Translatable pathways classification (TransPath-C) for inferring processes germane to human biology from animal studies data: example application in neurobiology

Abstract

How to translate insights gained from studies in one organismal species for what is most likely to be germane in another species, such as from mice to humans, is a ubiquitous challenge in basic biology as well as biomedicine. This is an especially difficult problem when there are few molecular features that are obviously important in both species for a given phenotype of interest. Neuropathologies are a prominent realm of this complication. Schizophrenia is complex psychiatric disorder that affects 1% of the population. Many genetic factors have been proposed to drive the development of schizophrenia, and the 22q11 microdeletion (MD) syndrome has been shown to dramatically increase this risk. Due to heterogeneity of presentation of symptoms, diagnosis and formulation of treatment options for patients can often be delayed, and there is an urgent need for novel therapeutics directed toward the treatment of schizophrenia. Here, we present a novel computational approach, Translational Pathways Classification (TransPath-C), that can be used to identify shared pathway dysregulation between mouse models and human schizophrenia cohorts. This method uses variation of pathway activation in the mouse model to predict both mouse and human disease phenotype. Analysis of shared dysregulated pathways called out by both the mouse and human classifiers of TransPath-C can identify pathways that can be targeted in both preclinical and human cohorts of schizophrenia. In application to the 22q11 MD mouse model, our findings suggest that PAR1 pathway activation found upregulated in this mouse phenotype is germane for the corresponding human schizophrenia cohort such that inhibition of PAR1 may offer a novel therapeutic target.

INTRODUCTION

Due to the limited ability to undertake in vivo experimental perturbations and measurements in humans, construction of models for human physiology and pathology is generally pursued in animal studies. A consequent challenge is how to translate observations from experimental animals to human subjects. Direct translation of observations from animals to humans is typically unsatisfactory because of significant differences in organismal biology at all levels of molecular and cellular properties. Use of computational frameworks for cross-species translation holds promise [1], and while most previous approaches have focused on features that are strongly shared between species [2–5], we have recently offered advances that are able to find features important for human biology that are not readily noticed as important for a non-human animal species [6–8]. Our earlier methods, however, have been demonstrated in studies where there are a substantive number of molecular components found in common between species, but now address yet more challenging applications where this is not the case. One such situation is neuropathologies, because of the especially major expansion of complexity inherent in neurobiology of humans.

As a prominent example, schizophrenia is a complex, chronic and debilitating psychiatric disorder that affects ~1% of the population worldwide [9]. Patients can present with heterogenous combinations of positive symptoms (auditory and visual hallucinations), negative symptoms (social withdrawal) and cognitive dysfunction (impaired executive function and working memory). The heterogenous presentation of the disease oftentimes makes diagnosis and proper treatment difficult. Additionally, schizophrenia can be driven by both genetic and environmental factors, many of which are only just being understood [10, 11].

The 22q11 microdeletion (MD) syndrome is associated with the deletion of chromosome 22q11 and is associated with a phenotypic expression of many diseases including schizophrenia. While only 1% of schizophrenia patients have 22q11 MD syndrome, between 20 and 30% of individuals with the syndrome have gone on to be diagnosed with schizophrenia or schizoaffective disorders [12]. Furthermore, individuals with 22q11 MD syndrome that go on to develop schizophrenia show impairments of cognitive and social function as well as display changes in neuroanatomical features similar to the broad characterization of schizophrenia patients [13].

In order to optimally study complex diseases such as schizophrenia and 22q11 MD syndrome, genetic preclinical animal models can be utilized. Recently, a Df1/+ mouse strain, a model of heterozygous 22q11 MD, was shown to display anatomical and behavioral alterations that can be related to schizophrenia symptoms [14]. Therefore, in this study, we aim to develop a new computational systems framework capable of inferring translatable pathway biology between the 22q11 MD mouse model of schizophrenia and the presentation of the disease in a human cohort. The results of this new framework, which we term Translatable Pathways Classification (TransPath-C), will identify shared dysregulation of targetable pathways that can then be used to prioritize novel therapeutic targets. Moreover, TransPath-C should be useful for a broad range of other cross-species translation applications.

