Explainable/interpretable artificial intelligence and knowledge graph neural network for synthetic lethality prediction: overview, challenges, and future directions Authors: Salsabil Felloussi, Fatima El Haoussi, Hamid Tairi, Noura Aherrahrou Published: 2026 ABSTRACT AbstractSynthetic lethality (SL) provides a powerful foundation for targeted oncology by identifying gene pairs whose simultaneous disruption is fatal to cancer cells while sparing healthy tissue. However, mapping these interactions experimentally across the human genome is restricted by massive combinatorial scale and high costs. To address this, computational approaches specifically graph neural networks (GNNs) have become the standard for modeling these complex biological relationships. Hence, by integrating heterogeneous knowledge graph (KG) and multi-omics data, GNNs can effectively capture the topological and molecular context required for accurate SL prediction. This review systematically evaluates the computational SL landscape across three core dimensions. First, we examine how diverse GNN architectures integrate multi-layered biological data. Second, we address model interpretability, contrasting intrinsically transparent architectures with post-hoc explanation methods, as a requirement for clinical trust. Finally, we explore the emerging role of Generative AI and large language models to leverage their zero-shot reasoning and broad biomedical knowledge.Keywords: synthetic lethality, knowledge graph, interpretability, explainability, LLMs, generative AI Introduction Synthetic lethality (SL) refers to a functional interaction between two genes (Fig. 1), where the simultaneous inactivation of both genes leads to cell death, whereas the inactivation of either gene alone produces a viable phenotype [1, 2]. The concept of SL was first discovered back in 1946 in Drosophila by Dobzhansky [3]. To date, it has been widely explored especially for cancer targeted therapeutics. In a cancerous cell, functionally disruptive mutation in one of the two genes of an SL pair leads to dependency on the second one for survival and malignant cells can be selectively killed by inhibiting the unmutated one. One of the most notable examples of SL is the selective vulnerability of cancers with mutations in the Breast Cancer Susceptibility Genes (BRCA1 and BRCA2) to inhibitors of polyADP-ribose polymerase (PARP). These BRCA-mutated tumors rely heavily on PARP-mediated base excision repair for DNA damage, a vulnerability that PARP inhibitors exploit [4]. Currently, PARP inhibitors have been approved for treating ovarian, breast, prostate, and pancreatic cancers that are associated with BRCA1/2 mutations [5, 6]. Following the example of the BRCA/PARP pair, other potential SL gene pairs were studied. Prior methods for SL are essentially wet-lab methods. However, common laboratory techniques including, yeast screening, drug screening, RNA interference (RNAi) screening, and clustered regularly interspaced short palindromic repeat (CRISPR) screening [7] are prone to some limitations [8]; yeast screening is restricted by the fact that only a small number of yeast genes have human counterparts; RNAi screening suffers from inconsistency across different cell lines and off target effects; and CRISPR-based genome editing, while powerful, is costly and also prone to off-target effects [9]. Moreover, given the vast number of possible gene combinations, it is impractical to screen all potential synthetic lethal pairs using solely laboratory-based methods [10]. Earlier computational approaches, including either machine learning (ML)-based like K-Nearest Neighbors (KNN) [11], Support Vector Machine (SVM) [12], or deep learning (DL)-based like GRSMF [13], DDGCN [14], and SLMGAE [15] can help overcome the challenges of wet-lab methods . Nevertheless, each method has its pros and cons. ML-based methods predominantly are restricted by complex feature engineering that are often handcrafted which hindered their capabilities [8, 16, 17]. Moreover, learning distinctive feature representation is still a complex and challenging task in artificial intelligence (AI)-based models [16]. To address these issues, more sophisticated AI-based methods that are based on graph-modeling and powered by biomedical knowledge have gathered an increasing interest among research community, and have emerged as promising tools very recently. Yet, these