The yeast, Saccharomyces cerevisiae, with its stable haploid state, well-annotated genome, and short generation times, has served for over 50 years as a powerful tool in the investigation of gene functions and interactions. In yeast, SL was traditionally discovered by the random mutagenesis of loss-of-function (LoF), gain-of-function (GoF), or overexpression (OE) of mutant query strains followed by a selection regimen, such as drug treatment (Albertini and Zimmermann, 1991; Stevenson et al., 2001). However, yeast-based SL screens are now routinely conducted using systematic screening of arrayed yeast strain collections or arrayed plasmid collections following either of two common methods (Figure 1A). The first approach involves transforming a collection of OE/LoF/GoF mutant strains into a collection of yeast with mutant strains to produce double mutants (Boone et al., 2007); the second involves crossing two sets of mutant yeast strains to obtain haploid double mutants (Segrè et al., 2005; Boone et al., 2007). The crossed mutant strain collections were subjected to synthetic genetic array analysis (SGA) (Tong et al., 2001; Kuzmin et al., 2016), diploid-based synthetic lethality analysis on microarrays (dSLAM) (Pan et al., 2004), and the ‘green monster’ (Suzuki et al., 2011). Recently, Charles Boone led researchers in generating a global SL network with more than 23 million double mutants that crossed 90% of the yeast genome, thereby identifying thousands of SL gene pairs and providing a diagram of the cell’s functional wiring (Costanzo et al., 2016). Based on SGA methodology, they developed trigenic-SGA (τ-SGA) to systematically screen and quantify trigenic interactions in yeast (Kuzmin et al., 2021a; Kuzmin et al., 2021b). The relatively recent introduction of CRISPR to study the yeast system has limited its usage, but Peccoud’s team has described more potential applications for its use in yeast (Adames et al., 2019).

In the yeast system, colony size is typically phenotyped to measure the effects of a single or double mutation on yeast growth/fitness (Baryshnikova et al., 2010), and other measurements, including microarray (Pan et al., 2004) and fluorescence (Suzuki et al., 2011), are also commonly used. Studies of SL in yeast have provided invaluable information regarding fundamental molecular processes that can be used for subsequent screens in higher-level organisms; Nielsen’s research team has summarized the advantages of yeast-based technologies in cancer biology (Ferreira et al., 2019).

RNAi- and CRISPR-based genome-editing technologies have greatly influenced SL screening capabilities. RNAi is a biological process in which an RNA molecule contributes to sequence-specific gene silencing via translational or transcriptional regression, and use of this process provided the first opportunity to knock down (KD) the expression of individual genes and allowed for high-throughput screening in human cells (Laufer et al., 2013). The recent discovery and adaptation of the CRISPR/Cas9 system also brings more flexibility to genetic perturbation. RNAi functions at the post-transcriptional level, but CRISPR/Cas9 has been engineered to introduce functional knock-out (KO) at the gene level. Easily programmable and highly effective, CRISPR-based gene editing has outperformed RNAi. Several research groups have compared RNAi- and CRISPR-based screening technologies (Haussecker, 2016; Housden and Perrimon, 2016; Morgens et al., 2016; Smith et al., 2017).

Conventionally, the SL screens in human cells were categorized based on the techniques employed. For example, Brough’s group (Brough et al., 2011) summarized three methods for identifying SL: 1) applying RNAi screens on cell lines with or without a mutated targeted gene, 2) using RNAi library screens in combination with chemical inhibitors, and 3) chemical library screens. Here, we group the SL screens—chemical inhibitor, RNAi, or CRISPR—based on the level of gene perturbation introduced during the screening. This categorization also lays the foundation for SL calculation (Section 2.2).

SL screenings involving single-gene perturbation compared to cell-line outcomes with or without perturbation of a targeted gene to identify SL partners of that gene. Both population- and individual sample-based screenings apply single-gene perturbation techniques (Figure 1B), but the two approaches differ in the number of cell lines used in the screening. Individual sample-based screenings examine a single cell line, whereas population-based screens utilize multiple cell lines with varied cancer backgrounds. In contrast, SL screenings facilitated by combinatorial gene perturbations involve simultaneous disruptions of two genes within a specified cell line. The SL gene pairs are then identified by determining significant differences between the observed and expected phenotypes. This type of screening is typically based on individual samples.