METHODS

Mouse and human datasets

The Df16(A)+/− mouse is a model of the 22q11-deletion syndrome (22q11DS) in humans. In a subset of the GEO GSE45935 dataset, 10 adult male Df16(A)+/− transgenic mice and 10 wild type (WT) littermates’ prefrontal cortex samples were reanalyzed for their gene expression using the Affymetrix Mouse Gene 1.1 ST Array microarray platform.

A subset of the human GEO dataset GSE53987, containing 18 control and 15 schizophrenia patients’ prefrontal cortex samples, was reanalyzed for gene expression using the Affymetrix Human Genome U133 Plus 2.0 microarray platform. Clinical information captured in the original dataset’s publication is reproduced in Supplementary Table 1 [15]. Of note, control and SZ groups did not differ in mean age, postmortem interval, brain pH, RNA integrity number or tissue storage time. The published records of the patients in this cohort had no documentation as to their 22q11 MD status.

Differential gene expression and pathway enrichment analysis

Within each dataset, genes for which one-to-one human–mouse orthologues exists were retained for analysis. Differentially expressed genes (DEGs) for the mouse and human datasets (False Discovery Rate [FDR] < 0.05) were calculated using the Limma statistical package in R. Additionally, a comparison of the resulting gene lists identified shared DEGs between the two datasets.

Within each dataset’s lists of differentially upregulated and downregulated genes, Panther Pathway enrichment analysis was performed using the PANTHER Classification System [http://www.pantherdb.org/] (FDR < 0.05).

Sample-based gene set enrichment analysis

Next, for each sample n, we ranked the genes based on their normalized expression value from this formula and performed ranked list gene set enrichment analysis (GSEA) [16, 17] using the list of pathways from the BIOCARTA Pathway database (MSigDB) [18]. BIOCARTA was chosen for our analysis from among other alternatives because we were able to find stronger relevance to neurobiological contexts. For each sample n, the pathway enrichment scores for the p number of pathways were recorded, resulting in a pathway enrichment matrix of p pathways x n samples.

Sparse principal component analysis

The mouse and human pathway enrichment matrices were individually transformed to have a mean of zero and variance of one for each pathway across each species’ individual samples using the PowerTransformer package in Python [19]. Next, sparse principal component analysis (sPCA) was performed on the mouse pathway enrichment matrix using the sparsepca package in R following penalty parameter optimization. In brief, the alpha penalty was varied between 10⁻⁶ and 0 using a sampling of 13 points. For each alpha penalty term, leave-one-out cross validation (LOOCV) was performed on the sPCA space to determine the mean and variance of the cumulative variation captured in the first three sparse principal components (sPCs). An optimal alpha penalty was determined to be one that had the largest penalty while still retaining at least 50% of the variance in the first three sPCs. Using the optimal penalty, an sPC space was constructed using the mouse data (Ms_sPC_Ms). Additionally, the human pathway enrichment matrix was projected into the mouse sPC space (Ms_sPC_Hu) by multiplying the transformed human pathway enrichment matrix by the eigen vectors of the Ms_sPC_Ms.

Regularized support vector machines of mouse and human phenotypes and comparison between models

Using the individual projections of the mouse and human samples in the mouse sPC space (Ms_sPC_Ms, Ms_sPC_Hu) with the corresponding phenotype labels (WT/22q11 and Control/Schizophrenia respectively), distinct L₁ regularized support vector machines (SVMs) were built for the mouse and human data using the LinearSVC function in scikit-learn package of Python. In brief, the L₁ penalty was varied between X and Y. Using LOOCV, the optimal L₁ penalty was determined by the minimization of the squared hinge loss function. Then, an SVM was trained on a species’ entire dataset. Using LOOCV with the optimal hyperparameter, the area under the receiver operating characteristic curve (AUC) was calculated. Next, non-zero weighted sPCs were analyzed for each species’ SVM model. Within each non-zero weighted sPC for each species, pathway loadings were determined and analyzed for the correlation with that species’ two phenotypes. Finally, pathways and their phenotype correlation were compared between the mouse and human models and pathways in common for the WT/Control or 22q11/Schizophrenia phenotypes.

