Mapping DNA damage-dependent genetic interactions in yeast via orgy mating and barcode fusion genetics

Condition-dependent genetic interactions can reveal functional relationships between genes that are not evident under standard culture conditions. State-of-the-art yeast genetic interaction mapping, which relies on robotic manipulation of arrays of double mutant strains, does not scale readily to multi-condition studies. Here we describe Barcode Fusion Genetics to map Genetic Interactions (BFG-GI), by which double mutant strains generated via en masse orgy mating can also be monitored en masse for growth and genetic interactions. By using site-specific recombination to fuse two DNA barcodes, each representing a specific gene deletion, BFG-GI enables multiplexed quantitative tracking of double mutants via next-generation sequencing. We applied BFG-GI to a matrix of DNA repair genes under ten different conditions, including methyl methanesulfonate (MMS), 4-Nitroquinoline 1-oxide (4NQO), bleomycin, zeocin and four other DNA-damaging environments. BFG-GI recapitulated known genetic interactions and yielded new condition-dependent genetic interactions. We validated and further explored a subnetwork of condition-dependent genetic interactions involving MAG1, SLX4, and genes encoding the Shu complex, and found a new role for the Shu complex as a regulator of the checkpoint protein kinase Rad53.


Introduction
The importance of condition-dependent genetic interactions Genetic interactions, defined by a surprising phenotype that is observed when mutations in two genes are combined [1], are powerful tools to infer gene and pathway functions [2,3]. Of the genetic interactions currently known in any species, the vast majority were found using Synthetic Genetic Array (SGA) technology in Saccharomyces cerevisiae [4][5][6][7] and these studies have yielded a rich landscape of genetic interactions. The sign of genetic interaction (defined to be negative when mutants are synergistically deleterious, and positive when the combination is less severe than would be expected from independent effects) provides clues about whether the genes act in parallel or in a concerted or serial fashion. Measuring similarity between genetic interaction profiles, both at the level of single genes and of clusters of genes, has revealed a hierarchical map of eukaryotic gene function [5,6]. However, the vast majority of genetic interaction mapping has been conducted under a single standard culture condition.
The importance and qualitative nature of gene function changes with environmental fluctuation, so that condition-dependent genetic interaction mapping is required. For example, pairs of DNA repair genes had different genetic interactions when cells were cultured in the DNA damaging agent methyl methanesulfonate (MMS) [4,8]. Further investigation revealed that there were 2-4 times more genetic interactions between DNA repair genes under MMS treatment compared with rich media alone. Different growth conditions are likely to reveal different genetic interactions [3], suggesting that a plethora of condition-dependent genetic interactions remain to be uncovered by tripartite gene × gene × environment studies.

Current genetic interaction discovery technologies
Essentially every large-scale genetic interaction mapping strategy in yeast uses a genetic marker system developed for the SGA technique, which works by mating a single-gene deletion query strain with an array of different single-gene deletion strains from the Yeast Knockout Collection (YKO) [9]. The SGA system provided genetic markers by which mated diploids can be subjected to a series of selections to ultimately yield haploid double mutants. In 'standard' SGA mapping, the fitness of the resulting double mutants is determined computationally by imaging each plate to measure cell growth of each separately-arrayed strain [10]. SGA has also been used to study genetic interactions in functionally selected gene matrices [11] and applied to detect environment-dependent interactions [4]. St Onge et al [8] used the SGA markers to generate all pairwise double mutants between 26 DNA repair genes in yeast. In that study, the authors cultured each double mutant individually in microplates and monitored a time course of cell density to infer the fitness of the double mutants and identify genetic interactions in the presence and absence of MMS.
Others have measured genetic interactions via competition-based fitness measurements in liquid cultures, adding fluorescent markers for tracking cell viability, and using robotic manipulation to inoculate and measure cell growth [12,13]. A recent technique called iSeq incorporated barcodes into single-mutant strains, such that pairs of barcodes identifying corresponding pairs of deleted genes could be fused by Cre-mediated recombination [14]. They demonstrated the method, showing that a pool corresponding to 9 gene pairs could be sequenced to monitor competitive growth of double-mutants en masse in different environments [14].
For each of the above methods, double mutants were generated by individual mating of two specific yeast strains, requiring at least one distinct location for each double-mutant strain on an agar or microwell plate and necessitating robotic strain manipulation to achieve large scale. By contrast, other methods to map genetic interaction generated double mutants in a one-vs-many fashion. For example, diploid-based synthetic lethality analysis on microarrays (dSLAM) [15] disrupted a single 'query' gene by homologous recombination via transformation of a marker into a pool of diploid heterozygous deletion strains bearing the SGA marker. After selecting for double-mutant haploids from such a 'one-by-many' haploid doublemutant pool, barcodes were PCR amplified from extracted double mutant DNA and hybridized to microarrays to infer the relative abundance and thus fitness of double mutants. Another method, Genetic Interaction Mapping (GIM) [16], generated a oneby-many pool of barcoded double mutants by en masse mating a single query strain to a pool of haploid gene deletion strains. Like dSLAM, GIM inferred strain abundance and fitness via barcode hybridization to microarrays. Despite the efficiency of generating one-by-many double-mutant pools, a matrix involving thousands of query strains would require thousands of such pools to be generated. Each of the above methods has advantages and disadvantages. For example, measuring growth time-courses of each double-mutant strain provides high resolution fitness measurements [8,13], but scalability is low. Standard SGA is highthroughput, but requires specialized equipment for robotic manipulation, and these manipulations must be repeated to test genetic interaction in a new environment.
The iSeq method shares the scaling challenge of SGA in strain construction, in that it requires many pairwise mating operations; however, once a double-mutant pool has been generated, it represents a promising strategy for measurement of competitive pools in different environments. The dSLAM and GIM methods allow generation of one-by-many pools, which reduces the number of mating operations, but both methods require customized microarrays as well as pool-generation and a microarray hybridization for every query mutation in the matrix.

