A yeast phenomic model for the influence of Warburg metabolism on genetic buffering of doxorubicin

Background The influence of the Warburg phenomenon on chemotherapy response is unknown. Saccharomyces cerevisiae mimics the Warburg effect, repressing respiration in the presence of adequate glucose. Yeast phenomic experiments were conducted to assess potential influences of Warburg metabolism on gene-drug interaction underlying the cellular response to doxorubicin. Homologous genes from yeast phenomic and cancer pharmacogenomics data were analyzed to infer evolutionary conservation of gene-drug interaction and predict therapeutic relevance. Methods Cell proliferation phenotypes (CPPs) of the yeast gene knockout/knockdown library were measured by quantitative high-throughput cell array phenotyping (Q-HTCP), treating with escalating doxorubicin concentrations under conditions of respiratory or glycolytic metabolism. Doxorubicin-gene interaction was quantified by departure of CPPs observed for the doxorubicin-treated mutant strain from that expected based on an interaction model. Recursive expectation-maximization clustering (REMc) and Gene Ontology (GO)-based analyses of interactions identified functional biological modules that differentially buffer or promote doxorubicin cytotoxicity with respect to Warburg metabolism. Yeast phenomic and cancer pharmacogenomics data were integrated to predict differential gene expression causally influencing doxorubicin anti-tumor efficacy. Results Yeast compromised for genes functioning in chromatin organization, and several other cellular processes are more resistant to doxorubicin under glycolytic conditions. Thus, the Warburg transition appears to alleviate requirements for cellular functions that buffer doxorubicin cytotoxicity in a respiratory context. We analyzed human homologs of yeast genes exhibiting gene-doxorubicin interaction in cancer pharmacogenomics data to predict causality for differential gene expression associated with doxorubicin cytotoxicity in cancer cells. This analysis suggested conserved cellular responses to doxorubicin due to influences of homologous recombination, sphingolipid homeostasis, telomere tethering at nuclear periphery, actin cortical patch localization, and other gene functions. Conclusions Warburg status alters the genetic network required for yeast to buffer doxorubicin toxicity. Integration of yeast phenomic and cancer pharmacogenomics data suggests evolutionary conservation of gene-drug interaction networks and provides a new experimental approach to model their influence on chemotherapy response. Thus, yeast phenomic models could aid the development of precision oncology algorithms to predict efficacious cytotoxic drugs for cancer, based on genetic and metabolic profiles of individual tumors.


Background
The Warburg effect refers to the phenomena of cancer cells undergoing a metabolic transition from respiration to aerobic glycolysis and has been documented for over 90 years, yet there remains a lack of consensus regarding how this contributes to cancer [1][2][3]. In humans, aerobic glycolysis is a cancer-specific metabolic transition; however, yeast normally represses respiration in the presence of adequate glucose [4][5][6]. Although not possible in a single cell organism to ascertain the role of the Warburg transition in oncogenesis, we wondered whether it might influence chemotherapeutic response, and particularly in the context of vulnerabilities created by genomic instability and unique to individual patient's cancers. Using doxorubicin as a model anti-cancer agent, we examined whether doxorubicin-gene interaction manifests differentially under glycolytic vs. respiratory conditions in yeast and how genetic insights from the yeast model might lead to predicting variable efficacy in killing cancer cells. It is also possible that the model could be informative regarding dose-limiting toxicity observed in cardiomyocytes, which have respiratory rates among the highest of all cell types [7].
Doxorubicin is used widely in oncology to treat both hematologic cancer and solid tumors [8]. Proposed mechanisms of doxorubicin cytotoxicity include topoisomerase II poisoning, DNA adduct formation, oxidative stress, and ceramide overproduction [8][9][10][11][12][13]. Topoisomerase II is an ATP-dependent enzyme that relieves the DNA torsional stress occurring with replication or transcription by catalyzing a double-stranded DNA (dsDNA) break, relaxing positive and negative DNA supercoiling, and finally re-ligating the DNA [14]. Inhibiting this activity can result in irreparable DNA damage and induction of apoptosis, selectively killing rapidly dividing proliferating cells [15][16][17]. Doxorubicin also causes histone eviction leading to chromatin trapping and damage [9,[18][19][20]. In addition to its potent anti-cancer therapeutic properties, doxorubicin is known for dose-limiting cardiomyocyte toxicity, causing cardiomyopathy and heart failure years post-treatment [21]. In this regard, topoisomerase IIB is highly expressed specifically in myocardiocytes, where tissue-specific deletion suppresses cardiac toxicity in mice [22]. Clinical guidelines recommend a maximum cumulative lifetime dose of 500 mg/m 2 ; however, doxorubicin toxicity is variable and has a genetic basis [23]. Thus, a detailed understanding of drug-gene interaction could advance the rationale for more precisely prescribing doxorubicin (among other cytotoxic agents) and also predicting toxicity, based on the unique genetic context of each patient's tumor genetic profile as well as germline functional variation.
To address these questions, this work establishes a yeast phenomic model to understand genetic pathways that buffer doxorubicin toxicity [24][25][26][27][28][29][30], and how the Warburg effect influences the doxorubicin-gene interaction network. We conducted yeast phenomic analysis of doxorubicin-gene interaction, consisting of quantitative high throughput cell array phenotyping (Q-HTCP) of the yeast knockout and knockdown (YKO/KD) libraries, using multiple growth inhibitory concentrations of doxorubicin in either dextrose-(HLD) or ethanol/glycerol-based (HLEG) media. Q-HTCP provided cell proliferation parameters (CPPs) with which to quantify doxorubicin-gene interaction and determine its dependence on respiratory vs. glycolytic metabolism [31][32][33]. The yeast phenomic model was used to predict causality underlying correlations between doxorubicin sensitivity and increased or decreased expression of the homologous human gene in pharmacogenomics data from cancer cell lines. Thus, the work details genetic pathways for buffering doxorubicin toxicity in yeast, including the influence of Warburg metabolism on the network, and applies the information to predict interactions between doxorubicin and functional genetic variation that could be present in cancers from different, individual patients.

Strains and media
The yeast gene knockout strain library (YKO) was obtained from Research Genetics (Huntsville, AL, USA). The knockdown (KD) collection, also known as the Decreased Abundance of mRNA Production (DAmP) library, was obtained from Open Biosystems (Huntsville, AL, USA). The genetic background for the YKO library was BY4741 (S288C MATa ura3-Δ0 his3-Δ1 leu2-Δ0 met17-Δ0). Additional information and lists of strains can be obtained at https://dharmacon.horizondiscovery.com/cdnas-and-orfs/non-mammalian-cdnas-and-orfs/ yeast/#all. Some mutants appear multiple times in the library and they are treated independently in our analysis. HL yeast media, a modified synthetic complete media [27], was used with either 2% dextrose (HLD) or 3% ethanol and 3% glycerol (HLEG) as the carbon source.
Quantitative high throughput cell array phenotyping (Q-HTCP) Phenomic data was obtained by Q-HTCP, a custom, automated method of collecting growth curve phenotypes for the YKO/KD library arrayed onto agar media [33]. A Caliper Sciclone 3000 liquid handling robot was used for cell array printing, integrated with a custom imaging robot (Hartman laboratory) and Cytomat 6001 (Thermo Fisher Scientific, Asheville, NC, USA) incubator. Three hundred eighty-four-culture array images were obtained approximately every 2 h and analyzed as previously described [28,33]. To obtain CPPs, image data were fit to the logistic equation, G(t) = K/(1 + e −r(t −l) ), assuming G(0) < K, where G(t) is the image intensity of a spotted culture vs. time, K is the carrying capacity, r is the maximum specific growth rate, and l is the moment of maximal absolute growth rate, occurring when G(t) = K/2 (the time to reach half of carrying capacity) [31]. The resulting CPPs were used as phenotypes to measure doxorubicin-gene interaction.
Interaction scores were calculated as previously described [28], with slight modifications, as summarized below. Variables were defined as follows: D i = concentration (dose) of doxorubicin R i = observed mean growth parameter for parental reference strain at D i Y i = observed growth parameter for the YKO/KD mutant strain at D i K i = Y i -R i , the difference in growth parameter between the YKO/KD mutant (Y i ) and reference (R i ) at D i K 0 = Y 0 -R 0 , the effect of gene KO/KD on the observed phenotype in the absence of doxorubicin; this value is annotated as "shift" and is subtracted from all K i to obtain L i L i = K i -K 0 , the interaction between (specific influence of) the KO/KD mutation on doxorubicin response, at D i For cultures not generating a growth curve, Y i = 0 for K and r, and the L parameter was assigned Y i max, defined as the maximum observed Y i among all cultures exhibiting a minimum carrying capacity (K) within 2 standard deviations (SD) of the parental reference strain mean at D i . Y i max was also assigned to outlier values (i.e., if Y i > Y i max).
The interaction was calculated by the following steps: 1) Compute the average value of the 768 reference cultures (R i ) at each dose (D i ) 2) Assign Y i max (defined above) if the growth curve is observed at D 0 , but not at D i , or if observed Y i is greater than is useful for standardizing against variance (Additional files 2, 3 and 4). 8) Calculate interaction z-scores (Fig. 1d): z-score (YKO/KD INT ) > 2 for L or ≤ 2 for K are referred to as gene deletion enhancers of doxorubicin cytotoxicity, and conversely, L interaction score ≤ 2 or K interaction scores > 2 are considered gene deletion suppressors (Fig. 1e).

