Yeast Augmented Network Analysis (YANA): a new systems approach to identify therapeutic targets for human genetic diseases

Introduction

Many disease-associated genes have been identified, yet most genetic diseases remain untreatable. One path to treatment is to develop extensive genetic networks in which human disease genes function (or dysfunction) and then target therapies to the genes identified in those networks. Eventually, an individual’s own genotype for proteins in the network may also be considered in the therapeutic options. Developing gene networks is already widely recognized as a powerful approach to identify new drug targets^1–6. Gene networks are based on the principle that networks contain proteins that interact (physically and/or functionally) and that these interactions govern most, if not all, cellular functions. Importantly, gene network interactions are often conserved in different organisms even though the output of the networks may differ⁷. Candidate networks can be generated from studies in model organisms and then extrapolated to human cells. One can expect that, for disease gene networks, some of the interacting genes might modulate the disease phenotype thus representing potential therapeutic targets.

The direct identification of disease networks from studies in human cells is, of course, challenging because of the large number of potential proteins and incomplete knowledge of how the activity of any one protein may affect the network output. Analysis of networks would benefit from studies in a tractable model organism amenable to high throughput methods. Given that network interactions may be conserved between humans and model organisms, even while the outputs may differ, we hypothesized that we can identify human disease networks in the fission yeast Schizosaccharomyces pombe, a simple, genetically tractable model organism, coupled to existing protein-protein interaction databases. We can then transfer that network knowledge to human cells with the goal of identifying therapeutic target genes.

How valid is the approach of using fission yeast networks to identify human disease gene networks? First, the proteins that control most core cellular functions are in fact evolutionarily conserved, underscoring the “deep homology” that exists between all living organisms^8,9. For example, in S. pombe >65% of the genome is orthologous to human (http://orthomcl.org/orthomcl/ and http://www.pombase.org). Second, many of the cellular processes implicated in human disease, e.g. vesicular transport, protein folding, metabolism, and RNA processing, are also evolutionarily conserved and highly interconnected¹⁰. Importantly, the value of a simple model organism for discovering “druggable” genetic pathways has recently been demonstrated for the budding yeast Saccharomyces cerevisiae¹¹, and thus extending the study of human disease genes to fission yeast seems promising.

Our approach (Figure 1) involves the expression of human disease associated genes in S. pombe and then the analysis of their effect on yeast fitness (growth). We performed high-throughput synthetic genetic array (SGA) screens to identify the fission yeast genetic modifiers that alter this effect, as measured by a simple yeast growth assay. The genetic modifiers are then assembled into human disease gene networks or clusters (using protein-protein interaction datasets from both S. pombe and humans). Any modifiers or genes in the networks or clusters represent potential therapeutic targets. This unbiased high-throughput approach could be widely and rapidly applied to many different disease-associated genes at relatively low cost.

Figure 1. YANA analysis of UBA1 genes. Numbers illustrate various steps in our analysis.

(1) Construction of query strains by integrating wildtype and mutant variants of the UBA1 gene into pombe. (2) Crossing the query strains (mutant or wildtype) to a complete non-essential deletion collection. (3) Bioinformatic analysis to generate the top ‘gene hits’. (4) Validation of the top hits in yeast. (5) Confirmation of the top hits in zebrafish.

To demonstrate the power of YANA, we considered spinal muscular atrophy, a common neurodegenerative disease affecting approximately 1/6000 births worldwide and the number one genetic cause of infantile death in the United States¹². Most cases (>90%) are caused by deletion of SMN1, a gene encoding the survival motor neuron (SMN), a protein involved in the assembly of spliceosomal small nuclear ribonuceoproteins (snRNPs)¹³. In contrast, X-linked spinal muscular atrophy is caused by mutations in UBA1, a gene encoding the ubiquitin activating (UBA) enzyme 1, an E1 ubiquitin ligase^14,15. Consistent with a role of ubiquitination in spinal muscular atrophy and SMN biology, it was recently shown that ubiquitin-mediated proteolysis regulates SMN stability^16,17. Therefore, we systematically investigated the genetic network for UBA1 to better understand the potential connections between UBA1, its modifiers and the spinal muscular atrophy phenotype. Using YANA we identified several potential therapeutic targets and validated one of these targets in a SMA vertebrate (zebrafish) model.

Methods

Results

We expressed the human wildtype UBA1 (wtUBA1) and the disease-causing variant 1617UBA1 (mutUBA1) in S. pombe under the control of a regulatable nmt1 promoter²⁷. We identified strains with stable integrated human UBA1 genes expressing equivalent levels of UBA1 proteins (Figure 2A). Under non-inducing conditions (+ thiamine) all UBA1-containing strains had a similar growth rate to the control non-UBA1-containing wildtype strain (Figure 2B). We then found that cells expressing human wtUBA1 (fission yeast cells grown in media lacking thiamine to induce UBA1 expression), experienced a decrease in cell growth. No growth defect was observed for yeast cells expressing mutUBA1 when compared to the control (Figure 2B).

Figure 2. Construction of S. pombe strains integrated with human UBA1.

