shown that pathogenic mutations in ZBTB18 are highly associated with neurodevelopmental disorders. In this work, we found that Zbtb18 haploinsufficiency leads to anxiety-like behaviors in mice, and snRNA-seq revealed different expression patterns on disease associated genes and pathways among cell populations. Within this framework, Forskolin, an activator of Adcy1, emerged as a top-ranked candidate in DG neurons, whereas Chlorpropamide (targeting Abcc8) and Osthole (targeting Parp1) were among the highest-priority hits in L2/3/4 CTX excitatory neurons. The works on mouse model further validate the power and applicability of cell-type specific networks in prioritizing the repurposed drugs. We also experimentally validated the expression levels of these targets before and after drug treatment. However, it should be noted that a single target cannot fully explain the mechanism of action for candidate drug since one drug can have different targets and complicated mechanisms.

Our pharmacological treatments on mouse model also have implications in molecular mechanisms and clinical indications. Forskolin is widely used to boost adenylyl cyclases and cAMP signaling pathway [47–50]. The success of Forskolin treatment in rescuing the aberrant behaviors suggested that the dysregulation of adenylyl cyclases and cAMP signaling pathway might be an important factor in Zbtb18^⁺/⁻ mice. Chlorpropamide is the FDA-approved drug initially indicated for diabetes, and its target SUR1, a subunit of ATP-sensitive K⁺ channel encoded by gene Abcc8, also plays important roles in neuronal survival and neuroinflammation. In multiple forms of CNS injury, SUR1-TRPM4 channels are markedly upregulated, and pharmacological blockade of SUR1 reduces cerebral edema and secondary neuronal death [51, 52]. The partial success of Chlorpropamide indicates that the dysregulated K_ATP pathway might also contribute to the phenotype in Zbtb18⁺/⁻ mice. Furthermore, it is often observed the comorbidity of diabetes and anxiety in the patients, and our work also implies that Chlorpropamide could play dual roles in the patients with both diabetes and anxiety. The Osthole target Parp1 is a key mediator of DNA damage responses and its excessive activation has been linked to neurodegeneration, neuroinflammation and depressive-like behavior. Consistent with these mechanistic considerations, we found that Osthole treatment rescued the anxiety-like behaviors of Zbtb18⁺/⁻ mice, thereby providing support for the role of PARP1 pathways in anxiety-like behaviors.

Our work has several limitations when using and interpreting the computational pipeline scPsyDrug. First, the performance of scPsyDrug is dependent on the quality and completeness of the underlying drug-target annotations and single cell data. Our framework can only prioritize the drugs with at least one known target or effector with prior knowledge, the common limitation in network based methods. The single cell sequencing depth and data quality also exert impact on the computational performance. The low cell number or data quality will result in less sensitivity in detecting cell-type specific networks and dysregulations, limiting the sensitivity in prioritizing repurposed drugs. Thus, the high quality of single cell data helps the better performance of scPsyDrug in rare cell populations, and the available scRNA-seq data processing methods can also be used to reduce the batch effect, cohort heterogeneity and technical variations to improve the performance of scPsyDrug. Second, the scPsyDrug is mainly used to prioritize or recommend drug candidates for further investigations instead of application to real cases, which is a common point in computational drug repurposing. We do not show that the prioritized drugs only act in the corresponding cell types since the drugs can have complicated mechanism of actions. Thus, the prior information on drug candidates, such as clinical trials and mouse works on similar disorders, can further improve the success rate of following experimental validations. Third, due to the lack of publicly available scRNA-seq data on other human psychiatric disorders or controls, we only limited our evaluations on MDD, SCZ, and PD. The scPsyDrug should be further evaluated when applying to other kinds of psychiatric disorders. Finally, although we used the Zbtb18^+/- mouse model to select and validate the prioritized compounds which can rescue anxiety-like behaviors, we did not perform dose-response optimization and longitudinal treatment paradigms. Thus, the current experimental evidence should still be considered as proof-of-concept rather than definitive preclinical validations.

Author contributions

YT developed the computational pipeline, performed data analysis and conducted the animal experiments. YZ performed the molecular experiments and participated in behavioral tests. ZW and YW participated in public data collection. CP conceived and administrated the project. YT and CP wrote the manuscript. All authors read the manuscript.

Funding

This work was supported by National Natural Science Foundation of China [32560156] and Yunnan Fundamental Research Project [202401AS070131]. Y.T was supported by the Scientific Research and Innovation Project of Postgraduate Students in the Academic Degree of Yunnan University and Scientific Research Fund Project of Yunnan Education Department [2025Y0060].

Data availability

The raw snRNA-seq reads have been deposited in the National Genomics Data Center (https://ngdc.cncb.ac.cn) under the BioProject accession number PRJCA053212, in which the raw reads were deposited under GSA accession number CRA035633 and the processed count matrices were under the OMIX accession number OMIX013646.

Code availability

The scPsyDrug is the R package with command lines and freely available from GitHub at https://github.com/CPenglab/scPsyDrug.

Ethics approval

The animal experiments were approved by the Institutional Animal Care and Use Committee of Yunnan University (approval number: YNU20240916), and all methods were performed in accordance with the relevant guidelines and regulations.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41398-026-04226-9.