Recursive expectation-maximization clustering (REMc) and heatmap generation
REMc is a probability-based clustering method and was performed as previously described [34]. Clusters obtained by Weka 3.5, an EM-optimized Gaussian mixture-clustering module, were subjected to hierarchical clustering in R (http://www.r-project.org/) to further aid visualization with heatmaps. REMc was performed using L and K interaction z-scores (Fig. 1f). REMc uses an expectation-maximization algorithm to define clusters probabilistically and is applied recursively to resolve gene interaction profile clusters. REMc terminates when a round of clustering reveals no new clusters. The cluster naming convention is "A-B.C.D-X", where "A" = the round of clustering, "B" = 0, and "C.D-X" indicates the cluster pedigree. For example, 1-0-0 refers to the first cluster of the first round, 2-0.0-3 the fourth cluster derived from 1-0-0 (in round 2 of REMc), 3-0.0.3-1 indicates the second cluster derived from 2-0.0-3 (in round 3), and so on [34]. The main effect of the gene KO or KD on cell proliferation, i.e., K i in the absence of doxorubicin (D 0 ) is also referred to as "shift." The shift was not subjected to REMc, but was included for hierarchical clustering and visualization by heatmaps after REMc. K i is termed shift, because this value is subtracted from the data series for each YKO/KD to obtain L i values, which are fit by linear regression for calculating drug-gene interaction. Additional file 5 contains REMc results in text files with associated data also displayed as heatmaps. In cases where a culture did not grow in the absence of drug, 0.0001 was assigned as the interaction score, and associated data were colored red ("NA") in the shift columns of the heatmaps.

Gene ontology term finder (GTF)
A python script was used to format REMc clusters for analysis with the command line version of the GO Term Finder (GTF) tool downloaded from http:// search.cpan.org/dist/GO-TermFinder/ [35]. GTF reports on the enrichment of Gene Ontology (GO) terms by comparing the ratio of genes assigned to a term within a cluster to the respective ratio involving all genes tested. Additional file 5 contains GTF analysis of all REMc clusters. GO-enriched terms from REMc were investigated with respect to genes representing the term and literature underlying their annotations [36].

Gene ontology term averaging (GTA)
In addition to using GTF to survey functional enrichment in REMc clusters, we developed GTA as a complementary workflow, using the GO information on SGD at https://downloads.yeastgenome.org/curation/literature/ to perform the following analysis: 1. Calculate the average and SD for interaction values of all genes in a GO term. 2. Filter results to obtain terms having GTA value greater than 2 or less than − 2. 3. Obtain GTA scores defined as |GTA value|-gtaSD; filter for GTA score > 2.
The GTA analysis is contained in Additional file 6 as tables and interactive plots created using the R plotly package https://CRAN.R-project.org/package=plotly. GTA results were analyzed primarily using the L interaction scores; however, GTA results with K interaction scores are included in Additional file 6: File D.

Validation of doxorubicin-gene interaction
We retested 364 YKO/KD strains having human homologs in the P-POD database [37] and L interaction scores greater than 2 or less than − 2 in at least one media type. Strains were struck to obtain four single colonies and arranged on replicate 384-well plates along with 20 reference strain controls and reanalyzed by Q-HTCP on HLD and HLEG, as in the genomewide experiment. Results are summarized in Fig. 2s, t, Additional file 2: Tables S5-S8, and Additional files 3 and 4: Files C-D.

Prediction of human homologs that influence tumor response to doxorubicin
PharmacoDB reports on pharmacogenomics data from cancer cell lines, including transcriptomics and drug sensitivity [38]. The PharmacoGx R/Bioconductor package [39] was used to analyze the GDSC1000 (https://pharmacodb.pmgenomics.ca/datasets/5) and gCSI (https://pharmacodb.pmgenomics.ca/datasets/4) datasets, which contained transcriptomic and doxorubicin sensitivity results. A p value < 0.05 was used for differential gene expression and doxorubicin sensitivity. For gene expression, the sign of the standardized coefficient denotes increased (+) or decreased (−) expression. The biomaRt R package [40,41] was used with the Ensembl database [42] to match yeast and human homologs from the phenomic and transcriptomic data, classifying yeast-human homology as one to one, one to many, and many to many.
(See figure on previous page.) Fig. 1 Experimental strategy to characterize differential doxorubicin-gene interaction, with respect to the Warburg metabolic transition. a The phenomic model incorporates treatment of individually grown cultures of the YKO/KD collection with increasing doxorubicin (0, 2.5, 5, 7.5, and 15 ug/mL) in "fermentable/glycolytic" (HLD) or "non-fermentable/respiratory" (HLEG) media. b Representative cell array images, treated and untreated with 15 ug/mL doxorubicin. c Time series of individual culture images, exemplifying gene deletion suppression (vps54-Δ0) and gene deletion enhancement (mms1-Δ0), relative to parental control ("RF1") in HLEG media with indicated concentrations (0, 5, and 15 ug/mL) of doxorubicin. d After image analysis, data time series are fit to a logistic growth function, G(t), to obtain the cell proliferation parameters (CPPs), K (carrying capacity), L (time at which K/2 is reached), and r (maximum-specific rate) for each culture. "ΔL" (left panel) indicates K i (see the "Methods" section). e Interaction is quantified by linear regression of L i (indicated "Delta_L" and "Delta_K" in right panels; see the "Methods" section) across the entire dose range, which is converted to a z-score by dividing with the variance of the parental reference control (see the "Methods" section). f Gene interaction profiles were grouped by recursive expectation-maximization clustering (REMc) to reveal deletion-enhancing and deletion-suppressing doxorubicin-gene interaction modules and the influence of the Warburg effect. Resulting clusters were analyzed with GOTermFinder (GTF) to identify enriched biological functions. g Gene Ontology Term Averaging (GTA) was used as a complement to REMc/GTF. h The model for genetic buffering of doxorubicin cytotoxicity incorporates primary and interaction effects involving glycolysis (green), and respiration (red), to explain the influence of Warburg context (blue) on doxorubicin-gene interaction (black)

Phenomic characterization of doxorubicin response genes
The workflow for analyzing doxorubicin-gene interaction and differential buffering of doxorubicin with respect to the Warburg effect is summarized in Fig. 1. Alternately, in a respiratory or glycolytic (HLEG or HLD media, respectively) context ( Fig. 1a), Q-HTCP technology was used for high throughput kinetic imaging of 384-culture cell arrays plated on agar media (Fig. 1b), image analysis (Fig. 1c), and growth curve fitting ( Fig.  1d) to obtain the CPPs, L (time to reach half-carrying capacity), K (carrying capacity), and r (maximum specific rate) [28,31,33], which were used to measure doxorubicin-gene interaction across the entire YKO/KD library. The departure of the observed CPP from the expected doxorubicin response for each YKO/KD strain was derived using distributions from many replicate reference strain control cultures and summarized across all doxorubicin concentrations by linear regression (Fig. 1e). Interaction scores with absolute value greater than two were considered as gene deletion enhancement (z-score_ L ≥ 2 or z-score_K ≤ − 2) or deletion suppression (z-score_L ≤ − 2 or z-score_K ≥ 2) of doxorubicin cytotoxicity. Gene deletion enhancement (e.g., mms1-Δ0) and suppression (e.g., vps54-Δ0) reveal functions that buffer or confer doxorubicin cytotoxicity, respectively. Doxorubicin-gene interaction profiles (selected if they contained L interaction scores with absolute value greater than 2, in either HLD or HLEG media) were analyzed by REMc and assessed for GO term enrichment (Fig. 1f). As a complement to clustering gene interaction profiles, functional enrichment was analyzed by GTA (see the "Methods" section), systematically querying all GO processes, functions, and components ( Fig. 1g and the "Methods" section) with respect to CPPs and Warburg status. Taken together, REMc and GTA reveal genetic modules that buffer doxorubicin, and how they are influenced by Warburg metabolism (Fig. 1h). Doxorubicin cytotoxicity was greater in HLEG than HLD media, evidenced by the reference strain being more growth inhibited ( Fig. 2a-l, Additional file 1: Figure S1). The "L" parameter was the most sensitive CPP, while K reported larger phenotypic effects (Fig. 2m, n) (Additional file 1: Figure S2). We noted a positive correlation between doxorubicin-gene interaction in HLEG and HLD; however, the interaction was mediaspecific and more abundant in the context of respiration, i.e., with HLEG media (Fig. 2o).
We compared our results with two prior studies of doxorubicin cytotoxicity in the yeast knockout collections [43,44]. One study was conducted in SC media with the haploid (BY4741) YKO library and identified 71 deletion enhancers of cytotoxicity [43]. A second study reported on the homozygous diploid (BY4743) YKO collection in YPD media, identifying 376 enhancers [44]. Overlap between these studies and ours is shown in Fig.  2p-r and in Additional file 7: Table S9-10. While many genes overlapped between the studies, differing results were also observed, possibly attributable to strain background, media conditions, and methods for scoring interactions [27,45]. To assess within-study reproducibility, we sub-cloned four colonies from glycerol stocks used in the first experiment and retested doxorubicin-gene interaction, revealing higher correlation and overall reproducibility within-study than between-study (Fig. 2s, t).

