SL-Cloud: A Cloud-based resource to support synthetic lethal interaction discovery

Implementation

SL-Cloud aggregates commonly used public data resources relevant for SL inference and integrates them with analysis workflows to infer SLIs by leveraging the cloud-based resources stored on the ISB-CGC platform (Figure 1). These three components (summarized below) represent an ecosystem that integrates software and data to enable the large-scale prediction of SLIs from existing cloud-hosted datasets.

Figure 1. Overview of the Synthetic Lethality Cloud (SL-Cloud) framework.

Schematic overview of the computational workflows and data resources aggregated in this framework to facilitate investigation of synthetic lethal interactions. Users with specific research questions (top-left) reuse inference workflows from the provided scientific computing notebooks to query public cancer genomics datasets from the vast resources provided by the ISB-CGC and additional datasets pre-processed in SL-Cloud. Three candidate synthetic lethal pair inference workflows were implemented, including the DAta-minIng SYnthetic lethality identification (DAISY), mutation-dependent synthetic lethality prediction (MDSLP), and conserved genetic interaction (CGI) workflows. Different data resources and data types were used for different inference workflows. For example, the DAISY and MDSLP workflows rely on statistical testing over the multiple omics data and functional screening data such as CRISPR and shRNA datasets for human cancers, whereas the CGI workflow is based on an ortholog mapping of the SL interactions identified in yeast.

Institute for Systems Biology Cancer Gateway in the Cloud (ISB-CGC).

ISB-CGC, a part of the National Cancer Institute (NCI) Cancer Research Data Commons, hosts derived data in Google Cloud Platform BigQuery tables, providing gene expression, mutation, copy number alteration, methylation, protein levels and other molecular characteristics for a broad range of cancers, such that these datasets can be rapidly analyzed using the power of cloud computing (Bleich, 2022; Reynolds et al., 2017). Google BigQuery is a columnar data warehousing solution that provides fast access through Structured Query Language (SQL)-based queries. The ISB-CGC resource contains over 1000 distinct derived data tables that are findable through both a BigQuery Table Search interface as well as through the native Google BigQuery interfaces. On this platform, users can analyze multiple large-scale datasets solely or in combination with their own datasets while circumventing the need to download and maintain a local copy of these petascale datasets, or to perform extensive data management tasks. We have designed SL-Cloud as a component of ISB-CGC so that we can leverage ISB-CGC’s broad availability, co-location of robust computational resources, and democratized access to raw and derived cancer research data. SL-Cloud allows cancer researchers access to key datasets and workflows to enable the identification of potential synthetic lethal gene pairs for a specific cancer of interest.

Large-scale datasets relevant for SL Prediction

A key feature of SL-Cloud is the aggregation of key cancer multi-omics datasets for SL inference in a single framework that facilitates integrative analysis. For SL-Cloud, we created and mined the relevant large-scale, publicly available multi-omics and functional screening datasets on ISB-CGC such as The Cancer Genome Atlas (TCGA) (Hutter & Zenklusen, 2018) patient-level data on somatic mutations, gene expression, and copy number alterations across 33 cancer types, (Figure 1) (for details see Table 1, which includes the URLs for specific resources). We identified additional public data resources that were pertinent for SL inference but previously unavailable on ISB-CGC. These datasets include a genetic interaction dataset TheCellMap derived from model organism interaction screens (Costanzo et al., 2016), and human pan-cancer cell line molecular characterization and functional screening datasets, primarily from the Cancer Cell Line Encyclopedia (CCLE) and the Cancer Dependency Map (DepMap) initiative (Dempster et al., 2019; Ghandi et al., 2019; McFarland et al., 2018; Meyers et al., 2017). Building on the infrastructure of the ISB-CGC, we established a new SL-focused cloud resource within ISB-CGC that incorporates the most relevant datasets for SLI prediction (Table 1).

