Computational pharmacogenomics screen identifies synergistic statin-compound combinations as anti-breast cancer therapies

Statins are a family of FDA-approved cholesterol-lowering drugs that inhibit the rate-limiting enzyme of the metabolic mevalonate pathway, which have been shown to have anti-cancer activity. As therapeutic efficacy is increased when drugs are used in combination, we sought to identify agents, like dipyridamole, that potentiate statin-induced tumor cell death. As an antiplatelet agent dipyridamole will not be suitable for all cancer patients. Thus, we developed an integrative pharmacogenomics pipeline to identify agents that were similar to dipyridamole at the level of drug structure, in vitro sensitivity and molecular perturbation. To enrich for compounds expected to target the mevalonate pathway, we took a pathway-centric approach towards computational selection, which we called mevalonate drug network fusion (MVA-DNF). We validated two of the top ranked compounds, nelfinavir and honokiol and demonstrated that, like dipyridamole, they synergize with fluvastatin to potentiate tumor cell death by blocking the restorative feedback loop. This is achieved by inhibiting activation of the key transcription factor that induces mevalonate pathway gene transcription, sterol regulatory element-binding protein 2 (SREBP2). Mechanistically, the synergistic response of fluvastatin+nelfinavir and fluvastatin+honokiol was associated with similar transcriptomic and proteomic pathways, indicating a similar mechanism of action between nelfinavir and honokiol when combined with fluvastatin. Further analysis identified the canonical epithelial-mesenchymal transition (EMT) gene, E-cadherin as a biomarker of these synergistic responses across a large panel of breast cancer cell lines. Thus, our computational pharmacogenomic approach can identify novel compounds that phenocopy a compound of interest in a pathway-specific manner. Significance Statement We provide a rapid and cost-effective strategy to expand a class of drugs with a similar phenotype. Our parent compound, dipyridamole, potentiated statin-induced tumor cell death by blocking the statin-triggered restorative feedback response that dampens statins pro-apoptotic activity. To identify compounds with this activity we performed a pharmacogenomic analysis to distinguish agents similar to dipyridamole in terms of structure, cell sensitivity and molecular perturbations. As dipyridamole has many reported activities, we focused our molecular perturbation analysis on the pathway inhibited by statins, the metabolic mevalonate pathway. Our strategy was successful as we validated nelfinavir and honokiol as dipyridamole-like drugs at both the phenotypic and molecular levels. Our pathway-specific pharmacogenomics approach will have broad applicability.


Significance Statement:
We provide a rapid and cost-effective strategy to expand a class of drugs with a similar phenotype. Our parent compound, dipyridamole, potentiated statin-induced tumor cell death by blocking the statin-triggered restorative feedback response that dampens statins pro-apoptotic activity. To identify compounds with this activity we performed a pharmacogenomic analysis to distinguish agents similar to dipyridamole in terms of structure, cell sensitivity and molecular perturbations. As dipyridamole has many reported activities, we focused our molecular perturbation analysis on the pathway inhibited by statins, the metabolic mevalonate pathway. Our strategy was successful as we validated nelfinavir and honokiol as dipyridamole-like drugs at both the phenotypic and molecular levels. Our pathway-specific pharmacogenomics approach will have broad applicability.
van Leeuwen et al 4