Identification of functional gene interaction modules
Gene interaction profiles were analyzed by REMc (Fig. 3, Additional file 1: Figure S3), as described previously (see the "Methods" section) [34]. GO TermFinder [35] was used to associate enrichment of biological functions with particular patterns of doxorubicin-gene interaction identified by REMc (Fig. 3, Table 1, Additional file 1: Figure S3, and Additional file 5: File C). We expect that clustering by gene interaction profiles should, in general, increase GO enrichment [34]. While true overall, as evidenced by the first two rounds of REMc revealed distinctive profiles of gene interaction in respiratory vs. glycolytic media (Additional file 1: Figure S3), later round clusters only sometimes exhibited greater GO term enrichment as (See figure on previous page.) Fig. 2 Q-HTCP provides cell proliferation parameters as phenotypes to quantify gene interaction. a, b Average pixel intensity and standard deviation for 768 reference strain cultures at indicated times after exposure to escalating doxorubicin concentrations in a HLD or b HLEG media. c, d Semi-log plots after fitting the data plotted above for c HLD or d HLEG to a logistic function (see Fig. 1d). e-l CPP distributions from data depicted in panels A-D for e-h HLD and i, j HLEG, including L (e, i), K (f, j), r (g, k), and (h, l) AUC. m, n Comparison of doxorubicin-gene interaction scores using the L vs. K CPP in the context of either m HLD or n HLEG media. Score distributions of knockout (YKO, green), knockdown/DAmP (YKD, red), and non-mutant parental (Ref, purple) strain cultures are indicated along with thresholds for deletion enhancement and suppression (dashed lines at ± 2). o Differential doxorubicin-gene interaction (using L as the CPP) for HLD vs. HLEG, classified with respect to Warburg metabolism as non-specific (NS), respiratory-specific (R), or glycolysis-specific (G) deletion enhancement (Enh) or deletion suppression GTA score revealed 129 GO terms, 39 of which were found by REMc/GTF (Table 2 and Additional file 6: Files A-C). GTA identifies functions of smaller GO terms, e.g., protein complexes. GTA with K interaction scores yielded only 35 GO terms (Additional file 6: File D), with only 3 being unique from GTA with L interaction; thus, we focused on L interaction for GTA analysis. Interactive scatter plots (html files in which points contain embedded information) were used to visualize significant GO terms from both REMc and GTA (Additional file 6: File B). GO termspecific heatmaps further aided visualization of relationships between genes and the GO terms (see Figs. 5, 6, 7, 8, 9 and 10 and Additional file 9) by systematically displaying, for all genes attributed to a parent term and its children, uniformity vs. pleiotropy of interaction effects across different conditions.
In summary, we used REMc, GTA, and GO termspecific scatterplots and heatmaps to discover genetic modules that alternatively buffer (i.e., deletion enhancing) or confer (i.e., deletion suppressing) doxorubicin cytotoxicity and to determine whether the Warburg-transition exerts influence on their effects (Fig. 4). Characterization of Warburg-differential, doxorubicin-gene interaction profiles. a The union of enhancers (L z-score > 2) or suppressors (L zscore ≤ 2) from the HLD and HLEG analyses totaled 2802 gene interaction profiles that were subjected to REMc (see the "Methods" section). b, c The column order is the same for all heatmaps; "+" indicates doxorubicin-gene interaction and "−" indicates "shift" (K 0 ; see the "Methods" section). Interactions by K are negative (brown) if enhancing and positive (purple) if suppressing, while the signs of interaction are reversed for L (see the "Methods" section). The heatmap color scale is incremented by twos; red indicates no growth curve in the absence of doxorubicin. b First round cluster 1-0-7 has a gene interaction profile indicative of HLEG-specific deletion enhancement. c Second round clusters (2-0.7-X) are ordered left to right by strength of influence. d The pattern of distributions for the different doxorubicin-gene interaction scores ("+" columns only from panel c) summarizes respective clusters from panel c. Deletion enhancement is considered to be qualitatively stronger if observed for K in addition to L The table headers are defined as follows: For the column, "Media," "Resp," "Glyc," and "Both" refer to whether the gene interaction type observed for the REMc cluster associated with the term was prominent in HLEG, HLD, or both media (see Additional file 1: Figure S3). For the column, "INT," "Enh," and "Sup" indicate deletion-enhancing or deletion-suppressing. The column "GTA" refers to GO term average. The column "Clust" refers to REMc ID

Warburg transition-dependent doxorubicin gene interaction modules
Despite both longstanding and renewed interest in the importance of the Warburg effect to oncogenesis, whether it influences cellular responses to chemotherapeutic agents is unknown. Thus, YKO/KD strains that display differential resistance to doxorubicin under respiratory (non-fermentable HLEG media) vs. glycolytic (fermentable HLD media) media provide new insight both into genes that function in pathways that may buffer or promote doxorubicin cytotoxicity and whether such pathways are potentially influenced by the Warburg transition. The phenomic assessments described below systematically quantify the contribution of each and every individual yeast gene to doxorubicin cell proliferation phenotypes. In addition, the influence of the Warburg effect on this network is detailed by differential doxorubicin-gene interaction on glycolytic (HLD) vs. respiratory (HLEG) media. In addition to direct implication of cellular pathways by the identification of genes annotated to their functions, functional enrichment among all genes was ascertained by GO term enrichment in gene clusters having similar gene-doxorubicin interaction profiles (REMc/GTF) or by systematic analysis of the average gene interaction value in Gene Ontology terms (GTA).

Chromatin organization and histone modification
REMc/GTF and GTA identified several chromatin-related processes that buffer doxorubicin toxicity in a respirationspecific manner, including DNA replication-independent nucleosome assembly, histone exchange, histone deacetylation, and histone methylation (Figs. 5 and 6).
(i) DNA replication-independent nucleosome assembly (HIR complex) REMc/GTF identified the HIR complex (HIR1-3 and HPC2), which functions as a histone chaperone in chromatin assembly and disassembly, in cluster 2-0.7-2 (Additional file 1: Figure S3 and Table 1) [46]. Along with Asf1 and Rtt106, the HIR complex is involved in DNA replication-independent (i.e., RNA transcriptional) histone deposition and regulates transcription of three of the four histone genes [46][47][48]. Furthermore, genes encoding for HTA1/HTB1, HHT1/2, and HHF1/2 were also respiratory-specific deletion enhancers. Asf1 and Rtt106 function in nucleosome assembly in both DNA replication and DNA replication-independent contexts. Asf1, which functions in the Rad53-dependent DNA damage response [49], enhanced doxorubicin toxicity in both respiratory The table headers are defined as follows: "gtaSD" refers to the standard deviation of GTA, "REMc related" refers to an REMc cluster ID if GTA-identified term was also found by REMc/GTF, and "p value" reports results from REMc/GTF. See Table 1 for other header definitions and glycolytic media, like other DNA repair genes (see below). In further contrast, genes associated with replication-dependent nucleosome assembly (RLF2, CAC2, MSI1) by the chromatin assembly factor complex, CAF-1, [50] were HLD-specific suppressors (Fig. 5a, b).
Prior studies have reported enhanced doxorubicin cytotoxicity due to nucleosome disassembly and "chromatin trapping" by the FACT complex, referring to binding and resulting damage to disassembled chromatin in the context of doxorubicin exposure [20]. POB3- Representative GO terms are listed, which were identified by REMc/GTF (orange), GTA (purple), or both methods, for HLD (left, red), HLEG (right, blue), or both media types (black), and for enhancement (above dashed line) or suppression (below dashed line) of doxorubicin cytotoxicity. Distance above or below the horizontal dashed line indicates the GTA value for terms identified by REMc or the GTA score if identified by GTA (see the "Methods" section). See Additional files 5 and 6, respectively, for all REMc and GTA results DAmP, the only member of the FACT complex represented in the YKO/KD library, resulted in suppression of doxorubicin cytotoxicity (Fig. 5b), presumably by suppressing its effect of trapping and damaging disassembled chromatin.
(ii) Histone exchange (Swr1 complex) The Swr1 complex (enriched in cluster 2-0.7-2) uses ATP hydrolysis to replace the H2A nucleosome with the H2AZ variant [51]. Swr1 complex genes showing respiration-specific buffering of doxorubicin toxicity Respiration increases the role for chromatin organization in buffering doxorubicin toxicity. a GO term-specific heatmaps for chromatin organization and its child terms (indicated by arrows) clarify related but distinct biological functions that buffer doxorubicin, with respect to Warburg status. b, c L-based doxorubicin-gene interaction scores associated with GO terms that were enriched in cluster 2-0.7-2. Dashed lines indicate z-score thresholds for enhancers (> 2) and suppressors (≤ 2). Sub-threshold gene interaction values are plotted, but not labeled Distinct histone modifications differentially influence doxorubicin cytotoxicity. a Rpd3L and Rpd3S complexes exert strong HLEG-specific doxorubicin-enhancing influence relative to other Sin3-type histone deacetylases and the HDA1 complex. b In contrast to histone deacetylation (panel a), histone acetylation exhibits deletion enhancement that is Warburg-independent. c Histone H3K4 methylation by the Set1C/COMPASS complex, which requires histone mono-ubiquitination of H2B by the Bre1/Rad6 complex, is opposed by Jhd2, a histone H3K4 demethylase. The respiration-specific deletion-enhancing interactions suggest the Warburg transition can protect tumors promoted by certain types of chromatin deregulation from doxorubicin included RVB1, SWC3, SWC5, ARP6, SWR1, VPS71, and VPS72 (Fig. 5c). Accordingly, the H2AZ variant, Htz1, which is enriched at most gene promoters in euchromatin [52][53][54], was also an HLEG-specific deletion enhancer. The Swr1 complex is recruited for repair of dsDNA breaks, where the H2AZ variant is incorporated [55]; however, the interaction profile of the Swr1 complex more closely resembles other respiratory specific enhancers involved in transcriptional regulation, whereas dsDNA-break repair by homologous recombination buffered doxorubicin toxicity independent of Warburg context (see cluster 1-0-6 from Additional file 1: Figure  S3 and Table 1). The Swr1 complex can also inhibit subtelomeric spread of heterochromatin by impeding SIRdependent silencing [56]. Consistent with knockout of Swr1 promoting silencing and having a deletionenhancing effect, deletion of SIR1, SIR3, or SIR4 (which disrupts chromatin silencing) also exerted respiratoryspecific suppression of doxorubicin toxicity (Fig. 5c).
The SAS acetyltransferase complex was deletion suppressing; SAS2 and SAS5 were HLEG-specific, and SAS4 was HLD-specific (Fig. 6b). The Sas2 acetyltransferase complex creates a barrier against spread of heterochromatin at telomeres by opposing Sir protein deacetylation via effects on histone H4K16 [61]. The deacetylating SIR proteins (SIR1, SIR3, SIR4) were also HLEG-specific suppressors (Fig. 5c), suggesting dynamic regulation of telomeric histones (not simply acetylation or deacetylation), or perhaps a function of Sas2 acetyltransferase that is independent of SIR protein functions, confers doxorubicin cytotoxicity in respiring cells.
Based on the findings above, it appears buffering of doxorubicin-mediated cellular toxicity by some transcriptionassociated chromatin regulators is alleviated by the transition from respiratory to glycolytic metabolism, whereas buffering by those more associated with DNA repair is relatively independent of metabolic context.

