Machine learning analysis of the orbitofrontal cortex transcriptome of human opioid users identifies Shisa7 as a translational target relevant for heroin-seeking leveraging a male rat model

Abstract

Background

Identifying neurobiological targets predictive of the molecular neuropathophysiological signature of human opioid use disorder (OUD) could expedite new treatments. OUD is characterized by dysregulated cognition and goal-directed behavior mediated by the orbitofrontal cortex (OFC), and next-generation sequencing could provide insights regarding novel targets.

Methods

Here, we used machine learning to evaluate human post-mortem OFC RNA-sequencing datasets from heroin-users and controls to identify transcripts predictive of heroin use. To determine a causal link to OUD-related behaviors, we examined the effects of overexpressing the top target gene in a translational rat model of heroin-seeking and behavioral updating. Additionally, we determined the effects of overexpression on the rat OFC transcriptome compared to that of human heroin users. Co-immunoprecipitation/mass-spectrometry from rat OFC elucidated the protein complex of the novel target.

Results

Our machine learning approach identified SHISA7 as predictive of human heroin users. Shisa7 is understudied but appears to be an auxiliary protein of GABAA or AMPA receptors. In rats, Shisa7 expression positively-correlated with heroin-seeking behavior. Overexpressing Shisa7 in the OFC augmented heroin-seeking and impaired behavioral updating for sucrose-based operant contingency. RNA-sequencing of rat OFC revealed gene co-expression networks regulated by Shisa7-overexpression similar to human heroin-users. Finally, co-immunoprecipitation/mass-spectrometry showed that heroin influences Shisa7 binding to glutamatergic and GABAergic receptor subunits. Both gene expression signatures and Shisa7 protein complex emphasized perturbations of neurodegenerative and neuroimmune processes.

Conclusions

Our findings suggest that OFC Shisa7 is a critical driver of neurobehavioral pathology related to drug-seeking behavior and behavioral updating, identifying a potential therapeutic target for OUD.

Introduction

Opioid use disorder (OUD) is associated with 350,000+ deaths annually worldwide; more than one death every 90-seconds.(1) Opioid-agonist medication for OUD helps reduce fatal overdose(2), but carries challenges limiting use and leaving millions of patients untreated. Although most opioid medication strategies and preclinical studies focus on opioid reward, it is known that cycles of misuse are driven by drug-seeking behavior linked to impaired top-down cortical control(3,4). The prefrontal cortex facilitates decision-making, value-encoding, and higher-order cognition generally, providing top-down glutamatergic regulation of subcortical regions such as the nucleus-accumbens, dorsal-striatum, and amygdala(5,6). The orbitofrontal cortex (OFC) processes sensorimotor and associative information to e.g., inhibitory responses, process prediction-errors, and credit-assignment during learning(6,7). Previous work has demonstrated the critical role of the OFC in opioid-seeking behavior(8,9), but neuropathological knowledge relevant to the human-OFC in OUD is limited.

Next-generation sequencing technologies have characterized molecular signatures in the brains of human opioid users, highlighting genes/networks related to neuro-, synapto- and axonogenesis, glutamate and GABA receptors, GPCR signaling, and various other neural processes(10–13). Moreover, machine learning of RNA-sequencing data has been leveraged to discover molecular targets for various disorders(14–16) but not for opioid use-associated datasets. Here, we used RNA-sequencing to profile the post-mortem human OFC transcriptome of opioid-users and leveraged machine learning to identify transcripts predictive of heroin-use. We then used a translational rodent heroin self-administration model to validate the causal impact of the top predictive gene, SHISA7, on heroin-seeking behavior, as well as behavior-updating, an OFC-related endophenotype(17,18). We determined the direct mechanistic relevance of SHISA7 to the rat-OFC transcriptome and its relationship to the molecular pathophysiology of heroin-users, which included specific biological processes related to neurodegeneration and neuroinflammation. Finally, we performed proteomics of the SHISA7 protein complex, identifying binding partners across excitatory, inhibitory, and glial systems impacted by heroin experience.

Methods and Materials

Detailed methods are described in the Supplementary Information.