Background
Triple-negative breast cancer (TNBC) is an aggressive subtype of breast cancer (BC) that has a poorer prognosis amongst the major breast cancer subtypes 1 . This poor prognosis stems from our limited understanding of the underlying biology, the lack of targeted therapeutics, and the associated risk of distant recurrence occurring predominantly in the first two years after diagnosis 2 . Cytotoxic anthracycline and taxane-based chemotherapy regimens remain the primary option for treating TNBC, with other classes of investigational agents in various stages of development. Therefore, novel and effective therapeutics are urgently needed to combat this difficult-to-treat cancer.
Altered cellular metabolism is a hallmark of cancer 3,4 and targeting key metabolic pathways can provide new anti-cancer therapeutic strategies. Aberrant activation of the metabolic mevalonate (MVA) pathway is a hallmark of many cancers, including TNBC, as the end-products include cholesterol and other non-sterol isoprenoids essential for cellular proliferation and survival [5][6][7] . The statin family of FDA-approved cholesterollowering drugs are potent inhibitors of the rate-limiting enzyme of the MVA pathway, 3-hydroxy-3methylglutaryl-CoA reductase (HMGCR) 5 . Epidemiological evidence shows that statin-use as a cholesterol control agent is associated with reduced cancer incidence 8 and recurrence [9][10][11][12][13] . Specifically, in BC, a 30-60% reduction in recurrence is evident amongst statin users, and decreased risk is associated with increased statin duration 9,12,14,15 . We and others have shown preclinically that Estrogen Receptor (ER)negative BC cell lines, including TNBC, are preferentially sensitive to statin-induced apoptosis 16,17 .
Moreover, three preoperative clinical trials investigating lipophilic statins (fluvastatin, atorvastatin) in human BC, showed statin use was associated with reduced tumour cell proliferation and increased apoptosis of high-grade BCs 18,19 . Thus, evidence suggests that statins have potential utility in the treatment of BC, including TNBC.
Drug combinations that overcome resistance mechanisms and maximize efficacy have potential advantages as cancer therapy. Blocking the MVA pathway with statins triggers a restorative feedback response that significantly dampens the pro-apoptotic activity of statins 20,21 . Briefly, statin-induced depletion of intracellular sterols, triggers the inactive cytoplasmic, precursor form of the transcription factor sterol regulatory element-binding protein 2 (SREBP2) to be processed to the active mature nuclear form, which induces transcription of MVA genes, including HMGCR and the upstream synthase (HMGCS1) 22 . We have van Leeuwen et al shown that inhibiting SREBP2 using RNAi, or blocking SREBP2 processing using the drug dipyridamole, significantly potentiates the ability of statins to trigger tumor cell death 21,23,24 .
Dipyridamole is an FDA-approved antiplatelet agent commonly used for secondary stroke prevention, and since statin-dipyridamole has been co-prescribed for other indications it may be safely used in the treatment of cancer. However, the exact mechanism of dipyridamole action remains unclear as it has been reported to regulate several biological processes. Moreover, the antiplatelet activity of dipyridamole may be a contraindication for some cancer patients. Thus, to expand this dipyridamole-like class of compounds that can potentiate the pro-apoptotic activity of statins, we employed a pathway-centric approach to develop a computational pharmacogenomics pipeline to distinguish compounds that are predicted to behave similarly to dipyridamole in the regulation of MVA pathway genes. Using this strategy, we identified several potential dipyridamole-like compounds including nelfinavir, an FDA-approved antiretroviral drug and honokiol, a compound isolated from Magnolia spp., which synergise with statins to drive tumour cell death by blocking the restorative feedback response. Correlation analysis of the statincompound combination synergy score, with basal mRNA expression across a large panel of BC cell lines, identified CDH1 expression as a predictive biomarker of response to these combination therapies. Taken together, we provide a new strategy to identify compounds that behave functionally similar to dipyridamole in an MVA pathway-specific manner, and suggest that this approach will have broad utility for compound discovery across a wide-variety of drug/pathway interactions.