Wildtype and mutant forms of UBA1 were integrated into fission yeast under the control of the nmt1 thiamine-repressible promoter. (A) Western blot analysis of protein expression levels shows that both forms of UBA1 (wildtype and mutant) are expressed to equivalent levels. (B) Expression of wild-type UBA1 inhibits S. pombe cell growth.

We set out to identify yeast gene mutations that either enhance or suppress cell growth dependent upon expression of either wt or mutUBA1 (i.e., epistatic modifiers). We used automated genetics to introduce the wt and mutUBA1 genes into a deletion strain collection corresponding to more than 90% of the non-essential genes in S. pombe (ca. 80% of the complete genome). We then tested each of the >7000 unique strains expressing either wt or mutUBA1 combined with a specific gene deletion, for growth properties. Each strain was scored for significance of growth difference compared to a control, and the top hits (P<0.05) were assembled into a database (Table 1). The categories of growth phenotypes were those that were synthetic lethal for wildtype or mutant (wt-SL, or mut-SL) or synthetically suppressed for wildtype or mutant (wt-SS, or mut-SS). We identified 173 UBA1 or mutUBA1 modifiers. Notably, 145 of the modifiers (83.8%) were orthologous to human genes, and thus are potential drug targets to modify the SMA phenotype.

To generate our candidate human disease gene networks (that we shall call network clusters) we relied on two complementary approaches. Both approaches rely on networks constructed from published protein-protein interaction data from either human (Figure 3 and Figure S3) or fission yeast (Figure 4, Figure S1 and Figure S2). By including two types of genetic interactions, synthetic lethals or synthetic suppressors, yeast augmented network analysis (YANA), yields the most complete unbiased list of potential modifiers. In a first approach, we converted all 145 S. pombe modifier genes to their corresponding human orthologs. All the data were combined to create network cluster diagrams. Individual datasets were delineated using a color key. (Note that we designate each interaction as a synthetic lethal or synthetic suppressor a priori, and include them both as it is difficult to predict which of these classes will have potential therapeutic value). Using the highest confidence (0.9) experimentally confirmed protein-protein interaction data (from www.string-db.org), we used cytoscape to draw network cluster modules that include all modifier genes. We identified 22 clusters using this approach (Figure S3). We then limited the number of networks by focusing on those modifiers that shared at least two interactions with other modifiers shrinking the number of clusters to 10 (Figure 3). Within our human primary network clusters, we identified several genes that represent compelling therapeutic targets. These included GSK3 (Glycogen Synthase Kinase-3), CUL3 (Cullin-3) and NEDD4 (E3 ubiquitin ligase) along with SUMO1 (Small Ubiquitin Modifier-1) and HDAC (Histon deacetylase), a known chromatin modifier. Interestingly, GSK3 inhibitors increase SMN levels in spinal muscular atrophy patient-derived fibroblasts and mouse motor neurons²⁸. The HDAC inhibitor Trichostatin A increases SMN expression and survival in a mouse model of spinal muscular atrophy²⁹. Therefore, evidence already exists that genes within our clusters affect SMA biology.

Figure 3. Human UBA1 primary gene interaction network.

Human orthologs of yeast genes identified from four independent SGA screens were used to map the corresponding protein-protein interaction network. Only proteins that have experimentally verified data (STRING) supporting a physical interaction with at least two additional genes/proteins identified in the UBA1-SGA screens are displayed. Of particular interest is the discovery of NEDD4, CUL3, GSK3, SUMO and HDAC in distinct network clusters. Proteins of interest are highlighted by a larger node size and indicated with an arrow.

Figure 4. S. pombe extended UBA1 gene interaction network (extracted from Figure S1 on the basis of identifying new primary hit interactions through a 1° neighbor).

Primary yeast gene hits from all four SGA screens were combined and mapped based on highest confidence experimentally confirmed protein-protein interaction data. For this extended analysis, we also included first-degree neighbors of modifiers that were not directly identified in the SGA screens. Only selected gene clusters are shown here; the complete extended network is shown in Figure S1. When extrapolated to human, the yeast extended network analysis identified Csn5 and Cul3, two modifiers also identified in our human UBA1 primary gene interaction network (boxed; see also Figure 3).

In our second approach, we constructed networks from just the S. pombe modifier genes using high confidence S. pombe protein-protein interaction data from STRING. For this analysis, we identified network clusters that were expanded to include all the so-called “first-degree (1°) interacting proteins”, proteins not identified in our SGA screens but that physically interact with a genetically identified modifier (Figure S1). The rationale underlying our approach to generate larger clusters comes from the facts that the 1° neighbor may be an essential gene (and therefore it is not present in our deletion library) or may fail to be detected in our screens for other, unknown reasons. This “Cluster with 1° neighbor” analysis reveals new interactions not observed in the yeast primary UBA1 gene interaction network (Figure S2). A few examples of these types of clusters are shown in Figure 4. The clusters include several genes all of which were uniquely identified in the wildtype screen, including both synthetic lethals and synthetic suppressors. Of particular interest is the cluster of proteins surrounding Skp1(S-phase kinase associated protein-1), a protein that interacts and stablilizes F-box proteins, additional F-box proteins, and proteins involved in cullin deneddylation and neddylation. In addition, we found that the RNA-binding protein Cwf2 interacts directly with three of our modifiers, and the plastin ortholog Fim1, interacts with two. Interestingly, Plastin 3 (PLS3) has been identified as a disease modifier in animal models of spinal muscular atrophy^30–32.