Barcode Fusion Genetics to map Genetic Interactions (BFG-GI)
Here we describe BFG-GI, which borrows elements from several previous approaches. Like iSeq, BFG-GI requires generation of barcoded single-mutant strains, with only minimal use of robotics. To generate double-mutant pools, BFG-GI uses the SGA marker system. It is similar to the GIM strategy in that it employs en masse mating. Unlike GIM and all other previous genetic interaction mapping strategies, BFG-GI employs many-by-many 'orgy mating' to generate all double mutants for a matrix of genes in a single mating step. All successive steps are also conducted en masse. We show that double mutants can be generated and monitored in competitive pools using BFG-GI. Like iSeq, BFG-GI infers double mutant fitness in competitively grown strain pools using next-generation sequencing of fused barcodes. Strain pools generated by orgy mating can be stored, and the aliquots can later be thawed and challenged under specific growth environments to detect condition-dependent genetic interactions without having to regenerate the double mutant strains. We assessed BFG-GI by mapping genetic interactions of DNA repairrelated genes under multiple DNA-damaging conditions, revealing many conditiondependent interactions and a new function for the Shu complex in regulating the Rad53 checkpoint protein.

BFG-GI experimental design overview
The first step in the BFG-GI process is generating uniquely barcoded donor and recipient strains from complementary mating types. Each donor and recipient contained loxP/2272 sites to mediate barcode fusion using the Cre/Lox system after the mating step. We created donors by crossing individual gene deletion strains from the YKO collection with proDonor strains that contained newly constructed We confirmed that barcode fusion occurred successfully using two neutralinsertion strains as controls (see Materials and methods for a definition of neutral loci). Specifically, we crossed a MATalpha Donor hoΔ::kanMX to a MATa Recipient ylr179cΔ::natMX and induced Cre/Lox recombination to fuse their barcodes. After sporulation and selection of the MATalpha haploid double mutant progeny (Materials and methods), we extracted genomic DNA, amplified barcode fusions by PCR and confirmed their integrity by Sanger sequencing (Fig 1C).
To scale up the BFG-GI process, we generated double mutants with unique fused barcodes en masse (detailed below). We selected hundreds of double mutants using a series of marker selection steps in a many-by-many fashion. Intermediate selection steps allowed us to fuse barcodes representing each donor and recipient parental pair within each double mutant cell (Fig 1D and Materials and methods).
Once we generated the pool of fused-barcode double mutants, aliquots were stored at -80°C for future experiments. Amplification and next-generation sequencing of fused barcodes in the pool allowed us to infer the relative abundance of each double mutant in each condition of interest (Fig 1D and

BFG-GI measures the strain abundance profile of a heterogeneous cell population
We first evaluated the ability of BFG-GI to accurately detect the abundance of double mutants. To generate reference data for this evaluation, we used the array-based SGA strategy to generate 2,800 double mutants by individual mating of barcoded BFG-GI strains, subsequently inducing barcode fusion via the Cre/Lox system. We recorded colony sizes, scraped plates to pool all double mutant cells, extracted genomic DNA, and sequenced the fused barcodes (Materials and methods). The resulting numbers of sequencing reads for each strain was strongly correlated with the corresponding colony sizes (R=0.92, Fig 2A). Importantly, very small or absent colonies correlated with double mutants with very few or no sequencing reads.
These results show that BFG-GI detects the abundance of specific double mutants in pools of cells, with results comparable to an array-based method.