Computational pharmacogenomic pipeline identifies dipyridamole-like compounds
We developed a computational pipeline that harnesses high-throughput pharmacogenomics analysis to identify dipyridamole-like compounds that synergise with statins by blocking MVA pathway gene expression to inhibit cancer cell viability (Figure 1). The LINCS-L1000 (L1000) 25 and NCI-60 26 datasets were chosen for these studies as they contain cellular drug-response data at the molecular and proliferative levels across a panel of cell lines, respectively. From these datasets we extracted drug structure, drug-induced gene perturbation data (gene expression changes after drug treatment) and drug-cell line sensitivity profiles for the 238 compounds common to both datasets. Treating each level of data as a separate layer, we restricted van Leeuwen et al the drug-gene perturbation layer from the L1000 dataset to only include the six MVA pathway genes present in the L1000 landmark gene set to enrich for compounds that phenocopy the MVA pathway-specific activity of dipyridamole (Supplemental Figure 1A). With dipyridamole as the reference input, we generated an MVA pathway-specific Drug Network Fusion (MVA-DNF) through the integration of 3 distinct data layers: drug structure, MVA-specific drug perturbation signatures, and drug-cell line sensitivity profiles. For each of the data layers incorporated into MVA-DNF, an 238x238 drug affinity matrix was generated, indicating drug similarity for a selected drug against all other drugs. Using the Pearson correlation coefficient, we computed the similarity for every pair of drug perturbation profiles and pairs of drug sensitivity profiles ( Figure 1B). From this, we identified 23 potential dipyridamole-like compounds that scored as significant (permutation test p-value <0.05); Methods; Figure 1B and Supplementary Table 1). Represented as a network, these hits display strong connectivity to dipyridamole as well as to each other.
We assessed the contribution of the different data layers (drug structure, drug-gene perturbation, and drug-cell line sensitivity) within the MVA-DNF for each of these 23 compounds ( Figure 1C). Drug perturbation played a significant role in the selection of novel dipyridamole-like compounds compared to drug sensitivity and drug structure. This reflects the specificity of the MVA-DNF towards the MVA pathway, in comparison to a 'global' drug taxonomy that is not MVA pathway-centric. Further assessment of the six MVA-pathway gene expression changes within the drug perturbation signatures highlights comparable expression profiles between dipyridamole and the novel dipyridamole-like compounds (Supplementary Figure 1B).
To prioritize and further interrogate the identified dipyridamole-like hits we annotated the 23 compounds by reported mechanism of action and potential clinical utility. Two compounds were excluded from further analysis as they were not clinically useful: Chromomycin A3, a reported toxin 27 , and cadmium chloride, an established carcinogen 28 . The remaining 21 compounds segregated into ten distinct categories, demonstrating that dipyridamole-like hits identified through our pharmacogenomics pipeline spanned a diverse chemical and biological space (Supplemental Figure 1C, Supplemental Table 1). We sought to validate the five hits that scored as most similar to dipyridamole, which belong to four different categories (RAF/MEK inhibitor, antiretroviral, anthracycline and natural product). Our lab had previously reported that the anthracycline doxorubicin potentiates lovastatin in ovarian cancer cells 29 confirming the reliability of our approach. Similarly, RAF/MEK inhibitors such as PD98059 and more recently Selumetinib (AZD6244) have been reported to synergise with statins to potentiate cancer cell death 30,31 . Of the top five hits, doxorubicin van Leeuwen et al was an existing BC chemotherapeutic agent 32 and therefore removed from further analysis. The molecular targeted compound (selumetinib) along with the novel three compounds were advanced for further evaluation (nelfinavir, mitoxantrone and honokiol) (Supplemental Table 1).

