Genome-scale networks link neurodegenerative disease genes to alpha-synuclein through specific molecular pathways

Abstract

Numerous genes and molecular pathways are implicated in neurodegenerative proteinopathies, but their inter-relationships are poorly understood. We systematically mapped molecular pathways underlying the toxicity of alpha-synuclein (α-syn), a protein central to Parkinson’s disease. Genome-wide screens in yeast identified 332 genes that impact α-syn toxicity. To “humanize” this molecular network, we developed a computational method, TransposeNet. This integrates a Steiner prize-collecting approach with homology assignment through sequence, structure and interaction topology. TransposeNet linked α-syn to multiple parkinsonism genes and druggable targets through perturbed protein trafficking/ER quality control and mRNA metabolism/translation. A calcium signaling hub linked these processes to perturbed mitochondrial quality control/function, metal ion transport, transcriptional regulation and signal transduction. Parkinsonism gene interaction profiles spatially opposed in the network (ATP13A2/PARK9, VPS35/PARK17) were highly distinct, and network relationships for specific genes (LRRK2/PARK8, ATXN2 and EIF4G1/PARK18) were confirmed in patient iPS cell-derived neurons. This cross-species platform connected diverse neurodegenerative genes to proteinopathy through specific mechanisms, and may facilitate patient stratification for targeted therapy.

Graphical abstract

INTRODUCTION

Common neurodegenerative diseases result in the loss of distinct neuronal populations and abnormal accumulation of misfolded proteins. Synucleinopathies – including Parkinson’s disease (PD), dementia with Lewy bodies and multiple system atrophy – are associated with abnormal intracellular aggregation of αsynuclein (αsyn). Alzheimer’s disease (AD) is associated with amyloid-β (Aβ) and tau accumulation, while amyotrophic lateral sclerosis (ALS) is associated with altered localization and accumulation of TAR DNAbinding protein 43 (TDP-43), and so forth. The richest source of hypotheses regarding the pathogenesis of these diseases has derived from neuropathology of postmortem brain. While providing pivotal insights, these observations are made decades after disease inception.

A revolution in human genetic analysis over the last twenty years has uncovered disease-causing mutations that connect protein-misfolding to the neurodegenerative process. For instance, point mutations and gene multiplication at the α-syn (SNCA) locus lead to rare but early-onset, highly penetrant forms of PD and dementia. Polymorphisms in regulatory regions of the SNCA locus that increase gene expression confer increased risk for late-onset PD(Fuchs et al., 2008; Nalls et al., 2014). These studies enabled the creation of animal and cellular disease models and enriched our understanding of disease mechanisms. But with this knowledge, a new set of challenges has emerged.

First, seemingly disparate genes have been tied to particular clinical phenotypes. For example, parkinsonism is characterized by slowness (bradykinesia), rigidity, tremor and postural instability. The most common form is PD, defined by α-syn pathology and loss of dopaminergic neurons. However, numerous other disease entities - tied to distinct genetic signatures and neuropathology – can lead to parkinsonism, demonstrating that there is not a simple correspondence between genotype, neuropathology and clinical presentation (Martin et al., 2011; Shulman et al., 2010; Verstraeten et al., 2015). Those few genetic loci with parkinsonism as the primary clinical phenotype have been given a numeric “PARK” designation (for example, SNCA/PARK1 locus itself and LRRK2/PARK8), but even mutations in the same gene can produce distinct neuropathology and diverse clinical presentations(Martin et al., 2011; Rajput et al., 2006; Shulman et al., 2010; Verstraeten et al., 2015). Understanding the inter-relationship between genetic risk factors for parkinsonism, and their relationship to α-syn itself, is vital for patient stratification and targeted therapeutic strategies.

Second, human genetic studies have sometimes produced ambiguous and controversial data. For rare variants, substantial recent genetic divergence of human populations may render traditional methods of cross-validation between different populations unfeasible (Nelson et al., 2012; Tennessen et al., 2012). Inconsistencies in the literature abound – for example, studies implicating UCHL1 as “PARK5” and the translation initiation factor EIF4G1 as “PARK18” have failed to reproduce. For common polymorphisms, the challenge is distinguishing between multiple candidate gene loci in linkage to a SNP. It is becoming clear that biological validation will be required to fully establish which genetic factors are causally related to disease processes, and how (Casals and Bertranpetit, 2012).

One approach to validating candidate gene variants, and understanding their relationship to proteinopathy, is to systematically screen the entire genome to identify every gene that modifies proteotoxicity when over-expressed or deleted. This is achievable in Baker’s yeast (Saccharomyces cerevisiae), a unicellular eukaryote of unparalleled genetic tractability. Yeast has proved highly informative for understanding the cytotoxicity induced by misfolded proteins(Khurana and Lindquist, 2010). This is not surprising because human genetic data for neurodegenerative diseases heavily implicate cellular pathways that are among the most highly conserved in eukaryotic evolution, including protein homeostasis and quality control, protein trafficking, RNA metabolism and mitochondrial function (Bras et al., 2015; Guerreiro et al., 2015).

Expressing toxic proteins relevant to neurodegeneration in yeast creates a robust and easily scorable growth/viability defect amenable to genome-wide phenotypic screening in yeast. Toxicities of α-syn, beta-amyloid and TDP-43 have been screened by individually over-expressing one of ~5500 ORFs that comprise the majority of the yeast genome(Khurana and Lindquist, 2010; H.-J. Kim et al., 2013; Treusch et al., 2011; Yeger-Lotem et al., 2009). These screens have guided the discovery of cellular pathologies in neurons and animal models(Cooper et al., 2006; Dhungel et al., 2014; Khurana and Lindquist, 2010; H.-J. Kim et al., 2013), shed important insights on the relationship of genetic modifier data to gene-expression analysis(Yeger-Lotem et al., 2009), and led to the identification of novel human disease genes(Elden et al., 2010). Recently, processes pinpointed by phenotypic screening in a yeast synucleinopathy model led to the discovery of cellular pathologies in induced pluripotent stem cell (iPSc)-derived neurons from patients with PD due to mutations at the α-syn locus(Chung et al., 2013). In that study, integrating high-throughput genetic and small-molecule screens identified genes and small molecules that could correct pathologies from yeast to neurons (Chung et al., 2013; Tardiff et al., 2013; 2014).

Here, we build genome-scale networks of α-syn and other proteotoxicities by combining a new computational approach with substantially broader yeast genetic screens. To discover meaningful molecular connections in yeast and patient-derived neurons, we develop a TransposeNet algorithm that: 1) maps yeast hits to their human homologs by considering sequence, structure and molecular interactions; 2) builds networks by linking yeast hits and hidden human genes through an optimization framework based on the prizecollecting Steiner forest algorithm (SteinerForest Ensemble); and 3) transposes molecular interactions across species from yeast to human, utilizing the unparalleled density of known molecular interactions in yeast to compensate for the relative sparseness of the human interactome. The networks linked many parkinsonism and neurodegenerative disease risk factors to α-syn toxicity through specific molecular pathways, most notably vesicle trafficking and mRNA metabolism.

RESULTS

SteinerForest Ensemble networks uncover biological connections between α-syn screen hits

One conventional approach to creating a network from a gene list is to connect them via known genetic or physical protein-protein interactions. To illustrate, we considered 77 genes that modify α-syn toxicity in our previous over-expression screen (Table S1 and Table S2). Even with the rich yeast interactome, 30 hits were not incorporated into the network (Fig. 1A, upper panel). Moreover, some genes (“hubs) occupied a central position in the network, not because of their importance to proteotoxicity, but because they were connected to more genes. For example, PMR1 is a hub that has 955 annotated interactions in BioGRID compared to the median of 70 interactions across the 77 modifiers (Fig. 1A, upper right; Table S2).

Figure 1. SteinerForest Ensemble builds proteotoxicity networks for yeast and uncovers druggable targets.

A) SteinerForest Ensemble methodology vs. conventional approach. 77 genetic modifiers (“hits”) from a previous over-expression screen against α-syn toxicity are mapped to the yeast interactome. The conventional approach misses 30 genetic hits and overemphasizes “hub” genes like PMR1. SteinerForest Ensemble includes all 77 hits and predicts additional nodes of biological relevance including the druggable targets Rsp5 and Calcineurin (Cnb1).

B) Left: Hits from three published over-expression screens for α–syn, Aβ and TDP-43 proteotoxicities in yeast. Venn diagrams indicate the numbers of genetic modifiers recovered. Right: A comparison of the output SteinerForest Ensemble networks generated from inputting these three sets of screen hits (empirical p-value for 1000 similarly connected random networks is shown for triple-wise comparison).

C) Growth curves demonstrating effects of a compound that activates Rsp5, NAB, on TDP-43 (left) and α-syn (right) toxicity. Yeast expressing either yellow fluorescent protein (YFP), TDP-43(TDP), or α-syn were treated with 20 μM (for TDP-43) or 10 μM (for α-syn) NAB. Growth was monitored over time by optical density (OD) at 600 nm. Results are representative of three experiments.

To build more inclusive networks, we adapted the “Prize-collecting Steiner Forest (PCSF) algorithm”, which connects gene or protein “nodes” through molecular interactions, or “edges” (S.-S. C. Huang and Fraenkel, 2009; Tuncbag et al., 2013; 2016) (Fig. 1A, lower). Edges can include genetic or physical interactions, or annotated pathways from curated databases(Szklarczyk et al., 2014) and are refined by minimizing “cost.” Costs increase 1) when a “prized” node (an original hit from a genetic screen) is excluded; 2) when an “edge” connecting two nodes derives from a low-confidence interaction; or 3) when edges connect to hubs. To ensure that our PCSFs were not dependent on specific parameterization, we generated an ensemble of 112 individual forests with different algorithm parameters, and created an averaged, or “collapsed”, representative network through a maximum spanning-tree algorithm (“Steinernet Ensemble”“; Fig. 1A, lower right).

To encompass the largest number of prized nodes while avoiding unlikely interactions, the PCSF method introduces “predicted nodes”, which are proteins or genes not part of the original prized hit list, (Fig. 1A, green triangles). Predicted nodes will occur between two nodes within the network. However, as the final network is a superposition of many different networks, these may be at the periphery in the final Ensemble output. Predicted nodes can add biological value because any high-throughput screen will miss many true biological connections.

When we applied SteinerForest Ensemble to our previous α-syn over-expression screen data, the fragmented networks became more coherently connected. All 77 modifier-genes were now incorporated in the network, (Fig. 1A, lower left; Table S1 and Cytsocape Data S20; Table S3). By penalizing the exclusion of genetic modifiers and the inclusion of hubs, the PCSF algorithm favored the biological context at the expense of hubs. To establish specificity of the network output, we generated ensembles of forests from 1000 sets of 77 genes randomly chosen from the yeast genome with identical connectivity (degree distribution) to the α-syn modifier list. An empiric p-value for each node (based probability of occurring in a network by chance) was significant (p=0.025, SD=0.021).

Importantly, predicted nodes (Fig. 1A, panel 4, green triangles) included genetic modifiers of α-syn toxicity not hit in the original screen but uncovered through other studies, including Sec14 (Phospholipase D)(Outeiro and Lindquist, 2003), and Pbp1 (yeast homolog of Ataxin-2 see below and Fig. 3). This network also identified two druggable targets: Cnb1 (Calcineurin subunit B) and Rsp5 (Fig. 1A, lower right). Cnb1 is targeted by FK506, a drug that ameliorates α-syn toxicity (Caraveo et al., 2014). Rsp5 is the target of a specific N-arylbenzimidazole (NAB) that protects against α-syn toxicity(Tardiff et al., 2013). The SteinerForest Ensemble methodology thus connects genetic screen hits through biologically relevant pathways, including druggable targets.

Figure 3. TransposeNet builds genome-scale molecular network for α-syn toxicity from genome-wide deletion and over-expression yeast screens.

A) Summary of genetic modifiers recovered in screens. 16 genetic modifiers (14 unique) from low-throughput investigations were also incorporated. Yeast homologs of genes linked to PD and other neurodegenerative disorders are listed. “y” preceding the human gene name indicates the “yeast homolog”.

B) A “humanized” network is generated from the 332 α-syn screen hits by TransposeNet. Genes of interest are enlarged, including multiple neurodegeneration-related disease genes (see also Supplemental Fig. 8 and Table S14). Gene ontology process enrichment within “stems” of the network are shown color-coded (full details in Table S12; gray portions were not enriched). Brown lines indicate extrapolated connections to VCP/Cdc48 through Vms1 (the yeast homolog of Ankzf1) and Hrd1 (the yeast homolog of Syvn1), and from VCP to Parkin/PARK2 and Pink1/PARK6. A target symbol marks two druggable nodes, Calcineurin(Caraveo et al., 2014) and Nedd4 (a target of NAB(Tardiff et al., 2013)). Inset: Network without transposition of yeast edges. LRRK2 and NFAT not included. Ontologically connected proteins (for example Rab proteins) are dispersed.

Cross-comparison of α-syn, TDP-43 and Aβ proteinopathies reveals distinct and shared mechanisms

To cross-compare different proteinopathies, we examined previous Aβ and TDP-43 over-expression screens (Fig. 1B; “yeast over-expression networks” in Table S1 and Cytoscape Data S21 and S22) and found virtually no overlap (Fig. 1B, left; Table S2). There was, however, reassuring overlap between the yeast genetic modifiers and disease genes associated with the human disorders including: putative parkinsonism genes recovered in the α-syn screen [ATP13A2 (PARK9) and EIF4G1 (PARK18)]; AD risk factors in the Aβ screen [PICALM, CD2AP, INPP5D and RIN3]; an ALS genetic risk factor in the TDP-43 screen (Elden et al., 2010).

SteinerForest Ensembles from these screen hits revealed more biological overlap between these proteinopathies including protein trafficking, mRNA translation, ubiquitination and cell-cycle genes (Table S3 and Table S4; Fig. 1B right). There was also a cross-over between genetic risk factors for distinct human diseases: the ATXN2 homolog was a predicted node in the α-syn network (confirmed as a modifier of α-syn toxicity; Figs. 3 and 4 below); the homolog of the parkinsonism gene VPS35 (PARK17) was a predicted node in the yeast Aβ network. VPS35 encodes a key component of the retromer complex, and defective retromer function has been identified in AD brain and animal models (Small et al., 2005). These overlaps were unrelated to increasing the number of genes. Empirical p-values for 1000 similarly connected random networks were statistically significant, whether considered pairwise (p<=0.002) or triple-wise (p<=0.001).

Figure 4. Genetic dissection of parkinsonism susceptibility genes reveals distinct biology.

A) Vesicle trafficking subnetwork within the α-syn map (from Fig. 3B) and location of PARK9 (ATP13A2). Green: trafficking proteins; brown: metal ion transporters.