The SL partners of a targeted gene (gene of interest) are identified by comparing gene essentialities between two groups of cell lines, either with or without perturbation of second-query genes (query strains in yeast). In population-based screening, these two groups are cells with target gene mutations and wild-type cells without the mutations. In cell-specific screening, the two groups are wild-type cells and cells with a mutated target gene or under certain perturbations using chemical inhibitors, RNAi, or CRISPR. SL calculation involves comparison of a gene’s essentiality between two groups, mainly by difference (Bommi-Reddy et al., 2008), fold-change or abundance ratio (Lord et al., 2008; Boettcher et al., 2014), Z-score (Turner et al., 2008; Martin et al., 2009; Steckel et al., 2012; Toyoshima et al., 2012; Shen et al., 2015), or t-test (Luo et al., 2009) of gRNA counts. Scoring by methods such as RNAi gene enrichment ranking (RIGER) (Luo et al., 2008), gene activity ranking profile (GARP) (Marcotte et al., 2012a), and observation of redundant siRNA activity (RSA) (König et al., 2007) is also commonly used for SL calculation.

The methods for calculating SL from data generated by double-perturbation experiments, such as double-mutant yeast, combinatorial RNAi, or CRISPR screening, can be placed into two categories. The first approach introduces growth phenotype and calculates the deviation of the observed growth phenotype from the expected growth phenotype for a specified gRNA-gRNA pair. The growth phenotype is measured by the change in frequency of the initial and surviving populations for single-gRNA (gRNA-safe gRNA pair) or gRNA-gRNA (Wong et al., 2016; Han et al., 2017; Horlbeck et al., 2018; Najm et al., 2018; Parrish et al., 2021), and the expected growth phenotype is then calculated by summation (Han et al., 2017; Parrish et al., 2021; Thompson et al., 2021) or quadratic fitting (Horlbeck et al., 2018) of the growth phenotypes of two single-gRNAs. These two sgRNAs will be identified as synthetic lethal partners if the observed phenotype of the paired gRNAs is significantly lower than the expected value, suggesting a notable shift in gene essentiality with the presence of the two gRNAs. The gene–gene interaction can be calculated from the average (Horlbeck et al., 2018; Parrish et al., 2021) or ranking (Han et al., 2017; Thompson et al., 2021) for corresponding sgRNA pairs.

A different approach models the combination effect of double perturbation as a two-way analysis of variance (ANOVA) with interaction. Two perturbations are considered synthetic lethal if they lead to significant decline in individual fitness compared with their combined additive effect (Shen et al., 2017; Zhao et al., 2018). Individual fitness can be measured either as cell viability, the change in frequency of inserted gRNA fragment in the surviving cells over time (Shen et al., 2017; Zhao et al., 2018), or, as discussed in 2.1, by colony size in yeast experiments (Baryshnikova et al., 2010; Costanzo et al., 2016). A variational Bayesian approach (Ito et al., 2021) or Dunnett’s test (Zhou et al., 2020) can also be used to calculate SL in combinatorial CRISPR screening.

Largely built on the assumptions derived from the SL concept, rule-based inferences use statistical tests to infer synthetic lethal gene pairs at the population level from single-gene-based genome-editing screenings and multi-omics profiling data. Table 2 presents the biological assumptions, sources of input data, and statistical tests.