Human OFC brain specimens and RNA-sequencing

Post-mortem OFC specimens were obtained from a brain-bank curated at the Icahn School of Medicine at Mount Sinai from a homogeneous cohort of Caucasian subjects who died from apparent heroin-overdose, and normal-controls (determined by self-report and ancestral informative marker analysis). Specimens were collected at autopsy within 24h of time-of-death as previously described(10). The control-cohort (n=28; age 15–65 (mean 36); 4 females) had not used illicit-substances and died of non-suicide causes such as accidents, cardiac-failure, or viral-infection. The heroin-cohort (n=34; age 19–46 (mean 26); 5 females) had positive toxicology of heroin-use with no other illicit substance-use.

Libraries were prepared from 1μg of total-RNA using Illumina’s TruSeq RNA-library preparation-kit V2, and sequencing was performed using an Illumina HiSeq2000. After de-multiplexing of each library pool according to its barcode-sequence, reads processed using the TopHat and Cufflinks pipeline(19).

Machine learning

RNA-sequencing data were normalized using the fragments-per-kilobase-per-million mapped fragments (FPKM) method(20), and inputted to a machine learning pipeline similar to previous work(21) (Supplementary Figure 1). Twelve samples (20%) were randomly-chosen as the test-set used during final model evaluation. The other 50 samples, the model development-set, were processed using a stratified cross-validation(22) and feature-selection(23) procedure.

In the five-fold stratified cross-validation on the development-set, each of the training-sets consisting of 4/5 folds were analyzed using three feature-selection measures: the analysis-of-variance (ANOVA) F-test(24), Gini importance(25), and Shapley values(26). Genes were ranked using each measure and the top-k (k=1–100) genes were used to fit 10 classification models(27) to the training folds and predict on the fifth validation fold. The maximum value of the F1-score (Fmax) and the area under the receiver operating characteristic curve (AUROC)(28) were used to rank combinations of model type, k, and feature-selection metric. These final three combinations (one per feature-selection metric) were trained on the entire development-set (n=50) to predict on the test-set (n=12), and we compared these approaches using performance metrics (AUROC, balanced-accuracy, F1-score, precision, recall).

Animals

Male Long-Evans rats (Charles River Laboratories, Raleigh, NC, USA) were pair-housed and maintained under a reverse dark/light-cycle (dark-cycle beginning at 7am) in a temperature-controlled vivarium. Animals acclimated to the facility for 1 week, during which they were handled daily. Food and water were available ad-libitum until behavioral testing, and animals were single-housed beginning after jugular-vein catheterization surgery. Testing and housing were in accordance with IACUC-approved protocols.

Heroin self-administration and seeking behavior

Animals underwent jugular-vein catheterization surgery for heroin self-administration, followed by 1-week recovery. Rats underwent 3h daily heroin self-administration sessions in operant boxes (Med-Associates, Fairfax, VT, USA) fitted with a syringe-pump, two levers, and cue-lights above the levers as previously described(10,29,30). Active-lever presses administered a cue-light and a 30μg/kg/infusion of diacetylmorphine-hydrochloride (NIDA) on a fixed-ratio-1 (FR1) reinforcement schedule. Inactive-lever responses had no programmed consequences. Active- and inactive-lever positions were counterbalanced across animals.

After the stabilization of self-administration, rats underwent a 2-week drug-free period in the home cage, followed by a cue-induced drug-seeking session in the same operant box context, without drug administration. For postmortem analyses, rats were sacrificed using carbon-dioxide and rapid-decapitation 1h after heroin-seeking. Brains were extracted, flash-frozen in ice-cold isopentane and stored at −80°C.

Lentiviral overexpression of Shisa7 in OFC

In a separate cohort of rats, after the stabilization of heroin self-administration behavior, we infused lentiviral-vectors in OFC to overexpress Shisa7, or Control virus, yielding four groups — Saline-Control (n=5), Saline-Shisa7 (n=7), Heroin-Control (n=13), Heroin-Shisa7 (n=13).

In lentiviral-overexpression experiments, rats underwent stereotaxic surgery at the beginning of the drug-free period (2 weeks prior to the drug-seeking session) for the administration of lentiviral constructs (pCDH-EF1-MCS-T2A-RFP (PGK-Puro); System-Biosciences) into the OFC to overexpress Shisa7 or control virus in both brain hemispheres.

Post-heroin behavioral updating and reward reorientation for sucrose operant behavior

In the overexpression experiments, 5-days after heroin-seeking, rats underwent a single 1h sucrose session to assess the impact of Shisa7 on behavioral updating, specifically how animals switch from responding to the cue of one reinforcer (drug) to a natural reinforcer (sucrose), a form of behavioral-updating. Rats were placed in the same context as the heroin self-administration experiments, with the active- and inactive-levers reversed, and the active-lever delivering sucrose on a FR1-schedule (45mg, BioServ, Flemington, NJ, USA). Animals were sacrificed 1h after the behavioral-updating session, brains were extracted, flash-frozen, and stored at −80°C.