B) Synthetic toxic interactions between trafficking genes and α-syn (spotting assays on agar plates). A-syn transgene is expressed from a galactose-inducible promoter (“on” in galactose, Gal; “off” in glucose, Glc). “y” ahead of the human gene name indicates the yeast homolog. ΔGAL2 strain (“nonspecific enhancer”) serves as a (+) control because it grows less well on galactose (regardless of α-syn expression). The (-) control, a deletion (ΔYMR191W), has no deleterious effect in presence of α-syn (“baseline toxicity”).

C) Expression of yeast VPS35 (yVps35), human VPS35 (hVps35), and human mutant (D620N) VPS35 in α-syn-expressing Vps35-deleted “IntTox” cells (yeast spotting assay, showing serial 5X dilution from left to right; transgenes are expressed from a galactose-inducible promoter).

D) – E) Cross-comparison of genetic interactors with similarly toxic HiTox α-syn, α-syn-ΔVPS35/PARK17, α-syn-ΔATP13A2/PARK9 strains. Spotting assay demonstrates relative levels of toxicity among these three strains (D; 5X serial dilution from left to right). In E), data is shown on dot-plots comparing the efficacy of 77 known α-syn modifiers (see Fig. 1) in HiTox α-syn (x-axis) versus ΔPARK17/α-syn (y-axis; D) ΔPARK9/α-syn (y-axis; E). Green: vesicle trafficking genetic modifiers, brown: metal ion transport modifiers. Axis scales represent growth relative to Mig1/Mig3 positive controls (= 100, black). Mig1/Mig3 over-expression represses the galactose promoter driving α-syn expression. Each spot assay in this figure was repeated 2–4 times. The dot plot is representative of two experiments performed on separate days with biological replicates. Transformants were plated in quadruplicate for each experiment.

One trafficking gene predicted to be a common node between all three proteinopathies was Rsp5, a ubiquitin ligase activated by NAB. Indeed, NAB was originally recovered in a small-molecule screen against TDP-43 proteinopathy in yeast. We utilized a sensitive bioscreen assay to test NAB on growth defects induced by these proteinopathies. Indeed, NAB rescued all three proteinopathies as predicted by the network. It was most effective for α-syn (Fig. 1C) and only rescued against Aβ toxicity synergistically in combination with other compounds known to protect from Aβ toxicity (manuscript in preparation). NAB failed to provide significant rescue for any 20 unrelated toxic yeast strains (Fig. S1).

TransposeNet generates “humanized” network

It would be desirable to identify connections between our yeast molecular networks to human genes, including human disease genes that have no straightforward homologs in yeast. We therefore developed TransposeNet, a suite of computational methods to “humanize” yeast molecular networks (Fig. 2A).

Figure 2. A “humanized”TransposeNet network incorporates LRRK2 into the α-syn proteotoxicity network.

A) A “humanized” network is generated from the 77 α-syn over-expression screen hits by TransposeNet. Each yeast gene (red triangle) is linked to its human homolog(s) (blue circle) by a weight proportional to the homology strength. Edges are weighted based on their experimental level of confidence. Certain nodes are enlarged for emphasis. LRRK2 is linked within network via NSF1 and STUB1. Inset: Density of known molecular interactions in yeast and human (Biogrid http://wiki.thebiogrid.org/doku.php/statistics). Abbreviations: DCA (diffusion component analysis); PARK (known “parkinsonism” gene). See Supplement for complete network.

B) The effect on the “humanized” network of withholding yeast edge augmentation.

C) Accumulation of Nicastrin in the endoplasmic reticulum (ER) in LRRK2^G2019S mutant iPSc-derived dopaminergic neurons compared to mutation-corrected control neurons. Endoglycosidase H (Endo H) removes post ER glycosylation and reveals the ER form of Nicastrin, an ER-associated degradation substrate. The post ER-to-ER ratio was calculated using the ratio of the mature form over the deglycosylated ER form. Data are represented as mean ± SEM (n=2 for patient 1 and n=3 for patient 2, ***; p<0.0001, two tail t-test).

The first step in Transposenet is assignment of yeast-to-human homology by considering sequence similarity (BLAST and DIOPT(Hu et al., 2011) scores), yeast-to-human structure alignments (using the HHpred tool)(Söding et al., 2005), and incorporating network topology (Fig. 2A, upper left). Network topology assesses neighborhoods of genetic and physical molecular interactions around a given protein, positing “guilt-by-association” logic that the topological place within a network relates to biological function(Cho et al., 2016). Thus sharing similar neighbors should be a factor in determining whether two proteins are homologs. The relative weight of each homology method was carefully tuned (STAR Methods and Fig. S2 for full details) providing a more comprehensive and unified protein homology score (Berger et al., 2013; Singh et al., 2008; Söding et al., 2005). The underlying framework that relates genes according to these different features is known as diffusion component analysis (DCA). DCA has also been used as the core algorithm in Mashup, a tool for integrating multiple hetergeneous interactomes. More information can be found in the Method Details section of the STAR Methods and in Cho et al., (2016).

Our method assigned 4923 yeast proteins to human homologs and conversely predicted yeast homologs for 15,200 human proteins, a substantial improvement over BLAST (4023 yeast to human and 7248 human to yeast) or BLAST with HHpred (4312 yeast to human and 9577 human to yeast). Additionally, our method improved predictions as determined by gene ontology (GO) accuracy and Jaccard similarity scores (STAR Methods; Fig. S3) and did not introduce false-positives for high confidence yeast-human proteins pairs (EnsemblCompara; STAR Methods).

These is high conservation of core eukaryotic biology from yeast to man, and pivotal complementation studies in yeast have determined the functions of many genes in other species, including human (Osborn and Miller, 2007) (Kachroo et al., 2015). On this basis, we used our homology tool to augment the human interactome with interactions inferred from the much better-curated yeast interactome. This was the key advance that enabled TransposeNet. Importantly, this cross-species “edge” transposition did not increase the rate of false-positive hits. Rather, it substantially improved network performance. In fact, for our screen hits the PCSF-based SteinerForest Ensemble out-performed two alternative network building methodologies, DAPPLE(Rossin et al., 2011) and PEXA(Tu et al., 2009) (STAR Methods and Fig. S4).

In our “humanized networks” (indexed in Table S1; Fig. 2A, right) each yeast gene (red triangle) is connected to one or more human homologs (circles) based on our homology tool-generated score. SteinerForest Ensemble then interconnects each resulting human gene/protein, through edges generated from the human interactome augmented with the “humanized” yeast molecular interactome. If a particular human homolog of a yeast genetic modifier had been implicated as a parkinsonism gene, a small inclusion weight is given. However, no special preference was given to any human disease genes other close homologs of our yeast hits.

Humanized network from over-expression screen connects α-syn to other human disease genes

We tested the humanized network approach on the 77 modifiers from the α-syn over-expression screen (“α-syn over-expression humanized network”; Table S1, Table S9 and Table S11; Cytoscape Data S23; Fig. 2A, right). Several predicted human nodes in the resultant humanize network had no obvious homologs in the yeast proteome, the most striking example being α-syn itself. α-syn was connected to ER quality control and protein trafficking modifiers through a predicted node Ap1b1 (Fig. 2A, right), a component of the clathrin adapter complex that localizes in the immediate vicinity of α-syn in neurons (accompanying manuscript; Chung et al. Cell Systems 2016). The emergence of α-syn in the humanized network strongly indicates that a functional, highly interconnected relationship between our original yeast genetic hits and α-syn is conserved from yeast to man.

LRRK2 and α-syn are connected through ER stress and vesicle trafficking

The kinase/GTPase LRRK2, another PD gene-encoded protein without an obvious yeast homolog, was centrally incorporated into the humanized network (Fig. 2A, right). We tested the robustness and specificity of this finding by computationally generating ensembles of humanized Steiner forests from 1000 lists of genes that were randomly selected (matching the size of our original α-syn genetic modifier list. LRRK2 and α-syn (SNCA) occurred together in 72% of humanized networks generated through PCSF from our input list (individually, SNCA appeared in 86% and LRRK2 in 76% of networks). Neither was incorporated in any of the humanized networks generated from Aβ or TDP-43 screen hits (“TDP-43”-and “Aβ”-”over-expression humanized networks” in Table S1 and Cytsocape Data S29 and S30). LRRK2 and α-syn appeared together in 0/1000 networks of the randomly generated ensembles. Without transposition of yeast interaction information into our networks, α-syn was peripherally placed and its connection to Ap1b1 (see above) was lost and LRRK2 was entirely absent (Fig. 2B). Thus, the inclusion of LRRK2 and α-syn is robust, specific, and dependent upon augmentation of human networks with yeast interaction data.

LRRK2 was related to the humanized α-syn network through proteins involved in ER-to-Golgi trafficking (Nsf1, Rab1a) and ER quality control (Stub1/Chip/Scar6, Sgk1, Syvn1), pathways previously implicated in LRRK2(Cho et al., 2014; G. Liu et al., 2012)- and α-syn(Chung et al., 2013; Cooper et al., 2006)-induced toxicity. Our data suggested they might be a key point of convergence. We previously showed that the A53T mutation and triplication of wild-type α-syn leads to pathologic accumulation of specific trafficked proteins in the ER of patient-derived neurons, including Nicastrin (Chung et al., 2013). Using previously generated LRRK2 mutant iPSc-derived neurons, we recapitulated this phenotype (Fig. S5). As early as 4 weeks after initiating differentiation, Nicastrin accumulated in the ER of LRRK2^G2019S cells compared to isogenic mutation-corrected controls (Fig. 2D) phenocopying the previously described pathology in neurons derived from patients with α-syn mutations. Thus, the humanized α-syn network correctly predicted a convergence of cellular pathologies in distinct forms of parkinsonism. A Nicastrin trafficking defect has also been demonstrated in LRRK2 knockout mouse fibroblasts(Cho et al., 2014), raising the possibility that the G2019S mutation may lead to deficiency of a LRRK2-related function in protein trafficking.

Genome-wide pooled overexpression and deletion screens against α-syn toxicity

For a more comprehensive view, we executed two additional genome-wide screens against α-syn toxicity (Fig. S6A number 1): 1) a genome-wide deletion screen identified nonessential genes that, when deleted, lead to an extreme sensitivity to low levels of α-syn that would otherwise not be toxic (Fig. S6A number 2; Fig. S6; Table S5); 2) a pooled screen in which the galactose-inducible over-expression library was transformed en masse into our α-syn HiTox strain (Fig. S6A number 3; Supp Fig. 7; Table S6). For pooled screens, we sequence plasmid DNA to identify genes specifically over- or under-represented under selective conditions comparing plasmid DNA sequences from a similarly transformed YFP control strain to identify α-syn-specific modifiers. These are putative suppressors and enhancers of toxicity, respectively. Pooled screens are far more rapid, and theoretically more sensitive, than individually transforming each library plasmid into the α-syn strain and measuring growth.

These screens encompassed tests of approximately 10,000 potential genetic interactions (~5500 overexpression, ~4500 deletion). After extensive validation of the hits (Fig. S6C and Fig. S7B), we recovered 318 genetic modifiers. Very little overlap existed between the specific genes recovered in the deletion versus the over-expression screens (Fig. 3B). However, we found considerable overlap in the biological pathways represented (see network analysis below). 16 modifiers have emerged in independent work from our laboratory(Caraveo et al., 2014; Chung et al., 2013) or were identified herein (Table S7). Fourteen of these were distinct from our screen hits, leading to a total of 332 genetic modifiers of α-syn toxicity (Fig. 3B).

Homologs of PARK and other neurodegeneration genes modify α-syn toxicity in yeast

Modifiers of α-syn toxicity included homologs of many known genetic risk factors for parkinsonism and other neurodegenerative disease phenotypes (Fig. 3A and Table S14). These included genes involved in vesicle trafficking (yRAB7L1, yRAB39B, ySORL1, ySYNJ1/PARK20, yVPS35/PARK17), mRNA translation (yATXN2, yEIF4G1/PARK18), mitochondrial quality control/function (yCHCHD2/10), metal ion transport (yATP13A2), transcriptional regulation (yATXN7), metabolism (yDHDDS) and signaling (yPDE8B, yPPP2R2B/ATXN12), among others. Many of these genes, including those at so-called PARK loci, have been implicated in neuronal pathologies quite distinct from the α-syn pathology that defines PD. Their recovery in our screens suggested that mechanisms of neurotoxicity related to diverse neurodegenerative disease genes may be shared.

Of the 19 PARK loci, 9 have clear yeast homologs (Table S8). Four of these emerged in our screens: yATP13A2 (PARK9) [YPK9 in yeast], yVPS35(PARK17) [VPS35], yEIF4G1(PARK18) [TIF4631, TIF4632] and ySYNJ1(PARK20) [INP53]. A fifth putative PARK gene, yRAB7L1(PARK16) [YPT7], emerged as a genetic modifier when tested as a candidate (see below). The probability of recovering homologs of these genes by chance is p=0.00013 (hypergeometric test. None of these yPARK genes were modifiers in the Aβ or TDP-43 over-expression screens (Table S2). These findings underscore the biological specificity of the α-syn screen hits in yeast.

TransposeNet generates a genome-scale “map” of α-syn toxicity

We applied TransposeNet to homologs of the 332 α-syn toxicity modifiers to generate a humanized network, or “map” (“Complete α-syn humanized network” in Table S1, Table S10 and Table S11; Fig. 3B; Fig. S8; Cytoscape Data S24). Multiple genes implicated in neurodegeneration emerged in this α-syn network by direct homology to yeast hits or as predicted network nodes (Fig. 3B; Fig. S8; Table S14).

We superimposed gene ontologies onto “branches” in our map (Fig. 3B) and various vesicle-mediated transport processes dominated. Genetic risk factors associated with typical PD (SNCA itself, LRRK2, RAB7L1, VPS35) were concentrated in the subnetwork enriched in vesicle trafficking genes (Fig. 3B). In contrast, the majority of neurodegeneration genes associated with non-Lewy neuropathology, atypical parkinsonism or non-parkinsonian neurodegenerative phenotypes (Table S14) were distant from the vesicle trafficking network. A full analysis of the biological processes enriched in the network “branches” is provided in Table S12. Notably, this humanized network elucidates the molecular context in which the previously identified druggable targets NEDD4(Tardiff et al., 2013) and calcineurin and NFAT (Caraveo et al., 2014) impact α-syn toxicity (Fig. 3B).

Furthermore, both α-syn itself and LRRK2 are predicted as nodes, just as in the over-expression network (Fig. 2A). In the ensemble of Steiner forests generated from our list of 332 modifiers, α-syn appeared in 100% and LRRK2 in 70%. In 1000 random sets of 332 genes, even when we forced the incorporation of five yPARK genes recovered in our genetic experiments (yPARK9, yPARK16, yPARK17, yPARK18, yPARK20), α-syn and LRRK2 appeared together in only 0.6% of humanized networks. Thus, yeast modifiers of α-syn toxicity generated a specific humanized network in which the PD-associated proteins α-syn and LRRK2 emerged as critical network nodes.