Mitochondrial functions
The greater number of deletion-enhancing doxorubicingene interactions in HLEG media, relative to HLD media (Fig. 2o), caused us to examine genes annotated to mitochondrial function more systematically. Many mitochondrial gene deletion strains grew very poorly on HLEG media and exhibited reduced carrying capacity on HLD media, as would be associated with petite mutants. Completely respiratory-deficient mutants clustered together in 1-0-0; however, many mitochondrial mutants maintained some or all respiratory capacity. For example, the respiratory chain complex III assembly and protein import into mitochondrial matrix terms were enriched in deletion-enhancing clusters, 1-0-7 and 1-0-8 (Table 1 and Additional file 1: Figures S3-4). Some of these strains appeared respiratory sufficient, yet the genes were required to buffer doxorubicin cytotoxicity under respiratory conditions. For example, evolutionarily conserved genes functioning in complex IV assembly (RCF1/YML030W and COA6) reached carrying capacity on HLEG media, yet exerted strong deletion enhancement of doxorubicin growth inhibition (Additional file 1: Figure S4A). In contrast, HLEG-specific deletionenhancing complex IV assembly components (COA2, CMC1, RCF2) and complex III assembly genes (FMP25, FMP36, QCR9, CBP4) were either not conserved in humans or exhibited strong respiratory defects (in absence of doxorubicin) (Additional file 1: Figure S4A-B). These findings appear to establish relevance of the yeast model to studies in cardiomyocytes, for which it was reported that doxorubicin toxicity is exacerbated by depletion of cytochrome c or cardiolipin, leading to reduced workload capacity, and accelerated aging [73,74]. Likewise, functionally conserved (TOM70, TIM10, TIM17, TIM23, and MGR2) and yeast-specific (TOM6 and TOM7) genes in protein import into mitochondrial matrix buffered doxorubicin cytotoxicity (Additional file 1: Figure S4C-E), perhaps relating to increased oxidative stress [75], which also enhances doxorubicin toxicity in mammalian cells [8,11].

(i) Protein folding in endoplasmic reticulum (ER membrane protein complex)
The ER membrane complex (EMC1-6, Fig. 7a) functions in protein folding in the ER [85] and together with the ER-mitochondria encounter structure (ERMES), the EMC enables ER-mitochondria phosphatidylserine transfer and tethering [86]. The EMC physically interacts with the mitochondrial translocase of the outer membrane (e.g., TOM5, 6, 7, 22, 70; described above) for the process of ER-mitochondria phosphatidylserine transfer [86]. The shared respiratory-specific, deletion-enhancing profiles suggest cooperative functions of the EMC and mitochondrial outer membrane translocase (Additional file 1: Figure S4D) in buffering doxorubicin cytotoxicity. In contrast to the EMC, genes involved in the ERMES complex (1-0-0; Additional file 5: File B-C) were essential for respiration, and thus, their influence on doxorubicin cytotoxicity could not be addressed with knockout mutants in HLEG media.

DNA topological change
DNA topological change, which refers to remodeling the turns of a double stranded DNA helix, was enriched in cluster 2-0.8-0 (Additional file 1: Figure S3 and Table 1). Representative genes were SGS1, TOP1, RFA1, RMI1, TOP3, MMS4, and MUS81 (Fig. 7b). Types I and II topoisomerases resolve supercoiling during replication and transcription [99,100]. Top1 is a type IB topoisomerase, which relaxes positive and negative supercoils [101,102], compared to Top3, a type IA topoisomerase that specifically acts on negative supercoiling [103]. The Mms4-Mus81 endonuclease has overlapping functions with Top3 and Sgs1 in DNA repair [104]; however, their respective influences on doxorubicin toxicity were quantitatively distinct in both respiratory and glycolytic contexts, with a greater requirement for the MMS4/MUS81 than SGS1, TOP3, RFA1, and RMI1 (Fig. 7b); the latter four, functioning together for decatenation and unknotting of dsDNA [105].

GTA reveals additional biological functions that buffer doxorubicin toxicity
GTA is a method complementary to REMc/GTF for discovering GO functions in Q-HTCP-derived phenomic data. Whereas GTF scores GO enrichment among genes within a cluster, GTA is independent of clustering and systematically assesses all genes in every GO term for interaction (see the "Methods" section).
GTA revealed 71 respiratory-specific deletion-enhancing GO terms, 24 of which were also found by REMc/GTF (see Additional file 6: File A). Strong enhancing terms (GTA value > 10) with functions relatively distinct from those identified above by REMc were tRNA (m1A) methyltransferase complex, MUB1-RAD6-UBR2 ubiquitin ligase complex, malonyl-CoA biosynthetic process, pyridoxal 5'phosphate salvage, maintenance of transcriptional fidelity during DNA-templated transcription elongation from RNA polymerase II promoter, RNA polymerase II transcription corepressor activity, pyruvate dehydrogenase activity, and eukaryotic translation initiation factor 2 complex (Fig. 7c). Most terms identified by GTA consisted of 2-3 genes and did not necessarily cluster together by REMc.

Respiration-specific gene deletion suppression of doxorubicin cytotoxicity
Clusters exhibiting respiration-specific gene deletion suppression revealed GO term enrichment for regulation of fatty acid beta-oxidation (cluster 2-0.3-1) and translation reinitiation (cluster 2-0.3-5) (Additional file 1: Figure S3 and Table 1). By GTA analysis, the EKC/ KEOPS complex and spermine biosynthetic process were additionally found to confer HLEG-specific deletion suppression ( Fig. 7d and Table 2).
Regulation of fatty acid beta-oxidation ADR1, OAF1, and PIP2 were grouped together in cluster 2-0.3-1 (Additional file 1: Figure S3 and Table 1), displaying HLEG-specific gene deletion suppression (Fig.  7d). The Pip2-Oaf1 complex binds to oleate response elements and, along with ADR1, regulates transcription of peroxisomal genes [106,107]. Doxorubicin inhibits betaoxidation of long-chain fatty acids in cardiac tissues, which is reversed by supplementing with propionyl-L-carnitine, and alleviates effects of doxorubicin cardiotoxicity [108]. Thus, the yeast model may be informative for investigating related gene networks in greater depth.

Translation reinitiation
In the respiratory-specific deletion suppressing cluster 2-0.3-5 (Additional file 1: Figure S3), TMA20, TMA22, and TIF34 represented enrichment for translation reinitiation, which is necessary after termination of short upstream open reading frames (uORFs) [109] (Fig. 7d). Some uORFs function in translational regulation of a downstream protein; for example, GCN4 expression is regulated in response to amino acid starvation [109]. However, using the Welsh two sample t test, we found no significant difference in means of interaction scores between the distribution of proteins regulated or not by uORFs [110] (p value = 0.8357) (Additional file 7: Table S12).

Spermine biosynthetic process
Loss of spermine biosynthesis, specifically SPE2 (S-adenosylmethionine decarboxylase) and SPE4 (spermine synthase), suppressed doxorubicin toxicity in HLEG media (Fig. 7d). The pathways of polyamine metabolism and their physiologic effects on cancer are complex [111,112], and although our data suggest spermine metabolism contributes to doxorubicin cytotoxicity, how this occurs mechanistically and specifically in respiring cells awaits further study [113].
Glycolysis-specific gene deletion enhancement of doxorubicin cytotoxicity HLD-specific deletion enhancement of doxorubicin cytotoxicity could represent lethal vulnerabilities that emerge when a tumor undergoes the Warburg transition. In this regard, several genes, but few enriched GO terms, were identified by REMc (Additional file 1: Figure  S3, clusters 1-0-5, 2-0.3-0, and 2-0.2-2; Additional file 5: File A). Ribonucleoprotein complex subunit organization was suggested (Table 1); however, the term-specific heatmap revealed doxorubicin-gene interaction within this cellular process to be pleiotropic (Additional file 1: Figure S7).

Glycolysis-specific deletion-enhancing terms identified by GTA
GTA analysis revealed HLD-specific deletion-enhancing genes encoding the Cul4-RING E3 ubiquitin ligase, the Dom34-Hbs1 complex, and the Ubp3-Bre5 deubiquitinase. GDP-Mannose Transport and dTTP biosynthesis were also revealed ( Fig. 8a and Additional file 6: File A). SOF1, HRT1, and PRP46 were computationally inferred to form the Cul4-RING E3 ubiquitin ligase complex [118]. Yeast Sof1 is an essential protein that is required for 40s ribosomal biogenesis, and overexpression of its human ortholog, DCAF13/WDSOF1, is associated with aggressive tumors and poorer survival in hepatocellular carcinoma [119]. DOM34/PELO and HBS1/HBS1L facilitate recycling of stalled ribosomes by promoting dissociation of large and small subunits through a process called no-go decay [120][121][122]. Knockdown by siRNA of either WDSOF1 or HBS1L was synthetic lethal in a KRAS-driven tumor model [123]. The Ubp3-Bre5 deubiquitination complex regulates anterograde and retrograde transport between the ER and Golgi [124,125]. Vrg4 and Hvg1 transport GDP-mannose into the Golgi lumen for protein glycosylation [126,127]. Reduced dTTP pools, evidenced by CDC8/DTYMK and CDC21/TYMS, can increase doxorubicin cytotoxicity in cancer cell lines [128]. The human homologs of UBP3, CDC8, and CDC21 were identified in genome-wide siRNA synthetic interaction studies in cancer cell line models [129][130][131].
For several examples above, like SOF1/DCAF13, genes could be targeted as both a driver of the tumor and as a sensitizer to doxorubicin. To systematically identify all candidate vulnerabilities specific to glycolytic tumor cells (not constrained by GO enrichment), we filtered the overall data set, limiting the list to genes with human homologs and to YKO/KD strains that were growth sufficient (low shift on HLD) (Additional file 1: Figure S8). The human homologs, along with functional descriptions, are provided in Additional file 10: Table S13.