Dipyridamole-like compounds induce apoptosis in combination with fluvastatin and block the sterol-regulated feedback loop of the MVA pathway
To investigate whether the dipyridamole-like compounds could potentiate fluvastatin-induced cell death similar to that of dipyridamole, we first investigated sensitivity to increasing statin exposure in combination with a sub-lethal concentration of the novel dipyridamole-like compounds (Supplemental Figure 2) in two breast cancer cell line models with differential sensitivity to fluvastatin as a single agent 16 . As seen with dipyridamole, we observed similar potentiation of fluvastatin (lower IC50) when combined with a sub-lethal concentration of selumetinib, nelfinavir, or honokiol, but not mitoxantrone (Supplemental Fig 3 and Supplemental Fig 4). Therefore, mitoxantrone was no longer pursued as a dipyridamole-like compound.
To determine the nature of the anti-proliferative activity of the statin-compound combinations, we evaluated cell death by fixed propidium iodide staining and PARP cleavage with selumetinib, nelfinavir, or honokiol.
Our data indicate that all three compounds, at concentrations that have minimal effects as single agents, phenocopy dipyridamole and potentiate statin-induced cell death (Figure 2A-C).
Mechanistically, statins induce a feedback response mediated by SREBP2 that has been shown to dampen cancer cell sensitivity to statin exposure. Moreover, blocking the SREBP2-mediated feedback response with dipyridamole enhances statin-induced cancer cell death 21,24 . We have shown that dipyridamole blocks the regulatory cleavage and therefore activation of SREBP2, decreasing mRNA expression of SREBP2-target genes of the MVA pathway. As expected, statin treatment induced the expression of SREBP2-target genes, INSIG1, HMGCR and HMGCS1 after 16 hr of treatment, which was blocked by the co-treatment with dipyridamole ( Figure 3A, Supplemental Figure 5A). Similarly, nelfinavir and honokiol both phenocopy dipyridamole and block the statin-induced expression of MVA pathway genes ( Figure 3A, Supplemental Figure 5A). By contrast, co-treatment with selumetinib did not block the fluvastatin-induced feedback response. Housekeeping gene RPL13A was used as a reference gene for normalizing mRNA between samples and was not altered in the presence of the compounds (Supplemental Figure 5B).
Because SREBP2 is synthesized as an inactive full-length precursor that is activated to the mature nuclear form upon proteolytic cleavage, we used western blot analysis to assess the protein levels of both van Leeuwen et al 8 full-length and mature SREBP2. Nelfinavir and honokiol, but not selumetinib, blocked fluvastatin-induced SREBP2 processing and cleavage similar to that of dipyridamole ( Figure 3B-C). This suggests that while selumetinib is a strong potentiator of statin induced cell death, it does not mimic the action of dipyridamole by blocking the restorative feedback response (Figure 3, Supplemental Figure 5).

Novel statin-compound combinations phenocopy synergistic activity of fluvastatin-dipyridamole in a breast cancer cell line screen
To investigate whether the potentiation of fluvastatin by nelfinavir and honokiol has broad applicability and examine the determinants of synergy, we further evaluated these statin-compound combinations across a large panel of 47 breast cancer cell lines. A 5-day cytotoxicity assay (sulforhodamine B assay; SRB) in a 6x10 dose matrix was used to assess fluvastatin-compound efficacy. As expected, dipyridamole treatment resulted in a dose-dependent decrease in fluvastatin IC50 value (Supplemental Figure 6A). Similarly, nelfinavir and honokiol treatment also resulted in a dose-dependent decrease in fluvastatin IC50 values similar to that of dipyridamole (Supplemental Figure 6A). This suggests that our computational pharmacogenomic pipeline predicts compounds that potentiate statin activity similarly to dipyridamole across multiple subtypes of breast cancer cell lines.
Next we evaluated statin-compound synergy using the Bliss Index model derived using  (Figure 5C-D). Overall, this data validates that our MVA-DNF pharmacogenomics strategy can successfully distinguish compounds that, like-dipyridamole, can synergize with statins to trigger BC tumour cell death.