RESULTS

Mouse genetic model of schizophrenia and human cohort have few shared DEGs

To better understand the complexity of schizophrenia, genetic preclinical models in mice have been developed, including the Df16(A)+/− transgenic line which models 22q11DS. In adults affected by this syndrome, 25–30% go on to develop schizophrenia [20, 21]. The goal of this analysis was to identify translatable biology between preclinical model and human cohorts with schizophrenia to (i) explore genetic dysfunctions in common between the two cohorts and (ii) shed light on potential therapeutics that could be successful in both the preclinical and clinical setting. To identify translatable biological findings, we first explored genes and pathways that were differentially expressed in the prefrontal cortex within the Df16(A)+/− transgenic mouse model (GSE45935, Ms22q11DS) and a cohort of control and schizophrenic patients (GSE53987, HuSZ). The Ms22q11DS cohort had a total of 1746 DEGs (Fig. 1A left, FDR < 0.05). In the HuSZ cohort, 73 genes were differentially expressed (Fig. 1A right, FDR < 0.05). Both sets of DEGs are provided in Supplementary Tables 2 and 3. In comparison to the Ms22q11DS cohort, the HuSZ DEGs had lower absolute fold change values with lower confidence via larger FDR values, which could be caused by inherent variability in a human versus transgenic mouse cohort. Additionally, the human cohort, unlike the mouse cohort, likely does not have the same genetic driver of schizophrenia across patients, resulting in higher variation in the human dataset, limiting the number of significant human DEGs. Between the Ms22q11DS and HuSZ cohorts, IGFBP6, FABP3 and PRKC1 were both upregulated in the WT and Control phenotypes compared with the 22q11DS and Schizophrenia phenotypes (which we will henceforth abbreviate as SZ), respectively (Fig. 1B intersection); however, inspection of these three genes indicated little biological pathway interaction between them. Next, to determine if there was greater shared pathway enrichment between the phenotypes of the two cohorts, we performed PANTHER Pathway enrichment analysis on the lists of up- and downregulated genes for the Ms22q11DS and HuSZ cohorts. In the Ms22q11DS cohort, the 22q11DS phenotype was significantly enriched for eight Panther pathways (Fig. 1C), including the ionotropic glutamate receptor pathway and metabotropic glutamate receptor groups I/III pathways. However, the HuSZ cohort had no enrichment of Panther pathways in their SZ phenotype so no direct comparisons could be made.

Translation-focused classifier can identify most germane dysregulated pathways from preclinical model to clinical cohort

Due to traditional approaches not lending well to identifying actionable translatable biology between preclinical and clinical models of schizophrenia, we next sought to develop a computational framework for accomplishing this. A previously demonstrated method from our group, Translatable Components Regression (TransComp-R), identified key molecular feature variations in a preclinical mouse model that were able to predict human responses to treatment in the context of inflammatory bowel disease and infer new therapeutic targets [8].

In our new study here, we have developed a novel method which we term Translatable Pathway Classification (TransPath-C) that can identify pathway variance in the preclinical Ms22q11DS cohort able to classify patients in the HuSZ cohort and consequently infer novel human pathway dysregulation. Figure 2 illustrates the computational pipeline that will be discussed further below. First, each species’ dataset is transformed from a gene expression matrix to a pathway enrichment score matrix. This is achieved by calculating a sample-based fold change for each gene and normalizing to the variance in expression for the control samples (Equation 1). Next, rank-based GSEA analysis is performed for each sample, resulting in a pathway enrichment score matrix for each species. Next, sPCA is implemented to identify axes, or sPCs, of variation within the pathway enrichment scores of the preclinical model. Ms22q11DS and HuSZ pathway enrichment data are then projected into the mouse sPC space. While it is possible that the unsupervised methods of sPCA may separate the phenotypes of the Ms22q11DS cohort, separation may be less apparent in the HuSZ cohort because the sPC space was built from the mouse cohort originally.