RNA-sequencing of rat OFC

RNA-extraction from isolated tissue was conducted as described previously(10,29,30). Using the SMART-Seq v4 Ultra-Low-Input RNA Kit (Clontech), reverse-transcription was performed on polyadenylated transcripts, along with second-strand synthesis of full-length cDNA and amplification of the product (Azenta Life-Sciences, Chelmsford, MA). FASTQ files were aligned to the mRatBN7.2 Rattus Norvegicus reference genome using the STAR aligner(31). Differential-expression analysis was performed using DESeq2.

Co-immunoprecipitation/mass-spectrometry (Co-IP/MS)

Protein-extraction from isolated tissue was conducted by sonicating OFC brain-punches in lysis-buffer provided in the Pierce MS-Compatible Magnetic-IP Kit (Protein A/G) (Thermo-Fisher; #90409). Following sonication, lysates were incubated on ice for 1h, centrifuged, and concentrations were quantified using the BCA-assay. Standardized protein amounts were processed through co-IP, IP samples were speed vacuum-concentrated, processed for mass-spectrometry and quantified using Progenesis-QI.

Statistical Analysis

Drug-seeking and behavioral-updating comparisons were conducted by repeated-measures-ANOVA for the active and inactive-levers, with experimental group as the between-subjects factor. Statistical significance was denoted by p<0.05 for behavioral and sequencing experiments. Post-hoc tests with Holm’s correction were conducted for statistically significant interaction-effects.

Results

Machine learning identifies Shisa7 OFC expression as predictive of human heroin-users

We began by conducting RNA-sequencing of the OFC transcriptome of human heroin-users and controls, identifying 1,411 differentially-expressed genes (648 upregulated, 763 downregulated; Supplementary File 2) that showed significant enrichment of immune-related ontologies using Enrichr(32) (Figure 1A,B). We leveraged a machine learning approach to accurately identify genes that predict heroin-users compared to controls. Upon splitting 62 OFC RNA-seq samples into development (n=50) and test (n=12) sets, a cross-validation and iterative feature-selection procedure identified best-performing combinations of predictive models and gene-sets from the development-set. For each of the three feature-selection techniques (ANOVA F-test, Gini importance, Shapley values), a final strategy was chosen consisting of a classifier, number of genes k, and probability threshold to translate the prediction probabilities into discrete classes (i.e., heroin-users/controls). For the ANOVA F-test, a gaussian-process model was chosen with 39 genes and a probability threshold of 50.97%. For Gini importance, a multilayer-perceptron was chosen with 11 genes and a probability threshold of 53.82%. For Shapley values, a multilayer-perceptron was chosen with 10 genes (Figure 1C) and a probability threshold of 52.72%. Figure 1C shows that the class-specific Shapley values were balanced, indicating equal predictive importance for the control and heroin-overdose classes. Final models (Table 1) were then trained using these combinations on the entire development-set and evaluated on the held-out test-set.

Figure 1.

Analysis of human post-mortem OFC RNA-seq data and the identification of predictive genes and models of heroin use from these data. A) Volcano plot showing differential expression of genes (DEGs) in the RNA-seq data. B) Gene ontology enrichment analysis of the human RNA-seq DEGs indicate immune-related perturbations. C) Top 10 genes ranked by the Shapley value feature selection method for the prediction of OFC sample class using machine learning. Blue reflects the Shapley value for predicting the Control class, and red reflects importance for predicting the Heroin class.

Table 1:

Ranked gene sets chosen using three feature selection techniques after training on the entire development set (n=50). Shisa7 is the top-ranked gene across techniques. GP: Gaussian Process; MLP: Multilayer Perceptron.