We then identified correlations between the extended S. pombe network clusters (Figure 4) and the human primary network clusters (Figure 3). In both networks, we identified components of the COPS complex, a known regulator of E3 cullins, and the E3 cullin Cul3. Note that Cul3 was not present in our primary S. pombe network (Figure S2) yet was present in the more extended network clusters that incorporate 1° neighbors; this illustrates the importance of adding 1° neighbors to our analysis of the S. pombe data.

Together, these data suggest a prominent role for E3 ubiquitin ligases in modulating UBA1 and potentially SMN1, making Cul3 the primary target for further analysis. The E3 ubiquitin ligase might inhibit SMN function, as ubiquitination often leads to protein degradation. If so, we reasoned that inhibition of E3 ubiquitin ligase might enhance SMN activity and thus suppress loss of function mutations in SMN1. To directly test the hypothesis that inhibition of E3 ubiquitin ligases can suppress a SMN1-spinal muscular atrophy phenotype, we turned to a vertebrate model of SMN deficiency previously reported in zebrafish³³. Knockdown of Smn in zebrafish embryos causes developmental defects in motor neuron axonal outgrowth that include truncations and abnormal branching of neurons²⁴. Motor neuron axon defects can be corrected by injection of mRNAs encoding wildtype human SMN²⁵. We first confirmed that injection of smn morpholino (MO) into wildtype zebrafish causes severe motor axon abnormalities as compared to control uninjected embryos (Figure 5A and 5B). We then tested whether an inhibitor of E3 ubiquitin ligases (MLN4294) would suppress the motor axon abnormalities. Addition of the Nedd8-E1 activating enzyme inhibitor MLN4294 (that blocks cullin-RING E3 ligases), at concentrations ranging from 10 µM to 15 µM, caused a concentration-dependent reduction in the degree of abnormal motor axon branching (Figure 5A and 5B). At 15 µM inhibitor we found motor axon abnormalities that were completely suppressed: defects were not significantly different to those observed in uninjected embryos (P=0.379). At higher concentrations of drug (20 µM), the rescue was less pronounced (Table 2) probably because higher levels of drug led to defects in development. Thus, E3 ligase inhibition can rescue neuronal defects caused by Smn protein depletion in zebrafish. In control experiments, MLN4294 failed to rescue the zebrafish mutant topped (Figure 6), a mutant defective in neuronal axon guidance³⁴. We also tested a drug that inhibits sumolyation corresponding to an unrelated target (SUMO) identified in our human primary network clusters (Figure 3); addition of this compound had no effect on the spinal muscular atrophy model (Figure S6) necessarily surprising since pmt3, which encodes the yeast ortholog of SUMO1, has the opposite effect of cul3 when deleted. Having been identified as a synthetic suppressor rather than a synthetic lethal, perhaps compounds that activate rather than inhibit sumolyation would be beneficial as a therapeutic.

Figure 5. MLN4294 rescues the abnormal neuron outgrowth in Smn-depleted zebrafish.

(A) Representative lateral views of motor axons in Tg(mnx1:GFP) zebrafish embryos expressing GFP in motor neurons and injected with control MO and then grown in 10 or 15 µM MLN4294. (B) Quantification of the effects of MLN4294 on motor axon development in zebrafish. Motor axons were scored in Tg(mnx1:GFP) embryos injected with control MO, and subsequently (10 hrs post-injection) incubated in 10 or 15 µM MLN4294. Embryos were classified as severe, moderate, mild, or no defects based on the severity of motor axon defects, and the percentage of each group is shown. Data in all graphs are represented as mean and SEM. (C) Western blot analysis of Smn and HuD protein following treatment with MLN4294. Quantification of the results is shown in Figure S5.

Figure 6. Treatment of zebrafish with MLN4294 fails to rescue defects in ventral motor axon guidance in the topped mutant.

(A) Quantification of the effects of MLN4294 on motor axon guidance in the zebrafish topped mutant. Motor axons were scored in embryos injected with DMSO, 10 or 15 µM MLN4294. Embryos were classified as severe, moderate, mild, or no defects based on the severity of motor axon defects, and the percentage of each group is shown. Data in all graphs are represented as mean and SEM. (B) Experimental data.

To determine how E3 ubiquitin ligases might influence Smn1 function, we examined the protein levels for both Smn and HuD. HuD is a known SMN interacting protein that binds RNAs controlling their translation and stability and functions in neural development and plasticity^34–36. We observed greater than a 2-fold increase in both Smn and Hud protein levels following treatment with MLN4294, a NEDD8 activating enzyme (NAE) inhibitor that prevents activation of E3 ligases (Figure 5C and Figure S5). It is therefore possible that rescue of the neuronal defects in our zebrafish model might involve stabilization of Smn as well as of additional proteins within the Smn network.