approaches mainly leverage the GNNs and their graph structure to generate insights that have been difficult to capture through traditional computational methods [16]. Notably, biomedical knowledge graph (KG) and AI have revolutionized SL research and our capabilities to understand complex diseases by incorporating data from diverse sources, hence, offering a powerful solution for capturing hidden relationships among entities (e.g. genes and diseases) and providing valuable insights for SL research [10, 18]. Besides, as AI-based technological innovations continue to evolve, the aim remains to leverage these advanced tools for more accurate, personalized, and effective precision medicine. In fact, LLMs have further revolutionized AI field, demonstrating superior performance in analyzing complex, and multidimensional data. However, alongside these powerful advances comes another growing concern: the models often act as “black boxes,” delivering impressive predictions with minimal insight into “why” and “how” those predictions are made [19]. In response to this, interpretability, the ability to explain the inner workings of AI-driven technologies, is of ultimate importance, not only for uncovering the biological mechanisms underlying traits and diseases, but also for contributing to the model’s trustworthiness. Additionally, explainable artificial intelligence (XAI) can help bridge this gap by incorporating “explainability” independently from the inner workings of the AI-based frameworks, by assuming that the model is black-box [20]. Thereafter, our contribution in this work consists of analyzing the explainability/interpretability of AI-based frameworks for SL prediction, focusing on those using KG. Thereby, bridging the gap between AI-based models and real-world applications of SL in precision medicine [21]. To the best of our knowledge, this is the first review that uncovers the use of KG as well as the explainability/interpretability aspects of the existing SL predictive models. Throughout our study: Synthetic Lethality interaction between Gene A and Gene B. Left panel : Inhibiting Gene A or Gene B individually is non-lethal; cells tolerate single-gene disruptions. Right panel : Combined loss of Gene A and Gene B induces cell death (Synthetic Lethality). We introduce existing SL KG databases. We introduce existing SL KG databases. We analyze the interpretability/explainability aspects of SL predictive models. We analyze the interpretability/explainability aspects of SL predictive models. We highlight the implication of KG, omics, and cell-line-specific data as a recent shift in SL research. We highlight the implication of KG, omics, and cell-line-specific data as a recent shift in SL research. We shed light on the importance of LLMs and Generative AI as promising solutions to boost the findings. We shed light on the importance of LLMs and Generative AI as promising solutions to boost the findings. We emphasizes the limitations of the reviewed approaches and suggest future possible directives to improve SL research. We emphasizes the limitations of the reviewed approaches and suggest future possible directives to improve SL research. Knowledge graph overview A KG is a semantic network that represents entities and their relationships, by integrating heterogeneous data into a structured ontology [22]. It applies reasoning to derive new insights, making it a powerful tool in AI for enhancing knowledge-based applications [23]. It is composed of entities that are representations of real-world objects and relationships modeling correlations between these entities [22]. Generally, building KG involves three steps: (i) extraction of knowledge from information sources, (ii) establishing appropriate representation models, and (iii) in-depth knowledge mining [24]. Knowledge extraction is a critical step of KG construction as it directly influences subsequent knowledge process [25]. It depends highly on the databases used for information extraction among other factors. The key processes for knowledge extraction involves entity extraction which identifies entities from different sources, relation extraction which discovers the relationships between these entities and attribute extraction that identifies descriptive characteristics of entities. Approaches to knowledge extraction integrate natural language processing (NLP), text mining, and ML. Recently, a