Generating a DNA repair-focused double-mutant strain pool
To test whether BFG-GI can accurately map genetic interactions, we generated a double mutant pool focused on DNA repair genes and compared BFG-GI results to those of other validated genetic interaction assays. We began by generating donor and recipient strains by crossing 35 YKO (yfg1Δ::kanMX, MATa) single gene deletion strains to BFG-GI proDonor strains, and 38 SGA query (yfg2Δ::natMX, MATalpha) single gene deletion strains to BFG-GI proRecipient strains (Fig 1). These strains included 26 DNA repair genes from a previous condition-dependent genetic interaction study [8], as well as 14 likely-neutral loci (e.g. the already-disrupted HO locus, pseudogenes, and other loci for which single-and double-mutant phenotypes have not been previously observed). Inclusion of neutral loci allowed us to infer single mutant fitness from pools of double mutants (Materials and methods).
To generate haploid double mutants, donor and recipient cells were scraped from plates and all subsequent steps in the BFG-GI pipeline were conducted en masse. First, the pools were combined for 'orgy mating'. Seven selection steps followed mating, including four that correspond to those in the standard SGA procedure: heterozygous diploid selection, sporulation, MATa progeny selection, and haploid double mutant selection. Additionally, before sporulation, we completed three selection steps to fuse barcodes and subsequently remove Cre to limit undesired recombination events (Fig 1C and S3 Fig). This generated a pool of 4,288 haploid double mutants, which was aliquoted and stored as frozen glycerol stock. Thawed samples were used to inoculate solid media appropriate for selecting haploid double mutant cells. The media was used alone, supplemented with dimethyl sulfoxide (DMSO) as a control, or supplemented with one of eight drugs targeting DNA repair pathways (S1 Table). We extracted genomic DNA, amplified and sequenced fused barcodes to infer the relative abundance of each double mutant in each condition.
To evaluate assay reproducibility, we ran all BFG-GI procedures in duplicate, starting from the mating step (technical replicates) and also barcoded multiple strains representing the same gene (biological replicates). Biological replicate strains had either the same or different parental strain origin (the parental strain for a given gene deletion might be from either the YKO or SGA query strain collection). Relative strain abundance was highly correlated between technical replicates (R > 0.95) and thus we decided to combine technical replicates for subsequent analyses. Next, we used a multiplicative model [1] to infer a genetic interaction score (GIS) from relative strain abundances (Materials and methods).
Correlation of GIS profiles between biological replicates representing the same gene were in general high, with 85% of replicates showing GIS R>0.5. We computationally excluded 20 biological replicates showing correlations below this cutoff from analysis and the remaining same-gene biological replicates showed correlations that were clearly distinct from strain pairs representing different genes ( Fig 2B). Most replicates showing GIS R<0.5 were recipients. To understand factors contributing to uncorrelated pairs we sequenced the genomes of 10 strain pairs with GIS R<0.5 and another 10 with GIS R>0.5. We found that all 10 strains with GIS R<0.5 had chromosome V duplicated, in agreement with the report of iSeq strains that show low reproducibility [14]. Chromosome V contains the CAN1 locus, which is where both BFI-GI recipients and iSeq strain constructs were inserted. In contrast, only 3 out of 10 strains with R>0.5 showed aneuploidies (also in chromosome V). All BFG-GI strains showing aneuploidies were recipients. This suggests that future versions of BFG-GI recipients for which selection markers are carried by plasmids may increase reproducibility, as we found for our Donor strains.
Our final dataset consisted of 3,360 double mutants, with 60 Donors and 56 Recipients, representing 39 genes (25 DNA repair genes and 14 neutral genes; S2 Table). Replicates representing SWC5 showed very low relative abundance in the sequencing results and were removed from subsequent analyses.
We evaluated the performance of BFG-GI measurements by contrasting our GIS for no-drug and MMS conditions (two conditions commonly used as gold standards in genetic interaction studies) against Epsilon scores for these same conditions [8]. We found that GIS and Epsilon scores correlated well with each other in both conditions: no drug R=0.57 and MMS R=0.75 (Fig 2C and Fig 2D).
Furthermore, at a false positive rate of 20%, BFG-GI showed a sensitivity of 50% for detecting positive genetic interactions and 70% for detecting negative genetic interactions ( Fig 2E).
Finally, we assessed the ability of BFG-GI to predict the relative abundance of three classes of double mutant strains. First, we measured the abundance of double mutant strains with mutations in the same gene. Heterozygous diploid double mutants for the same gene (e.g. MMS4/mms4Δ::kanMX mms4Δ::natMX/MMS4) can survive in media supplemented with selective antibiotics, but haploids should not survive because they have only one locus for each gene and thus are not expected to carry both antibiotic resistance markers. Thus, haploid strains for same-gene pairs are expected to behave like synthetic lethal combinations and be depleted from the pools. Our BFG-GI sequencing results agreed with this hypothesis (Fig 2F). Second, we assessed the abundance of double mutants representing pairs of linked genes (<30 kbp apart). Independent segregation is reduced between linked genes, and as expected our BFG-GI quantitation measurement by sequencing indicated these double mutants were depleted from the pools ( Fig 2F). Last, we analyzed double mutants representing unlinked genes and we found that their GIS distribution is clearly distinguishable from same-gene and linked gene pairs ( Fig 2F).
Taken together, these results provide evidence that BFG-GI is a powerful tool to generate double mutants by mating en masse and to monitor strain abundance to infer condition-dependent genetic interactions.