Synthetic lethal interaction prediction workflows

We implemented SLI inference workflows and distributed them as a set of Jupyter notebooks that use functions from Python scripts provided with this resource. The notebooks offer code optimization and integration with the ISB-CGC through the BigQuery interface to access the relevant pre-processed large-scale cancer genomics datasets described above through embedded SQL queries. Importantly, the workflows can be tailored to individual cancer researcher needs by copying the existing code and optimizing it to their specific use case. We present here a brief overview of three example workflows implemented (for technical details see accompanying documentation on the project GitHub page, the url is available in the Data and software availability section).

The DAta-minIng SYnthetic lethality identification workflow (DAISY): This previously published workflow is re-implemented using up-to-date, large-scale data resources as described above (Figure 1; Table 1) (Jerby-Arnon et al., 2014). DAISY applies multiple inference procedures that include:

Inference of synthetic dosage lethality (SDL), whereby overactivation of one gene causes its interaction partner to become essential for cell viability, is also implemented in DAISY. DAISY SL predictions are gene pairs that are found by all three inference modules. Each individual module can also provide evidence of SL potential independently. The workflow, as we have implemented it, enables users to list predicted synthetic lethal pairs from each workflow, for each dataset, and aggregate them or use them independently. The DAISY workflow also enables users to perform pan-cancer or tissue type-specific analyses by tuning to the specific biological question being examined.

Mutation-dependent synthetic lethality prediction (MDSLP): This workflow combines mutation and functional screening data to infer SL pairs from cancer cell line data. The MDSLP workflow is based on the rationale that, for tumors with mutations that have an impact on protein expression or structure (functional mutation), the knockout effects or inhibition of a partner target gene show conditional dependence for the mutated molecular entities (Figure 1). Leveraging the public cancer cell line datasets including gene mutation data from CCLE, and functional screening data generated by either shRNA or CRISPR technology from DepMap (Dempster et al., 2019; Ghandi et al., 2019; McFarland et al., 2018; Meyers et al., 2017), we integrated these data modalities to evaluate mutation-based conditional dependence. This workflow enables users to statistically test whether the knockout or knockdown effects for one gene will be altered if another gene is mutated in specific contexts, such as in pan-cancer, or tumor type-specific cell lines. The increase in gene knockout or knockdown sensitivity provides evidence to support potential SLIs.

Conserved genetic interaction (CGI) workflow: We implemented this workflow based on published methods described in (Srivas et al., 2016). The CGI workflow leverages cross-species conservation to infer experimentally derived SLIs in yeast to predict relevant synthetic lethal pairs in humans. For SL-Cloud, we downloaded and preprocessed TheCellMap dataset, the most comprehensive S. cerevisiae genetic interaction network inferred from synthetic genetic array (SGA) screens from (Costanzo et al., 2016) (see details in Table 1). Genetic interactions are inferred if the combined effect of a double mutant on cell viability differs from that of the combination of single mutant effects. SLIs are defined in this context as negative genetic interactions in which the cell viability of a double-mutant yeast colony is lower than that of the respective single-mutant colonies. We provide the inferred SLIs in humans by yeast-to-human ortholog mapping.

Operation

SL-Cloud workflows are implemented in a set of Python notebooks that can be edited and run via a Jupyter notebook interface. All the requirements are specified in the respective Jupyter notebooks. Users can also access and run the SL-Cloud workflows in mybinder.org, a platform that allows users to run the implementations without installing any libraries or downloading code from a GitHub repository (see setup and operation instructions here).