TransposeNet generated a coherent network: 295 out of 332 of yeast modifier genes in a single tree network (Table S10) with biologically intuitive “stems” comprising genes of similar ontology (Fig. 3B). Networks generated from these 332 modifiers without transposition of yeast interactome data (Fig. 3B, inset; Cytoscape Data S25) produced three fragmented networks comprising 136, 2 and 122 yeast genes, respectively (Fig. 3B, inset). Genes that should be related biologically through involvement in common cellular processes were not (Fig. 3B). Moreover, LRRK2 and NFAT were not incorporated. Testable hypotheses, such as the relationship of EIF4G1 and ATXN2 (Fig. 5 and Fig. 6, below), did not emerge because these proteins landed in different networks. DAPPLE(Rossin et al., 2011) and PEXA(Tu et al., 2009) also produced highly fragmented or dense “hairball” networks useless for hypothesis generation (Fig. S9) and, strikingly, did not include critical nodes like LRRK2 (Fig. S10). Thus, transposition of yeast networks to augment the human interactome created a coherent, biologically meaningful α-syn network.

Figure 5. mRNA translation factors impact α-syn toxicity from yeast to patient-derived neurons.

A) mRNA translation subnetwork in α-syn toxicity (from Fig. 3B), including ATXN2, EIF4G1 (PARK18) and PABPC1.

B) Effects of yPABPC1, yAtaxin2 and yEIF4G1-1 on α-syn toxicity (left: quantitative PCR; right: bioscreen growth assay).

C) Bulk mRNA translation in mutant α-syn^A53T iPSc neurons compared to isogenic mutation-corrected control neurons as measured by ³⁵S-cysteine and ³⁵S–methionine incorporation over time (phosphorimager scan). Commassie staining shows loading of protein. Two subclones of the mutation-corrected line were compared to α-syn^A53T cells (n=4).

D) TALE-TFs were designed to elevate the endogenous levels of ATXN2 or EIF4G1 genes. These bind to the 5′ UTR of the target genes, and recruit a transcriptional activator(Sanjana et al., 2012). Q-PCR indicates transcript levels after AAV-mediated TALE-TF delivery into A53T iPS neurons.

E) Effect of increasing endogenous EIF4G1 or ATXN2 levels on bulk translation in A53T neurons (n=3). Data are represented as mean ± SEM. *; p<0.05 **; p<0.01 two tail t-test).

An endocytic and retrograde trafficking subnetwork in the α-syn toxicity map incorporates yeast homologs of RAB7L1 (PARK16) and VPS35 (PARK17)

In the α-syn map, homologs of some parkinsonism genes coalesced in a sub-network around YPT6, the yeast homolog of RAB6A (Soper et al., 2011)(Fig 4A). Included were YPT7, VTH1 and VPS35, which encode proteins involved in endosomal trafficking. YPT7 is a close homolog of RAB7L1, a leading candidate for the PARK16 locus(Macleod et al., 2013; Nalls et al., 2014), and also of the Mendelian parkinsonism risk factor RAB39B(Wilson et al., 2014). VTH1 is a close yeast homolog of SORL1, an established AD risk modifier(Rogaeva et al., 2007) that encodes a protein involved in intracellular sorting(Nykjaer and Willnow, 2012). VPS35 is homologous to the Mendelian risk factor for classic PD, VPS35 (PARK17)(Zimprich et al., 2011). VPS35, with VPS26 and VPS29, comprise the retromer complex that transports cargo from endosomal to golgi compartments. In our accompanying study (Chung et al. Cell Systems 2016), we show that deletion of the VSP26 and VPS29 core retromer components strongly enhances α-syn toxicity. A fourth gene involved in golgi-to-endosome and endocytic trafficking, INP53, is homologous to the Mendelian parkinsonism gene SYNJ1(PARK20)(Olgiati et al., 2014). Deletion of any one of these genes was not toxic in a wild type strain. However, deletion of any one of these genes in a strain expressing low (nontoxic) levels of α-syn produced a strong and synergistic growth defect (Table S5, Fig. 4B and Fig. S10A). Importantly, ectopic expression of yeast or human VPS35 rescued the toxicity induced by deleting VPS35, but expression of a disease-causing mutation (VPS35-D620N) did not (Fig. 4C). Finally, yRAB7L1 enhanced α-syn toxicity when deleted, but rescued from this toxicity when over-expressed (Fig. S10B).

The α–syn map predicts diverging genetic interaction profiles in ΔPARK9 (ATP13A2) and ΔPARK17 (VPS35)-sensitized yeast models

To test functional consequences of being located in distinct subnetworks of our α-syn map, we compared VPS35 (PARK17) and ATP13A2 (PARK9). ATP13A2 is a type 5 P-ATPase implicated in cation transport and metal ion homeostasis(Kong et al., 2014; Park et al., 2014; Ramonet et al., 2012; Tsunemi and Krainc, 2014). Mutations in ATP13A2 lead to juvenile-onset parkinsonism or Kufor-Rakeb syndrome, which is distinct from PD (Schneider et al., 2010). yATP13A2 suppressed α-syn toxicity in our over-expression screen (Fig. 1C) and deletion of yATP13A2 strongly enhanced α-syn toxicity (Fig. 4B). In our humanized network, ATP13A2 was spatially distant from VPS35 lying well outside the vesicle trafficking subnetwork (Fig. 3C and Fig. 4A). We asked whether this spatial separation reflected differences in underlying biology.

We generated three strains with similar toxicities (Fig. 4D). In one strain toxicity resulted from overexpression of α-syn (HiTox). In two other strains, mild toxicity induced by intermediate levels of α-syn expression was enhanced by deletion of yeast ATP13A2 (hereafter, ΔATP13A2/α-syn) or VPS35 (hereafter, ΔVPS35/α-syn). These three yeast strains thus modeled cellular pathologies related to three forms of familialpParkinsonism: two with typical α-syn pathology (PD related to α-syn multiplication and VPS35 (PARK17)-associated parkinsonism) and one with strikingly different pathology, PARK9 (ATP13A2).

While ΔATP13A2 sensitizes cells to metal ion stress(Kong et al., 2014), ΔVPS35 strains exhibit retrograde trafficking defects (Seaman et al., 1997) suggesting that ΔATP13A2 and ΔVPS35 strains were differentially sensitized to α-syn toxicity. We asked whether our 77 α-syn over-expression screen hits affected the toxicity of our ΔVPS35/α-syn and ΔATP13A2/α-syn models.

We expressed these α-syn toxicity modifiers in each of the yeast models and monitored growth. For the α-syn HiTox and ΔVPS35/α-syn models, 69/77 genes overlapped (Fig. 4E, left), correlating well with the similar pathology associated with these genetic forms of parkinsonism. Notably, the overlapping modifiers were enriched in vesicle trafficking genes (Table S13). In contrast, there were only 3/77 modifiers in common between α-syn HiTox and ΔATP13A2/α-syn (Fig. 4E, right). These were involved in iron and manganese homeostasis (CCC1) and actin cytoskeleton rearrangements (ICY1, AFI1), respectively. Notably, metal ion homeostasis is strongly implicated in Kufor-Rakeb syndrome (Schneider et al., 2010) and its mammalian models (Park et al., 2014). Thus, neurodegenerative diseases that are genetically, clinically and neuropathologically distinct may nonetheless share some key molecular pathologies.

mRNA translation subnetwork links α-syn to PABPC1, EIF4G1 and ATXN2

In our over-expression screen against α-syn toxicity, TIF4632 (hereafter, yEIF4G1-2) emerged as a suppressor. TIF4632 is a yeast homolog of the of translational initiation factor EIF4G1. The genome-wide deletion and pooled over-expression screens identified additional genetic modifiers related to mRNA translation, including initiation factors and multiple ribosomal subunits (Fig 3B and Fig 5A; Table S5 and Table S6). These included PABPC1 (cytoplasmic poly(A)-binding protein-encoding), the homolog PAB1; the ATXN2 homolog PBP1; and the second EIF4G family homolog in yeast, TIF4631 (hereafter, yEIF4G1-1). These hits were confirmed by quantitative PCR (Fig. 5B, left), and overexpression of these genes suppressed α-syn toxicity in bioscreen (Fig. 5B, right) and/or spot (Fig. S11) growth assays. Genetic experiments in different proteinopathy models revealed that the effects of these modifiers on α-syn toxicity were specific (Fig. S11). Thus, perturbation of mRNA translation was not simply a generic proteotoxic response.

Protein translation is perturbed in PD patient-derived neurons

Because we recovered numerous genetic modifiers in the mRNA translation and mRNA processing pathways (Fig. 3 and Fig. 5), we asked whether protein translation was perturbed in cellular syucleinopathy models, including PD patient-derived neurons. Bulk changes in protein production were initially assessed by determining the rate at which S³⁵-radiolabeled methionine and cysteine is incorporated into protein over time after a brief “pulse labeling”. Over-expression of α-syn in HEK (human embryonic kidney) cells and primary rat cortical neurons reduced the accumulation of S³⁵-Met/Cys without changing amino acid uptake (Fig. S12). Similarly, in 6–8 week-old iPSc neurons harboring the α-syn^A53T mutation, S³⁵-Met/Cys incorporation into protein was reduced compared to isogenic mutation-corrected controls (Fig. 5C). Thus, our α-syn screens and network analysis identified a strong effect of α-syn toxicity on bulk mRNA translation in cellular models of synucleinopathy. This effect was not attributable to a canonical ER stress response, because phosphorylation of EIF2A (Fig. S12D; Fig. S13A) or XBP1 splicing (Fig. S13B) was not altered in these neurons.

Conserved genetic interactions of ATXN2 and EIF4G1 from yeast to patient neurons

We next tested whether human homologs of two translation factors that suppressed α-syn toxicity when overexpressed - ATXN2 and EIF4G1 - could similarly reverse the protein translation defect in neurons. We generated TALE-TF constructs to transcriptionally upregulate neuronal isoforms of EIF4G1 and ATXN2 (Sanjana et al., 2012) (Fig. 5D, left), These constructs were then delivered with an adeno-associated viral vector to differentiated α-syn^A53T iPSc-derived neuronal cultures.

Ten days after transduction, endogenous EIF4G1 and ATXN2 mRNA levels increased by approximately 4-fold, as measured by quantitative PCR (Fig. 5D, right). This increased expression substantially reversed the defect in bulk protein translation we had observed in α-syn^A53T neurons (Fig. 5E). Over-expression of EIF4G1 increased translation in A53T neurons but not in mutation-corrected controls. ATXN2 over-expression equally increased bulk translation in mutant and control cells (Fig. 5F).

Thus, our cross-species molecular network predicted a biological interaction between α–syn and mRNA translation factors in PD patient-derived neurons. These data strengthen the argument that perturbed mRNA translation is an important aspect of α-syn toxicity. Interestingly, we identified a strong signature of decreased translation of mRNA translation-related transcripts in ribosomal footprinting experiments of α-syn^A53T iPScderived cortical neurons at 4 and 12 weeks of differentiation compared to isogenic mutation-corrected control neurons (Fig. S14A and 14B, Fig. S15, Tables S16 through S19). Indeed, mRNA-related translation factors, ribonucleoproteins and ribosomal proteins were not only enriched in our genetic and translational maps of α-syn toxicity, but also among proteins in the immediate vicinity of α-syn in neurons (Chung et al, Cell Systems 2016, accompanying manuscript Fig. 4; “spatial α-syn map”, Fig. S15A). Moreover, a number of mRNA translation proteins directly complexed with α-syn. This convergence of genetic, translational and spatial maps suggests the connection between α-syn toxicity and mRNA metabolism is deeply rooted in α-syn biology.

Discussion

We describe a coherent, systems-level analysis of how α-syn misfolding and mistrafficking perturbs cell biology. Genome-wide screens identified modifiers of the toxic consequences of α-syn expression in Baker’s yeast S. cerevisiae. Our key computational contribution, TransposeNet, coupled richly annotated molecular interactions in yeast with a Steiner prize-collecting algorithm and a sophisticated cross-species homology tool to visualized the screen hits as a “humanized” molecular network. TransposeNet revealed that α-syn pathology is deeply connected to human genetic risk factors for parkinsonism, and parsed out the molecular pathways through which these connections occur. We envisage TransposeNet as a valuable resource for the community, easily generalizable to the modeling of any physiologic or pathologic process in genetically tractable organisms

A pressing challenge in neurodegeneration is to determine whether genes associated with highly distinct pathologies, but that nevertheless converge on similar clinical phenotypes, are related at a molecular level or not. Our network tied α-syn not only to genes that cause classical PD (Ogaki et al., 2015), but also genes that cause parkinsonism with different pathologies, and genes associated with other neurodegenerative phenotypes altogether (Table S14). The relationships were highly specific to α-syn. Moreover, genes tied to classical PD or α-syn pathology (like RAB7L1, VPS35 and LRRK2) were concentrated in a vesicle trafficking-associated subnetwork, while genes tied to “atypical” parkinsonism (like ATP13A2 and ATXN2) were in separate subnetworks. For a few examples, our network revealed convergent and divergent molecular pathologies related to the spatial location on the map. Thus, LRRK2 and α-syn pathologies were connected at the level of perturbed protein trafficking, confirmed in patient-derived neurons. In another example, VPS35 and ATP13A2^{^} exhibited highly distinct genetic modifier profiles in yeast. Other network and model-organism studies provide important support for our results, including connections between α-syn and LRRK2(Cho et al., 2014; G. Liu et al., 2012), RAB7L1 and LRRK2 (Macleod et al., 2013) (Beilina et al., 2014) and between VPS35 and α-syn (Dhungel et al., 2014).

For some genes connected to α-syn toxicity by our network, including EIF4G1(PARK18) and CHCHD2, human genetic data is limited or controversial (Funayama et al., 2015; Z. Liu et al., 2015; Ogaki et al., 2015) (Chartier-Harlin et al., 2011; Nuytemans et al., 2013). Another gene, RAB7L1, is one of two candidates in linkage with a parkinsonism-associated common gene variant (PARK16). Our analysis affirms a connection between such genes and α-syn proteinopathy and provides the biological context in which to place these interactions. For example, we make no claim that the translation factor EIF4G1 should be designated a “PD gene.” However, EIF4G1 and α-syn toxicity are connected in the context of an important and previously unrecognized direct effect of α-syn on mRNA biology and protein translation. This was confirmed by multiple hits in our genetic analysis (Fig. 3) and in our mRNA translational profiling of neurons (Fig. S14). Interestingly, an emerging connection is emerging between mRNA translation and other parkinsonism-related genes (Gehrke et al., 2015) (Martin et al., 2014). Moreover, a connection to mRNA translation and metabolism was also confirmed in our spatial mapping of α-syn in neurons (Chung et al., Cell Systems 2016, accompanying manuscript). This map revealed that α-syn is in the immediate vicinity and complexed to proteins involved in mRNA translation and protein trafficking, suggesting that these perturbations may be upstream or proximal events in a-syn toxicity.