Histone deubiquitination
Histone deubiquitination was identified by GTA and includes SUS1, SGF11, SGF73, UBP8, and SEM1 (Fig. 8b); all except SEM1 are part of the DUBm complex, which mediates histone H2B deubiquitination and mRNA export [132]. Loss of histone H2B ubiquitination resulting in HLEG-specific enhancement (Fig. 6c) is consistent with loss of the DUBm deubiquitinase being suppressing. Together, they implicate regulation by histone ubiquitination as a mechanism of doxorubicin response. The human homologs of UBP8, USP22, and USP51 were identified in an RNAi screen for resistance to ionizing radiation [133].

Meiotic chromosome condensation
SMC2, SMC4, YCG1, and YCS4 constitute the nuclear condensin complex, which functions in chromosome condensation and segregation. The condensin complex associates with chromosomal sites bound by TFIIIC and the RNA Pol III transcription machinery [137], where it facilitates clustering of tRNA genes at the nucleolus [138] (Fig. 8b). The condensin complex has been suggested as a potential therapeutic target for cancer [139], and human homologs YCG1/NCAPG2, YCS4/NCAPD2, and SMC4/SMC4 are synthetic lethal with the Ras oncogene [123].

Warburg transition-independent doxorubicin geneinteraction modules:
Since cancers may have both respiratory and glycolytic cell populations, targeting Warburg-independent interactions could be especially efficacious, as described below.

DNA repair
Warburg-independent, deletion-enhancing pathways included homologous recombination and break-induced replication repair (Fig. 9a), along with the Ino80 complex (Fig. 9b), the latter explained by its role of histone acetylation in the recruitment of DNA repair machinery to dsDNA break sites [51]. The Ino80 complex influences doxorubicin response in fission yeast [140,141], further suggesting evolutionary conservation of this interaction, and thus potential relevance to mammalian systems [142]. DNA repair pathways, such as those involving RAD52 and INO80, are evolutionarily conserved, involved in genome instability and tumorigenesis [143], and predictive of therapeutic response in some cancers [144], thus representing potential tumor-specific biomarkers for chemotherapeutic efficacy.

Media-independent deletion suppression
Loss of genes functioning in processes that augment doxorubicin toxicity results in suppression of its growth inhibitory effect. This was suggested in both respiratory and glycolytic contexts for sphingolipid homeostasis, telomere tethering at nuclear periphery, and actin cortical patch localization (Additional file 1: Figure S3, clusters 2-0.4-1 and 2-0.3-3). Conversely, their overexpression in cancer could potentiate toxicity and therapeutic efficacy.
Since the GARP genes and fatty acid elongase activity genes function together in sphingolipid metabolism, we searched all genes annotated to this term and found Fig. 9 Warburg-independent deletion enhancement of doxorubicin cytotoxicity. Gene interaction profiles showing deletion enhancement in both respiratory and glycolytic context included: a double-strand break repair via homologous recombination, and its child terms (indicated by arrows), and b the Cul8-RING ubiquitin ligase, Ino80 complex, Lst4-7 complex, and MCM complex other media-independent suppressors to include TSC3, LIP1, SUR1, SUR2, IPT1, and SKN1 (Fig. 10a). Doxorubicin treatment induces accumulation of ceramide [12,13], which mediates anti-proliferative responses and apoptosis in yeast and human and appears to mechanistically underlie the influence of this gene group [153] (Additional file 1: Figure S10). These findings were further supported by the deletion enhancer, SCH9, which negatively regulates ceramide production by inducing ceramidases and negatively regulating ISC1 (Fig. 10a) [154]. Multidrug-resistant HL-60/MX2 human promyelocytic leukemia cells are sensitized to doxorubicin by N, N-dimethyl phytosphingosine [155].
Taken together, the model provides genetic detail regarding how disruption of sphingolipid metabolism increases resistance to doxorubicin and that this occurs in a Warburg-independent manner, seemingly by reducing apoptosis associated with doxorubicin-induced ceramide overproduction [12,156,157].
Telomere tethering at nuclear periphery Enrichment for telomere tethering at nuclear periphery in cluster 2-0.4-1 was comprised of NUP60, NUP170, MLP1, and ESC1. Although growth deficient on HLD media, NUP84, NUP120, and NUP133 also exerted deletion suppression in HLEG (Fig. 10b). Fig. 10 Warburg-independent deletion suppression of doxorubicin cytotoxicity. Doxorubicin-gene interaction profiles and L-interaction plots for genes associated with deletion suppression in HLEG or HLD media, including a cellular sphingolipid homeostasis, along with its parent term, lipid homeostasis, and related term sphingolipid metabolism and b actin cortical patch localization and telomere tethering at nuclear periphery Nuclear pore functions include coordinating nuclearcytoplasmic transport and localizing proteins and/or chromosomes at the nuclear periphery, which contributes to DNA repair, transcription, and chromatin silencing [158]. Thus, deletion of nuclear pore genes could influence doxorubicin resistance by multiple potential mechanisms involving altering chromatin states, transcriptional regulation, maintenance of telomeric regions, and DNA repair. Doxorubicin-gene interaction profiles for all nuclear pore-related genes are provided in Additional file 1: Figure S11A.

Respiratory-deficient doxorubicin-gene interaction modules
From cluster 1-0-0, we noted that respiratory deficient YKO/KD strains (those not generating a growth curve on HLEG) also had low K and/or increased L "shift" values on HLD, as would be expected of petite strains [164]. Strains in this category tended to display deletion enhancement (Additional file 1: Figure S3) and function primarily in mitochondrial processes (Additional file 5: File C; see GO enrichment for cluster 1-0-0 and derivative clusters), including mitochondrial translation, mitochondrion-ER tethering, protein localization into mitochondria, mitochondrial genome maintenance, respiratory chain complex assembly, and proton transport. Compromise of mitochondrial respiration leading to sensitization of cells to doxorubicin is of interest given recent findings that some glycolytic cancers are respiratory deficient [165,166].

Phenomics-based predictions of doxorubicin-gene interaction in cancer cell lines
We next investigated how measures of enhancing and suppressing interactions from the yeast phenomic model could serve to predict and prioritize candidate effectors of cancer cell line sensitivity and transcriptomic data [167,168]. Differential gene expression, by itself, has been clearly shown to be a poor predictor of whether protein function affects proliferative response to a particular drug [169]. Yeast doxorubicin-gene interaction was matched by homology to differential gene expression in doxorubicin-sensitive cancer cell lines, using PharmacoGx [39] and biomaRt [40,41]) in conjunction with the GDSC1000 [170,171] or gCSI [172,173] databases (Fig. 11). Differential gene expression analysis was performed for individual tissues and for data aggregated for all tissues. Yeast gene deletion enhancers were used to predict causality for human homologs underexpressed in doxorubicin-sensitive cancer cell lines, termed "UES." Conversely, yeast gene deletion suppressors were matched to human homologs overexpressed in doxorubicin sensitive cells, termed "OES" (Additional file 11).
There was higher correspondence for yeast gene interactions with differential gene expression in the gCSI vs. the GDSC database, partially explained by the greater number of genes reported in gCSI than GDSC (Additional file 12). Such differences are possibly due to sensitivity for measuring gene expression arising from distinct platforms used to measure gene expression and cell cytotoxicity, and different sample sizes in the respective data (https://pharmacodb.pmgenomics.ca/ drugs/273). The gCSI data reported more UES and OES genes than GDSC (Additional file 11: Files B-E and Additional file 12) and consequently greater overlap with the yeast phenomic data.
As detailed in Tables 3 and 4, we expanded the analysis to genes representative of GO term enrichments revealed by the yeast phenomic model having human homologs differentially expressed across all cancer tissues, but without restricting by Warburg-independence or gCSI/GDSC co-evidence. Results for individual tissues are also provided in Additional file 11: File A. We considered whether correlations between doxorubicin-gene interaction in yeast with pharmacogenomic results could be enhanced by particular combinations of the data from this study and prior yeast studies. To briefly summarize, there was a higher correlation for deletion enhancers under respiratory (27%) than glycolytic (19%) conditions, and considering both media, there was higher correspondence of deletion suppressors (43%) than deletion enhancers (25%). The complete data are provided in Additional file 7 and Additional file 11 to enable comparisons between data sets, and example comparisons are summarized in Additional file 12.