The data-mining synthetic-lethality-identification pipeline (DAISY) algorithm (Jerby-Arnon et al., 2014) is the first rule-based statistical-inference approach to identify and evaluate SL gene pairs from The Cancer Genome Atlas (TCGA) of the National Cancer Institute (NCI) and the National Human Genome Research Institute (NHGRI), the Cancer Cell Line Encyclopedia (CCLE), and DepMap. The first rule is called the genomic survival of fitness. Specifically, for a SL gene pair, A and B, it is assumed that if A is inactive (based on SCNA, gene expression, and somatic mutation), B will not be deleted or have a high copy number. The Wilcoxon rank test is used to compare the Gene B copy number between clinical samples with the active or inactive Gene A. The Wilcoxon rank test is also used with the second rule, to test whether cancer cells with inactive Gene A (based on SCNA and gene expression) are more likely than those with active A to have essential Gene B (based on shRNA screenings). For the third rule, a positive Spearman’s correlation coefficient is applied to investigate the co-expression of synthetic gene pairs A and B. Several published RNAi experiments in human cell lines have demonstrated the utility of DAISY.

Srihari et al. (2015) proposed a statistical method based on the concept of mutual exclusivity that assumed the likelihood that combinations of genes that exhibit mutual exclusivity in genetic events are synthetic lethal. They considered six key DNA-damage response (DDR) genes that are frequently altered across four cancer types (breast, prostate, ovarian, and uterine). The assessment of SCNA and gene expression of TCGA-identified genes that were altered in a mutually exclusive manner was based on a hypergeometric test with these six DDR genes as synthetic lethal partners. This model was validated against GARP essentiality scores from in vitro studies (Marcotte et al., 2012b; Vizeacoumar et al., 2013).

The identification of clinically relevant synthetic lethality (ISLE) (Lee et al., 2018) utilizes three criteria, employing a large initial pool of laboratory-identified candidate SL pairs determined either by double-knockout screens or guilt-by-association using large-scale single gene knock-out experiments as inputs. First, gene expression and SCNA data were used to identify candidate gene pairs whose co-inactivation was less frequent than expected as calculated using a hypergeometric test of their individual inactivation frequencies. Second, a gene pair was selected if its co-inactivation led to better predicted patient survival in TCGA samples according to the Cox proportional hazards model. In the last step, ISLE considered the tendency of functionally interacting genes to co-evolve and calculated phylogenetic similarity across 86 species in a tree of life structure using non-negative matrix factorization to select SL pairs comprising genes with strong phylogenetic similarity. Initial candidate gene pairs satisfying all three conditions were validated on other datasets (Barretina et al., 2012; Costello et al., 2014; Gao et al., 2015; Menden et al., 2019), and prediction performance was tested by phenotypic drug response screens in vivo.

Analysis of synthetic lethality by comparison with tissue-specific disease-free genomic and transcriptomic data (ASTER) (Liany et al., 2020a) predicted SL gene pairs for both cancerous tissues using SCNA from TCGA and disease-free tissues using gene expression data from the genotype-tissue expression project, GTEx, of the NHGRI. The main consideration of ASTER is whether cancer samples are tissue-specific when the gene-pair, A-B, exhibits a pattern of mutual exclusivity. Using disease-free tissues from GTEx as reference, in Test 1, ASTER selected disease samples with a high copy number in Gene A, and then compared the expression of Gene A in these samples to that in the reference GTEx samples. In Test 2, a subset of disease samples with a low copy number for Gene A and high copy number for Gene B was selected, and the expression of Gene B in this sample subset was then compared to that of the reference GTEx samples. Test 3, the final test, assessed whether the expression levels of Gene A were significantly higher than those of Gene B between two diseases. The Wilcoxon rank sum test was performed for each of the three tests and followed by Fisher’s method to combine the p-values. ASTER utilizes fewer datasets and has a simpler framework for hypothesis testing than DAISY and ISLE and outperformed those methods.

Synthetic lethal identification in R (SLIdR) (Srivatsa et al., 2019) is a statistical framework for identifying SL pairs from large-scale perturbation screens, including essentiality profiles from Project DRIVE with corresponding mutation and SCNA from CCLE. RSA was used to compute the gene-level essentiality score of each cell line (König et al., 2007) with a cutoff value of ˗3 in more than 50% of the cell lines. Driver genes for each cancer type were defined using the Broad Institute’s mutation significance file (MutSig MAF) from TCGA (Lawrence et al., 2014), focusing on genes demonstrating significant mutations in cancer samples; cell lines were sorted into mutated or WT groups based on the mutation status of the driver genes. SLIdR aims to identify synthetic lethal interactions between a driver gene and another perturbed gene based on a statistical test ranking essentiality scores across all perturbed genes for each mutated versus WT cell line. SLIdR identifies synthetic lethal gene pairs based on the assumption that a mutation in the driver gene in combination with knockdown of the perturbed gene yields lower essentiality scores compared to scores in the WT group. A one-sided statistical test based on the Irwin-Hall distribution is used to determine statistical significance. Of the potential synthetic lethal pairs identified by SLIdR, only one of the top synthetic lethal pairs, AXIN1 and URI1, was validated in vitro in this paper.