In silico validation of known or suspected SLIs

Synthetic lethality between BRCA1/2 and PARP1/2 is well documented and is the rationale behind the design of PARP inhibitors such as olaparib, rucaparib, and niraparib (Ashworth & Lord, 2018). These agents are approved for treating BRCA-mutated ovarian cancer and advanced breast cancer. To perform an in silico validation analysis of this well-established SLI, we applied MDSLP to gene mutation and functional screening data from pan-cancer cell lines (Dempster et al., 2019; Ghandi et al., 2019; McFarland et al., 2018; Meyers et al., 2017). As shown by the MDSLP-shRNA workflow and consistent with our expectations, functional mutations of BRCA2 showed significant sensitivity to PARP1 knockdown (two-sided t-test, P < 0.01, FDR < 0.1). We applied the same workflow to gene essentiality data derived from CRISPR screens, but did not find this expected interaction. The MDSLP workflow using CRISPR-derived datasets revealed that the functional mutation of BRCA2 shows a synthetic lethal partnership with PARP2 (P < 0.01, FDR < 0.05). shRNA-derived and CRISPR-derived BRCA2 synthetic lethal pairs showed limited overlap (Figure 2A). Only 6.6% (48 out of 729) of the BRCA2-related synthetic lethal pairs nominated from shRNA-derived inference with a threshold of FDR < 0.1 were also predicted using CRISPR essentiality screens. Of the 1433 potential partner genes predicted by any of the resources, only 48 partner genes were predicted by two resources (Figure 2B). Of these, WRN, TSC2, RPL22L1 showed the most significance with both CRISPR-derived and shRNA-derived inference (FDR < 0.01 for both inference procedures).

Figure 2. Predicted synthetic lethal interaction partners of BRCA2 based on cancer cell line dependency datasets.

A. Network-based representation of predictions generated by the mutation-dependent synthetic lethality prediction (MDSLP) using either shRNA (upper panel, orange) or CRISPR (lower panel, blue) functional screening datasets. B. Potential synthetic lethal partners of BRCA2 as predicted by both the CRISPR and shRNA functional screening data. Each node represents a gene, and edges represent potential synthetic lethal interactions. The node color encodes overlap between the gene-dependency assay type, with yellow representing synthetic lethal pairs predicted in both the CRISPR data and the shRNA functional screening data. The node size and font size indicate the strength of the statistical relationship, with high-confidence pairs having larger node sizes or font sizes. FDR levels: < 0.01 (largest), < 0.05 (median), and <0.1 (smallest).

Interestingly, we did not predict any BRCA2-related synthetic lethal pairs from the other two workflows implemented in SL-Cloud. BRCA2 has no yeast homolog and, therefore, conserved interactions could not be inferred by the CGI workflow. DAISY nominated no synthetic lethal partners for BRCA2 with its default settings but predicted potential BRCA1-PARP1/2 SLIs across all three of its component inference modules with non-default parameter settings. Both gene pairs, BRCA1-PARP1 and BRCA1-PARP2, showed statistically significant co-expression, with their correlation coefficients ranging from 0.26 to 0.59 across patient-derived and cancer cell line datasets [Figure 3A(i,ii) and 3B(i,ii)] (P < 0.01). In addition, we found statistical support for a BRCA1 and PARP1/2 SLIs by the SoF inference procedure [Figures 3A(iii) and 3B(iii,iv)] (P < 0.05), whereas the FunEx module found statistical support for a BRCA1-PARP1 SLI [Figure 3A(iv)] (P < 0.01), but not for a BRCA1-PARP2 SLI, based on the cancer cell line gene-dependency CRISPR or shRNA datasets. In summary, the BRCA1-PARP1 interaction was supported by all three DAISY inference modules, whereas only two modules supported the BRCA1-PARP2 interaction.

Figure 3. Evidence for BRCA1-PARP1/2 synthetic lethal relationship based on the DAta-minIng SYnthetic lethality identification (DAISY) workflow.

A. Evidence for a BRCA1-PARP1 synthetic lethal relationship by pairwise co-expression in i) Pan-Cancer Atlas and ii) CCLE datasets respectively; iii) by survival of the fittest in in CCLE data; and iv) by functional examination in DepMap CRISPR. B. Evidence for a BRCA1-PARP2 SL relationship by pairwise co-expression in i) Pan-Cancer Atlas and ii) CCLE datasets and by survival of the fittest in iii) Pan-Cancer Atlas and iv) CCLE datasets. R, Spearman correlation coefficient; p, P value by the one-sided Wilcoxon rank-sum test; * P < 0.05; ** P <0.01; **** P < 0.0001.