Discussion
By blocking the statin-induced restorative feedback response, dipyridamole potentiates statin efficacy to drive tumor cell death 21,24 . However, due to the polypharmacology of dipyridamole and the potential contraindication of this platelet-aggregation inhibitor for some cancer patients, it was essential to identify additional dipyridamole-like compounds and expand this class of agents to provide synergistic statin+compound treatment options for cancer therapy. To this end, we developed a novel computational pharmacogenomics pipeline that distinguished compounds that are similar to dipyridamole at the level of structure, MVA pathway gene expression perturbation, and anti-proliferative activity. We identified 23 potential dipyridamole-like compounds and then evaluated several of the top hits for their ability to phenocopy dipyridamole. Through this approach, we validated that nelfinavir and honokiol sensitize breast cancer cell lines to statin-induced cell death by blocking the statin-induced restorative feedback loop.
Analysis of basal RNA and protein expression identified the canonical EMT gene CDH1 (E-cadherin) as a biomarker of the synergistic response to both statin+nelfinavir and statin+honokiol treatment. Thus, the computational pharmacogenomics screen described here identified synergistic statin-compound drug combinations as novel anti-breast cancer therapies.
The integration of a computational pharmacogenomics pipeline and cellular validation to identify novel compounds with similar biological activities provides a rapid and inexpensive strategy that has potential broad applicability as it is also adaptable. For example, one issue we had to overcome in identifying dipyridamole-like compounds was the polypharmacology of dipyridamole itself. Dipyridamole was originally identified for its anti-platelet aggregation activity and thus the mechanism of action remains unclear. Several activities of dipyridamole have been described including an inhibitor of phosphodiesterases (PDEs) 38 , nucleoside transport 39 and glucose uptake 40 . The complexity associated van Leeuwen et al with this polypharmacological activity beyond the mevalonate pathway was circumvented by restricting the gene perturbation layer of the DNF to MVA pathway genes. This shows that the computational pharmacogenomics pipeline described here is likely tunable to drug-specific structural features, activities and signaling pathways.
The new statin-sensitizing agents identified here using MVA-DNF include nelfinavir and honokiol, which like dipyridamole, inhibit statin-induced SREBP2 cleavage and activation 21,24 . To date, a number of SREBP2 inhibitors have been identified that block SREBP2 processing from its precursor to mature form, including fatostatin, betulin, and xanthohumal (ER-Golgi translocation), PF-429242 (site-1 protease (S1P) cleavage), and nelfinavir and 1,10-phenanthroline (site-2 protease (S2P) cleavage). Additional SREBP2 inhibitors include BF175 and tocotrienols that target SREBP2 transcriptional activity and protein stability, respectively. However, other than nelfinavir, these agents have many reported targets and are only used as tool compounds for research purposes.
The S2P protease inhibitor nelfinavir was approved for use in 1997 as an antiviral for the treatment of HIV, and in recent years has begun to be evaluated for its utility as an anti-cancer agent 41 . While combination studies of statins and nelfinavir have not been previously reported or investigated in the context of cancer, open-label, multiple-dose studies have been performed to determine the interactions between nelfinavir and two statins (atorvastatin and simvastatin) in healthy volunteers. It was stated that coadministration of nelfinavir and simvastatin should be avoided while atorvastatin should be co-administered with caution. It should be noted that the family of statin drugs are metabolized by different enzymes.
Therefore, these interactions of nelfinavir with atorvastatin and simvastatin were likely due to drug-drug interactions leading to the inhibition of CYP3A4. By contrast, fluvastatin is metabolized by CYP2C9 providing additional rationale for our use of fluvastatin in statin-drug combinations as the probability of drugdrug interactions is significantly reduced.
To the best of our knowledge, this is the first study to report honokiol to synergize with statins in the context of cancer. Honokiol is a natural product commonly used in traditional medicine and has a number of reported mechanisms of action. How honokiol inhibits SREBP2 remains unknown, however this is the first study to interrogate its activity in SREBP2 translocation and gene expression alone and in combination with statins. As honokiol and its derivatives are presently under development, these data can van Leeuwen et al now be incorporated into future structure activity relationship analyses to enrich or lessen this new feature of honokiol. Two additional predicted dipyridamole-like compounds tested in this study include selumetinib and mitoxantrone, which did and did not sensitize breast cancer cells to statin-induced apoptosis.
Selumetinib functions through an SREBP2-independent mechanism, suggesting that not only is the identification of feedback-dependent mechanisms beneficial for cancer treatment but also shows that additional feedback-independent classes of statin-sensitizers can be identified. This is particularly important as some multiple myeloma and prostate cancer cell lines have been shown to lack the feedback response.
The data presented here has important clinical implications for statins as anti-cancer agents.
Despite some positive results from window-of-opportunity clinical trials in breast cancer using statins, a modest effect was seen from statins alone 18,19 . Therefore, discovery of novel therapeutic combinations will be necessary to achieve significant clinical impact. Since nelfinavir is poised for repurposing and statins have demonstrated anti-cancer activity in early-phase clinical trials 18,19,[42][43][44][45][46] , clinical studies to further evaluate the therapeutic benefit of this combination could proceed swiftly. Furthermore, consideration of available gene and protein expression across our large collection of breast cancer cell lines identified a mesenchymal-enriched gene expression profile as highly predictive of sensitivity to all three statin+compound (dipyridamole, nelfinavir or honokiol) combinations. We further showed that CDH1 expression levels served as a biomarker of synergistic response. This reinforces the dipyridamole-like behaviour of nelfinavir and honokiol, identified by our pharmacogenomics pipeline, and creates opportunities for biomarker-guided clinical studies. CDH1 expression as a biomarker of predicted response to the combination of fluvastatin+nelfinavir could be used to identify those patients most likely to benefit.
We also observed this synergistic response to the combination therapies across multiple subtypes of breast cancer. Previously we had identified the basal-like breast cancer subtype as more sensitive to statins alone; here, we have expanded the scope of statin treatment to encompass the wider breast cancer population.
These findings can also be explored beyond breast cancer as CDH1 is expressed in most cancers, for example sarcomas which are fixed in a mesenchymal state and have previously been reported as responsive to statins as single agents 37,47 .
Taken together, our computational pharmacogenomics pipeline reveals that starting with compounds that act within or on a specific pathway, it is possible to identify additional compounds to van Leeuwen et al 13 increase a class of inhibitors and/or better help understand compound mechanism of action. Our study also provides a strong preclinical rationale to warrant further investigation of the fluvastatin+nelfinavir combination, as well as the CDH1 biomarker ( Figure 5E). The ready availability of these well-tolerated drugs as well as simple methods for assessing CDH1 expression could enable rapid translation of these findings to improve breast cancer outcomes.