Classifier	Gene list
F-test, GP, k=39	SHISA7, KCNH5, ELK1, FBXL14, ARHGDIB, JUN, TMEM204, GRID1, PPP4R2, CPEB4, LRRC57, A2M, CHPT1, STK17A, ECH1, SON, GIMAP7, YPEL1, NET1, SSX2IP, TGFBR2, RFX3, PCDH10, SEC22C, CNDP1, C1orf96, BUD13, CD34, RAMP2, SMARCD2, JAKMIP2, LHX2, MOBP, CCNG2, RTN4RL1, HS2ST1, CYR61, METTL3, GNAQ
Shapley, MLP, k=10	SHISA7, SMARCD2, KCNH5, RTN4RL1, DNAJB4, ELK1, PCDH10, FKBP15, ZNF362, ISOC1
Gini, MLP, k=11	SHISA7, SMARCD2, KCNH5, RTN4RL1, ELK1, DNAJB4, PCDH10, ZNF576, ZNF362, NEURL1B, ISOC1

The models built using Gini importance and the Shapley value had higher performance metrics (Table 2) than the one using the ANOVA F-test. Examining the genes used in these three models (Table 1) showed that all three feature-selection methods ranked SHISA7, a putative auxiliary-subunit of GABAA receptors(33–35) or AMPA receptors(36,37), as the most predictive gene. SHISA7 showed a trending reduction in our OFC RNA-sequencing (log₂ FC=−0.1495, p=0.0549).

Table 2:

Performance metrics for three machine learning models based on different feature selection techniques on the test set. GP: Gaussian Process; MLP: Multilayer Perceptron. Positive indicates the heroin use class; negative indicates the control class.

	F-test, GP, k=39	Gini, MLP, k=11	Shapley, MLP, k=10
AUROC	0.912	1.0	1.0
Bal. acc.	0.67	0.92	0.92
F1_pos	0.75	0.92	0.92
F1_neg	0.5	0.91	0.91
Prec. Pos.	0.6	0.86	0.86
Prec. Neg.	1.0	1.0	1.0
Recall Pos.	1.0	1.0	1.0
Recall Neg.	0.33	0.83	0.83

Shisa7 is reduced in the rat OFC after heroin experience and correlated with heroin-seeking

SHISA7 being predictive of heroin-use, based on our machine learning experiments, raised questions of its causal relation to heroin-use. Thus, we studied Shisa7 expression in a heroin self-administration model in rats that acquired stable FR1 heroin intake followed by a cue-induced drug-seeking session. Shisa7 expression was significantly reduced in the OFC of rats with heroin experience compared to saline (Figure 2A,B; t(16)=2.6, p=0.0195), and the infralimbic cortex (Supplementary Figure 3; t(16)=3.02, p=0.014), matching the trend seen in human heroin-users. Additionally, OFC Shisa7 expression positively-correlated with active-lever presses in heroin-seeking (Figure 2C; r=0.73, p=0.018), indicating that while Shisa7 expression decreased after heroin-experience, expression was positively-associated with drug-seeking.

Figure 2.

Heroin self-administration and seeking is associated with a decrease in OFC Shisa7 expression. A) Rats showed clear heroin self-administration and drug-seeking behavior. B) Heroin reduced Shisa7 expression (qPCR) in the OFC of rats with heroin self-administration experience. C) Shisa7 expression was positively associated with heroin-seeking behavior. Shisa7 expression is quantified as fold change in qPCR relative to a housekeeping gene. Samples sizes ranged from 8–10/group.

In separate animals, a single acute heroin-exposure did not affect Shisa7 expression in OFC (Supplementary Figure 4; t(14)=−0.17, p=0.863), indicating that Shisa7 perturbation occurs primarily with chronic exposure. Infralimbic cortical Shisa7 expression after heroin-experience was not associated with heroin-seeking.

Shisa7-overexpression augments heroin-seeking and blunts behavioral updating for sucrose operant behavior

Given the effect of heroin-exposure on Shisa7 expression, we tested the direct behavioral relevance of Shisa7 on heroin-related behaviors using lentiviral overexpression (Figure 3A).

Figure 3.

Shisa7 causally impacts heroin-seeking and post-heroin behavioral updating for sucrose operant behavior. A) Heroin self-administration behavior prior to lentiviral overexpression. B) Heroin seeking is augmented by Shisa7 overexpression in comparison to animals with heroin experience that received control vector (lever*virus-F_1,24=5.51, p=0.028; active presses: Cohen’s d: 1.311, p_holm=0.007; inactive lever presses: Cohen’s d: 0.099, p_holm=0.802). C) Post-heroin behavioral updating for sucrose self-administration is impaired by Shisa7 overexpression in comparison to animals with heroin experience that received control vector. (lever*virus-F_1,23=4.41, p=0.047; active presses: Cohen’s d: 0.805, p_holm=0.201; inactive lever presses (Cohen’s d: −0.442, p_holm=0.55). Sample sizes: Sal-Con (n=5), Sal-Shisa7 (n=7), Her-Con (n=13), Her-Shisa7 (n=13).