Finally, by identifying connections between druggable targets and gene networks, our approach provides a glimpse of how treatments might in the future might be targeted to specific genetic lesions. We envisage that the inflexibility of a single clinical or pathologic diagnosis will yield to a more nuanced molecular diagnosis. In this scenario, genetic lesions will be matched to compound targets, and confirmed in “personalized” cellular models in which combinatorial genetic lesions are introduced to reflect specific genetic risk and biology. Emerging genome-editing technologies will enable such models to be developed in patientderived cells, and genome-wide screening to be carried out as well(Hasson et al., 2013; Khurana et al., 2015; Shalem et al., 2014; Wang et al., 2014). These will unquestionably be welcome advances, but impressive developments will continue in simple model organisms. Variomic libraries in yeast, for example, now enable mutagenesis at single-amino acid resolution across the entire yeast proteome(Z. Huang et al., 2013), unlocking enormous potential for target identification in phenotypic screens. We envision multi-faceted, cross-species approaches will continue to evince critical insights into many complex diseases, and perhaps fulfill therapeutic promises in the post-genomics era.

STAR METHODS

CONTACT FOR REAGENT AND RESOURCE SHARING

Further information and requests for resources and reagents should be directed to the Lead Contact, Vikram Khurana, by email at vkhurana@bwh.harvard.edu or by post at:

• Ann Romney Center for Neurologic Diseases, Brigham and Women’s Hospital

• Harvard Stem Cell Institute

• Building for Transformative Medicine, 10^th Floor

• 60 Fenwood Road, Boston MA 0211

EXPERIMENTAL MODELS AND SUBJECT DETAILS

Yeast Strains

For the deletion screen, strains were in the BY4741 background and have been described in detail elsewhere(Baryshnikova et al., 2010; Tong and Boone, 2006)

For all experiments except the deletion screen and validation, the yeast strains used were in the w303 background (MATa can1-100, his3-11,15, leu2-3,112, trp1-1, ura3-1, ade2-1). The vector control strain contained empty vector at the trp and ura loci (pAG304Gal, pAG306GAL). The NoTox α-syn strain contained α-syn fused to green fluorescent protein (α-syn-GFP) inserted at the his locus (pAG303Gal-α-syn-GFP). IntTox and HiTox α-syn strains contained multiple tandem copies of α-syn-GFP inserted at this and trp loci (pRS303GAL-α-syn-GFP, pRS304GAL-α-syn-GFP). IntTox strains have 4–5 copies of α-syn while HiTox cells have >6 copies of α-syn. The ΔPARK17/α-syn and ΔPARK9/α-syn were generated by replacing the PARK17/VPS35 or PARK9/SPF1 gene loci in IntTox α-syn strains with a kanamycin resistance cassette (VPS35::kanMX or SPF1::kanMX).

Human iPSc lines

iPSCs from control individuals and PD patients carrying G2019S LRRK2 along with isogenic gene-corrected controls were generated as previously described(Reinhardt et al., 2013). Skin biopsy, human dermal fibroblast culture, iPS cell generation and mutation correction for the patient harboring the A53T mutation (α-syn^A53T) have been described previously(Cooper et al., 2006; Soldner et al., 2011). In that previous publication the A53T iPS line was referred to as WIBR-IPS-SNCA^A53T. For all iPSc lines, informed consent was obtained from patients prior to cell donation using a written form, and the protocol was approved by the relevant institutional review board: for LRRK2 iPSCs this was the Ethics Committee of the Medical Faculty and the University Hospital Tübingen (Ethik-Kommission der Medizinischen Fakultät am Universitätsklinikum Tübingen); for the A53T line, the IRB of the Boston University Medical Campus and the MIT Committee on the Use of Humans as Experimental Subjects.

Human iPSC generation and differentiation into midbrain dopaminergic (DA) neurons for LRRK2 mutant lines

iPSCs were differentiated into mDA neurons using a floor plate-based protocol with minor modifications(Kriks et al., 2011; Schöndorf et al., 2014). Differentiation was based on exposure to LDN193189 (100 nM, Stemgent) from days 0–11, SB431542 (10 mM, Tocris) from days 0–5, SHH C25II (100 ng/mL, R&D), purmorphamine (2 mM, EMD) and FGF8 (100 ng/mL, Peprotech) from days 1–7 and CHIR99021 (CHIR; 3 mM, Stemgent) from days 3–13. Cells were grown for 11 days on Matrigel (BD) in knockout serum replacement medium (KSR) containing DMEM, 15% knockout serum replacement, 2 mM L-glutamine and 10 μM β-mercaptoethanol. KSR medium was gradually shifted to N2 medium starting on day 5 of differentiation. On day 11, media was changed to Neurobasal/B27/L-Glut containing medium (NB/B27; Invitrogen) supplemented with CHIR (until day 13) and with BDNF (brain-derived neurotrophic factor, 20ng/ml; R&D), ascorbic acid (0.2 mM, Sigma), GDNF (glial cell line-derived neurotrophic factor, 20 ng/ml; R&D), TGFβ3 (transforming growth factor type β3, 1 ng/ml; R&D), dibutyryl cAMP (0.5 mM; Sigma), and DAPT (10 μM; Tocris,) for 9 days. On day 18, cells were dissociated using Accutase (Innovative Cell Technology) and replated under high cell density conditions on dishes pre-coated with 15 μg/ml polyornithine and 1 μg/ml laminin in differentiation medium (NB/B27 + BDNF, ascorbic acid, GDNF, dbcAMP, TGFβ3 and DAPT). At DIV30, cells were collected and, after centrifugation, cell pellets were stored at −80°C until further analysis.

Human pluripotent stem cell culture for α-syn mutant lines

Skin biopsy, human dermal fibroblast culture, iPS cell generation and mutation correction for the patient harboring the A53T mutation (WIBR-IPS-^A53T) have been described previously(Cooper et al., 2006; Soldner et al., 2011). In that previous publication the A53T iPS line was referred to as WIBR-IPS-SNCA^A53T.

Our pluripotent stem cell lines were initially maintained (5%0₂, 3%C0₂) on mitomycin C inactivated mouse embryonic fibroblast (MEF) feeder layers in hES medium [DMEM/F12 (Invitrogen) supplemented with 15% fetal bovine serum (FBS) (Hyclone), 5% KnockOut Serum Replacement (Invitrogen), 1 mM glutamine (Invitrogen), 1% nonessential amino acids (Invitrogen), 0.1 mM β-mercaptoethanol (Sigma) and 4 ng/ml FGF2 (R&D systems)]. Cultures were passaged every 5 to 7 days either manually or enzymatically with collagenase type IV (Invitrogen; 1.5 mg/ml). At around 50 passages prior to differentiation, lines were passaged to plates pre-coated with growth factor-reduced matrigel (BD Biosciences; 1:30 in DMEM:F12) and cultured (21% 0₂, 5% CO₂) in mTESR-1 medium (Stem Cell Technologies), thereafter being passaged every 5 to 7 days enzymatically with dispase (Invitrogen; 1mg/mL) until differentiation (at passage 40–90). For karyotyping, standard G-banding chromosomal analysis of cell lines was performed every 10–20 passages (Cell Line Genetics, Inc). We confirmed mycoplasma-negative status of our cultures every 2–4 weeks (MycoAlert, Lonza).

Primary rat cortical cultures

All animal work was approved by the MIT Committee on Animal Care. Embryos were harvested by cesarean section from anesthetized pregnant Sprague-Dawley rats at embryonic day 18. Cerebral cortices were isolated and dissociated with Accumax (Innovative Cell Technologies, Inc) digestion for 20 min at 37 °C and triutration with Pasteur pipette. Poly-ornithine and laminin-coated 96 well plates were seeded with 4×10⁴ cells respectively in neurobasal medium (Life Technologies) supplemented with B27 (Life Technologies), 0.5 mM glutamine, 25 μM β–mercaptoethanol, penicillin (100 IU/ml) and streptomycin (100 μg/ml). One third of the medium was changed every 3 to 4 days.

METHOD DETAILS

Yeast-to-human homology

Since yeast and human are evolutionarily distant species, to identify human homologs for yeast proteins, we developed a four-tiered meta-analysis pipeline. Our meta-analysis started at the sequence level, in which we first identify genes/proteins that are similar across yeast and humans. We then extend this analysis to the structural level, where we investigate the proteins that are structurally, and thus more distantly, similar across the species. Next, we identify proteins that are similar within each species by using a network-topology based approach. Finally, we introduce an approach to integrate similarity across sequence, structure and network topology. Details are as follows:

1) Sequence Similarity

To compute the sequence similarity between a yeast protein and a human protein, we used NCBI protein BLAST with the BLOSUM62 substitution matrix(Altschul et al., 1990; 1997). Sequence similarity was computed for all pairs of yeast proteins and human proteins. We used an E-value threshold =1E-5 to determine significance. We also used DIOPT(GTEx Consortium, 2013; Hu et al., 2011; Reinhardt et al., 2013; Söding et al., 2005), an integrative ortholog prediction webserver, to predict human orthologs for each yeast protein. We stored all filtered yeast-human protein pairs together with their BLAST E-values, bit scores and DIOPT scores.

2) Evolutionary and Structural Similarity

For each yeast and human protein, we applied PSI-BLAST to construct a multiple sequence alignment and build a profile hidden Markov model to encode a remote evolutionary signature. We then applied HHpred(Kriks et al., 2011; Robinson and Oshlack, 2010; Schöndorf et al., 2014; Söding et al., 2005; Voevodski et al., 2009), with the profile hidden Markov models and secondary structure annotations as input, to compare all pairs of yeast proteins and human proteins. As with the sequence similarity calculation, we also used an E-value=1E-5 threshold. We stored all filtered yeast-human protein pairs with their HHpred E-values and bit scores.

3) Network Topology (Diffusion Component Analysis; DCA)

The central idea behind our network topology approach is to try to capture functionally-related modules at the protein level, so that each node can be represented with a low-dimensional vector, instead of a single score, that captures homologous proteins in the network, along with conserved patterns of interactions. The eventual goal (see Integrative Approach, below) is to be able to compare low-dimensional representations of node vectors across species to yield information in other organisms. However, if we follow a straightforward PageRank-like approach(Cho et al., 2015.; Tuncbag et al., 2016; Voevodski et al., 2009) to compute each node’s vector, we get inaccuracies in functional similarity prediction due to network noise. Thus, using the intuition that compression decreases noise, we reduce the dimensionality of the vectors using sophisticated machine learning techniques. Our approach has been shown to reduce noise and be better able to extract topological network information such as functional similarity(Bailly-Bechet et al., 2011; Cho et al., 2015). The approach has recently been generalized into a method called Mashup (Cho, H. et al 2016).

More formally, let A denote the adjacency matrix of a (weighted) molecular interaction network G = (V; E) with n nodes, each denoting a gene or a protein. Each entry B_i,j in the transition probability matrix, which stores the probability of a transition from node i to node j, is computed as B_i,j = A_i,j/Σ_k A_i,k. The diffusion algorithm is then defined as

$$s_i^{t+1}=(1-p)s_i^{t}B+p{e}_i$$

until convergence, where p is the probability of restart, controlling the relative influence of local and global information in the network; e_i is a binary vector with e_i(i) = 1 for node i itself and e_i (j) = 0 for other nodes j. When the diffusion patterns of two nodes are similar to each other, it implies that they are in proximal locations in the network with respect to other nodes, which potentially suggest functional similarity. In practice, diffusion vectors obtained in this manner are still noisy, in part due to their high dimensionality as well as the noise and incompleteness of the original high-throughput network data. With the goal of noise and dimensionality reduction, we approximate each diffusion vector with a multinomial logistic model based on a latent vector representation of nodes that uses far fewer dimensions than the original vector. Specifically, we compute the probability assigned to node j in the diffusion vector of node i as:

$$\stackrel{\sim }{s_{ij}}=exp(w_i^T{x}_j)/{\mathrm{\sum }}_{k}exp(w_i^T{x}_k)$$

where superscript T denotes vector transposition; w_i and x_i are low-dimension vectors. Each node is given two vector representations, w_i and x_i. We refer to w_i as the context feature and x_i as the node feature of node i, both capturing the intrinsic topological properties in the network. This multinomial logistic regression model is applied to model the relevance between a node and other nodes in a network, which can be modeled as a discrete distribution over all nodes in a network. To obtain w and x vectors for all nodes, we optimize the KL-divergence (or relative entropy) between the diffusion vectors s_i and the model vectors $\stackrel{\sim }{s_{\iota }}$:

$${\mathit{min}}_{w,x}C(s,\stackrel{\sim }{s})={\mathrm{\sum }}_i{D}_{KL}(s_i\Vert \stackrel{\sim }{s_i})$$

Akin to PCA, which reveals the internal low-dimensional linear structure of matrix data that best explains the variance, this approach computes a low-dimensional vector-space representation for all genes such that the connectivity patterns in the network can be best explained. Comprehensive experiments showed that these low-dimensional vectors w and x are more accurate at identifying functional association within the network (Cho et al., 2016.; Tuncbag et al., 2013).

4) Integrative Approach

To compare proteins from yeast and human, we extended the above DCA method to consider the topology of both interactomes as well as the sequence/structural similarity between them. We converted the sequence and structure similarity scores to a probability distribution, and feature vectors of all pairs of nodes, including the sparse vector representations ones, were jointly computed by minimizing the Kullbeck-Leibler (KL) divergence between the relevance vectors and the parameterized multinomial distributions.

Formally, here we have two interactomes, G_Y for yeast and G_H for human. To capture the topological similarity within interactomes, we perform the described diffusion algorithm on G_Y and G_H separately and then obtain diffusion vectors $s_i^Y$ for yeast protein i and $s_j^H$ for human protein j. Similar to DCA on a single network, we also assign vectors $w_i^Y,x_i^Y$ for each yeast protein, and vectors $w_i^H,x_i^H$ for each human protein. To the sequence/structural similarity between obvious homologs, we normalize the BLAST bit scores between each yeast protein i and its human homologs j into a probability distribution as $b_{ij}^Y={\mathit{bit}}_{ij}/{\mathrm{\sum }}_k{\mathit{bit}}_{ik}$. Similarly we also normalize the BLAST bit scores between each human protein j and its yeast homologs i into a probability distribution as $b_{ji}^H={\mathit{bit}}_{ij}/{\mathrm{\sum }}_k{\mathit{bit}}_{ik}$. We likewise do the same normalization for HHpred bit scores as $h_{ij}^Y$ and $h_{ji}^H$, and $h_{ij}^Y$ and $d_{ji}^H$ for DIOPT scores. Between each yeast protein i and human protein j, we approximate each normalized bit score distribution vector with a multinomial logistic model as:

$$\stackrel{\sim }{t_{ij}}=exp(w_i^T{x}_j)/{\mathrm{\sum }}_{k}exp(w_i^T{x}_k)$$

Similar to the definition of $\stackrel{\sim }{s_{ij}}$ for genes in the same molecular network, $\stackrel{\sim }{t_{ij}}$ captures the homologous similarity between a yeast gene and a human gene. In this way, although in different networks, yeast and human genes are represented in the same vector space.