This example demonstrates how SL-Cloud facilitates the exploration of the SLIs for a particular gene by using orthogonal SLI prediction workflows and multiple datasets to assess the stability and reproducibility of particular SLIs of interest. For the established SLI between BRCA deficiency and PARP1/2 enzymes, we saw variation in the output of multiple prediction approaches and datasets in confirming this bona fide SLI. These analyses highlight some of the challenges related to SL prediction, including unaccounted for variation resulting from differences in the technology and size of the datasets used to make the SL prediction, and the implicit or explicit assumptions made by the underlying analytical approaches. This example shows how SL-Cloud can be applied to enable researchers to explore a particular SLI of interest in different datasets or using different prediction approaches.

Pathway-based SL analysis

SLI partners tend to form functional interaction networks (Costanzo et al., 2016; Jerby-Arnon et al., 2014). For example, Ku et al. reported that synthetic lethal screen hits are more robust at the pathway rather than at the gene level (Ku et al., 2020). To demonstrate pathway-based SLI discovery, we analyzed SLI-related genes in the DNA damage and repair (DDR) pathway. DDR deficiency due to loss-of-function alterations by mutation, deletion, or epigenetic silencing is prevalent across lineages affecting approximately 33% of all cancers in TCGA (Knijnenburg et al., 2018). Impaired DDR leads to genomic instability, and tumors exhibiting DDR loss are prone to DNA-damaging agents and, therefore, potentially vulnerable to inhibitors that target compensatory DDR pathways via a synthetic lethal mechanism (Lord & Ashworth, 2012). Using a well-curated set of 276 genes annotated for involvement in DNA damage repair from (Knijnenburg et al., 2018) we predicted synthetic lethal partners from the three workflows described above (Figure 1).

Consistent with our expectations, different SL prediction approaches led to a diverse set of predicted SLIs. Each workflow identified more than 1,000 synthetic lethal/synthetic dosage lethal partner genes, except for CGI, which identified only 67 synthetic lethal partner genes. Predicted SLI gene sets largely did not overlap; however, functional enrichment analysis showed shared pathway involvement in the interactions identified (Figure 4A). In particular, we found significant KEGG pathway enrichment in synthetic lethal partner genes for the cell cycle, RNA metabolism, splicing machinery, chromatin organization, and transcriptional regulatory pathways (FDR < 0.05). Interestingly, several of these results are broadly related to genomic stability maintenance, and as such, confirm previously published reports from Ku et al., 2020 and others that genes involved in SLIs tend to belong to related pathways. Gene ontology biological processes (GOBPs) enriched by synthetic lethal partner genes vary, but a clustering analysis based on hierarchical structure and semantic similarity of GOBPs showed the synthetic lethal partner genes identified via the different approaches were associated with similar biological processes (Figure 4B). A clustering analysis summarized 350 GOBPs that were initially identified via four approaches into 25 representative GOBP groups. The 25 GOBP groups are associated with synthetic lethal partner genes identified by at least two workflows, and their biological functionality is mirrored by the pathway enrichment results presented above, with enrichment in genes involved in the cell cycle, transcriptional regulation, chromatin organization, and the DNA damage response. In summary, we have demonstrated that pathway-based SL prediction is easily implemented in this framework and can quickly generate useful insights beyond the single-gene level.

Figure 4. Pathway enrichment of predicted synthetic lethal partners of DNA damage repair (DDR) genes.

Heatmaps depicting A. the KEGG or REACTOME pathway and B. Gene Ontology Biological Process (GOBP) enrichment for predictions made using four different approaches (columns). Increasing color intensity represents increasing statistical significance (P < 0.05, calculated by a hypergeometric test) for enrichment. Pathways or GOBPs were labeled if they were enriched by synthetic lethal partner genes identified by at least two prediction approaches. The redundant GOBPs were further reduced by REVIGO. The 398 GOBPs enriched by at least one approach were reduced to 172 GOBPs based on their semantic similarities, and then summarized into 27 representative groups whose enrichment significance is represented in the heatmap. Clustering analysis was performed for GOBPs inside and outside of the 27 representative groups, separately. The red line down the left side of panel B indicates the separation between the clustering analyses. DAISY, data mining synthetic lethality identification workflow; MDSLP-CRISPR, mutation-dependent synthetic lethality prediction workflow with CRISPR; MDSLP-shRNA, mutation-dependent synthetic lethality prediction workflow with shRNA; CGI, conserved synthetic lethal interactions from yeast screens.