We conducted a two-way-ANOVA of Shisa7 expression with drug and virus as fixed-factors. We observed a significant main effect for virus (Supplementary Figure 5; F_1,33=8.32, p=0.007) but not drug (F_1,33=2.93, p=0.096) or the interaction of virus and drug (F_1,33=0.91, p=0.348).

We performed a 2×2 ANOVA to assess heroin-seeking, using drug and virus as between-subject factors and lever as a within-subject factor. We observed significant main effects for lever (F_1,34=9.92, p=0.003) and group (F_1,34=9.65, p=0.004), but not virus (F_1,34=2.74, p=0.107). The only significant interaction was lever*group (F_1,34=7.6, p=0.009), driven by active lever presses in the heroin animals. While the lever*virus*group interaction was not statistically significant (F_1,34=2.2, p=0.147), the Heroin-Shisa7 group showed increased heroin-seeking compared to the Heroin-Control group, driven by active presses (Mean difference=99.69, t-stat=3.97, Bonferroni-corrected p=0.005) and a one-way ANOVA showed a significant lever*virus interaction (F_1,24=5.51, p=0.028; Figure 3B).

Given the OFC’s role in behavioral updating, highly relevant to addiction(38,39), we examined the effects of Shisa7-overexpression in heroin-experienced rats using a behavioral updating for sucrose operant behavioral paradigm. Shisa7-overexpression impaired reward reorientation for sucrose self-administration in rats with heroin experience compared to Heroin-Control (Figure 3C; lever*virus-F_1,23=4.41, p=0.047). Specifically, the Heroin-Control showed greater active presses (Cohen’s d: 0.805, p_holm=0.201) and discrimination (Cohen’s d: 1.537, p_holm=0.009), whereas the Heroin-Shisa7 group failed to discriminate (Cohen’s d: 0.289, p_holm=0.55). There was a trend difference in rewards obtained (F_1,23=3.305, p=0.082). These results indicate that Shisa7-overexpression increased drug-seeking and impairs reward reorientation.

Overlap between RNA-sequencing pattern induced by Shisa7 OFC overexpression and transcriptome signature of human heroin OFC

To understand the potential role of Shisa7 in the OFC transcriptome of human heroin-users, we conducted RNA-sequencing on the rat OFC following Shisa7-overexpression and behavioral experiments (Supplementary Figure 6).

Downstream analyses examined three comparisons of these four groups: the effects of heroin (Saline-Control vs. Heroin-Control), Shisa7-overexpression in the drug-naïve condition (Saline-Control vs. Saline-Shisa7), and Shisa7-overexpression in the drug condition (Heroin-Control vs. Heroin-Shisa7). Examining the effect of heroin on the OFC (Figure 4A), gene-ontology analyses showed enrichment for genes related to oxidative phosphorylation, electron transport, and multiple neurocognitive diseases (Huntington’s, Parkinson’s, amyotrophic lateral sclerosis, and Alzheimer’s diseases). Shisa7-overexpression in the non-drug condition (Figure 4B) resulted in DEGs enriched for ontologies related to intracellular signaling, the spliceosome, synaptic membrane, axonogenesis, and dendrite. Shisa7-overexpression in the drug condition (Figure 4C) induced DEGs enriched for immune-related, axonal/neuronal, and microglia related ontologies, strongly similar to the signature seen in the human RNA-seq data (Figure 1B). Shisa7-overexpression in both the non-drug and heroin drug conditions was associated with increased expression of multiple genes associated with the AMPA and GABA complexes, indicating the direct relevance of Shisa7 expression to both neurotransmitter systems (Supplementary Figure 7).

Figure 4.

RNA-sequencing of OFC from rats that underwent heroin self-administration and Shisa7 overexpression. A-C) Gene ontology analyses of differentially expressed genes between Saline-Control/Heroin-Control, Heroin-Control/Heroin-Shisa7, and Saline-Control/Saline-Shisa7 groups. D-F) RRHO plots showing the pairwise overlap of genes from A-C. G-I) RRHO plots showing the overlap of genes from A-C with the human OFC RNA-seq data. J) Schematic overview of RRHO plot layout. Color scales of D-F and G-I are the same.