Deletion enhancers with UES homologs
Concordance between deletion-enhancing doxorubicingene interaction in yeast and UES observed for the corresponding human homologs in cancer cells suggests synergistic targets for inhibition and biomarkers of genetic vulnerabilities that may increase therapeutic efficacy for doxorubicin (Table 3 and Fig. 11c). Many of these genes function in processes identified by the yeast phenomic model (Tables 1, 2, 3, and 4) and have annotated roles in cancer biology. Doxorubicin-enhancing interactions that were UES in both gCSI and GDSC included ACTL6B, identified as a candidate tumor suppressor gene in primary hepatocellular carcinoma tissue [174]; PPM1L, which regulates ceramide trafficking at ER-Golgi membrane contact sites [175] and exhibits reduced expression in familial adenomatous polyposis [176]; RPS6KB2, which was UES in the breast, ovary, and bone in gCSI, while RPS6KA1, A2, A5, and A6 were UES in select tissues in both databases (Additional file 11: File A); SEC11/SEC11C, which is upregulated in response to hypoxia in non-small cell lung cancer tissue [177] and for which deletion enhancement was stronger in glycolytic media (Additional file 1: Figure S8); SEC7/ARFGEF2 (alias BIG2) which exhibits increased gene and protein expression in pancreatic cancer [178], and shRNA knockdown of ARFGEF2 can reduce Burkitt's lymphoma cell survival [179].
MGR2/ROMO1 is involved in protein import into the mitochondrial matrix and overexpression of ROMO1 has been associated with poor prognosis in colorectal [181] and non-small cell lung cancer patients [182]. MAK3/ NAA30, a component of the NatC complex (Fig. 8a), induces p53-dependent apoptosis when knocked down in cancer cell lines [183]. MUS81 knockdown enhances sensitivity of colon cancer lines to epirubicin (doxorubicin analog) [184], cisplatin, and other chemotherapy agents by activating the CHK1 pathway (Fig. 7b) [180].
The glycolysis-specific deletion enhancer, SFA1, has seven human homologs, of which three (ADH4, ADH1A, and ADH6) were UES in gCSI data (Additional file 1: Figure S8). High expression of ADH1A or ADH6 was predictive of improved prognosis for pancreatic adenocarcinoma [185], and high expression of ADH1A or ADH4 had improved prognosis for nonsmall cell lung cancer [186]. The ERG13 homolog, HMGCS1, has been suggested as a synthetic lethal target for BRAF V600E -positive human cancers [187], and HMGCS2 plays a role in invasion and metastasis in colorectal and oral cancer [188]. These data suggest doxorubicin treatment may have anti-tumor efficacy specifically in glycolytic tumors with reduced expression of SFA1 and ERG13 homologs.
(See figure on previous page.) Fig. 11 Use of the yeast phenomic model to predict doxorubicin-gene interaction in cancer cells. a BiomaRt was used to assign yeast-human gene homology for the GDSC and gCSI datasets. b PharmacoGx was used to retrieve differential gene expression for doxorubicin sensitive cell lines from the gCSI and GDSC databases, searching data from individual tissues or across data aggregated from all tissues. Human genes that are underexpressed in doxorubicin sensitive cell lines (UES) with yeast homologs that are deletion enhancers are predicted to be causal in their phenotypic association. Similarly, human genes that are overexpressed in doxorubicin sensitive cancer cell lines (OES) would be predicted to be causal if the yeast homolog was a deletion suppressor in the phenomic dataset. c, d Boxes inside of Venn diagrams indicate the genes for which gene interaction profiles are shown in the heatmaps below. Gene names are to the right of heatmaps, with blue labels indicating genes identified in both the GDSC and gCSI databases and black labels indicating genes found only in the gCSI dataset. The category of homology (see panel a) is indicated in the left column of each heatmap. c Deletion enhancement by yeast genes predicts human functions that buffer doxorubicin cytotoxicity, and thus, reduced expression of homologs in cancer cell lines is predicted to increase doxorubicin sensitivity. d Deletion suppression by yeast genes predicts functions that mediate cytotoxicity and is shown for human homologs having significant association of overexpression in cancer cell lines with increased doxorubicin sensitivity. e, f Genes representing enhancing or suppressing modules from REMc or GTA that are e UES or f OES in at least one of the two databases. Red labels indicate genes found only in the GDSC database. Additional file 11 reports all results from the analysis described above, including assessment of individual tissues

Deletion suppressors with OES homologs
Genes that promote toxicity of a drug could lead to increased sensitivity if overexpressed in cancer cell lines or, correspondingly, deletion suppression by yeast phenomic analysis. Choosing chemotherapeutic agents for patients based on their tumors exhibiting high expression of genes known to increase sensitivity represents a targeted strategy to increase therapeutic efficacy and could be particularly effective if the sensitizing overexpressed genes happen to also be drivers [189]. Human genes that were OES and homologous to yeast deletion suppressors are highlighted in Table 4 and Fig. 11d. ARP2/ACTR2 is a member of the Arp2/3 protein complex (see Additional file 1: Figure S11C), and silencing of the Arp2/3 protein complex reduces migration of pancreatic cancer cell lines [190]. EEF2 protein is overexpressed in multiple cancer types, where shRNA knockdown inhibits growth [191]. CDK7 overexpression in breast [192,193] and gastric [194] cancer is predictive of poor prognosis. RPL34 overexpression promotes proliferation, invasion, and metastasis in pancreatic [195], non-small cell lung [196], and squamous cell carcinoma [197], while RPL32 was also overexpressed in a prostate cell cancer model [198]. In contrast to Rps6k family members being UES/deletion enhancing, Rps6 was OES/deletion suppressing in ovarian tissue. RPS6 overexpression portends reduced survival for patients with renal carcinoma [199] and hyperphosphorylation of Rps6 confers poor prognosis in non-small cell lung cancer [200]. Overexpression of UBE2D1 is associated with decreased survival in lung squamous cell carcinoma tissue [201], and numerous additional ubiquitin-conjugating enzyme family members were OES in analysis of individual tissues (Additional file 11: File A). We expanded the analysis, similar to the way described above for the deletion enhancers, by relaxing the matching criteria in order to identify additional deletion suppressing pathways revealed by the yeast model (Additional file 11). The extended analysis identified yeast-human conserved functions in metabolism (SPE2, SPE4, VPS53, ELO2, ELO4), histone demethylation (JHD1, JHD2), translation reinitiation (TMA22, TIF32), the condensin complex (YCG1, YCS4, SMC2), and telomere tethering at the nuclear periphery (NUP170) ( Table 4 and Fig. 11f). SPE2/AMD1 is required for spermidine and spermine biosynthesis, and upregulation of AMD1 by mTORC1 rewires polyamine metabolism in prostate cancer cell lines and mouse models [202]. VPS53, a component of the GARP complex involved in sphingolipid homeostasis, is a tumor suppressor in hepatocellular carcinoma [203][204][205]; additionally, transfection with VPS53 transcript induces apoptosis and sensitizes cervical cancer cells to doxorubicin [206], consistent with our model. Inhibition of ELOVL6 For column "DB": "gCSI," "GDSC", or "Both" indicate UES in the gCSI, GDSC, or both databases. Column " Fig." refers to specific figures. Columns "HLD L|K" and "HLEG L|K" contain the L and K interaction scores for HLD and HLEG media, respectively. "GDSC pval" and "gCSI pval" refer to the significance of differential gene expression in the respective databases. "Ref" refers to relevant literature citations. "H" refers to homology type: "1," "2," and "3" indicate 1:1, 1:many, and many:many, respectively (homologous to yeast ELO2 and ELO3) in mice reduces tumor growth and increases survival [207]. The histone demethylase, JHD1/KDM2B, is overexpressed in pancreatic cancer [208] and is associated with poor prognosis in glioma [209] and triple-negative breast cancer [210]. A second homolog, JHD2/JARID2, is required for tumor initiation in bladder cancer [211]. The yeast model also predicts causality underlying OES associated with genes involved in translation reinitiation, TMA22/DENR (translation machinery associated) and TIF32/EIF31. DENR-MCT-1 regulates a class of mRNAs encoding oncogenic kinases [212][213][214], and its overexpression in hepatocellular carcinoma is associated with metastasis [215]. TMA22/DENR also exerts evolutionarily conserved influence on telomeric function and cell proliferation [216]. YCG1/NCAPG and SMC2/SMC2 are components of the condensin complex, which are overexpressed in cancer [139]. NUP170/NUP155, which functions in telomere tethering at the nuclear periphery (Fig. 10b), is hyper-methylated in association with breast cancer [217,218], where its reduced expression contributes to a signature for bone metastasis [219].