BFG-GI reveals condition-dependent genetic interactions
Having determined that BFG-GI can accurately detect genetic interactions, we analyzed the DNA-repair focused double mutant pool under 10 culture conditions (S1 Table) to identify condition-dependent genetic interactions. We applied an absolute Z-score of GIS=1 as a cutoff to identify positive and negative genetic interactions in each of the 10 conditions (S1 Table and S2 Table). We found that although almost all genes showed at least one genetic interaction, some genes showed markedly more interactions than others. For example, we found that the DNA helicase gene SGS1 paired with MMS4, MUS81 or SLX4 (all of which participate in template switching during break-induced replication) yielded negative interactions in all 10 conditions ( Fig 3A). Another DNA helicase gene, SRS2, interacted negatively with both SGS1 and the DNA translocase gene RAD54 in all 10 conditions. By contrast, both SGS1 and SRS2 showed positive genetic interactions in most conditions with RAD5 and RAD57, which are involved in error-free DNA damage tolerance and recombinational repair of double-strand breaks, respectively.
These findings coincide with previous reports showing SGS1 and SRS2 centrality in DNA repair pathways in both unperturbed and MMS-induced stress conditions [8].
We identified condition-dependent genetic interactions by comparing GISs between each pair of conditions. For example, we found mus81Δ/rad5Δ displayed a negative GIS in DMSO and a positive GIS in MMS. This change is shown as a red edge in Fig 3B, panel i, and agrees with a previous report [8]. By contrast, most changes in genetic interactions between DMSO and MMS were from neutrality in one condition to either positive or negative GIS in the other (Fig 3B, panels i and iv). In fact, when we extended this analysis to include all pairwise condition comparisons we found that that the vast majority of genetic interaction changes (>95%) were from neutrality in one condition to either positive or negative GIS in the other condition; less than 5% of genetic interactions changed from positive in one condition to negative in the other ( Fig 3C).
As expected, the two conditions most similar to each other were no-drug and DMSO, which showed only eleven sign changes, all from neutrality to either positive or negative GIS (S1 Table). Among the eight drugs, cisplatin showed the least changes compared with DMSO or no drug conditions (53 and 50 sign changes, respectively), followed by bleomycin and zeocin (each exhibiting 52 sign changes), which are members of the same family of glycopeptides that intercalate into DNA to induce double strand breaks [17]. In contrast, condition pairs showing the highest number of sign changes include camptothecin vs. either 4NQO or MMS (each with 84 sign changes). These data are consistent with the fact that these drug pairs have different mechanisms of action and cause DNA lesions that are repaired by different pathways.
Our BFG-GI results indicated that genes encoding all four members of the Shu complex showed negative genetic interactions with both MAG1 and SLX4 during exposure to MMS. Additionally, the Shu complex genes interacted negatively with SLX4 during treatment with 4NQO, bleomycin and zeocin ( Fig 4B). Mag1 is a 3methyladenine DNA glycosylase that removes alkylated bases from DNA to initiate base-excision repair (BER), thereby protecting cells against alkylating agents like MMS [23,24]. Slx4 promotes the activity of three structure-specific endonucleases [25][26][27][28] and, upon exposure to MMS, plays a key role in down-regulating phosphorylation of the checkpoint kinase Rad53 [29,30]. We generated double mutants for each Shu complex member with either MAG1 or SLX4 and tested their fitness on media containing DMSO