Finally, we optimize an extended DCA objective function as:

$$\begin{array}{l}{\mathit{min}}_{w^Y,w^H,x^Y,x^H}\sum _{i\in V_Y}{D}_{KL}(s_i\Vert \stackrel{\sim }{s_i})+\sum _{j\in V_H}{D}_{KL}(s_i\Vert \stackrel{\sim }{s_j})+{\alpha }_{\mathit{Blast}}\sum _{i\in V_Y}{D}_{KL}(b_i\Vert \stackrel{\sim }{t_i})+{\alpha }_{\mathit{HHpred}}\sum _{i\in V_Y}{D}_{KL}(h_i\Vert \stackrel{\sim }{t_i})\\ +{\alpha }_{\mathit{Dipot}}\sum _{i\in V_Y}{D}_{KL}(d_i\Vert \stackrel{\sim }{t_i})+{\alpha }_{\mathit{Blast}}\sum _{j\in V_H}{D}_{KL}(b_j\Vert \stackrel{\sim }{t_j})+{\alpha }_{\mathit{HHpred}}\sum _{j\in V_H}{D}_{KL}(h_j\Vert \stackrel{\sim }{t_j})\\ +{\alpha }_{\mathit{Diopt}}\sum _{j\in V_H}{D}_{KL}(d_j\Vert \stackrel{\sim }{t_j})\end{array}$$

where α_Blast, α_HHpred and α_Diopt are parameters to tune the importance of each similarity component. Importantly, by optimizing these vectors, we integrate both molecular network connectivity and sequence similarity information into the same vector space for the purpose of comparison. Here we used a greedy method to select these parameters. Specifically, we incrementally added each term and find the optimal or reasonable weight for the term, according to the functional concordance between the predicted yeast-human homology pairs. The details of the parameter selection procedure can be found in the “Parameter Tuning” section below. On the basis of the analyses included therein, we chose α_Blast = 10, α_HHpred = 5 and α_Diopt = 5.. Finally, we computed the integrated homologous association $p_{ij}=(\stackrel{\sim }{t_{ij}}+\stackrel{\sim }{t_{ji}})/2$ between yeast protein i and a human protein.

To find significant homology pairs, we computed p_ij for all yeast-human protein pairs and constructed the empirical background distribution. We used 0.0005 as the empirical p-value cut-off to predict putative human homologs for yeast proteins and remove the homolog j if p_ij < 0.5 max _k{p_ik}. The background distribution is generated by randomly pairing human and yeast genes. Utilizing this cutoff, there were 4923 yeast proteins with predicted human homologs, greatly improving the coverage of BLAST (4023 yeast proteins) and HHpred (4312 yeast proteins) (Fig. S3A).

Preprocessing of interactomes

We downloaded both yeast and human interactomes from the STRING v9.1 (string-db.org). In STRING, q_ij are the confidence values assigned for each edge in the interactome. We removed predicted interactions and recalibrated the confidence for each interaction pair, such that $q_{ij}=1-\left(1-q_{ij}^{\mathit{experiment}}\right)\ast (1-q_{ij}^{\mathit{database}})$ with only “experimental” and “database” channels included. We also removed interaction pairs with low confidence q_ij < 0.2. After the preprocessing, we obtained a yeast interactome with 372026 interactions and 6164 proteins and a human interactome with 643822 interactions and 15317 proteins.

For the human networks, we also included two recently published high-quality binary human interactome datasets (11045 from high-quality re-curated binary interactions extracted from 7 public repositories; and 13944 from a recent yeast-2-hybrid experimental dataset)(Geetha et al., 1999; Hu et al., 2011; Rolland et al., 2014). Since these interactions were unweighted, we needed to assign confidence scores for them. To estimate a good confidence value, we extracted all physical binary interactions from the BIOGRID database (v3.2.116) and computed the statistics of STRING confidence scores of these interactions. Since interactions from BIOGRID are mostly from high-throughput experiments and they are binary, we used the mean or median statistics to assign confidence scores for new binary interactions. The quantile statistics of STRING confidence scores of BIOGRID interactions were 25%: 0.391, 50%: 0.620 and 75% 0.717. The average value of STRING confidence scores of BIOGRID interactions was 0.588. We thus considered it reasonable to assign a 0.6 confidence score for each unweighted binary interaction in these datasets.

As we were modeling neurodegenerative proteinopathies in the current work, we further pruned the human interactome to be brain-specific. To do so, we took GTEX gene expression dataset to only include genes appreciably expressed in brain (GTEx Consortium, 2013; Hu et al., 2011; Söding et al., 2005). Specifically, we normalized 357 GTEX brain RNA-seq datasets by the RPKM method(Robinson and Oshlack, 2010; Söding et al., 2005; Voevodski et al., 2009). We then filtered our human interactome such that only proteins with normalized brain expression level greater than (in at least one of 357 RNA-seq datasets) were included. In the end, our brain-specific interactome contained 369634 interactions and 10365 proteins.

Augmentation of human interactome with yeast-to-yeast edges (for humanized networks only)

Since genetic interactions are sparse in the human interactome, we used inferred homology to augment the human interactome by transferring edges from the yeast interactome. To do so, we added an edge between human proteins j and k if there exist a pair of yeast proteins i and l such that the integrated homologous association p_ij and p_kl satisfy p_ij * p_kl > 0.2 (see definitions above). This threshold was chosen to make the augmented brain interactome attain a similar density (~0.018) to that of yeast interactome (~0.019) with 751282 interaction pairs transferred.

Prize-collecting Steiner forest algorithm

We used the prize-collecting Steiner forest (PCSF) construction to analyze yeast networks and the augmented human-yeast network described above(Cho et al., 2015; Tuncbag et al., 2013; 2016.; Voevodski et al., 2009). For a network G(V, E, c, p) of node (gene) set V and edge (interaction) set E (where p(v) ≥ 0 assigns a prize to each node v ∈ V, and c(e) ≥ 0 assigns a cost to each edge e ∈ E), the goal of PCSF is to find a set of trees F(V_F, E_F ) to minimize the following cost function:

$$f(F)=\sum _{v\notin V_F}(\beta ·p(v)-\mu ·d(v))+\sum _{e\in E_F}c(e)+\omega ·\kappa$$

where k is the number of connected components or trees in the forest F; β is a parameter quantifying the trade-off between node prize and edge cost; d(v) is the degree of node v; μ is a parameter to penalize hub nodes with a large number of neighbors in the network. In this way, the algorithm searches for a network of relatively high-confidence edges linking the experimental data.

To optimize the objective function f(F), we introduced an extra root node v₀ into the network connected to each node v ∈ V by an edge (v, v₀) with cost ω. This step transforms the PCSF problem into a Prize-collecting Steiner Tree problem (PCST), which can be solved by a previously published messagepassing- algorithm(Bailly-Bechet et al., 2011; Cho et al., 2015). After the tree solution was obtained, we removed node v₀ and all edges that point to it from the tree solution and obtained the forest solution. It is not hard to show that the tree solution is optimal for the above PCST if and only if the forest solution is optimal for the original PCSF. Although the message-passing algorithm is not guaranteed to find the optimal solution, it works very well in practice(Cho et al., 2015; Tuncbag et al., 2013), and more importantly, it is substantially faster than linear programming approaches, which cannot handle large networks such as the yeast and human interactomes.

A computational difficulty of PCSF is how to tune the parameters β, ω and μ. Since β controls the scale of the prize values for nodes, we assigned a constant prize value (100) to each gene from our screens in our experiments. A perturbation of any parameter can potentially change the topology of the network structure, making the choice of parameters critical.(Altschul et al., 1997; 1990; Ashburner et al., 2000; Tuncbag et al., 2013) Thus, instead of choosing a single set of parameters, we developed an ensemble approach to obtain the consensus network from multiple reasonable parameter settings.

To decide the range of parameters, we set the upper and lower bounds such that: the network solution of PCSF contained sufficient number of predicted proteins (which is half of the number of input prize genes); the network solution did not introduce hub nodes with more than 1000 neighbors in the input network. We discretized the range of the parameters into a grid and enumerated all possible parameter combinations for PCSF runs. For the yeast network, the range of β was {1,2,4,6,8,10,12}; the range of ω was {1,2,3,4,5,6,7,8}; the range of μ was {0.001,0.003}. For the humanized network, the range of β was {4,6,8,10,12,14,16}; the range of ω was {3,4,5,6,7,8,9,10}; the range of μ was {0.003,0.005}. We also injected edge noise for PCSF runs to test for robustness, using the default Gaussian noise setting in the PCSF program. After obtaining the solutions for each PCSF parameter setting, we computed the frequency of each possible edge appearing in the ensemble of all solutions. The frequency of an edge is a surrogate for the robustness of the edge across different parameter settings. Finally, we took as input the edges and their frequencies in the ensemble of all solutions and applied a maximum spanning tree algorithm to find the most robust, representative network.

To evaluate the significance of the selected nodes in the solution, we constructed a background distribution for each node by simulating the same PCSF and ensemble process using a random selection of the same number of yeast genes as input. We computed background distributions using random gene sets with identical degree distribution to that of the prize node lists. Specifically, we binned all yeast genes into four categories, each containing genes with degrees [1–5], [5–10], [10–100] and [>100] respectively. Random gene sets are then sampled without replacement from these categories such that the statistics of the degree distribution were identical to those of a prize node list. We then performed PCSF and generated 10000 random ensembles of forests from 1000 random sets to compute the empirical distributions of each node in the background.

To evaluate the significance of the overlaps of the forests relating to different proteinopathies (Fig. 1D), we also calculated pairwise and triple-wise intersections of these random sets as background distributions. For example, we randomly paired the random ensembles generated for α-syn and random ensembles for tdp-43 and computed the distribution of the sizes of their overlaps. In this way, we constructed background distributions to evaluate the significance of the overlaps compared that simply caused by the increased size of the networks. Empirical p-values are also computed. Similar to our previous results, all the pairwise overlaps were statistically significant (p<=0.002). For the triple-wise intersections, the p-value was even more significant (p<=0.001).

Node and edge setup for yeast and humanized Steiner networks

Aside from differences in parameterization (noted above), there were some important differences between the yeast networks and the “humanized” networks.

For the yeast networks (Fig. 2), “prize nodes” were modifier hits from yeast genetic screens. Each prized node was assigned “100” as the arbitrary prize value. Edges for yeast networks were derived from STRING experimental and database edges. As described above, each edge was assigned a weight q_ij.

For the humanized networks (Figs. 3 and 4), “prize nodes” were similarly defined as modifier hits from yeast genetic screens. Yeast-to-human edges were weighted by the strength of homology (p_ij above) between yeast proteins and their human homologs. On the humanized networks, these are the first-order links seen between the red triangles (which are hits from the screen) and blue circles (human homologs). If one of the clear human homologs of a yeast modifier was a known parkinsonism or neurodegenerative gene – for example, a PARK locus gene – an arbitrary reward of 0.5 was added to p_ij to favor inclusion of that node over other potential homologs. Finally, edges between human proteins in the humanized networks were derived from STRING, but also from other sources, as described in “Pre-processing of interactomes” and “Augmentation of human interactome with yeast-to-yeast edges” above.

Parameter Tuning for Computational Pipelines

Here, we provide analyses and guidelines for the parameters used in our paper.

Weights for BLAST, HHpred and Diopt in the DCA homology tool

Since it is impossible to select the optimal parameters without enumerating all possible combinations, we performed a greedy analysis for the parameter selection for the extended DCA objective function. Specifically, we incrementally added each term and found the optimal or reasonable weight for the term.

Since BLAST is the most sensitive method for sequence homology detection, we first explored a reasonable parameter interval for BLAST. We only retained the two network topology terms and the BLAST terms in the extended DCA objective function and enumerated alpha_BLAST from the set of {1,2,5,10, 20,100}. To evaluate the performance, we computed the average accuracy of Gene Ontology (GO) of the top 5 homologs predicted by our method, as outlined in the Methods section in the manuscript. In Supp Fig. 2A it is readily seen that when the BLAST weight was too small (<10), our method was not able to fully exploit the homology information from BLAST. When this weight was greater than or equal to 10, the predictive performance became saturated and only provided slight performance improvement over the original BLAST method. When the weight became too large (=100), the predictive performance dropped and was identical to that of BLAST. This is because that the effect of network topology is diminished and our method simply reconstructed BLAST’s results. Thus, on the basis of the analysis we simply fixed the BLAST parameter to 10, although there might be better choices at extra computational cost by enumerating a larger and more refined set of possible values.

After we fixed the BLAST weight, we added the HHpred terms and performed the same analysis for HHpred weight. From the performance curve (Supp Fig. 2B), we observed that the optimal HHpred weight was around 5. This weight is smaller than BLAST weight, presumably at least in part because BLAST already captures most relevant homology information, while HHpred’s results extend BLAST by including extra remote sequence and structural homologs.

Finally, we fixed both BLAST and HHpred weights and performed the analysis for Diopt weights (Fig. S2C). For Diopt, the performance difference was very small as long as the weight was not too large (<20). This was consistent with the Diopt database only providing a few additional sequence homologs missed by both BLAST and HHpred. For simplicity, we chose its weighting equal to 5 as well.

Significance threshold for BLAST and HHpred in the DCA homology tool

We chose 1E-5 because it is a reasonably stringent threshold that is typically used for sequence homology or structure prediction(Geetha et al., 1999). Other choices of the threshold are possible but we believe that the results are not appreciably different from our setting. The following website and paper indicates 1E-5 is a reasonably stringent cutoff for protein BLAST.

Reward to homologs of known Parkinson genes

The major reason why we added reward values to homologs of known Parkinson genes is that the prizecollecting Steiner forest (PCSF) algorithm is not guaranteed to include all prize nodes in the final network. In addition, our homology tool can sometimes assign similar scores to two homologs, one with known literature support, the other without. Although the PCSF algorithm itself is able to distinguish most correct homologs by considering the connectivity, we found that by rewarding well-known homologs the noise can be further reduced. The reward parameter 0.5 is chosen such that existing homologs of well-known Parkinson’s genes from our screens are included in the final networks. It is obvious that larger reward values can have also the similar effect, but we didn’t explore those choices because we hoped to not to overtune the effect of this reward heuristic in our pipeline.