A L₁ regularized support vector machine classifier was then built for each species, using the projection of the species’ samples into the mouse sPC space as features for the model. Non-zero-weighted sPCs of each species’ classifier were then investigated for shared pathways with regard to phenotype between the mouse and human classifiers. Using this computational approach, we can identify, in an unsupervised fashion, variation in pathway enrichment of the Ms22q11DS cohort that is predictive of both the 22q11DS and SZ phenotypes in the mouse and human cohorts, respectively.

TransPath-C predicts mouse and human disease status from mouse pathway enrichment metrics

In order to reduce the Ms22q11DS and HuSZ cohorts from gene expression matrices to pathway enrichment matrices, we used the previously described sample-based preranked GSEA analysis using the BIOCARTA pathway collection from MSigDB (289 gene sets in total). When performing sPCA on the Ms22q11 pathway enrichment matrix, we first need to determine an optimal penalty value for sparsity. To do so with this unsupervised clustering approach, we identified the maximal penalty value that retained at least 50% cumulative variance within the first three principal components (Fig. 3A), as empirically we expect the majority of variance to be captured within the first few PCs in a non-sparse scenario. The resulting sPC space for the Ms22q11DS cohort was defined by 18 sPCs with at least one non-zero-weighted pathway comprising each sPC (Fig. 3B). Projection of the Ms22q11DS pathway enrichment matrix within the mouse sPC space (Ms_sPC_Ms) showed good separation of WT and 22q11DS samples across sPC2 which captured 9.2% of the total variance in the pathway enrichment dataset (Fig. 3C). However, when the HuSZ pathway enrichment matrix was projected into the mouse sPC space (Ms_sPC_Hu), there was no definitive separation of the Control and SZ phenotypes within the space of the first two sPCs (Fig. 3D).

With the scores matrix of the Ms_sPC_Ms and Ms_sPC_Hu projections, we next built two individual L₁-regularized SVM classifiers for each species, SVM_Ms and SVM_Hu, respectively. Using LOOCV to determine the optimal penalty parameter and overall model performance, we determined that the SVM_Ms had an AUC = 1 (Fig. 4A), with perfect classification of the 10 WT and 10 22q11 phenotypes (Fig. 4B). Additionally, sPC2 was the only non-zero-weighted sPC used by the final model built on all 20 samples (Fig. 4C), which was the same sPC that was able to successfully cluster the phenotypes in the unsupervised sPCA analysis (Fig. 3C). When analyzing model performance using LOOCV for the SVM_Hu model, we obtained an AUC = 0.75 (Fig. 4D), with the error primarily coming from incorrectly classifying some of the Control patients as the SZ phenotype (Fig. 4E). Additionally, the SVM_Hu model had a greater number of non-zero sPCs used to build the final model using all 33 patient samples (Fig. 4F); however, unlike the SVM_Ms model, sPC2 was not included.

TransPath-C identifies upregulated PAR1 pathway signaling in 22q11DS and schizophrenia phenotypes

While there were no shared sPCs used between the SVM_Ms and SVM_Hu models, we wanted to determine whether there were shared pathways contained within the sPCs used that correlated similarly between phenotypes of both models. Because sPC2 and sPC8 possess the greatest coefficient values in the SVM_Ms and SVM_Hu models, respectively, we sought their potentially shared pathways. Indeed, we found that enrichment for the PAR1 pathway correlated with the 22q11DS and SZ phenotypes, respectively (Fig. 4G and H). These results suggest that while methods used in Fig. 1 could not find translatable pathway biology, TransPath-C identified variance in the 22q11DS pathway enrichment, due in part to PAR1 pathway activity, that was predictive of both the Ms22q11DS and HuSZ cohorts. Furthermore, analysis of other pathways can help gain insight into each cohort’s biology that drives phenotype classification. For instance, while not shared between the mouse and human cohorts, MET1 pathway enrichment is associated with and predictive of the 22q11 mouse phenotype. Similarly, upregulation of the CCR5 and hSWI/SNF pathways are associated with and predictive of the SZ human phenotype.