The algorithm for mining synthetic lethals (MiSL) (Sinha et al., 2017) extracts human pan-cancer data for 12 specific types of cancer from the TCGA dataset to identify mutation-specific synthetic lethal partners. Its underlying assumption is that a mutated gene’s synthetic lethal partners will be amplified more frequently or deleted less frequently in samples that harbor the mutation and concordant changes in expression across multiple cancers. MiSL aims to identify partners of gene B that have more copies in the presence of mutated Gene A based on Boolean implications of either preferred amplification in the presence of the mutation or deletion only in the absence of the mutation as determined using Fisher’s exact test and maximum likelihood estimation. Two filtering steps are applied afterward to increase accuracy. First, candidate genes serving merely as passengers are excluded. An example of a passenger is a deletion in Gene A that is not differentially down-regulated in samples with deletions in A compared to the rest of samples. Second, only genes that are differentially overexpressed in the presence of the mutation versus the WT based on a t-test are retained to form the final candidate SL partners. MiSL’s successful identification of SL interaction between mutation in IDH1 and ACACA in leukemia was validated by gene targeting and patient-derived xenografts.

Network models select single genes or gene combinations as potential drug targets based on the network’s topology. Though the criteria for selecting gene combinations are technically irrelevant to the concept of synthetic lethality, many gene pairs selected from network models are potentially SLs. Table 2 shows biological assumptions, input data sources, and network statistics for network models.

Virtual inference of protein activity by enriched regulon analysis (VIPER) (Alvarez et al., 2016) evaluates the functional relevance of genetic interactions in regulatory proteins based on gene-expression data from TCGA. VIPER requires accurate cellular networks that are highly dependent on tumor context (Margolin et al., 2006). Based on a probabilistic framework that includes target status (activated or repressed, with or without overlapping) and statistical confidence, VIPER applies an optimized rank-based analysis to compute the enrichment of a protein’s regulon in differentially expressed genes. VIPER is frequently used to infer aberrant protein activity from gene expression, and the correlation between regulator and target genes generated from the probabilistic framework in VIPER provides valuable information for context-specific gene–gene interactions and has potential use in SL prediction.

The optimal control node (OptiCon) algorithm (Hu et al., 2019) is a network controllability-based method to identify synergistic key regulators as candidate targets for combination therapy. OptiCon constructs a gene regulatory network from three pathway databases−the reactome pathway knowledge base (Jassal et al., 2020), the Kyoto Encyclopedia of Genes and Genomes (KEGG) (Kanehisa and Goto, 2000), and the NCI-nature pathway interaction database (Krupa et al., 2007)−and obtains gene-expression data of tumor tissues with matched normal tissues from TCGA for calculating gene deregulation scores. OptiCon first assesses a disease-perturbed gene regulatory network (DRN) to identify a set of optimal control nodes (OCNs) in a specified disease that controls the transition of the network between any two conditions. The identification of OCNs is formulated as a combinatorial optimization problem and is solved through a ‘greedy search’ algorithm. OptiCon then identifies synergistic OCN pairs by defining a synergy score that captures both enrichment of recurrently mutated genes (mutation score) and density of crosstalk between pathways (crosstalk score) controlled by a pair of OCNs. The synergistic pairs of OCNs predicted by OptiCon are supported by synthetic lethal interactions from the SynLethDB and the study by Shen’s group (Shen et al., 2017). The top predictions were validated experimentally by CRISPR screening (Han et al., 2017).