Discussion
A relatively comprehensive and quantitative phenotypic model of the gene-drug interaction network underlying the cell proliferative response to doxorubicin is presented. We demonstrate how a yeast phenomic model is predictive of genetic vulnerabilities to a cytotoxic agent that arise in cancer cells due to differential gene expression. Although an unbiased, experimental model of genetic interaction is largely descriptive with respect to molecular mechanisms, we propose it to nevertheless be a unique and powerful resource to model genotypephenotype networks at the cellular and organismal level. Genes that promote or buffer a cell proliferative response to defined perturbations report on molecular networks that potentially influence a variety of phenotypes [220,221]. For disease phenotypes involving evolutionarily conserved cellular processes, such as cell cycle and DNA repair (e.g., cancer) or folding of biogenesis of polytopic membrane proteins (e.g., cystic fibrosis), phenomic models can account for biological context and potential genetic modifiers of disease [25,28], serving to further inform the disease literature while also generating in an unbiased experimental manner new hypotheses that can be tested across species [26,[222][223][224][225].
To create the doxorubicin-gene interaction model for cancer, we used advances in Q-HTCP for phenomic analysis of the YKO/KD library and developed customized GO tools to mine cancer pharmacogenomics data via homology information. The information resulting from the model generates new hypotheses, further integrating the yeast and cancer literature, thus providing new insights for investigators with complementary expertise to further clarify the mechanistic basis of the observed gene-drug interaction network [27]. We propose that it may be further possible to leverage yeast phenomics to advance precision oncology models somewhat independently of molecular mechanisms; i.e., if genetic interactions that determine the response to cytotoxic chemotherapy can be successfully predicted, it may not be necessary to detail all interactions mechanistically in order for For column "DB": "gCSI," "GDSC," or "Both" indicate UES in the gCSI, GDSC, or both databases. Column " Fig." refers to specific figures. Columns "HLD L|K" and "HLEG L|K" contain the L and K interaction scores for HLD and HLEG media, respectively. "GDSC pval" and "gCSI pval" refer to the significance of differential gene expression in the respective databases. "Ref" refers to relevant literature citations. "H" refers to homology type: "1," "2," and "3" indicate 1:1, 1:many, and many:many, respectively the model to be a useful tool for decision-making in precision oncology. Many genes are implicated in oncogenesis and in chemotherapeutic response, with varying degrees of tissue-specific influence and yeast-human homology. The ability to assess mutation, differential gene expression, and other molecular correlates of cancer and chemotherapeutic efficacy is growing, but the direct assessment of drug-gene interaction (i.e., phenotypic/cell proliferative responses) remains a challenge due to the complex genetics and tissuespecific aspects of cancer. In contrast to this daunting complexity, yeast is a single-cell eukaryotic organism that is uniquely amenable to precise and genome-wide measures of drug-gene interaction, and from which fundamental contributions to our understanding of human disease are well established [226][227][228][229][230]. We show that yeast, which naturally exhibits aerobic glycolysis, can be used to explore the potential of the Warburg effect to influence the anticancer efficacy of doxorubicin, and thus potentially other chemotherapeutic agents [1,231]. From an unbiased systems perspective, we observed that a less extensive genetic network is required to buffer doxorubicin in glycolytic vs. respiring cells. The doxorubicin-gene interaction network suggested GOenriched genetic vulnerabilities that may occur in cancer cells due to genomic instability, including defective chromatin regulation, protein folding and modification, mitochondrial function, and DNA topology; while impeding the ability to buffer doxorubicin toxicity in a respiratory context, such vulnerabilities can be relieved of by the Warburg transition to glycolytic metabolism. Also specific to respiratory conditions, the yeast model predicts that doxorubicin is less toxic to cells if functions for fatty acid beta-oxidation, spermine metabolism, and translation reinitiation are compromised by mutation ( Fig. 12 and Tables 1 and 2). On the other hand, cells that transition to glycolytic metabolism need dTTP biosynthesis and protein complexes including the Cul4-RING E3 ubiquitin ligase, and the Ubp3-Bre5 deubuiquitinase, as well as Dom34-Hbs1, which functions in 'no-go' mRNA decay, in order to buffer doxorubicin ( Fig. 12 and Table 2). These findings indicate that the metabolic status of cancer cells further influences the vulnerabilities to chemotherapy that may emerge from genetic alterations occurring in cancer. Thus, along with monitoring Warburg status in conjunction with cancer genetic profiling, yeast phenomic models that predict potential influences of Warburg status on chemotherapy response and cancer vulnerabilities emerging from somatic mutations unique to each individual patient, may help better predict therapeutic outcomes and thus be useful to develop more efficacious treatment algorithms.
Since the Warburg metabolic status is not monitored clinically, we thought the most relevant genes from the yeast phenomic model for predicting genetic modifiers of doxorubicin cytotoxicity in pharmacogenomics data would be those influencing doxorubicin cytotoxicity similarly in either metabolic context. Functionally enriched genes in this category represented DNA repair and histone H3-K56 acetylation, along with deletion suppressing pathways, including sphingolipid Processes that influence doxorubicin cytotoxicity in a more Warburg-independent manner are unshaded. Arrowheads indicate processes for which genes predominantly transduce doxorubicin toxicity, based on their loss of function suppressing its growth inhibitory effects. Conversely, a perpendicular bar at the line head indicates a process that buffers doxorubicin toxicity, as genetic compromise of its function enhances the growth inhibitory effects of doxorubicin homeostasis, actin cortical patch localization, and telomere tethering at the nuclear periphery ( Fig. 12 and Table 3). We expanded the analysis to genes that were not GO-enriched, because genes may have evolutionarily conserved influences on phenotype independent of prior functional annotation and also independent of enrichment all together. Thus, we examined homologs systematically (regardless of GO enrichment) for yeast phenomic-PharmacoDB correlation and if evident specifically in either the glycolytic or respiratory context (Fig. 11, Tables 3 and 4, and Additional file 11). As discussed in the "Results" section, many genes not identified solely by GO-enrichment were nevertheless representative of enriched buffering processes from the phenomic model. The supplemental data files provided enable mining the pharmacogenomics data with the yeast model, filtering on combinations of desired criteria, including metabolic status, tissue type, and pharmacogenomics data set (Additional file 11).
Regarding the Warburg influence on chemotherapy response highlighted by the yeast phenomic model, work with cancer cell lines, mice, and acute myeloid leukemia blast cells from patients have suggested histone eviction, increased mutation rates at active promoter sites are important mechanisms of doxorubicin toxicity [18,19,232], including accumulation of damage from chromatin trapping by the FACT complex [20]. Further support of the importance of chromatin regulation was suggested by transcriptional control and assembly of histones, as well as histone modifications, as also suggested by differential gene expression from the pharmacogenomics data. The yeast model suggests that most of these effects are particularly important in a respiratory context; thus, from a precision medicine perspective, tumors that are promoted by genetic compromise in chromatin regulation [233,234] would be potentially more susceptible to treatment, but only if they have not undergone the Warburg transition to glycolysis. Analogously, patients with germline variation resulting in functional compromise of chromatin regulation may have normal tissue (e.g., cardiac muscle) that is susceptible to doxorubicin and thus may suffer greater toxic side effects of cancer treatment.
The genetic and phenotypic resolution of the yeast model can help resolve differential buffering by related complexes or pathways. In the example of histone deacetylase complexes, the class I (RPD3L and 3S complexes) and class II (HDA1 complex) HDAC genes interact differentially with doxorubicin. The Sin3-type class I HDAC complex exerts stronger deletion enhancement that is respiration specific, while the Class II (HDA1) complex shows weaker deletion enhancement that is relatively independent of Warburg status (Fig. 6a). These observations suggest that stratifying cancers based on their Warburg metabolic status could be informative for clarifying the clinical efficacy of different HDAC inhibitors in combination with doxorubicin. Consistent with the yeast model, pan-HDAC inhibitors have been shown to enhance the anti-cancer efficacy of doxorubicin, as well as its cardiotoxicity [235,236]. Interestingly, shRNA-mediated inhibition of HDAC6 enhanced doxorubicin cytotoxicity in transformed cells [237], but protected against doxorubicin induced cardiotoxicity [238]-findings which could relate to the reduced toxicity associated with loss of class II vs. class I histone deacetylase function. Given that cancers can be driven by epigenetic plasticity [233,234], such as could occur by loss of histone deacetylase function, information about the Warburg metabolic status could help clarify the likely impact of mutations in Sin3-type (class I) vs. HDA1-like (class II) histone deacetylase complexes. While speculative, this example illustrates the possible utility of yeast phenomic models to generate unbiased, systems-level experimental insights and may be of interest given the availability of HDAC6-specific inhibitors [239].
The examples of integrating yeast phenomic data with cancer cell line pharmacogenomics data to predict therapeutic efficacy are not limited to doxorubicin and/or the Warburg phenomenon. Analogous phenomic models could be generated for other cytotoxic agents and/or metabolic states, so long as the corresponding targets and buffering networks are conserved. Consistent with prior studies in yeast examining the question [169], we found the global correlation of human UES and OES with yeast deletion suppressors and enhancers to be low, further indicating the value of phenomic models for interpreting associations of gene expression with actual traits that are directly subject to natural selection. We anticipate that future integrative studies and ultimately clinical trials will further clarify how yeast phenomic studies can contribute to personalizing therapeutic efficacy for patients.
Although we focused the yeast model on predicting causality among differentially expressed genes that were associated with doxorubicin sensitivity in pharmacogenomic experiments, it can also be directly informative even if not correlated with gene expression. For example, genes required for DNA recombinational repair can be functionally regulated relatively independently of transcription [169]. Accordingly, these were detected more strongly from the yeast phenomics than the pharmacogenomics data.
In summary, we envision yeast phenomic drug-gene interaction models as a complement to existing cancer pharmacogenomics, providing an experimental platform to quantitatively derive drug-gene interaction network knowledge that can be integrated with DNA, RNA, protein, epigenetic, metabolite profiling, and/or cell proliferation data collected from tumors. Examples of experimental validation of the yeast model in cancer cell lines, as described through the manuscript, are summarized in Table 5. As a future step, predictions regarding treatment response of cancer to specific cytotoxic agents could be tested prospectively with patient samples, in vitro or in patient-derived xenograft models. Such a strategy could also be extended to before and after treatment(s) to understand how cancers evolve to buffer the drug's toxicities. Analyses of patient-derived tumor organoids, for example, could include predictive modeling and experimental validation for the development of treatment strategies, both initially and with recurrence [240][241][242]. Though we have focused on a single cytotoxic agent for demonstrating the principle here, yeast phenomics would also accommodate modeling of combination chemotherapy, both for anticancer efficacy and host toxicity [243]. The influence of the Warburg effect or other influences of metabolic or nutrient status could also be integrated into such personalized models of cancer chemotherapy efficacy [244]. Thus, yeast phenomic models can be tailored to examine increasingly complex interactions: also including background genetic factors such as homologous recombination deficiency [144]. Yeast phenomics provides the experimental capabilities and genetic tractability to model genetic buffering networks relevant to human disease at high precision and resolution. However, advanced strategies for applying yeast phenomics to predict genetic influences on human disease biology remain to be developed. A major premise of precision medicine should be to comprehensively and quantitatively account for the contribution of genetic variance to phenotypes as well as influential interacting factors such as cell energy metabolism, age, drugs, or other environmental factors. This is an overwhelming challenging in humans, as functional genetic variation, as exemplified in cancer, is essentially too abundant to resolve at a systems level, particularly with respect to higher-order interactions as undoubtedly occur with combination chemotherapy. Thus, yeast phenomics, which can define gene interaction networks and genetic buffering in a systematic and global way [28,245,246], offers the potential to help resolve gene interaction networks that contribute to disease and therapeutic response [24,247].