or various genotoxins using spot dilution assays ( Fig 4C). Our results validated the MAG1-Shu complex interaction in MMS that we detected with BFG-GI, and are consistent with a previous study [31]. The negative interactions between MAG1 and Shu complex members are explained by the simultaneous loss of Mag1-mediated BER that directly removes alkylated bases and decreases in error-free lesion bypass, a major pathway used during MMS-induced DNA replication blocks [32], in the double mutants [31] (Fig 4A). Our spot dilution assays also confirmed that MAG1 interacts negatively with SLX4 during MMS treatment, in agreement with a previous study showing that BER is unlikely to be the major function of SLX4 [25]. Of particular interest, we validated the BFG-GI interactions between Shu complex members and SLX4 during treatment with MMS, 4NQO, bleomycin, or zeocin ( Fig 4C). As the nature of the SLX4 interactions with Shu complex genes is unknown, we decided to study them in more detail.
The negative genetic interactions between SLX4 and Shu complex members in MMS were unexpected, given that the Shu complex promotes error free lesion bypass [18,21,31,33] and SLX4 shows epistatic relationships with error-free lesion bypass genes during MMS treatment [25]. A major role for Slx4 during MMS treatment is in the down-regulation of Rad53 phosphorylation and activation, which occurs by Slx4 competing with Rad9 for binding to Dpb11 in order to limit the formation of Rad9-Dpb11 complexes that activate Rad53 [29,30,34,35]. Cells deleted for SLX4 or PPH3, which encodes the catalytic subunit of the protein phosphatase PP4 complex that binds and dephosphorylates Rad53 during MMS treatment [36], display hyperactivation of Rad53 upon MMS treatment and sensitivity to MMS that is suppressed by expression of a hypomorphic rad53-R605A allele defective for full Rad53 activation [29,30,34]. In slx4Δ pph3Δ cells, Rad53 hyperactivation is further elevated and these double mutants display synergistic sensitivity to MMS [29]. To determine whether the genetic interactions between SLX4 and Shu complex members reveal an unanticipated role for the Shu complex in regulating Rad53-P levels (Fig 4D), we tested the sensitivity of pph3Δ/Shu complex double mutants to MMS using spot dilution assays. Combining pph3Δ with deletion of any of the Shu complex genes resulted in a dramatic increase in MMS sensitivity relative to the single mutants (Figs 4C and 4E), indicating negative genetic interactions similar to those seen between SLX4 and Shu complex members (Fig 4C), or between SLX4 and PPH3 [29].
To assess MMS-induced Rad53 activation in Shu complex mutants more directly, we monitored Rad53 phosphorylation (which is a proxy for Rad53 activation) using western blot assays. Consistent with the role of SLX4 in dampening Rad53 activation [29,30,37], slx4Δ cells challenged with MMS showed an increase in Rad53-P levels relative to wild type ( Fig 4F). Interestingly, three of the Shu complex mutants (csm2Δ, psy3Δ, and shu1Δ) also showed an increase in Rad53-P levels upon treatment with MMS (Fig 4F), indicating that these Shu complex mutants, like slx4Δ and pph3Δ cells, display hyperactivated Rad53 under exposure to MMS. We asked whether the MMS sensitivity of Shu complex mutants could be suppressed by expression of the rad53-R605A allele. Expression of rad53-R605A, which is not effectively hyper-activated, suppresses the MMS sensitivity of slx4Δ and pph3Δ [29,30]. Similarly, the MMS sensitivity of csm2Δ, psy3Δ, shu1Δ and shu2Δ mutants was partially suppressed by rad53-R605A ( Fig 4G). Together, our data indicate that the Shu complex, like Slx4 and Pph3, regulates Rad53 activation in response to MMS treatment, as revealed by unique condition-dependent genetic interactions detected by BFG-GI.