Next, we used the rank-rank hypergeometric overlap (RRHO) test (Figure 4D–F) to measure the transcriptional overlap of the different conditions of our rodent model in a threshold-free manner. There were significant concordant gene expression changes between the effects of heroin and Shisa7-overexpression in the non-drug condition (Figure 4D), indicating that Shisa7-overexpression without heroin induces strongly similar transcriptional changes to heroin alone. Consistently, there were significant concordant gene expression changes between Shisa7-overexpression in both the non-drug and drug conditions (Figure 4E), showing that Shisa7 induces similar gene expression programs in both the drug and non-drug contexts. Agreeing with these findings, there was statistically significant concordant gene expression between Shisa7-overexpression in the drug condition and the effects of heroin alone (Figure 4F).

We then measured the overlap of transcriptomic changes between our three rodent model conditions and the human heroin-users (Figure 4G–I). There was limited overlap between the human data and the rat heroin condition (i.e., saline vs. heroin; Figure 4G), which may be due to the time points that the brains were studied. Interestingly, there was statistically significant discordant up- and down-regulation of genes between the human data and Shisa7-overexpression in the non-drug condition (Figure 4H), indicating that Shisa7-overexpression alone induces opposing transcriptional programs to those seen in heroin overdose in humans. Given that heroin consumption reduces Shisa7 expression, it follows that overexpressing Shisa7 would result in opposing changes. These two contrasting results may be a consequence of an acute administration of heroin that leads to fatal overdose in a chronic user versus the withdrawal state associated with heroin-seeking and Shisa7-overexpression. Finally, there was statistically significant concordant downregulation of genes between the human data and Shisa7-overexpression in the drug condition (Figure 4I), indicating that Shisa7-overexpression in the context of a history of heroin experience shares a downregulated transcriptional program with heroin overdose in humans.

Gene co-expression modules are associated with rodent heroin-seeking and urine opioid toxicology in humans

We used multiscale embedded gene co-expression network analysis (MEGENA) to identify modules of gene co-expression in human heroin-users, and in rats with heroin experience and Shisa7-overexpression. We identified modules with greatest enrichment for DEGs in three conditions (i.e., effects of heroin in humans, Shisa7-overexpression in the non-drug condition, Shisa7-overexpression in the drug condition) and their associated gene ontologies. Full descriptions of top modules and their key drivers are in Supplementary Figures 8–11.

Next, we examined co-expression modules (Supplementary Figures 9–11) in the rat associated with heroin-seeking, behavioral updating for sucrose, or Shisa7 expression (Figure 5). To this end, we used only the modules with statistically significant enrichment of DEGs to perform eigengene analysis using WGCNA which takes the first principal-component of the expression matrix (i.e., expression of all genes in the module across all samples within the group) and correlated these eigengenes with heroin-seeking and behavioral updating for sucrose, along with Shisa7 expression. Examining the effects of heroin in the rat (Figure 5A), inwere negatively-associated with heroin-seeking (measured as the percentage of active-lever presses) and/or Shisa7 expression, with top ontology enrichments for the mitochondrial complex, regulation of ATPase activity, and eukaryotic translation elongation.

Figure 5.

Heatmaps of correlation coefficients between module eigengenes and heroin-seeking and post-heroin behavioral updating for sucrose self-administration in the rat (effects of heroin only) (A), urine toxicology in the human (B), and SHISA7 expression in both species. Rows indicate all modules enriched for DEGs, annotated with unique colors and their top gene ontology terms. *P < 0.05

In the human OUD condition (Figure 5B), 5 DEG-enriched modules were also negatively associated with Shisa7 expression and/or positively associated with urine levels of 6-monoacetylmorphine, an active metabolite of heroin, and their top ontology enrichments related to RNA binding, intracellular protein transport, IL4/IL13 signaling, and regulation of cell death. Overall, the results using an unbiased approach to select gene co-expression modules enriched for DEGs in the OFC transcriptome associated with heroin experience in humans and the translational rat model, demonstrate that these modules negatively-associate with Shisa7 expression, accompanying the group-level reduction that initially identified Shisa7 as a target.