Conclusions
A yeast phenomic model for the influence of Warburg metabolism on doxorubicin cytotoxicity revealed that glycolysis reduces the cellular reliance on genetic buffering networks. The model reports gene deletionenhancing and deletion-suppression pathways and leverages yeast phenomic results to predict differentially expressed human genes that are causal in their association with doxorubicin killing from cancer cell line pharmacogenomics data. As such, this yeast model provides systems-level information about gene networks that buffer doxorubicin, serving as example of how Q-HTCP applied to the YKO/KD enables experimental designs to quantify gene interaction globally at high resolution; in this case, resolving how gene networks buffer doxorubicin cytotoxicity differentially with respect to Warburg metabolic status. Understanding cytotoxicity in terms of differential gene interaction networks has the potential to inform systems medicine by increasing the precision and rationale for personalizing the choice of cytotoxic agents, improving anti-tumor efficacy, and thereby reducing host toxicity. Yeast phenomics is a scalable experimental platform that can, in principle, be expanded to other cytotoxic chemotherapeutic agents and metabolic states, singly or in combination, thus providing versatile, tractable models to map drug-gene interaction networks and understand their complex influence on cell proliferation.
Additional file 1: Figure S1. Doxorubicin dose responses of the YKO/ KD parental strains, BY4741a, BY4742alpha, and BY4743a/alpha diploid. Figure S2. Correlation between interaction scores based on L vs. other CPPs (K, r, and AUC), for both HLD and HLEG media. Figure S3. A summary of the first and second rounds of REMc. First round clusters are at the left end of each row of heatmap thumbnails; second round clusters derived from each first round cluster are ordered to the right by relative strength. Rows are grouped into panels by similarity in their gene interaction profiles. The columns in each heatmap have the same order from left to right (see inset panel), with K to the left and L to the right. Within the K and L groups, HLD is to the left and HLEG to the right.  Figure S4. Doxorubicin-gene interaction profiles for selected mitochondrial GO terms. Figure S5. Deletion of mitochondrial genes tends to influence doxorubicin-gene interaction in a respiratory (HLEG media) more so than a glycolytic (HLD media) context. Figure S6. Heatmaps for GO terms comprised of overlapping gene sets. Figure S7. Pleiotropic phenotypic influences from genetic perturbation of ribonucleoprotein complex subunit organization. Figure S8. HLD-specific deletion enhancement of doxorubicin toxicity by evolutionarily conserved genes. See also Additional file 10: Table S13. Figure S9. GO term-specific heatmaps for mRNA 3' end processing and mRNA cleavage gene interaction profiles. Figure S10. Suppression of doxorubicin cytotoxicity by perturbation of sphingolipid and ceramide metabolism. Figure S11. Deletion suppressing doxorubicin-gene interaction for nuclear pore and actin cortical patch functions is relatively Warburg-independent.
Additional file 2. Doxorubicin-gene interaction data; Tables S1-S8. Tables S1-S4 are the genome-wide experiment: Table S1. YKO/KD strains in HLEG. Table S2. Reference cultures in HLEG. Table S3. YKO/KD strains in HLD. Table S4. Reference cultures in HLD. Tables S5-S8 are the validation study: Table S5. YKO/KD strains in HLEG. Table S6. Reference cultures in HLEG. Table S7. YKO/KD strains in HLD. Table S8. Reference cultures in HLD. Additional file 5. REMc results with doxorubicin-gene interaction profile heatmaps and Gene Ontology enrichment (GO Term Finder; GTF) results. File A contains REMc results and associated gene interaction and shift data. File B is the heatmap representation of each REMc cluster after incorporating shift values and hierarchical clustering. File C contains the GTF results obtained for REMc clusters for the three ontologiesprocess, function, and component.
Additional file 6. Gene Ontology Term Averaging (GTA) results and interactive plots. File A contains all GTA values, cross-referenced with REMc-enriched terms. File B displays GTA values associated with abovethreshold GTA scores (see note below) plotted for HLD vs. HLEG. GTA values for REMc-enriched terms are also included (regardless of whether |GTA score| >2). File C displays a subset of File B, containing only GO Terms with above-threshold GTA scores and that were enriched by REMc/GTF. File D reports GTA value using the K parameter. Files B-D should be opened in an Internet web browser so that embedded information from File A can be viewed by scrolling over points on the graphs. Subsets in each of the plots can be toggled off and on by clicking on the respective legend label. In the embedded information, X1 represents HLEG and X2 represents HLD information. Note: The GTA score threshold (for L) indicates that GTA-gtaSD > 2 for enhancers or GTA+gtaSD < -2 for suppressors, in at least one media. Additional file 7. Systematic comparisons involving genome-wide studies of doxorubicin-gene interaction. Table S9. Genes with deletionenhancing doxorubicin-gene interaction from Xia et al. 2007 and Westmoreland et al. 2009 . Table S10. Summary of experimental details associated with Table S9. Table S11. Test of enrichment for doxorubicingene interaction among genes encoding proteins predicted as substrates of the NatC complex. Table S12. Test of enrichment for doxorubicingene interaction among genes predicted to be regulated by conserved uORFs (Cvijovic et al. 2007).
Additional file 8. Quantitative summaries of REMc clusters. File A depicts REMc results, in terms of cluster distributions of L and K interaction ('shift' is not used for REMc and thus is not displayed), as a way to visualize cluster differences quantitatively. File B is organized by first round clusters and plots the change in p-value for significant terms with respect to round of clustering. Clusters derived from one another and sharing enrichment of the same GO term are connected by a line. Only GO terms with a background size of 500 or smaller are included. Scroll over a symbol to see embedded detail about each GO term. The square root of the p-value is used on the y-axis to evenly distribute data.
Additional file 9. GO term-specific heatmaps for REMc/GTF-enriched clusters. GO term-specific heatmaps for significant GO process terms were generated as described in methods and Figs. 3 and 4. Any related child terms are presented in subsequent pages of the parent file name. GO terms with more than 100 children, with 2 or fewer genes annotated to the term, or a file size over 300KB are not shown. All heatmaps are generated with the same layout (see Figs. 3

and 4).
Additional file 10: Table S13. HLD-specific gene deletion enhancement, not associated with 'shift' / growth deficiency. Data were selected for yeast-human homologs if the respective YKO/KD strains generated growth curves in both HLD and HLEG media (in the absence doxorubicin), and either of the following two sets of criteria were met: (1) HLD L interaction > 2 and HLEG L interaction < 2; these data were further filtered for HLD L Interaction -HLD L Shift > 4, and are presented in Additional file 1: Figure S8A.; or (2) HLD L Interaction -HLEG L interaction > 4 and HLEG K interaction > -10; these data were further filtered for HLD L Interaction -HLD L Shift > 4, and are presented in Additional file 1: Figure S8B. Data included in Additional file 1: Figure S8 are indicated in the last column.
Additional file 11. Integration of yeast phenomic and cancer cell line pharmacogenomic data to predict human genes that modify doxorubicin toxicity in cancer cells. (A) Tables of UES and OES human genes and whether their yeast homologs were found to be deletion enhancing or deletion suppressing, respectively. (B-C) Overlap between the gCSI and GDSC1000 databases with regard to UES and OES human genes (B) across all tissues or (C) for individual tissues. Note: the intersection of UES with OES between gCSI and GDSC was used as a negative control for assessing UES and OES overlap. (D-E) Yeast phenomic doxorubicin-gene interaction profiles for homologs of human UES or OES genes, sub-classified according to interaction type (deletion enhancing or suppressing) and Warburg-dependence of the interaction, for the (D) gCSI or (E) GDSC1000 databases. Similar to Fig. 11, yeast-human homology relationships are shown to the left of heatmaps (blue -one to one; green -one to many; red -many to many). (F-I) Interactive plots for yeast-human homologs, comparing the p-value of human genes to L interaction scores for yeast counterparts in (F, G) HLD or (H, I) HLEG from (F, H) gCSI or (G, I) GDSC1000. For the standardized coefficient ('estimate'; color gradient), a negative value (purple) indicates UES, while a positive value (orange) indicates OES. Thus, the model would predict causality for a human gene if its yeast homolog has a positive L interaction (deletion enhancing) and is colored purple (UES), or a negative L interaction (deletion suppressing) and colored orange (OES). Genes are only plotted if the human homolog was significant (p-value < 0.05).
Additional file 12. Comparisons between yeast studies of doxorubicin in the context of integrating cancer pharmacogenomics data. Overlapping and unique sets of genes reported from the different studies of doxorubicin, using the YKO/KD libraries, are assessed with regard to correlation with cancer pharmacogenomics data.

Availability of data and materials
All data generated or analyzed during this study are either included in this published article and supplementary files or will be freely supplied upon request.
Ethics approval and consent to participate Not applicable

Consent for publication Not applicable
Competing interests JLH has ownership in the Spectrum PhenomX, LLC, a shell company that was formed to commercialize Q-HTCP technology. The authors declare that they have no other competing interests.

Glossary of terms CPPs
Cell proliferation parameters: parameters of the logistic growth equation used to fit cell proliferation data obtained by Q-HTCP. The CPPs used to assess gene interaction in this study were K (carrying capacity) and L (time required to reach half of carrying capacity, K/2) [27, 28, 31, 33] DAmP Decreased abundance of mRNA production: refers to the method of making the yeast knockdown alleles, where the 3' UTR of essential genes is disrupted, reducing mRNA stability and therefore gene dosage [