Confidence threshold for existing interactomes and predicted links

The choice of confidence threshold for STRING is indeed a trade-off between false-positives and true-positives. A stringent threshold, e.g. 0.8, can reduce the number of false-positives but the truncated yeast and human interactomes appeared to be too sparse and disconnected. Such thresholds may work well for signaling pathways or other well-studied and localized biological pathways but we did not feel this was an appropriate approach for complex proteinopathies, where mechanisms are poorly understood (and casting a “broader net” seems more appropriate) and where the connections between seemingly disparate disease-relevant genes are not well understood. Thus, we selected 0.2 to only exclude very low-confidence interactions and still maintain the major connectivity of the interactomes.

Confidence score for new high-throughput binary interactomes

Since the new high-throughput binary interactomes are unweighted, we need to assign an appropriate score to merge them with STRING interactions. To estimate an appropriate confidence value, we extracted all physical binary interactions from the most recent BIOGRID database and computed the statistics of STRING confidence scores of these interactions. Since interactions from BIOGRID are mostly from high-throughput experiments and they are binary, we can use the mean or median statistics to assign confidence scores for new binary interactions. The quantile statistics of STRING confidence scores of BIOGRID interactions are 25%: 0.391, 50%: 0.620 and 75% 0.717. The mean value of STRING confidence scores of BIOGRID interactions is 0.588. We thus assigned 0.6 since it closely related to both the median and mean statistics, judging it a reasonable assignment for incorporating new high-throughput binary interactions into existing STRING database.

Parameters for prize-collecting Steiner forest algorithm (PCSF)

As noted above, we used an ensemble approach to avoid the problem of parameter selection. There is no obvious way to determine the effectiveness of a set of parameters for PCSF. Furthermore, since there are several parameters, enumeration of all combinations becomes computationally infeasible. To address this issue, as noted above, we selected a wide-range of possible parameters, ran PCSF with all parameter combinations and made an ensemble network from single networks generated from each parameter combinations. These parameters are chosen such that the final network can connect 80% prize nodes in the network. Our parameter range also excludes networks that are overly distorted by “greedy” hyperconnected hubs like ubiquitin. As noted in our methods section, we further tested robustness by injecting noise into the edge distribution. There is no question that there is an element of subjectivity here, as with any parameterized model but we have taken great pains to be as broad as we feel we possibly can. Ultimately, the purpose is to generate tenable hypotheses or to predict biologically meaningful interactions.

Spotting assays

Yeast were cultured in synthetic media consisting of 0.67% yeast nitrogen base without amino acids (Fischer Scientific) supplemented with amino acids (MP Biomedicals) and 2% sugar. For most experiments, cells were first grown to mid-log phase in synthetic media containing glucose and then re-cultured overnight in synthetic media containing 2% raffinose. Mid-log phase cells were then diluted in synthetic media containing galactose. Typically, cells were induced for six hours at 30ºC.

Each strain was diluted to a starting OD₆₀₀ = 1.0 and serially diluted five-fold and then spotted on agar plates containing galactose (inducing) or glucose (control) plates.

Screening against known α-syn modifiers in ΔPARK17/α-syn and ΔPARK9/α-syn strains

The standard lithium acetate transformation protocol was adapted for use with 96-well plates(Cooper et al., 2006; D. Gietz et al., 1992; R. D. Gietz et al., 1995). Following transformation, cells were grown to saturation in synthetic media with raffinose lacking uracil for selection of yeast transformed with the desired plasmid. Once at saturation, they were spotted onto synthetic media plates with either glucose or galactose. Following two days of growth, galactose and glucose plates were photographed and analyzed by eye. In parallel experiments, transformed yeast were rediluted to OD₆₀₀ = 0.01 in 35 μL of galactose media in 384-well plates. Growth in 384-plates was monitored by measuring the OD₆₀₀ after 18, 24, and 48 hours of growth (Tecan safire²) giving a quantifiable measure of growth.

Small molecule (NAB2) treatment

Control, TDP-43 or α-syn yeast strains were grown to log-phase (OD₆₀₀ ~0.5) in complete synthetic media containing raffinose (non-inducing). Cultures were then diluted to an OD₆₀₀ of 0.01 (TDP-43 experiment) and 0.025 (α-syn experiment) in complete synthetic media containing 2% galactose to induce expression of the toxic protein. For NAB treatment, 10 μM (for α-syn) or 20 μM (for TDP-43) were added to the cultures and incubated in a Bioscreen instrument with intermittent shaking at 30ºC for two days.

Pooled α-syn overexpression screen

Pooled genetic screens were carried out in a YFP control strain and an α-syn strain. The yeast FLEXgene library representing most yeast open reading frames(Hu et al., 2007) was pooled from an arrayed bacterial library stock and grown to saturation in deep 96 well plates at 37°C. Cultures were pooled and plasmids isolated using Qiagen maxi prep kits. The pooled FLEXgene library was then transformed en masse into either control YFP or α-syn-expressing yeast strains and selected on five square 15 cm solid agar plates lacking uracil for plasmid selection. Approximately 10⁶ CFUs were obtained, representing an approximate 200-fold coverage of the ~6,000 yeast genes. Colonies were rinsed off of each plate, pooled, brought to 20% glycerol, aliquoted to individual use tubes (~100 μL) snap forzen in liquid nitrogen, and stored at −80°C.

Pooled screens were executed as follows. An aliquot of pooled yeast library was thawed on ice and diluted at three different concentrations into 3 × 30 mL flasks with SRafUra (~0.025, 0.05, and 0.1). After shaking at 30°C overnight, the culture with an OD₆₀₀ between 0.4 and 0.8 was selected to begin the pooled screen. Cultures were then diluted to and OD₆₀₀ of 0.1 in SGal Ura to induce expression of either YFP or α-syn 50 OD units were kept as time zero and centrifuged, washed with water, and frozen. Cultures were then maintained in log phase growth for 24 hours, making appropriate dilutions when needed to maintain and OD₆₀₀ under 0.8. After this time, 50 OD units worth of culture were centrifuged, washed with water, and pellets frozen.

Plasmids were then isolated from yeast using Qiagen minipreps with the following adaptations. Five minipreps were done per 50 OD units. Cell pellets were resuspended in buffer and lysed by bead beating with small acid-washed beads. Beads were removed and the lysate then taken through the conventional miniprep protocol. The purified plasmids from the five preps were then pooled. The yeast ORFs contained on the FLEXgene plasmids were then amplified using PCR primers that annealed to the attR Gateway sequences flanking the ORFs. HiFidelty Platinum Taq was used for amplification. 5 uL DNA was used per 50uL reaction and four reactions were performed per sample. 30+ cycles with a ~6′ extension time was used to ensure amplification of longer ORFs. PCR product was purified using Qiagen PCR columns. Two micrograms of PCR product was then sonicated, purified on Qiagen Minelute PCR columns, and the OD₂₆₀ re-analyzed. This product was then used as input for library generation and sequencing by the Whitehead Institute Genome Technology Core. Illumina HiSeq platform was used to sequence approximately 120 million 40 bp single end reads.

Reads were mapped to the yeast ORFs sequences with bowtie(Langmead et al., 2009). We made a bowtie index with the DNA sequences of the yeast ORFs reported in Hu et al.(Hu et al., 2007), plus 903 ORFs that were present in SGD but were not included in the list of sequences from in Hu et al. Reads were mapped allowing 2 mismatches (-n 2) in the seed, seed length of 40 (-l 40), suppressing all alignments that map to more than one place (-m 1) and using “--best” and “--strata”. Unmapped reads were trimmed with fastx_trimmer (http://hannonlab.cshl.edu/fastx_toolkit/commandline.html) to remove the first 20 nt, and remapped with bowtie using the following parameters: “-n 0 -l 20 --best --strata -m 1 “. The number of reads mapping to each ORF was obtained parsing the output sam files. Differential expression analysis was done with the R package Noiseq(Tarazona et al., 2011). NOISeq is a nonparametric method to identify differentially expressed genes from count data. NOISeq calculates fold change values and probability of differential expression. The probability (P-val) of differential expression for each gene is derived from the joint distribution of fold-change differences (M)- absolute expression differences (D) values for all the genes within the Table Set.

A gene was selected for validation if it was: (A) up or down consistently in the two pooled α-syn screens (|log2 fold change| >0.8 in both screens) except when neither experiment was associated with a P-val of >0.5); (B) had an average fold change with absolute value of >2.5 (regardless of P-val); (C) known modifiers from previous experimentation that had a fold-change in the pooled screen consistent with that source. Any gene with an |log2 fold change|>1.0 in the YFP control (in the same direction as the putative suppressor or enhancer) was excluded, as well as genes associated with galactose metabolism that would be expected to alter expression of gal-inducible transgenes. Thresholds were guided by knowledge gained from our previous extensive characterization of the arrayed α-syn over-expression screen hits (see Fig. 1). Put another way, our previous over-expression screen was used as a “gold standard” to analyze the pooled over-expression data.

Pooled Screen-QPCR Verification

Transformed cells generated from the pooled screen (“Pooled α-syn overexpression Screen* method) were thawed on ice and diluted in SRaf-Ura to resulting ODs of approximately 0.03, 0.05 and 0.1. Cultures were grown at 30°C overnight and cultures with an OD of 0.4–0.8 were chosen for induction. These cultures were diluted to an OD of 0.1 in SGal-Ura. 50 OD units were stocked representing the time zero time point. Induced cultures were grown for 24 hours and 50 OD units were stocked representing the 24hr time point. Plasmids were isolated using the Qiagen miniprep kit (27106) splitting the 50OD units for each time point in to 5 samples. Following cell resuspension in P1 buffer cells were lysed by bead beating using acid-washed beads. Following bead beating, beads were removed from samples and lysates subjected to the standard miniprep kit protocol. Resulting plasmids were pooled and used for QPCR analysis. The standard attF primer was used in combination with an orf specific reverse primer (sequence generated by Primer3 such that the product <150bp in size) for QPCR analysis. Multiple negative controls used to normalize samples and positive controls were run on all QPCR plates. QPCR analysis was performed using technical triplicates of biological triplicates on the Applied Biosystems (7900HT) using the SYBR green fluorescence detection system (Applied Biosystems). The program for amplification comprised 40 cycles of 95°C for 15 seconds and 60°C for 1 minute.

Pooled Screen-Growth Curve Analysis

Each individual putative modifier was overexpressed in the α-syn strain using the Flexgene overexpression library. Three independent Ura⁺ transformants were grown in SRaf-Ura at 30°C overnight. Cultures were subcultured in SRaf-Ura and at an OD of 0.4–0.8 were diluted in Sgal-Ura for induction. Each isolate was set up in triplicate and growth was monitored every 15 mins for approximately 60 hours.

Genome-Wide Deletion Screen (Synthetic Gene Array Methodology)

The method used was essentially as described previously(Baryshnikova et al., 2010; Tong and Boone, 2006). Briefly, deletion strains were pinned on to YPD+G418 plates. Query strains (α-syn and wild-type control) were grown in 5ml overnight cultures in YPD at 30°C and spread on YPD plates and grown overnight. Deletion strains were mated to each query strain by pinning together on YPD and grown for 48hrs at 30°C. Resulting diploids were pinned to SD/MSG-Ura+G418 and grown for 2 days at 30°C. Cells were pinned to sporulation media plates and incubated at 23°C for 7days. Spores were pinned to SD-His/Arg/Lys+canavanine+thialysine and grown for 2 days at 30°C. Cells were pinned to fresh SD-His/Arg/Lys+canavanine+thialysine and grown for 1 day at 30°C. Cells were pinned to SD/MSG-His/Arg/Lys+canavanine+thialysine + G418 and grown for 2 days at 30°C and then pinned to SD/MSG-His/Arg/Lys/Ura +canavanine+thialysine+G418 and grown for 2 days at 30°C. For the initial screen, cells were pinned both to SD/MSG-His/Arg/Lys/Ura +canavanine+thialysine+G418 and to Sgal/MSG-His/Arg/Lys/Ura +canavanine+thialysine+G418 and spot growth was monitored. For validation studies, cells were pinned to liquid SD/MSG-His/Arg/Lys/Ura +canavanine+thialysine+G418 and grown overnight at 30°C and then pinned both to SD/MSG-His/Arg/Lys/Ura +canavanine+thialysine+G418 and to Sgal/MSG-His/Arg/Lys/Ura +canavanine+thialysine+G418 and spot growth was monitored. Stock solutions (1000X) were prepared as follows: G418 200mg/ml, canavanine 50mg/ml, thialysine 50mg/ml. The method above was used for the initial screen and repeated, in duplicate, using 96-well plate format for validation of the initial screen hits.

Human iPSC generation and differentiation into midbrain dopaminergic (DA) neurons for LRRK2 mutant lines

iPSCs from control individuals and PD patients carrying G2019S LRRK2 along with isogenic gene corrected controls were generated as previously described(Reinhardt et al., 2013). iPSCs were differentiated into mDA neurons using a floor plate-based protocol with minor modifications(Kriks et al., 2011; Schöndorf et al., 2014). Differentiation was based on exposure to LDN193189 (100 nM, Stemgent) from days 0–11, SB431542 (10 mM, Tocris) from days 0–5, SHH C25II (100 ng/mL, R&D), purmorphamine (2 mM, EMD) and FGF8 (100 ng/mL, Peprotech) from days 1–7 and CHIR99021 (CHIR; 3 mM, Stemgent) from days 3–13. Cells were grown for 11 days on Matrigel (BD) in knockout serum replacement medium (KSR) containing DMEM, 15% knockout serum replacement, 2 mM L-glutamine and 10 μM β-mercaptoethanol. KSR medium was gradually shifted to N2 medium starting on day 5 of differentiation. On day 11, media was changed to Neurobasal/B27/L-Glut containing medium (NB/B27; Invitrogen) supplemented with CHIR (until day 13) and with BDNF (brain-derived neurotrophic factor, 20ng/ml; R&D), ascorbic acid (0.2 mM, Sigma), GDNF (glial cell line-derived neurotrophic factor, 20 ng/ml; R&D), TGFβ3 (transforming growth factor type β3, 1 ng/ml; R&D), dibutyryl cAMP (0.5 mM; Sigma), and DAPT (10 μM; Tocris,) for 9 days. On day 18, cells were dissociated using Accutase (Innovative Cell Technology) and replated under high cell density conditions on dishes pre-coated with 15 μg/ml polyornithine and 1 μg/ml laminin in differentiation medium (NB/B27 + BDNF, ascorbic acid, GDNF, dbcAMP, TGFβ3 and DAPT). At DIV30, cells were collected and, after centrifugation, cell pellets were stored at −80°C until further analysis.