Supervised machine-learning models learn associations between input features and known SL data to predict novel SL gene pairs using multi-omics data. Table 3 shows how population features and omics features were generated for these machine-learning models.

The Mashup algorithm (Cho et al., 2016) involves the topological integration of multiple network types through graphic representation. For SL prediction, Mashup uses the STRING network for protein interactions (Szklarczyk et al., 2021), the Cancer Genome Project for drug-response profiles in cancer cell lines (Garnett et al., 2012), and the gene ontology (GO) (Ashburner et al., 2000) and Munich Information Center for Protein Sequences (MIPS) (Mewes et al., 2002) databases for functional annotation. A random walk with restart is employed to calculate the diffusion and connectivity of each data node within an individual network (Tong et al., 2006). During their integration, calculated gene features are minimized across networks to represent the topology of all networks. Afterward, the generated features along with graphic representation of the data networks are used in machine learning to predict the synthetic lethal interactions specified by Jerby-Arnon et al. (2014). The interactions are defined by the mean and absolute difference between the calculated feature representations across gene pairs and fitted by a support vector machine (SVM) using a programming library (LIBSVM) (Chang and Lin, 2011). The model’s prediction efficacy was validated by data on fifty cancer drugs with single-gene targets in over 639 cell lines obtained from the Cancer Genome Project (CGP).

Collective matrix factorization (CMF) (Liany et al., 2020b) is an unsupervised method that utilizes low rank factorization on design matrix inputs. The datasets used in CMF are represented in the matrix and include protein complex co-memberships from the comprehensive resource of mammalian protein complexes (CORUM) (Giurgiu et al., 2019), human PPI from the human integrated protein–protein interaction reference (Hippie) (Alanis-Lobato et al., 2017), co-expression profiles from the search tool for the retrieval of interacting genes/proteins database (StringDB) (Szklarczyk et al., 2021), and pathway co-membership scores calculated from Broad Institute’s molecular signatures database (MSigDB) (Subramanian et al., 2005). These datasets were factorized together to target the SL interactions data from the research groups of Laufer et al. (2013), Vizeacoumar et al. (2013), Shen et al. (2017), and Zhao et al. (2018), and from the SynLethDB database. Each input was integrated through similarly annotated rows and columns, and the CMF methods were implemented in three ways: by CMF (low-rank), gCMF (group-sparse CMF using group-sparse prior on columns), and dCMF (deep-learning CMF utilizing multiple auto-encoders for dimensionality reduction). The gCMF method performed the best for tasks inferring synthetic lethal interaction using principal component analysis and graphic features from DAISY. CMF was not validated further; the only reported validation was on the held-out datasets from five datasets used for training.

Synthetic lethality to logistic matrix factorization (SL²MF) (Liu et al., 2020) uses logistic matrix factorization to obtain latent protein factors for the prediction of SL pairs. The model’s design adopts a similarity-based drug-target interaction model named BLM-NII (-neighbor-based interaction-profile inferring) (Mei et al., 2013). SL²MF uses GO semantic similarity (Ashburner et al., 2000) and PPI data from the human protein reference database (HPRD) (Keshava Prasad et al., 2009) to bolster the model’s predictions in the networks’ topologies. The representative latent gene factors were used in a logistic function to predict SL pairs. The SL data were obtained from SynLethDB and used in 5-fold cross-validation. The SL pairs used as positive training samples exclude the pairs predicted from DAISY and high-scoring SynLethDB pairs. The model’s performance achieved an AUC of 0.85 and an AUPRC of 0.24. The model was validated in silico and by comparison with DAISY in overlapping SL predictions within the SynLethDB.

The graph-regularized self-representative matrix factorization (GRSMF) (Huang et al., 2019) model represents a matrix by a linear combination of its rows and columns using SL interactions as the input matrix. In the process, the model is bolstered by graph regularization with GO semantic similarity (Ashburner et al., 2000) and uses a majorization-minimization objective function (Yang and Oja, 2011) in its training The model is applied to SL data from SynLethDB with 5-fold cross-validation and its performance is compared with BLM-NII (Mei et al., 2013), SL²MF (Liu et al., 2020), and SMF (GRSMF without graph regularization support). AUC scores demonstrated that GRSMF (0.92) and SMF (0.89) outperformed both SL²MF (0.85) and BLM-NII (0.74). The model was validated in silico.