Examining the effects of Shisa7-overexpression in the heroin drug condition (Figure 6), 30 modules were significantly associated with Shisa7 expression (27 positively; 3 negatively), and/or heroin-seeking, with top ontology enrichments related to the synapse (post-synapse organization, synaptic signaling, cortical cytoskeleton, synapse structure/activity), epigenetic modifications (chromatin modifying enzymes, SWI/SNF complex), among other cellular processes. Notably, the top ontology for a module negatively associated with both Shisa7 expression and heroin-seeking was “Immune effector process,” and immune-related ontologies were very strongly represented in the DEGs for the human heroin condition (Figure 1B). These results indicate the direct impact of Shisa7-overexpression in the context of a history of heroin exposure on multiple biological processes related to synaptic plasticity and epigenetic reprogramming, known critical factors in various neuropsychiatric disorders(40–43), and immune-related processes that were strongly implicated in the human heroin OFC RNA-seq data. None of the modules were, however, associated with behavioral updating for sucrose.

Figure 6.

Heatmap of correlation coefficients between module eigengenes and heroin-seeking, post-heroin behavioral updating for sucrose self-administration, and Shisa7 expression, for Shisa7 overexpression in the context of a history of heroin experience. Rows indicate all modules enriched for DEGs, annotated with unique colors and their top gene ontology terms. *P < 0.05

We also analyzed MEGENA co-expression network topology metrics (betweenness centrality, degree centrality, and clustering coefficient) between the two heroin conditions to assess the effect of Shisa7-overexpression on global expression-architecture phenotypes (Supplementary Figure 13). Shisa7-overexpression in the context of a history of heroin exposure did not alter the centrality of genes or their number of edges in the OFC co-expression network, but it did increase the tendency of genes to cluster with one another, and the relationship clustering among the neighbors of individual nodes, along with the distributional properties all three metrics.

Co-IP/MS shows differential Shisa7 protein binding to GABAergic and glutamatergic subunits after heroin self-administration

The conflicting results from the limited SHISA7 literature indicate it may function as an auxiliary subunit of AMPA receptors(36,37) or GABAA receptors(33–35). To obtain direct insights regarding SHISA7 complex in relation to heroin exposure, we leveraged co-immunoprecipitation coupled with mass-spectroscopy. This was studied in a separate cohort of rats that underwent the same heroin self-administration, withdrawal-period, a 1h heroin-seeking session, and sacrificed 1h later. OFC tissue was processed through co-IP with an anti-Shisa7 antibody and mass-spectroscopy to quantify Shisa7 protein-protein interactions (PPIs). 1352 proteins were detected in the Saline group and 1354 proteins detected in the Heroin group (895 overlapping, with 961 total isoforms). There were 282 differentially expressed proteins (DEPs p<0.05; Figure 7A). Shisa7 binding to subunits of both GABA (Gabbr1, Gabrb2) and glutamate receptors (Grm3, Gripap1) was apparent in the Heroin group supporting both lines of prior research suggesting binding between Shisa7 and GABA or glutamate receptor subunits(34,36).

Figure 7.

Heroin affects Shisa7 protein-protein interactions. A) Volcano plot of Shisa7 protein complex co-IP data. The x-axis shows the Log2-fold change of protein expression between the Saline and Heroin groups; the y-axis shows the Log10-pvalue. B) Pathway enrichment analysis for DEPs.

Heroin also affected Shisa7 binding to other proteins relevant to GABA and glutamatergic transmission (e.g., Cnksr2, Gls) and various other proteins critical to neuronal/glial/synaptic signaling such as Syngap1 (mediator of AMPA/NMDA synaptic plasticity), Gfap (astrocyte marker), and Ntrk2 (neurotrophic receptor tyrosine kinase). Biological pathways enriched in DEPs (Figure 7B) included neurocognitive-disease similar to the RNA-sequencing data from human heroin-users and rat heroin self-administration/Shisa7-overexpression, along with expected processes related to basic protein, RNA, and cellular functions.

In separate pathway/ontology analyses on up- and down-regulated DEPs, we observed that down-regulated DEPs also represented a neurocognitive signature (Supplementary Figure 12A–B). Along with assessing protein-expression changes, we examined changes in protein-protein regulatory relationships via differential-correlation analysis. 894 differentially-correlated protein-protein pairs, and pathway/ontology analyses on the unique proteins across pairs again highlighted neurocognitive pathways (Supplementary Figure 12C). Finally, we conducted pathway/ontology analyses of proteins expressed exclusively in either group, and these analyses further highlighted neurodegenerative disease pathways (Supplementary Figure 12D–E). Overall, heroin-induced changes to the Shisa7 protein complex in terms of expression, correlation, and mutually exclusive expression showed strong associations with neurocognitive disease pathways, in agreement with gene-expression signatures of human heroin overdose, heroin self-administration in the rat, and Shisa7-overexpression.