few DL methods have also been used to automate the process of knowledge extraction, such as BERT [26], BiLSTM-CRF [27], DeepKG [28], and CNN [29]. Other KG construction methods include knowledge fusion that integrates information from different sources, primarily through entity alignment, and ontology construction [23, 30]. Entity alignment also known as entity resolution or entity matching determines if different entities refer to the same real-world object [23]. This involves textual similarity and structured similarity functions and often requires external information. Ontology construction involves creating a taxonomy, building hierarchical structures, and adding metadata from various data sources [15]. KG modeling is also part of basic techniques for KG construction. It involves selecting appropriate representations, such as graph databases (e.g. Neo4j), that can efficiently handle the complex relationships and large volumes of data typical in KGs. This storage solution allows for rapid querying and retrieval of interconnected data, facilitating dynamic updates and scalability [31]. In addition to KG completion, that is the process of predicting new or missing facts based on existing facts, ensuring that the KG is complete [32]. Thereafter an overview of some SL KG databases used for Knowledge extraction, summarized in Table 1. Table 1KG databases: data quality, heterogeneity, and noise analysis.KG/DatabaseKG/Data qualityHeterogeneityNoiseSynLethDB − 4 source classes; 293 PubMed papers− 35 943 human SL pairs, 5 species− High-throughput entries score 0.50 only− MySQL relational; no graph reasoning− Weighted confidence score per method− Gene-ID harmonization via NCBI Entrez − SL, non-SL, synthetic rescue types− Multi-species, cell line as metadata− No pathway or drug entity layer − KRAS hub-gene enrichment bias− Method-agnostic CRISPR scoring− No cell-line context structuring SynLethKG − 11 entity types, 27 relation types− Neo4j,37 341 nodes; 1.4 M edges− High sparsity− Extends SynLethDB with Hetionet/DrugBank− GO, pathway, anatomy links added− No independent SL pair validation − Rich schema: clinical + molecular− Enables drug–gene–cancer multi-hop paths− SDL and rescue as distinct edges − Inherits SynLethDB noise− KG layer integrity unvalidated− Cell line not a graph dimension SLKG − 3 entity types; 4 relation types− Includes SDL− Application-specific − Gene, drug, cancer only− 19 987 SL, 3039 SDL pairs− No GO, pathway, or anatomy layer − Inherits SynLethDB SL noise− Drug name inconsistency across DBs− SDL evidence too sparse, weak score SLKB − 11 CDKO studies; raw sgRNA counts kept− Corrects SynLethDB cell-line errors− 16 059 SL, 264 424 non-SL pairs − 22 cell lines, diverse CRISPR platforms− No ontological entities beyond gene pairs− CDKO-only, excludes RNAi and biochemical − 1.21% top-10% overlap across 5 methods− 22% SL pairs conserved cross-cell-line− Method choice dominant noise source NexLeth − 707 pairs; GPT-4 + expert annotation− 79.5% hypothetical (Mhypo) unvalidated− Augmented via ChatGPT paraphrasing − Linguistic tiers (Mfact/Mhypo) only− No drug, cancer, or pathway entities− 707 pairs coverage not comprehensive − LLM confabulation risk− Keyword filter insufficient for validity− Inherits all SynLethDB biases KG databases: data quality, heterogeneity, and noise analysis. − 4 source classes; 293 PubMed papers − 4 source classes; 293 PubMed papers − 35 943 human SL pairs, 5 species − 35 943 human SL pairs, 5 species − High-throughput entries score 0.50 only − High-throughput entries score 0.50 only − MySQL relational; no graph reasoning − MySQL relational; no graph reasoning − Weighted confidence score per method − Weighted confidence score per method − Gene-ID harmonization via NCBI Entrez − Gene-ID harmonization via NCBI Entrez − SL, non-SL, synthetic rescue types − SL, non-SL, synthetic rescue types − Multi-species, cell line as metadata − Multi-species, cell line as metadata − No pathway or drug entity layer − No pathway or drug entity layer − KRAS hub-gene enrichment bias − KRAS hub-gene enrichment bias − Method-agnostic CRISPR scoring − Method-agnostic CRISPR scoring − No cell-line context structuring − No cell-line context structuring − 11 entity types, 27 relation types − 11 entity types, 27 relation types − Neo4j,37 341 nodes; 1.4 M edges − Neo4j,37 341 nodes; 1.4 M edges − High sparsity − High sparsity − Extends SynLethDB