DISCUSSION

A crucial task in better understanding the biology and targeted interventions for complex neurological diseases, such as schizophrenia, is the development of robust, translatable animal models. Genetically manipulated models, such as the 22q11 MD mouse model, discussed here offer phenotypic insight to a subset of genetic causes of the polygenic disease of schizophrenia. While humans with 22q11 MD syndrome represent only 1% of all adult schizophrenia cases, ~25–30% of these 22q11 MD syndrome patients go on to be diagnosed with schizophrenia [20, 21]. Given the polygenic nature of schizophrenia and the fact that genetic mouse models only capture a singular genetic factor to the disease, we aimed to determine whether computational species translation methods could identify translatable and actionable targets in the 22q11 MD mouse model that could be applied to a broader schizophrenia population [22]. Previous methods to achieve this have included the identification of shared genes between a human schizophrenia GWAS study and a recombinant inbred mouse panel BXD. Results identified a singular gene, APBB1IP, that was significantly associated with prepulse inhibition in the recombinant inbred mouse model and elevated risk score of schizophrenia in humans [23]. To date, however, the authors here have not identified any studies that investigate beyond identifying shared genes between preclinical models of schizophrenia and human patient cohorts.

Limiting variables in species translation efforts are (i) omic technologies used between species, (ii) cross-representation of features (genes, proteins, etc.) within these technologies within human and animal species and (iii) shared biology between the animal model and human cohort under consideration. Prior approaches have relied on shared technologies (i.e. mouse and human DNA microarrays) probing both preclinical animal and human cohorts to identify shared significantly expressed genes. However, when we performed this analysis between the 22q11 MD mouse model and human schizophrenia cohort, we were left with only three genes that did not lead to any actionable targets or shared pathway involvement. Previous computational translation models, including TransComp-R, have allowed for varying high throughput-omic technologies to be used to probe each species, increasing the number of comparative studies that could be done. However, even with TransComp-R, one was still confined to investigate genes within the model that were shared one-to-one orthologues within each species’ dataset. In the event that there is fewer overlap in one-to-one orthologues between the two species’ datasets, much of the data are not used in the resulting species translational model. In the TransPath-C approach presented here, genes included in pathway analysis are limited only to one-to-one orthologues that are present in each dataset individually, not those shared between the two datasets. This allows us to maximize the proportion of each species’ data that can be used in analysis. Additionally, previous gene-to-gene comparisons between mouse and human datasets have resulted in contradictory conclusions [24], likely due to filtering techniques for DEGs or technical variability between species cohorts and technology platforms used within each cohort. By collapsing our transcriptomic datasets to pathway activation datasets, we could use all one-to-one orthologues in each species and not simply those shared between the datasets when calculating pathway enrichment scores. Additionally, transforming transcriptomic datasets into pathway activation datasets reduces the dimensionality of our data, which is crucial when handling so few observational datapoints. Furthermore, the endpoint results of our TransPath-C approach are presented in the form of shared differentially regulated pathways between the preclinical animal model and human cohort. Since single target inhibitors have been shown to typically result in resistance and be ineffective for complex disease [25], identifying novel targets from a pathway perspective may be advantageous. Additionally, pathway-based screening techniques have shown increased hit rates for identifying novel therapeutics [26], further motivating our approach to identify shared pathway dysregulation in our schizophrenia species translation model. Overall, we believe that the output of our TransPath-C translational model will allow for the greatest possible success in the identification of novel therapeutics due to its pathway-based approach in the representation of the datasets. Using additional datasets from other mouse models and human cohorts could expand the scope of translational findings although disparities among the various within each species will require careful attention to the extent of generalization possible versus the degree of restriction—i.e. the notion of ‘precision medicine’.