Discussion

Very limited molecular data is available about the human OFC in relation to heroin-use despite the critical role of the OFC in drug-seeking behavior that perpetuate the cycle of addiction. Leveraging a machine learning-based target-identification approach with explainable artificial intelligence techniques, accurately predicted the group assignment of human controls and heroin-users based on the OFC transcriptome signature. This approach identified SHISA7, a relatively understudied auxiliary-subunit of GABAA and/or AMPA receptors, as the gene most predictive of group classification. SHISA7 is selectively expressed in the human and rodent brain on the gene and protein level, with high expression across the cortex and lower expression in subcortical regions(44,45). Consistent with the trend observed in the human heroin users, Shisa7 expression in the rat OFC was downregulated following repeated heroin self-administration. However, Shisa7 expression positively-correlated with heroin-seeking, suggesting an association with relapse-related behavior. This may indicate that Shisa7 downregulation is a homeostatic-like effect, and that a lack of this homeostasis leads to greater heroin-seeking. Critically, we confirmed this relapse-related association by overexpressing Shisa7 in the OFC after heroin self-administration, leading to augmented heroin-seeking and impaired post-heroin behavioral updating for sucrose. These results highlight the direct behavioral relevance of OFC Shisa7 expression for heroin-seeking/craving and compromised behavioral updating, endophenotypes characteristic of OUD(17,18).

The transcriptomic signature induced by OFC Shisa7-overexpression in the drug-naïve state had a significant concordant overlap with the transcriptome induced in the heroin condition, indicating that Shisa7 without heroin exposure induces similar transcriptomic changes as repeated heroin use. This supports the AI strategy identifying Shisa7 as predictive of the OFC signature in heroin users.

Expanded insight about Shisa7 was obtained from the co-immunoprecipitation/mass spectrometry that illuminated the SHISA7 protein complexes as containing both GABA- and glutamatergic proteins, thus unifying previous conflicting reports(34,36). Furthermore, the results highlight SHISA7 protein-protein interactions within other biological systems critical to neuronal, glial, and synaptic signaling. A clear neurobiological phenotype of interest relates to neurocognitive disease pathways. Indeed, a consistent and key finding in multiple analytical assessments that was evident in the human heroin-users, both for the gene-expression signatures and Shisa7 protein-complex, as well as with Shisa7-overexpression in the rat model was the significant associations to neurodegenerative and neuroimmune processes. This has important implications for neurological conditions observed in OUD, and other potential neuropsychiatric risks(10,46–48). Our studies coincide with other transcriptomic studies of OUD highlighting heroin-induced effects on pathways related to neuroinflammation and cytokine responses(49,50), synaptic remodeling, and neurodegeneration(51). Currently no specific pharmacological agents exist for Shisa7, but the current results provide a strong foundation for developing future novel medication strategies.

A limitation of our study was the small human OFC RNA-seq dataset used for the machine learning experiments, though this is common in human studies due to a lack of available tissue and has provided impactful discoveries in the field of psychiatric genomics(52). Another limitation is that the rodent experiments were conducted only in males (the majority of the human specimens), and we will expand these studies to. A further limitation relates to the differences between the human and rodent sample acquisition—the humans died from a fatal heroin-overdose, whereas the rats were sacrificed weeks after their last heroin exposure.

In summary, the recent use of AI strategies to identify potential targets based on sequencing data generated from the human brain is a promising approach to identify novel biomarkers with mechanistic and clinical relevance to expand neurobiological insights underlying addiction. Leveraging machine-learning identified OFC Shisa7 as a promising target in human heroin-users that directly induced a similar transcriptional signature in the OFC in animals that never consumed heroin and to impact drug-seeking and behavioral updating in a translational OUD model. Overall, SHISA7’s multifunctional role in the brain, along with its direct impact on addiction- and cognition-related behaviors, point to its importance in the molecular pathophysiology of OUD. Moreover, its causal link to neurodegenerative and immune biological processes suggests significant relevance for other neuropsychiatric conditions. Overall, the findings emphasize the role of OFC Shisa7 expression in addictive and cognitive behaviors, its effects on the transcriptome, and its multifaceted protein complex that may be targetable for future medication development.