with Hetionet/DrugBank − Extends SynLethDB with Hetionet/DrugBank − GO, pathway, anatomy links added − GO, pathway, anatomy links added − No independent SL pair validation − No independent SL pair validation − Rich schema: clinical + molecular − Rich schema: clinical + molecular − Enables drug–gene–cancer multi-hop paths − Enables drug–gene–cancer multi-hop paths − SDL and rescue as distinct edges − SDL and rescue as distinct edges − Inherits SynLethDB noise − Inherits SynLethDB noise − KG layer integrity unvalidated − KG layer integrity unvalidated − Cell line not a graph dimension − Cell line not a graph dimension − 3 entity types; 4 relation types − 3 entity types; 4 relation types − Includes SDL − Includes SDL − Application-specific − Application-specific − Gene, drug, cancer only − Gene, drug, cancer only − 19 987 SL, 3039 SDL pairs − 19 987 SL, 3039 SDL pairs − No GO, pathway, or anatomy layer − No GO, pathway, or anatomy layer − Inherits SynLethDB SL noise − Inherits SynLethDB SL noise − Drug name inconsistency across DBs − Drug name inconsistency across DBs − SDL evidence too sparse, weak score − SDL evidence too sparse, weak score − 11 CDKO studies; raw sgRNA counts kept − 11 CDKO studies; raw sgRNA counts kept − Corrects SynLethDB cell-line errors − Corrects SynLethDB cell-line errors − 16 059 SL, 264 424 non-SL pairs − 16 059 SL, 264 424 non-SL pairs − 22 cell lines, diverse CRISPR platforms − 22 cell lines, diverse CRISPR platforms − No ontological entities beyond gene pairs − No ontological entities beyond gene pairs − CDKO-only, excludes RNAi and biochemical − CDKO-only, excludes RNAi and biochemical − 1.21% top-10% overlap across 5 methods − 1.21% top-10% overlap across 5 methods − 22% SL pairs conserved cross-cell-line − 22% SL pairs conserved cross-cell-line − Method choice dominant noise source − Method choice dominant noise source − 707 pairs; GPT-4 + expert annotation − 707 pairs; GPT-4 + expert annotation − 79.5% hypothetical (Mhypo) unvalidated − 79.5% hypothetical (Mhypo) unvalidated − Augmented via ChatGPT paraphrasing − Augmented via ChatGPT paraphrasing − Linguistic tiers (Mfact/Mhypo) only − Linguistic tiers (Mfact/Mhypo) only − No drug, cancer, or pathway entities − No drug, cancer, or pathway entities − 707 pairs coverage not comprehensive − 707 pairs coverage not comprehensive − LLM confabulation risk − LLM confabulation risk − Keyword filter insufficient for validity − Keyword filter insufficient for validity − Inherits all SynLethDB biases − Inherits all SynLethDB biases NexLeth NexLeth [33] is the newest SL KG dataset. Based on fine-tuning of pretrained large language models (LLMs), and built up on prompt-based queries from New Bing, NexLeth generates natural language explanations of SL gene pairs [34]. It comprises 707 gene pairs and their explanations. Accordingly, NexLeth offers a standard for evaluating SL explanation methods and supports the deeper understanding of SL mechanisms [33]. Synthetic lethality knowledge base SLKB is another SL knowledge database [35]. Although it does not explicitly integrate a KG, it specifically incorporates combination gene double knockout (CDKO) data from 11 published experiments across 22 cell lines, totaling 16 059 SL gene pairs and 264 424 non-SL gene pairs. These experiments use various CRISPR-Cas9-based methods to explore synthetic lethal interactions between genes, with data sourced from in vitro studies and computational analyses. SLKB stores the sgRNA counts, library sequences, and original gene-pair SL scores of the original studies. It uses five different SL scoring methods: Median with/without background normalization, sgRNA-derived with/without background normalization, Horlbeck score, GEMINI score, and MAGeCK score with 1,21% prediction overlap among them. SLKB houses a relational MySQL database containing both original and new scores. It is hosted as an R Shiny web application using R v.4.2.1 and Shiny v.1.7.3. The knowledge base offers an overview of the SLKB and the latest updates, a table for browsing SL-pair data from CDKO studies, various network visualization formats, a page for downloading both processed and raw data files, a references page, and the option to compare SL pairs across different SLKB scoring methods using Venn diagrams. Synthetic lethality knowledge graph The synthetic lethality knowledge graph (SLKG) [36] integrates data from multiple biomedical