Discussion
We developed a new technology, called BFG-GI, in which pools of double mutant yeast strains corresponding to a matrix of target genes are generated en masse through many × many 'orgy mating'. These pools are induced to form doublemutant-identifying chimeric barcodes by intra-cellular site-specific recombination, and assayed for growth via next-generation sequencing. Aliquots of these pools can be stored, and later cultured with different drugs to identify condition-dependent genetic interactions. To our knowledge, BFG-GI is the first method to generate haploid double-mutant strains en masse for a many × many matrix of genes without the requirement for multiple mating steps, thus enabling large-scale conditional genetic interaction mapping without extensive use of robotics.
BFG-GI showed good agreement with previous methods in mapping genetic interactions commonly used to benchmark genetic interaction technologies.
This false positive rate is conservative, in that potentially novel true interactions are treated as false positives. We detected and validated unanticipated interactions between the Shu complex and SLX4, and found that the Shu complex dampens Rad53 activation during MMS treatment. Thus, our results provide evidence for a previously uncharacterized role of the Shu complex in the cellular response to DNA damage by MMS.
We calculated similarity between the genetic interaction matrices of different drugs, and found that those with similar mechanisms of action, like zeocin and bleomycin, are considerably more similar to each other than those with different mechanisms of action, like MMS and camptothecin. This suggests the potential of BFG-GI to shed light on drug mechanisms through measurement of gene-geneenvironment interactions.
One advantage of BFG-GI is its cost-effectiveness. BFG-GI uses fewer reagents and less robotic assistance than other technologies to map genetic interactions because it is a pool-based technology. Pool-based technologies require less media, plates, and drugs than array-based technologies, a substantial cost advantage particularly when the price of drugs is factored in. For example, the amount of media used in 1536 spot arrays on OmniTrays is reduced 50-fold by studying the same number of gene-pairs in 100 OD pooled cultures in 143 cm 2 Petri dishes (optimal cell density for pooled double mutant selections). BFG-GI is also more costeffective than other barcode-sequencing technologies because in BFG-GI, strains are pooled at the mating step, rather than generating strains using robotically manipulated strain arrays.
The reproducibility of BFG-GI indicates that it is a robust technology.
Technical replicates in BFG-GI are highly reproducible, and 85% of the biological replicates correlated well with each other (GIS R>0.5). The remaining 15% of biological replicates showing low correlations could be identified and removed computationally. We concur with the iSeq study [14] that aneuploidies in chromosome V are the main factor contributing to the replicates with low reproducibility. Chromosome V carries both CAN1 and URA3 loci, which were replaced by selection markers in the iSeq protocol [14], while CAN1 was replaced by the recipient constructs in BFG-GI. Thus, de novo mutations around these loci during strain construction could explain the low correlation between some pairs of biological replicates. This possibility is supported by our observation that almost all BFG-GI strains showing GIS R<0.5 were recipients, whereas donors -for which constructs are carried on plasmids-showed GIS R>0.5. In the BFG-GI protocol, once the donor and recipient barcodes are fused, the relic donor plasmid is counterselected with 5-FOA to reduce the chance of undesired recombination events. We concur with Jaffe et al. [14] who suggest that future protocols using constructs located on plasmids, such as the one we used with the proDonor strains, or in other chromosomal loci may serve to eliminate this issue. Notwithstanding this issue, however, the BFG-GI method proved to be highly accurate in comparisons with previous benchmark studies.
We took several steps to reduce the chance of undesired strains in BFG-GI from taking over pooled cultures, including optimization of both mating and sporulation, and adapting protocols and constructs reportedly improving the selection of the MATa double mutant progeny in SGA. For example, mating and sporulation are the two primary population bottlenecks when generating haploid double mutants by meiotic segregations, and they must be optimized to maintain a pool complexity that is large enough to interrogate all desired gene-gene combinations. Optimizing these two processes is also important to reduce potential jackpot effects in the pool cultures (i.e. to avoid strains with genetic anomalies to take over the entire pool growth). We elaborated on previous studies to optimize mating [38] and sporulation [39] for our culture pools. We found that cell density was a key factor for mating efficiency (3% using 300 ODs vs. 22% using 30 ODs, in the same mating area, Materials and methods). Similarly, the time allowed for sporulation dramatically affected its efficiency (4% at 5 days vs. 18% at 12 days).
Furthermore, we used the STE2 and STE3 promoters currently used for SGA to select for haploid cells which have been reported to perform better than earlier alternatives (e.g. MFA/MFalpha promoters) [40]. We used these constructs to first However, once generated these proDonor and proRecipient "toolkits" can be used many times to create donor and recipient strains representing different genes with minimal robotic manipulation. We anticipate that BFG-GI will be a valuable technology to map condition-dependent genetic interactions in yeast and, as nextgeneration sequencing costs continue to decrease, BFG-GI can be expanded to interrogate pools of double mutants representing bigger sets of gene pairs, including full genome combinations, across multiple conditions.