The ensemble-based machine-learning model (ESML) (Lu et al., 2015) uses multiple classifiers and multi-omics datasets, including RNA sequencing data generated by the Broad Institute’s firehose suite of tools and pipelines (Tomczak et al., 2015) and SCNA data from the cBioPortal for cancer genomics (Gao et al., 2013), to define gene-pair interaction features, namely homozygous co-loss, heterozygous co-loss, mixed co-loss, co-underexpression, and expression up-down signals. The co-loss signals are derived from gene deletion profiles in SCNA data, whereas co-expression profiles are computed from the RNA-seq data. The model consists of seven different classifiers: the adaptive boosting (AdaBoost) algorithm (Freund and Schapire, 1997), the J48 algorithm (Salzberg, 1993), JRip, a Java-based implementation of the RIPPER algorithm (Cohen, 1995), the Logit function (Lu et al., 2015), the LogitBoost boosting algorithm (Friedman et al., 2000), partial decision trees (PART) (Frank and Witten, 1998), and the random forest algorithm (Breiman, 2001). The same gene-pair interactions were fed into each of the classifiers, and the outcomes showing greatest agreement across classifiers were chosen. The framework is then applied to the synthetic lethal pairs from Laufer’s (Laufer et al., 2013) and Vizeacoumar’s groups (Vizeacoumar et al., 2013) to generate population-based and genome-wide-scale SL interactions. Under a probability threshold of at least 0.81, the model achieved a precision score of 0.67 and recall score of 0.10.

The dual-dropout graph convolutional network (DDGCN) (Cai et al., 2020) is the first graph neural network (GNN) model to predict SL gene pairs. DDGCN proposes a novel dual-dropout mechanism to solve the problem of overfitting associated with the sparsity of SL. Known SL gene pairs are used to construct a synthetic lethal interaction network in which each gene is a node and SL interactions form edges, which allows the prediction of novel SL to be cast as a link prediction task to complete missing edges in the interaction network. The dual-dropout consists of a coarse-grained node dropout that randomly drops some gene nodes during each training iteration, and a fine-grained edge dropout that randomly removes some edges for further fine-tuning. DDGCN has been theoretically justified and validated utilizing the SynLethDB database, with a predicted AUC of 0.85 and AUCR of 0.90.

The graph contextualized attention network to synthetic lethality (GCATSL) (Long et al., 2021) is another GNN-based model that incorporates various biological data sources utilizing graph attention network (GAT) architecture. Compared to a basic GNN model, GAT can effectively distinguish and preserve differences among neighbors by assigning different weights. In GCATSL, three feature graphs were constructed using as input features, biological processes (BPs) and cellular components (CCs) from GO as well as the PPI network from the biological general repository for interaction datasets (BioGRID) (Stark et al., 2006), and a dual-attention mechanism (node- and feature-level attention) that is designed to learn node representations from multiple feature graphs. Specifically, node-level attention was used with GAT to learn preliminary representations for each input feature graph, and feature-level attention was then implemented to integrate these three feature graphs and generate the final representation for each gene node. Prediction performance was validated on the SynLethDB database, with prediction AUC of 0.94 and AUCR of 0.95.

The knowledge graph for synthetic lethality (KG4SL) (Wang et al., 2021) is a GNN-based method that incorporates a knowledge graph (KG) into the prediction of SL. The authors highlighted that existing methods often regarded each SL pair as an independent sample and failed to consider the underlying biological mechanisms; whereas some shared biological factors might latently imply dependency among SL pairs. In contrast, KG4SL considers knowledge graphs involving biological processes, diseases, and compounds. Given the heterogeneous input graph, KG4SL utilizes an attention mechanism to handle the passing of messages of different types of nodes and edges. The inner product between the representations of the gene pair is regarded as the probability of being SL. KG4SL has shown excellent performance in the SynLethDB database, yielding an AUC of 0.95 and AUCR of 0.96.