Human pluripotent stem cell culture for α-syn mutant lines

Skin biopsy, human dermal fibroblast culture, iPS cell generation and mutation correction for the patient harboring the A53T mutation (WIBR-IPS-^A53T) have been described previously(Cooper et al., 2006; Soldner et al., 2011). In that previous publication the A53T iPS line was referred to as WIBR-IPS-SNCA^A53T.

Our pluripotent stem cell lines were initially maintained (5%0₂, 3%C0₂) on mitomycin C inactivated mouse embryonic fibroblast (MEF) feeder layers in hES medium [DMEM/F12 (Invitrogen) supplemented with 15% fetal bovine serum (FBS) (Hyclone), 5% KnockOut Serum Replacement (Invitrogen), 1 mM glutamine 15% fetal bovine serum (Invitrogen), 1% nonessential amino acids (Invitrogen), 0.1 mM β-mercaptoethanol (Sigma) and 4 ng/ml FGF2 (R&D systems)]. Cultures were passaged every 5 to 7 days either manually or enzymatically with collagenase type IV (Invitrogen; 1.5 mg/ml). At around 50 passages prior to differentiation, lines were passaged to plates pre-coated with growth factor-reduced matrigel (BD Biosciences; 1:30 in DMEM:F12) and cultured (21% 0₂, 5% CO₂) in mTESR-1 medium (Stem Cell Technologies), thereafter being passaged every 5 to 7 days enzymatically with dispase (Invitrogen; 1mg/mL) until differentiation (at passage 40–90). For karyotyping, standard G-banding chromosomal analysis of cell lines was performed every 10–20 passages (Cell Line Genetics, Inc.). We confirmed mycoplasma-negative status of our cultures every 2–4 weeks (MycoAlert, Lonza).

Human neural induction by embryoid body (EB) formation

A previously published protocol was used without modification(Chung et al., 2013; Hu et al., 2007; J.-E. Kim et al., 2011). This protocol has been repeated here for completeness.

To initiate differentiation, on day 0 human ES or iPS cell colonies were pretreated for 30–60 min with 5 μM Y-27632/ROCK inhibitor (Calbiochem), single cell-dissociated after 5–10min exposure to accutase (StemPro Accutase; Life Technologies) and then re-suspended in neural base (NB) medium, which is DMEM/F12 (Gibco/Life Technologies) supplemented with N2 and B27. N2 and B27 supplements from Life Technologies and used at ½-1% and 1–2%, respectively. Cells were plated in AggreWell 800 microwells (StemCell Technologies; priming and plating per manufacturer’s protocol; 2.4×10⁶ cells were well) in NB medium supplemented with dual SMAD inhibitors(Chambers et al., 2009; Langmead et al., 2009) recombinant human Noggin (R&D Systems) at 200ng/mL and 10 M SB431542 (Tocris Bioscience), as well as 5μM Y-27632. Noggin and SB431542 remained in the medium at these concentrations throughout the neural differentiation protocol.

On day 1 medium was ½-changed. By day 2, well-formed neuralized EBs (NEBs) were typically observed in the AggreWells and transferred to Petri dishes (4 AggreWell wells/Petri dish) overnight, in NB medium. On day 4, NEBs were transferred to a dish coated with growth factor-reduced Matrigel (1:30 in DMEM:F12; BD Biosciences) for attachment. Y-27632 was omitted from this day onward. From day 5 to day 10, attached NEBs were additionally exposed to 20 ng/mL FGF2 (R&D Systems) and recombinant human Dkk1 at 200 ng/mL (R&D Systems). On day 10, neural rosettes were dissected (P20 pipette tip), incubated in accutase supplemented with DnaseI (Sigma Aldrich) for 10 min at 37°C and gently dissociated to small cellular clumps and single cells. After washing, the rosettes were re-plated on plastic dishes pre-coated with poly-L- ornithine and laminin (BD Biocoat) at high density (200,000/cm²) in neural progenitor cell (NPC) medium, which is NB medium supplemented with 20 ng/mL FGF2. (Life Technologies), supplemented overnight with 10μm Y-27632. Typically, one Aggrewell 800 well provided enough NPCs for at least 1–2 6-wells at passage 0.

Thereafter, the surviving NPCs proliferated. Medium change was daily. They could be passaged up to 10 times before neural differentiation, and could successfully be freeze/thawed at early passage (p1 to p5) without compromising differentiation potential. Freezing medium was NPC medium with 10% FBS (Hyclone).

Human cortical neural differentiation

A previously published protocol was used without modification(Chung et al., 2013; Hu et al., 2007; J.-E. Kim et al., 2011). This protocol has been repeated here for completeness.

To begin neural differentiation, NPCs were dissociated with accutase and re-plated on matrigel-coated T75 flasks (CytoOne). The next, day medium was fully changed to Neural Differentiation (ND) medium, which is NB medium supplemented with recombinant human BDNF and GNDF (both at 10ng/mL; R&D Systems) and dibutyryl cyclic AMP (Sigma; 500μM), and without FGF-2. Thereafter, media was ½-changed every other day. On day 7–9, differentiating neurons were gently dissociated to single cell, resuspended in pre-chilled Hank’s balanced salt solution (HBSS; Gibco/Life Technologies) supplemented with 0.1% bovine serum albumin (Gibco/Life Technologies). After a wash step, cells were plated on 6- or 24-well plastic plates pre-coated with poly-ornithine and laminin (BD Biocoat) for biochemical assays. Medium was ½-changed every 3 days for up to 12 weeks.

Primary rat cortical cultures

Embryos were harvested by cesarean section from anesthetized pregnant Sprague-Dawley rats at embryonic day 18. Cerebral cortices were isolated and dissociated with Accumax (Innovative Cell Technologies, Inc) digestion for 20 min at 37 °C and trituration with Pasteur pipette. Poly-ornithine and laminin-coated 96 well plates were seeded with 4×10⁴ cells respectively in neurobasal medium (Life Technologies) supplemented with B27 (Life Technologies), 0.5 mM glutamine, 25 μM β–mercaptoethanol, penicillin (100 IU/ml) and streptomycin (100 μg/ml). One third of the medium was changed every 3 to 4 days.

AAV-1 transduction of iPS neurons

Plasmids containing verified TALE-TFs were purified endotoxin-free (Qiagen) and packaging into adenoassociated viruses serotype 1 (AAV-1) was conducted by the Gene Transfer Vector Core, Massachusetts Eye and Ear Infirmary/MEEI, Harvard Medical School (mini-scale production). A53T and mutation-corrected cortical neurons were aged for 4–7 weeks at a plating density of 0.25–0.75 × 10⁶ cells/cm². Cells were transduced with 30 μl of the mini scale produced MEEI AAV-1 titer, containing a single TALE-TF or the TALE cloning backbone alone, in 500 μl ND medium. ND medium was changed 12–16 hours post-transduction.

Antibodies

Mouse anti- Carboxypeptidase Y	Life Technologies A66428	Western blot	1:10 000
Rabbit anti-Nicastrin	Cell Signaling 3632	Western blot	1:1000
phospho eIF2A	Cell Signaling 9721	Western blot	1:1000
total eIF2A	Cell Signaling 2103	Western blot	1:1000
LRRK2	Abcam Ab133474	Western blot	1:500

Protein labeling with ³⁵S-methionine/-cysteine

A53T and mutation-corrected cortical neurons were aged for 4–8 weeks at a plating density of 0.25–0.75 × 10⁶ cells/cm². Prior to the protein labeling the cortical neuronal cultures were kept in Neural Differentiation (ND) medium without methionine and cysteine for 90 min. ND medium was DMEM complemented with 1% (v/v) B-27, 0.5% (v/v) N-2 and 1% (v/v) GlutaMAX supplement, 1% (v/v) MEM non-essential amino acids, 1% (v/v) Penicillin-Streptomycin (all Life Technologies) as well as 10 ng/ml BDNF and GDNF (both R&D Systems) and 500 μM cAMP (Sigma-Aldrich). For protein labeling the neuronal cell cultures were incubated in ND medium supplemented with ³⁵S-methionine and -cysteine (Perkin Elmer) at a final concentration of 100 μCi/ml for various duration. After a quick wash with cold PBS, cells were lysed in a buffer containing 50 mM Tris-HCl and 2% (w/v) SDS, supplemented with protease inhibitor cocktail (Sigma-Aldrich). The samples were boiled at 100°C for 5 min and spun down at 10,000 g for 15 min. The supernatant was collected and the protein concentration was determined using BCA assay (Pierce, Thermo Fisher Scientific). ³⁵S labeled samples were run in 4–12% Nupage Bis-Tris gel (Life Technologies). As a loading control, gels were stained with SimplyBlue SafeStain (Life Technologies), and destained by incubation in water. Thereafter, the gels were incubated in 11.2% (v/v) salicylic acid and 10% glycerol (v/v) for 15 min. The gels were dried and exposed to a phosphor screen (Fujifilm) for a minimum of 48 hours. The screen was scanned using the phosphorimager BAS-2500 (Fujifilm) and ³⁵S incorporation was determined by measuring the intensity of each lane (MultiGauge Analysis Software v2.2, Fujifilm).

Free ³⁵S-methionine/-cysteine in the cytosol

Rat primary neurons overexpressing either GFP or αSyn-GFP were incubated with ³⁵S-methionine and – cysteine at 100 μCi/ml for various durations. After a quick wash with cold PBS, cells were lysed in RIPA buffer for 20 min on ice and the debris was removed by centrifugation. Proteins in the lysates were precipitated by adding 1 volume 100% TCA to 4 volume of lysate and incubate 10 min at 4°C. After centrifugation at 14K rpm for 10 min, supernatant was collected to measure a cytosolic pool of free ³⁵S-methionine/-cysteine. ³⁵S incorporation was determined by quantifying using an LS 6500 liquid scintillation counter (Beckman Coulter) with 5 μl of the sample being immersed in 7 ml scintillation cocktail (National Diagnostics).

Cell lysis and endoglycosidase H digestion

Cells were lysed in a buffer containing 20 mM HEPES, 150 mM NaCl, 10% (v/v) glycerol, 1 mM EGTA, 1.5 mM MgCl₂, 1% (v/v) Triton X-100, pH to 7.4, protease inhibitor cocktail (Sigma-Aldrich), and protein phosphatase inhibitor cocktail 1 and 2 (Sigma-Aldrich), and incubated in an ice/water slurry for 20 min, followed by 2 freezethaw cycles (-80°C/37°C, ~ 1 min each). Supernatant was collected after ultracentrifugation at 100,000 g, 4°C, for 30 min. Protein concentration was determined using BCA assay (Pierce, Thermo Fisher Scientific). Endoglycosidase (Endo) H (New England Biolabs) digestion was performed based on the manufacturer’s instructions. Briefly, 20–40 μg bulk protein was assembled in 15.3 μl reaction volume; 1.7 μl denaturing buffer was added and samples were boiled for 10 min at 100°C. Then 2 μl of G5 buffer and 1 μl of Endo H or 1 μl H₂O were added to the denatured reaction and incubated for 2 hours at 37°C.

Western Blotting

For protein trafficking after Endo H digestion, protein samples were denatured in sample buffer (20 mM Tris-Cl pH 6.8, 4% (v/v) glycerol, 180 mM 2-mercaptoethanol, 0.0003% (v/v) bromophenol blue and 2% (v/v) SDS), run in 10% Tris-glycine gel, and wet transferred with 20% methanol onto PVDF membranes (BioRad). Blots were blocked in a 1:1 dilution of Odyssey blocking buffer (Li-Cor Biosciences) and PBS for 1 hour at room temperature, followed by incubation with primary antibodies in a 1:1 dilution of Odyssey blocking buffer (Li-Cor Biosciences) and PBS containing 0.1% Tween 20 (PBST) at 4°C overnight with gentle rocking. After three 5 min washes with PBST, blots were incubated with secondary antibodies such as anti-mouse or -rabbit IgG conjugated to IRDye 680 or 800 (1:10,000, Rockland) in a 1:1 dilution of Odyssey blocking buffer and PBST for 2 hours at room temperature. After three 5 min washes with PBST and two with water, blots were scanned using the Odyssey quantitative fluorescent imaging system (Li-Cor Biosciences) and bands were quantitated using Odyssey Software v2.1 (Li-Cor Biosciences).

For other Western blots, samples were lysed in RIPA buffer and run in either 8 or 10% Nupage Bis-Tris gel (Life Technologies) and transferred using iBlot (Life Technologies). Blocking was in 5% nonfat dry milk in PBST. As for the secondary antibodies and chemiluminescent detection, anti-mouse, -rabbit IgG or avidin conjugated to HRP was used with SuperSignal West Pico chemiluminescent substrate (Thermo Fisher Scientific).

TALE-TF design

TALE-TFs were designed to target between 200bp upstream (5′) and 50bp downstream (3′) of the transcription start site (TSS) of ATXN2 or EIF2G transcripts. Within these regions near the TSS, we identified DNAseI hypersensitive regions from human ventromedial prefrontal cortex samples (Thurman et al., 2012, PMID: 22955617). Within these DNAseI HS regions, we designed 5 TALE-TFs for each transcript.

Each TALE-TF was designed to target a 14bp genomic sequence consisting of an initial thymidine (T) plus 12 full repeats and 1 half repeat. For each TALE-TF, the TALE repeats were cloned into an rAAV transfer plasmid using a PCR-based, Golden Gate cloning strategy as previously described(Konermann et al., 2014; Sanjana et al., 2012; Tarazona et al., 2011). The rAAV transfer plasmid contained the TALE backbone fused to the synthetic VP64 activator domain along with a 2A-linked EGFP that is cleaved during translation.

TALE-TF assembly

14-mer transcription activator-like effector transcription factors (TALE-TFs) were constructed using Golden Gate cloning as described previously (Sanjana et al. 2012). For each gene, ATXN2 and eIF4G1 (transcript variant 7), five different TALE-TFs were designed with the 14 bp long target loci being located in the proximal promoter region (ATXN2 TALE-TF #1: 5′-TGTCCAGATAAAGG-3′, #2: 5′-TGAACCTATGTTCC-3′, #3: 5′-TGCCAGATTCAGGG-3′, #4: 5′-TGGAGCGAGCGCCA-3′, #5: 5′-TAGCTGGTCATGGT-3′; eIF4G1 TALE-TF #1: 5′-TGTCACGTGACGGG-3′, #2: 5′-TGTGGCTGTCACGT-3′, #3: 5′-TCAAAGTTCGGGAG-3′, #4: 5′-TCGCGGAACAGAGA-3′, #5: 5′-TCTCCTGCCTCAGC-3′). For each TALE-TF the correct sequence of the DNA-binding domain was verified by Sanger sequencing and all TALE-TF clones with non-silent mutations were excluded.