Methods
Our analysis design encompasses both computational identification and refinement of dipyridamole-like compounds, as well as experimental validation of the most promising candidates.

MVA-specific Drug Network Fusion (MVA-DNF).
We developed a computational pharmacogenomic pipeline (MVA-DNF) that facilitates identification of analogues to dipyridamole, by elucidating drug-drug relationships specific to the mevalonate (MVA) pathway. MVA-DNF briefly extends upon some principles of the drug network fusion algorithm we had described previously 48 , by utilizing the similarity network fusion algorithm across three drug taxonomies (drug structures, drug perturbation, and drug sensitivity). Drug structure annotations and drug perturbation signatures are obtained from the LINCS-L1000 dataset 25,49 , and drug sensitivity signatures are obtained from the NCI-60 drug panel 26 . Drug structure annotations were converted into drug similarity matrices by calculating tanimoto similarity measures 50 and extended connectivity fingerprints 51 across all compounds, as described previously 48 . We extracted calculated Z-scores from drug-dose response curves for the NCI-60 drug sensitivity profiles, and computed Pearson correlation across these profiles to generate a drug similarity matrix based on sensitivity 26 . We used our PharmacoGx package (version 1.6.1) to compute drug perturbation signatures for the L1000 dataset using a linear regression model, as described previously 52 .
The regression model adjusts for cell specific differences, batch effects and experiment duration, to generate a signature for the effect of drug concentration on the transcriptional state of a cell. This facilitates identification of gene expression which has been significantly perturbed due to drug treatment. These signatures indicate transcriptional changes that are induced by compounds on cancer cell lines. We further refined the drug perturbation profiles to a set of six MVA-pathway genes (Supplementary Figure 1A) that had been obtained from the literature as well as repositories of pathway-specific gene sets including MSigDB 53 , HumanCyc 54 and KEGG 49,55 . These gene sets include 'mevalonate pathway' and 'superpathway of geranylgeranyldiphosphate biosynthesis I (via mevalonate)' from the HumanCyc 56 , and 'Kegg Terpenoid Backbone Biosynthesis' from KEGG 55,57 . The filtered drug-induced gene perturbation signatures were subsequently used to generate a drug perturbation similarity matrix that elucidates drug-drug relationships based on common transcriptional changes across the six MVA-pathway genes. We calculated similarity between estimated standardized coefficients of drug perturbation signatures using the Pearson correlation van Leeuwen et al coefficient. Finally, we used the similarity network fusion algorithm 58 to integrate drug structure, drug sensitivity, and MVA-pathway specific drug perturbation profiles, to generate an MVA-pathway specific drug taxonomy (MVA-DNF) spanning 238 compounds.