These individual sample-based models are all supervised approaches. Table 3 shows how their input features are defined.

The multiple network decision tree (MNDT) model (Wong et al., 2004) utilizes a decision tree classifier to predict SL interactions. Prior to training on the decision trees, the gene pairs are given manually curated features to depict their genetic interaction networks. Data for the networks were obtained from various sources, including MIPS (Mewes et al., 2002) for functional relations and Goldberg and Roth’s physical interaction network (Goldberg and Roth, 2003). A total of 123 gene-pair characteristics, comprising common upstream regulators, gene co-occurrence, and chromosomal distances between genes, were compiled in the genetic network. Interactions of these gene-pairs were then extended to a third gene, designated 2hop that interacts with the other two. For instance, if gene C has a physical interaction with Gene A in one network and synthetically lethal interaction with Gene B in another, then the A-B gene pair is assigned a 2hop-physical-SL characteristic. According to the curated features, the gene pairs were fed into the decision tree to decide which leaf node the input should land on and the location of the node containing the SL classification prediction. Then, the features were trained on SGA-analyzed data from Tong et al. (Tong et al., 2001) and fitted to SL data obtained from an early version of Tong et al. (2004) for validation.

The multi-network and multi-classifier (MNMC) (Pandey et al., 2010) model was extended from the MNDT (Wong et al., 2004) model utilizing multiple classification techniques, including k-nearest neighbors, neural network, random forest, and SVM in addition to decision trees. Individual-network and overlaid-network features, including gene expression, protein–protein interaction, transcription factor binding, and functional annotation profiles from GO and KEGG, were generated from the yeast dataset to predict synthetic lethal interactions. A total of 152 individual network features were identified, 62 of which hinted at stronger connections, such as physically interacting genes within the PPI network. Additionally, 90 overlaid features were generated via 2hop interactions across the networks as described in MNDT. This allowed the model to capture repeated and similar interactions across different network types and to create an integrated representation input dataset. The Kolmogorov-Smirnov test was applied to determine the top features against distinguishing synthetic lethal and non-synthetic-lethal gene pairs, and the model was then trained using synthetic lethal interactions from the Saccharomyces genome database (Cherry et al., 1997), as described by Wong et al. (2004). MNMC was validated in silico within the training datasets.

The DiscoverSL (Das et al., 2019) model uses clinical and multi-omics data from TCGA (Tomczak et al., 2015) and data from the MSigDB pathway annotation database (Subramanian et al., 2005) to predict synthetic lethal interactions in the SynLethDB. Feature sets of differentially expressed genes, expression correlations, mutual exclusivity, and pathway information were calculated, and a random-forest classifier created with these four features was trained on SL interactions. After the model was trained on SynLethDB, the derived patient-specific SL interactions were validated in silico by visualizing shRNA essentiality screens, SCNA targetability, cell-line drug sensitivity data, and Kaplan-Meier survival curves against different gene expression profiles. This model was further validated in silico on cell lines from the genomics of drug sensitivity in cancer (GDSC) database (Yang et al., 2013). Other studies, such as that of Origanti and associates regarding CHEK1 and p21 (Origanti et al., 2013), confirmed some of the predicted SL interactions.

Expression to synthetic lethality prediction, EXP2SL (Wan et al., 2020), is a machine-learning network based on a semi-supervised neural network. EXP2SL was used to extract gene-expression profiles from the L1000 project of the library of integrated network-based cellular signatures (LINCS) of the National Institutes of Health’s Common Fund (Subramanian et al., 2017), apply a multi-layer fully connected neural network to individually encode the profiles for each input gene pair, and then concatenate the encoded representations to make the final prediction. Because synthetic lethal labels for an individual sample are limited, EXP2SL designs a semi-supervised Bayesian personalized ranking (BPR) loss into the objective function to incorporate a large amount of unlabeled data. Testing of the model on the combinatorial CRISPR SL datasets in three different cell lines (Shen et al., 2017; Najm et al., 2018; Zhao et al., 2018) demonstrated its competitive prediction ability.