The results of our mouse 22q11 MD-to-human schizophrenia translation model identified upregulation of the PAR1 pathway in each species’ diseased phenotype. The most decisive DEGs for this inference were GNA13 and PTK2B upregulation and PIK3CA downregulation, with GNA13 showing the strongest relationship with pathology in both mouse and human contexts. Previously, it has been shown that PAR1 can enhance neuron excitability associated with NMDA receptor-mediated neuronal damage in ischemia and intracerebral hemorrhage [27, 28]. Additionally, thrombin-mediated overactivation of the PAR1 pathway has been linked to neurodegenerative diseases including Alzheimer’s and Parkinson’s diseases [29]. Furthermore, PAR1 activation has been shown to suppress myelin gene transcription and lead to demyelination in the adult murine spinal cord [30]. Postmortem prefrontal cortex analyses of schizophrenia patients indicate impaired myelination in schizophrenia patients [31], which could therefore be linked to the overactivation of the PAR1 pathway indicated by our species translational model. Finally, GNA13 has been recently identified in a new machine learning approach to genome-wide association study data [32]. Therefore, we propose the PAR1 pathway mechanism shown in Fig. 5 as a translatable, targetable pathway between the 22q11 MD mouse model and human schizophrenia.

While direct experimental validation of a role for PAR1 pathway activity in the mouse model could be conducted, the most important proof will reside in human studies. Currently, PAR1 inhibitors are being tested in clinical trials for patients with percutaneous coronary interventions (PCIs). PAR1 drives platelet aggregation and thrombosis, and the PAR1 inhibitor, Vorapaxar, has undergone Phase II and recruitment for Phase IV clinical trial evaluation for myocardial infarction. Phase II results indicate that Vorapaxar was not associated with increased risk of bleeding during PCI and trended toward reduced incidence of ischemic events including peri-procedural myocardial infarction [33]. Furthermore, preclinical knockout models of PAR1 in mice undergoing myelin injury showed improved replenishment of myelinating cells and increased remyelinated nerve fibers [30]. Finally, treatment of rat Schwann cells with low levels of PAR1 agonist resulted in neural survival and elongation, while treatment with high levels of PAR1 agonist resulted in inhibition of neurite extension [34]. Therefore, due to neurodegeneration preclinical models and human clinical trials, PAR1 inhibitors, such as Vorapaxar, could be viable drug candidates for schizophrenia patients, as called out by our species translational model.

More broadly, we suggest that TransComp-C may be useful for translating biological insights across studies where minimal overlap in molecular features is apparent. This approach offers a capability for inferring pathway activities important in an application context of primary interest, such as human subjects, even when they are not directly evident in the experimental context of an animal model.

ACKNOWLEDGEMENTS

The authors would like to thank David Prober of the California Institute of Technology for biological insights and members of the Lauffenburger lab, in particular Brian Joughin and Doug Brubaker, for helpful technical suggestions. Figure 5 was created using BioRender. The use of trade, product or firm names in this report is for descriptive purposes only and does not imply endorsement by the US Government. Permission was granted by the Chief of Engineers to publish this information. The findings of this report are not to be construed as an official Department of the Army position unless so designated by other authorized documents.

FUNDING

This work was partially supported by the Institute for Collaborative Biotechnologies UARC contact W911NF-19-D-0001, Cooperative Agreement W911NF-19-2-0026 from the Army Research Office (DAL) and Joint Program Committee Military Operational Medicine Research Program (E.J.P., N.G.).

CONFLICT OF INTEREST STATEMENT

The authors have no conflicts of interest to declare.

References