Identification of analogues to dipyridamole
We interrogated the MVA-DNF taxonomy using a variety of approaches to identify a candidate set of dipyridamole-like compounds. Using MVA-DNF similarity scores, we first generated a ranking of all compounds closest to dipyridamole. We then conducted a perturbation test, to assess the statistical relationship of each ranked drug against dipyridamole. Briefly, drug fusion networks were generated 1000 times across perturbation, sensitivity, and drug structure profiles, each time using a random set of six genes to generate a 'pathway-centric' drug perturbation similarity matrix. Z-scores and p-values were calculated to determine the statistical relevance of a given dipyridamole-like analog in MVA-DNF, compared to the randomly generated networks. From this, we further ranked a list of dipyridamole-like candidate compounds by their statistical significance within MVA-DNF (p-value<0.05), resulting in identification of 23 candidate dipyridamole analogs.
For each of the dipyridamole analogues we identified, we conducted a similar assessment of significance to identify the relationships of these compounds to dipyridamole and to themselves. A network of dipyridamole-like analogues was rendered using iGraph R package 59 . Using MVA-DNF similarity scores, we further computed the contribution of each of the drug layers (structure, sensitivity and perturbation) in the identification of dipyridamole-like compounds.
We assessed the regulation of gene expression for genes involved in the mevalonate pathway across all of the top-selected dipyridamole analogues, by analyzing the drug-induced transcriptional profiles (described above) of the selected analogues. To prioritize the dipyridamole analogues, the candidate compounds were categorized, and compounds that were known toxins or carcinogens were excluded from the analysis (Supplemental Table 1, Supplemental Figure 1C). Top hits from the largest categories were selected for further validation.

Cell culture and compounds
All cell lines were cultured as described previously 16,24 . Briefly, MDA-MB-231 and HCC1937 cells were cultured in Dulbecco's Modified Eagle's Medium (DMEM) and Roswell Park Memorial Institute medium (RPMI), respectively. All media was supplemented with 10% fetal bovine serum (FBS), 100 units/mL penicillin and 100 μg/mL streptomycin. Cell lines were routinely confirmed to be mycoplasma-free using the

Breast cancer cell lines panel
The breast cancer cell line 34 panel was a generous gift from Dr. Benjamin Neel. RNAseq quantification was done using Kallisto pipeline 60 using human transcriptome reference hg38.gencodeV23 61 . RPPA processed data was downloaded from 34 . SCMOD2 62 breast cancer subtypes of these cell lines were obtained using genefu R package 63 .

Cell death assays
Cells were seeded at 2.5x10 5 cells/plates and treated the next day as indicated. After 72 hours, cells were fixed in 70% ethanol for >24 h, stained with propidium iodide and analyzed by flow cytometry for the subdiploid (% pre-G1) DNA population as a measure of cell death as previously described 6 . van Leeuwen et al

Drug combinations synergy analysis
Viability scores were calculated using standard pipelines from PharmacoGx R package 52 and synergy