Ribosomal Footprint Profiling

For ribosome footprint profiling, 12-week old cells were treated with cycloheximide (100 ug/mL) for 5 min at 37 °C to stop translation elongation. Cells were washed twice with ice-cold 9.5 mM PBS, pH 7.3, containing 100 μg ml⁻¹ cycloheximide, and lysed by adding lysis buffer (10 mM Tris-HCl, pH 7.4, 5 mM MgCl2, 100 mM KCl, 2 mM dithiothreitol, 100 μg ml⁻¹ cycloheximide, 1% Triton X-100, 500 U ml⁻¹ RNasin Plus, and protease inhibitor (1× complete, EDTA-free, Roche)), scrapping cells from the plate, and then triturating four times with a 26-gauge needle. After centrifuging the crude lysate at 1,300g for 10 min at 4 °C, the supernatant was removed and flash-frozen in liquid nitrogen. The lysate was thawed on ice, after which ribosome profiling and mRNA-seq were performed as described previously(Subtelny et al., 2014) using a detailed protocol available at http://bartellab.wi.mit.edu/protocols.html. The 4-week old cells were washed twice with 37 °C growth media, then after removing the media by aspiration the plates were sealed and then plunged into liquid nitrogen. Cells were then lysed with lysis buffer as described above, but cycloheximide was excluded from all solutions including the sucrose gradients. After thawing on ice, a small amount of cycloheximide-free zebrafish RPF lysate was spiked into the 4-week old cell lysates (10-fold less based on A₂₆₀) prior to digestion with RNase I.

RPF and RNA-seq tags were mapped to the ORFs, as described previously(Subtelny et al., 2014). To account for the zebrafish reads present in the 4-week old samples, indexes comprising both the zebrafish and human genomes or transcriptomes were created and these data were mapped to the combined indexes. Only reads mapping uniquely were considered, and those mapping to zebrafish were excluded from the analysis.

Enriched pathways in the translational profiling for the 4-week and 12-week datasets were computed with the Gene Set Enrichment Analysis tool, available at the Broad Institute website (http://software.broadinstitute.org/gsea/index.jsp).

QUANTIFICATION AND STATISTICAL ANALYSIS

Comparison with existing homology prediction approaches

To evaluate the functional association between yeast proteins and the predicted human homologs, we computed the average accuracy of Gene Ontology (GO) of the top 5 homologs predicted by our method, HHpred and BLAST(Altschul et al., 1997; 1990; Ashburner et al., 2000; Tuncbag et al., 2013) (Fig. S3B). We chose the top 5 homologs since yeast proteins often have more than one good human homolog. The accuracy of a homolog was calculated as the percentage of overlapped GO labels between the yeast protein and the putative homolog. We noted that the number of assigned GO labels per gene varied considerably between yeast and human proteomes, so that the GO accuracy metric favored predicted homologs with a large number of labels and query proteins with a small number of GO labels, potentially biasing the analysis. Furthermore, false positives were not considered by this metric. To address these issues, we computed the widely used Jaccard similarity score, which is the number of overlapping GO labels divided by the total number of unique GO labels of the yeast (or human) gene and its human (or yeast) homolog. BLAST’s accuracy for 4023 yeast proteins was 31.1%. HHpred in conjunction with BLAST achieved of 32.6% for accuracy for 4312 yeast proteins. Our method obtained 31.6% accuracy for a significantly greater number, 4923, of yeast proteins. It also outperformed BLAST on 4023 yeast proteins with BLAST output (32.0% vs 31.1% accuracy and 25.2% vs 24.3% Jaccard similarity) and HHpred on 4312 proteins with HHpred output (34.1% vs 32.6% accuracy and 26.9% vs 24.9% Jaccard Similarity). The improvements over BLAST and HHpred were significant (paired t-test p-values < 0.01).

We then tested our method on finding yeast homologs for human proteins (Fig. Ss 3C and 3D). The improvement of the coverage over BLAST and HHpred was even more substantial than for generating human homologs from yeast proteins. Our method predicted homologs for 15200 proteins but BLAST and HHpred only covered a relatively small portion of human proteome (7248 and 9577 respectively). Accuracy metrics also favored the DCA method. Our method improved the predictive power over BLAST (57.6% vs 57% accuracy and 26% vs 26.6% Jaccard similarity) and HHpred (56% vs 54.9% accuracy and 25% vs 24.2% Jaccard similarity) on proteins which BLAST or HHpred can find yeast homologs on both GO accuracy and Jaccard similarity score. These comparisons were all statistically significant (all p-values < 0.01 by paired t-test).

We also compared our homology tool to the state-of-the-art Ensembl Compara method. Ensembl Compara identifies high confidence homolog pairs through phylogenetic tree-based clustering and analysis across multiple species. This sequence-based method sacrifices coverage for accuracy, and these pairs are considered a gold standard for traditional analyses (Vilella et al., 2009). We downloaded the Ensembl Compara v85, and mapped gene ids to the gene names used in our homology tool, identifying 5093 high-confidence yeast/human pairs for 2409 yeast genes. Among these pairs, there are three major categories: “one-to-one”, “one-to-many” and “many-to-many”. To evaluate our DCA homology tool, we checked whether it performed at least as well for high-confidence yeast/human pairs, whether predicted as one-to-one, one-to-many or manyto- many by Ensembl Compara. Since orthology relationships between human and yeast genes can be ambiguous due to their remote evolutionary distance, DCA and Ensemble Compara may predict different putative homologs, especially for the many-to-many case. For such cases, we also computed the GO accuracy as the percentage of overlapping GO labels between a yeast protein and the predicted homolog. For clear one-to-one pairs by Ensembl Compara, DCA differed in only 25 of 1040 genes. Of those 25 genes that differed, our method achieved comparable accuracy in ontology prediction (0.394) as compared to Ensembl Compara (0.388) based on ontology matching. There were 1518 entries in the “many2many” prediction category. For these, our method achieved a correct pairing (0.414) equivalent to Ensembl Compara (0.412). Finally, for the yeast genes in which a one-to-many correspondence was predicted, there were 2535 entries. Again, our method identified homologs by gene ontology (0.391) similar to Ensembl Compara (0.390). Among the differences, we observed most of them to be similar genes within the same family; moreover, these differences are not statistically significant. Thus, our approach does not disrupt homology prediction for highconfidence orthology pairs, a surrogate for false-positivity in the absence of any other gold standard yeast-tohuman homolog pairing. From these results, we demonstrated that DCA provides comparable yeast-to-human accuracy as Ensembl Compara for the same input yeast genes.

Recently, Kachroo et al.(Kachroo et al., 2015)carefully tested 414 essential yeast genes for complementation by homologs that were clear by sequence. Thus, for each of these 414 yeast/human gene pairs, the complementation assay provided a binary and experimentally strong readout of homology. Kachroo et al. developed a method to predict which of these high confidence pairs were likely to be actual positive complementation pairs. They utilized more than 100 features, including careful manual curation of sequence properties, network features, transcriptional and translational features, and expression abundances, to establish a predictive tool. They showed that this predictive tool could be trained on a subset of the experimentally tested yeast/human pairs to correctly identify functional replaceability in a separate test set. To demonstrate the effectiveness of integration of DCA, we trained a DCA-based classifier, using only sequence and network information, to predicted true yeast-human complementation pairs. In particular, to check the predictive power of our DCA pipeline, we built a classifier based on the low-dimensional gene vectors obtained from our joint DCA learning pipeline. For each pair of yeast/human genes, we built features based on the gene vectors to consider their sequence similarity and topological roles in their molecular networks. These features, including element-wise product and difference and sum of two gene vectors, were used as input to a gradient boosted forest classifier. We tested whether this classifier, for our more elemental, automated DCA tool based on only sequence/network features, could be tuned to also predict the functional complementation between yeast and human. When we trained our DCA classifier via 5-fold cross-validation on the yeast/human pairs from Kachroo et al. we achieved a high rate of prediction accuracy (AUC=0.82, SD=0.08). This was comparable to the intricate, manual integrated method of Kachroo et al., demonstrating that our automated homology tool, based on only sequence and network topology, is sufficient for training a classifier for this specific homology task. It is worth noting that methods utilizing sequence-similarity alone, including BLAST and HHpred, performed considerably worse than DCA (0.70 and 0.69, respectively). It is clear that our DCA-based classifier, which effectively integrates network topology and sequence similarity, is just as effective as the method in Kachroo et al. that utilizes more than 100 features, thus overcoming the barrier of major timeconsuming manual feature curation.

Evaluation of PCSF and humanized Steiner networks

We tested PCSF on two separate datasets and demonstrate vastly superior performance when compared to existing methods. For comparison, we identified two popular algorithms, DAPPLE (Rossin et al., 2011) and PEXA (Tu et al., 2009), and implemented them. Both methods take seed genes and identify subnetworks that span the seed genes to reveal possible functional interconnectedness of these genes. The first algorithm, DAPPLE, identifies significant direct and one-hop indirect edges in the human interactome to connect as many seed genes as possible. The second algorithm, PEXA, utilizes existing pathway annotations, such as KEGG or Reactome, to cover seed genes. Merging and pruning are then applied to link connected components and remove hanging genes. For these comparisons, we provided each algorithm with yeast-to-human homology links and injected yeast interaction edges into the human network, just as we provide for our PCSF method. For DAPPLE, we used the predicted dense network with significant one-hop indirect edges, since the sparse direct network is not able to identify hidden genes. We curated hits from 15 complete screens in yeast (Tong, 2004). In these screens, a gene is deleted as well as its genetic interactors or modifiers. We used these genetic modifiers as input for the network algorithms. The inactivated gene was hidden from the algorithm, and was used to evaluate the predicted network. Taking cues from previously-published methods (Yeger-Lotem et al., 2009), here we considered an algorithm successful in discovering the cellular response if the predicted hidden human genes were significantly enriched for specific gene ontology biological process terms attributed to the hidden inactivated yeast gene (hypergeometric test; p-value < 0.01). We generated humanized networks with PCSF, and two alternative methods: DAPPLE (Rossin et al., 2011) and PEXA (Tu et al., 2009). For these screens, the success rate of PCSF was 47%, as compared to DAPPLE and PEXA which were 6.6% and 13%, respectively. These results suggest superior performance of PCSF over DAPPLE and PEXA.

To better understand the relevance of genes and predicted pathways recovered by PCSF, DAPPLE and PEXA, we designed a well-controlled simulation. To mimic genetic screens of perturbed pathways, we selected individual pathways from the well-known human pathway database KEGG and identified all genes in each pathway (Supplemental Table S15). We then identified yeast homologs via stringent Ensembl one-to-one mapping. We treated those human genes with clear yeast homologs as “perturbed” and picked their homologs’ genetic interaction neighboring genes as hits from a “virtual yeast genetic screen”. Virtual screens like these minimize experimental noise as a confounding factor and enable cleaner evaluation of algorithm performance. Since we know the “true” pathway information, this method can be used to test the sensitivity and specificity of algorithms by quantifying how often “relevant” genes in the original KEGG pathway are recovered as predicted (non-seed) genes. We chose 50 KEGG pathways that had at least 5 human genes with clear yeast homologs and created 50 associated “virtual” screens for testing (Table S15). We used two performance metrics: precision, i.e. the percentage of predicted hidden genes shown in the original KEGG pathway, and recall, i.e. the percentage of the original KEGG genes shown as hidden nodes in the predicted pathway. Ideally, these values would be 100% for perfect predictions. For PCSF, the average precision and recall values are 63% and 74% resp. In contrast, for DAPPLE, the average precision and recall values are 6% and 47% resp., whereas for PEXA, they are 8% and 83% resp. The differences between three precision values are substantial: PCSF has much higher precision within very compact subnetworks, while both DAPPLE and PEXA predict huge “hair ball” networks with low precision. It is worth noting that PEXA has a very high recall value likely because it uses the KEGG pathways to build networks, and thus predictably has high recall (because the simulated screens here are generated from KEGG pathways); however, its precision metric is very low.

Further, we tested the effectiveness of injected yeast genetic interactions into networks through the simulated yeast genetic screens we generated, and cross-compare our PCSF method with the other algorithms, DAPPLE and PEXA. First, we tested performance by removing all injected yeast interactions. For PCSF, the average precision and recall values are 37% and 54% resp. For DAPPLE, the average precision and recall values are 8% and 27% resp. Compared to the precision and recall results (i.e., 63% and 74% for PCSF versus 6% and 47% for DAPPLE), it is clear that both PCSF and DAPPLE have much lower recall if yeast interactions are excluded. This analysis thus confirms with data that injection of yeast interactions into “humanized” networks provide key connections between genetic modifiers to the perturbed genes. For PEXA, the average precision value is 9%, similar to that with yeast injection, whereas the recall rate is again predictably very high. Secondly, we tested the effects of randomly removing a portion of injected genetic interactions over 10 trials. The average precision and recall values are shown in Fig. S4, demonstrating the relationship between the accuracy of these methods and the percentage of injected yeast interactions. A notable observation is that the performance becomes reasonable when > 40% of interactions are injected. The performance of PEXA remains relatively unchanged because it utilizes the human KEGG pathway information in its algorithm, as noted above. In terms of false-positives and -negatives, there is clearly a trade-off between the different methods. PCSF works best for our current work, as PCSF identifies a small set of relevant genes for cost-effective experimental explorations.

Statistical methods and data analysis for cell-based assays

Sample sizes for all experimentation were chosen based on our previous extensive experience with the methods and assays in these studies. For most experiments in mammalian cells, robustness and consistency of the results are typically established after three biological replicates are analyzed. Unless otherwise stated in the figure legends, this was the standard number of replicates required for all experiments. For all human and rat cellular experiments, significance was then determined by appropriate statistical tests that are standard in the field. The two-tail t-test was applied when there were only two conditions to compare within the experiments. One-Way ANOVA with a multiple comparisons post-hoc test was performed when experiments include multiple conditions. Data points were excluded based on the following pre-established criteria: 1) errors were introduced to the particular sample while performing the experiments, 2) the values are greater or less than two standard deviation from the mean. For yeast spot assays, results were considered significant when three biological replicates (unless otherwise stated) demonstrated the same trend by eye. Methods used for Fig. 5E are outlined in the figure legend. For the pooled screen yeast assay (Fig. 4A, Fig. S6) detailed statistical methods for reads and cutoff thresholds are supplied above in the methods. The statistical methods for the computational analysis are described in detail in the methods sections above.

DATA AND SOFTWARE AVAILABILITY

All cytsocape files depicting networks are provided in the supplement and referred to in the text. The TransposeNet pipeline is described at http://transposenet.csail.mit.edu. The DCA/Mashup web portal is http://mashup.csail.mit.edu. The PCSF web portal is http://fraenkel-nsf.csbi.mit.edu/omicsintegrator/.