Tumor-specific SL analysis

An overlooked factor that can affect reproducibility of SLIs is context dependence. Genetic background, epigenetic cell state, and tissue type can influence synthetic lethal genetic interactions (Nijman & Friend, 2013; Ryan et al., 2018). We show that the MDSLP workflow and our re-implementation of the DAISY algorithm can be applied to restricted subsets of the underlying data that represent samples or cell lines arising from the same cancer type. The rationale behind this type of analysis is that samples from the same cancer type could represent similar cellular origins, having a characteristic genetic interaction network that is tissue-type specific.

To illustrate this principle, we investigated the previously reported SLI between ARID1A and ARID1B. Functional loss of ARID1B is a specific vulnerability in ARID1A-mutated cancers, as it affects the composition of the SWI/SNF complex (Helming et al., 2014). We applied MDLSP and DAISY to predict synthetic lethal partners for ARID1A. Via the MDLSP workflow, we found statistical evidence of differential dependency for ARID1B between ARID1A-mutated and wild-type cell lines in various cancer types, suggesting the potential for a SLI between the two genes across tissue types (Figure 5A). Similar to our findings with the BRCA-related synthetic lethal partners, we also saw differences in the strength of the relationship based on whether ARID1B was knocked down via shRNA or knocked out using the CRISPR-Cas9 system. As there is strong and compelling evidence for this SLI, we find support for the interaction occurring across multiple cancer types, even if the evidence comes from shRNA-derived or CRISPR-derived dependency datasets alone.

Figure 5. SL-Cloud enables cancer type-specific synthetic lethal inference for ARID1A and ARID1B.

A. Evidence for synthetic lethality generated from the mutation-dependent synthetic lethality prediction (MDSLP) workflow as applied to cancer cell line dependency datasets when comparing the gene dependency scores (effects) for the shRNA and CRISPR datasets between the ARID1A-mutated group and wild-type group for different cancer types and across all cell lines (pan-cancer analysis). The threshold for statistical significance is FDR < 0.05. B. Statistical evidence for a synthetic lethal relationship between ARID1A and ARID1B from the DAta mIning SYnthetic lethality identification workflow (DAISY), with the results for each inference module being represented in the columns. Column heatmaps summarize the datasets used for each procedure. An asterisk (*) indicates that a test passed the FDR threshold (0.05); gray shading represents an invalid test or a lack of data availability. Gray shading represents an invalid test. C. Heatmap visualization depicting the Spearman correlation between ARID1A and ARID1B for the annotated cancer type for a patient derived sample (Pan-Cancer Atlas) or cancer cell line (CCLE) (rows) across different cancer types (columns).

DAISY does not predict an SLI between ARID1A and ARID1B when applied strictly, that is, when the requirement is set for statistically significant evidence across all three DAISY inference modules (Figure 5B). However, when considering each module individually, we see strong support for the interaction, with strong positive correlation (Spearman ρ in the range [0.3 to 0.77]) between these two genes across almost all cancer types considered (Figure 5B, C). Similar to the findings with MDSLP, we found evidence for the interaction between these two genes in ovarian cancer by using the functional examination module applied on shRNA-derived dependency dataset. This is unsurprising, as the underlying rationale for the DAISY functional examination inference procedure and MDSLP inference strategy are quite similar, and both approaches are applied to the same dataset. We found no statistical support for the interaction via genomic SoF inference, which may be explained by the fact that neither of those genes is inactivated by recurrent focal deletions that underpin that inference module.

This illustrative example showcases the flexibility of the SL-Cloud framework in that it is relatively easy to compare the results of different SL prediction approaches, while varying algorithmic parameters, using the same or different data types, or to restrict analysis to a given tumor type for further elucidation of context specificities in SL.