Selected DNA repair and neutral gene strains
We retrieved strains representing 26 DNA repair genes whose null mutants were sensitive to MMS [8]

Donor toolkit construction
We constructed donor strains by generating two DNA fragments with overlapping ends which were co-transformed into yeast where they recombined generating pDonor constructs (S1 Fig). The first fragment, called preD1, contained the hygromycin resistance gene (HygR) driven by the Schizosaccharomyces pombe TDH1 promoter and terminator, a unique barcode, loxP/2272 loci, and flanking primer sites. First, we used a Gibson assembly [41] to produce plasmid pFR0032 with the PspTDH1-HygR-TspTDH1 backbone. Then, we used three consecutive PCRs to add barcodes, priming sites, loxP/2272 loci, and in-yeast recombination adapters (S1A and preD2 fragments were co-transformed into yeast strain RY0771 (derived from BY4742) and merged by in-yeast assembly to generate pDonor plasmids (S1C Fig). We arrayed transformant strains to extract DNA and sequenced the preD1 loci, and proceeded with those strains containing confirmed preD1 loci. We mated selected MATalpha proDonors with MATa deletion strains of interest (i.e. DNA repair or neutral genes) from the YKO collection (S1D Fig). A series of selective passages (S1D

Recipient toolkit construction
We constructed recipient strains using a method based on the delitto perfetto construct [42] to enhance homologous recombination of contructs as follows. First, we used consecutive PCRs to produce a fragment preR1, containing the

Mating optimization for en masse BFG-GI
We focused on optimization of cell density for en masse orgy mating because previous evidence shows cell density influences mating efficiency [38]. We

Generation of MATa haploid double mutants with fused barcodes
We used cultures recovered from 5-FOA counter selection to inoculate liquid PRE5 pre-sporulation media for 2 hrs at 30°C to induce exponential growth, then spun down the cells and transferred them to SPO2 sporulation media [39] supplemented with histidine, leucine, methionine and uracil to mask BFG-GI strain auxotrophies at concentrations used in the SGA sporulation protocol [10]. We incubated sporulation cultures at 21°C for 12 days. This resulted in ~18% sporulation efficiency, as evaluated by counting CFU's in non-selective and selective media and tetrad visualization. Shorter incubation periods reduced the sporulation efficiency (~4% at 5 days, ~13% at 7 days). Finally, we selected the MATa haploid progeny from sporulation cultures, followed by haploid double mutant selection. Aliquots were stored in glycerol at -80 degrees for future use.

Exposure of pooled cultures to drugs
Before challenging haploid double mutant pools to drugs we identified the appropriate drug concentration for our experiment by exposing a neutral BFG-GI haploid double mutant (hoΔ::kanMX/ ylr179cΔ::natMX) in growth assay liquid cultures to various drug concentrations. We selected drug doses corresponding to 20% of the minimal inhibitory concentration for the neutral test strain (S1 Table).
To expose mutant strains to drugs we thawed frozen haploid double mutant pools, allowed the pools to recover for 2 hrs in haploid double mutant liquid media at 30°C, and then used 100 ODs of this culture to inoculate 143cm 2 petri dishes containing solid media supplemented with each DNA repair drug. We cultured pools at 30°C for 24 hrs and then collected samples to sequence fused barcodes and thus infer each double mutant abundance.

Generation of BFG-GI double mutants in an array format
Mating and selecting donor and recipient strains for an array format was similar to the pool-based en masse orgy assay described above, but in this case we used robotic assistance to pairwise mate each donor with an array of recipients. We completed all steps, including sporulation, on solid media, and imaged the final haploid double mutant selection plates. We scraped cells from the final selection plates to sequence the fused-barcode population which allowed us to compare colony sizes with numbers of sequencing reads.

Next-generation sequencing and mapping of fused barcode pairs
The BFG-GI technology relies on the Cre/Lox system to recombine the complementary donor and recipient loxP/lox2272 sites that flank the barcodes (Fig   1). We sequenced the fused barcodes from pools of cells using the following steps:

Whole-genome sequencing and detection of chromosome duplications
Genomic DNA from 20 strains (S2 Table) was extracted by cell wall disruption with Zymolyase 100T 10mg/ml (Amsbio) and purified using AMPure beads (Agilent).
gDNA was quantified with Quant-it Picogreen dsDNA assay kit (Invitrogen) and normalized to 2ng/ul for DNA fragmentation and library normalization with a Nextera XT DNA Library Prep Kit, using a transposase (Tn5) for tagmentation. A limited-cycle PCR was used to add Illumina sequencing adapters and indices i5 and i7. PCR amplicons with size between 400 and 800 bp were gel purified using a 2% E-Gel EX agarose 2% (Invitrogen) and MiniElute Gel Extraction kit (Qiagen). Whole genome sequencing was conducted on an Illumina NextSeq 500 using a HighOutput 150 cycles v2 kit with 40x coverage. Sequencing results were mapped against the reference genome UCSC sacCer3 (SGD vR64.1.1), corrected for GC content, and chromosomal duplications detected with the HMMcopy R package [43].

Retesting double mutant construction and spot dilution assays
We generated double mutant strains for retesting in spot dilution assays by mating single mutant MATalpha SGA queries with MATa YKO collection strains, the exceptions being the MATa RAD53 (MBS1437) and rad53-R605A (MBS1440) strains with the RAD53 loci linked C-terminally to a 6xHis-3xFLAG-kanMX6 tag and resistance marker [30]. Next, we induced sporulation of heterozygous diploid double mutants as we did for BFG-GI strains. To confirm segregation of kanMX and natMX markers we manually dissected haploid double mutants from tetrads and verified segregation using both selective media and PCR. Sanger sequencing confirmed residue 605 of RAD53 and rad53-R605A strains. We grew strains overnight to saturation in liquid media, diluted them 1:10, and then used 1:5 serial dilutions for the spot assays. All cultures used YPD media supplemented with indicated drug concentrations.

Calculating relative abundance of strains from fused-barcode sequencing counts
We counted the total number of reads (C) for the donor (i) and recipient (j) barcode pairs (ij) in each condition-specific pool (k), and then divided by the total number of barcode counts in each k : We also calculated the marginal frequencies (M) for each barcode:

Inference of single mutant abundance
First we contrasted the relative abundance of each barcode with measurements from the heterozygous diploid pool (h), which we used as a 'time zero' reference control.
β is a pseudocount-based regularization parameter such that: and α is the number of pseudocounts used to avoid overestimating poorly measured barcode pairs. Values for α between 1 and 10 were tested and no major differences were found in terms of benchmarking against previously published datasets [8] (described below), therefore α=1 was used.
Strains grown in selective pools of haploid double mutants need the two gene deletion markers kanMX and natMX to survive, therefore, we approached the single mutant and wild type relative abundance by using measurements for neutral genes (S4 Fig) in k and h pools. For the wild type inference we used: Finally, to infer each single mutant fitness (W), barcode relative abundance of each strain was contrasted with the relative abundance inferred for wild type: !",! and !" = !" !",!

Inference of double mutant abundance
Similar to the single mutant metrics, we used the double mutant relative abundance to calculate strain changes over time: Furthermore, we compared the normalized double mutant values to wild type values to derive double mutant fitness:

Genetic Interaction Score
Our genetic interaction score was inspired by the multiplicative model that is now commonly used to score genetic interactions [1]: GIS values smaller than zero represent negative genetic interactions, whereas those above zero represent positive genetic interactions. To contrast genetic interactions between condition pairs in Fig 3, we obtained a Z-score for GIS within each condition and used an absolute GIS = 1 as cutoff. Onge et al. dataset [8]. (F) Density plot comparing the GIS distribution for samegene pairs (which are expected to behave like synthetic lethals given the SGA double-mutant selection process) with that for linked and unlinked gene pairs.  Table). Sporulation was conducted in flasks with liquid media shaking at 200rpm. We used the following reagent concentrations: G418=200 µg/mL, clonNat=100 µg/mL, canavanine=100 µg/mL, thialysine=100 µg/mL, hygromycin=200 µg/mL, 5-FOA=1 mg/mL. Amino acid concentrations were as described in [10].

S4 Fig. Strains and Genes in BFG-GI Pools
Sixty Donors, representing 34 genes and 56 Recipients, representing 38 genes were crossed all-vs-all as a pool. The first number in the parentheses is the total number of strains and the second number in the parentheses is the number of genes. Single mutant fitness was inferred from double mutant fitness measurements corresponding to one DNA repair-and one neutral gene. Similarly, the neutralneutral double mutants were used to infer the wild type fitness.

S1 Table. BFI-GI Tested Conditions
Description of conditions tested, including drug names, concentrations and vendor codes.       Figure S3 Figure S1       Figure S4