databases to comprehensively represent SL and synthetic dosage lethality (SDL) interactions. It consists of three primary entity types: are genes, drugs, and cancer types, in addition to four relationship types, including SL/SDL gene pairs, cancer-mutant gene associations, drug–target gene interactions, and drug indications. Data are collected from large-scale biomedical databases, including SynLethDB [37], DAISY [38], DRIVE DATA PORTAL [39], and DepMap [40] provide SL/SDL pairs. For cancer–gene relationships DisGeNET [41], ONGene [42], TSGene [43], and COSMIC [44] are solicited; while TTD [45], DrugBank [46], and DGIdb [47] are employed for drug-related information extraction. Entities are preprocessed, and their nomenclature is standardized using PanDrugs [41] and PubChem Identifier Exchange Service. SLKG implements a drug repositioning scoring schema, integrating SLScore/SDLScore, DrugScore, and CancerScore. These metrics evaluate SL/SDL interaction reliability, drug development status, and cancer–gene associations, respectively. Scores undergo min-max normalization with thresholds applied to identify top candidates. Moreover, SLKG supports drug combination identification using an integer programming algorithm, optimizing therapeutic potential while minimizing drug antagonism. The KG is implemented on the Neo4j platform, facilitating efficient querying and analysis of complex biological relationships. SynLethKG SynLethKG is a KG integrated into the SynLethDB 2.0 [48] database. The database includes 50 868 SL pairs where 35 943 pairs are of humans, 381 of mice, 439 of fruit flies, and 14 000 are derived from yeast. Data are collected from multiple sources. By reviewing 293 recent research papers published on PubMed, SL pairs were manually curated focusing on SL interactions identified through wet-lab experiments. Data are also extracted from established public databases such as GenomeRNAi [49] and BioGRID [50]. Finally, by applying computational predictions using methods like GEMINI [51], SL pairs from combinatorial CRISPR screens are identified. SynLethKG incorporates 11 biomedical entities like genes, drugs, cancers, pathways, and GO terms among others. It also integrates 27 types of relationships including gene–gene interactions, gene–disease associations, and drug–target interactions (DTI). Entities and relationships are assembled using Python scripts from open-source projects like Hetionet [52]. Relationships from various sources are integrated into the Neo4j graph database to store SL pairs and related biomedical knowledge. Relationships in the KG are stored as edges between nodes, with properties such as species, PubMed references, and confidence scores attached to these edges. Confidence scores for SL pairs are derived through a quantification and integration process. First SL pairs are assigned individual scores based on their identification method, with higher weights given to more robust techniques (e.g. CRISPR interference) compared with high-throughput screening approaches. Then, the obtained weights are combined to produce a normalized confidence score. SynLethKG provides a nuanced assessment of SL pair reliability, facilitating more informed analysis and interpretation of the data. Interpretability/explainability The AI-based models, specifically deep neural networks, have introduced solutions to many challenging problems [53]. However, with these high-level solutions arises the black-box issue that translates the complexity of the mechanisms underlying the models’ prediction. Thus, in spite of solutions that offer impressive accuracy, AI-based models often struggle with transparency and interpretability. In response to this, several emerging approaches and computational methodologies aiming to make AI-based models more understandable for humans and trustable, are systematically organized through different taxonomies in recent surveys [54]. Within this context, Kakkad and his co-workers have distinguished between post-hoc methods (explaining pretrained models) and self-interpretable (ante-hoc) methods that build explainability directly into the architecture [55]. In the same line, Adadi and her collaborators have differentiated between complexity-related methods, scoop-related methods, and model-related methods [56]. The former ones improve the interpretability of models, either by designing an algorithm that is intrinsically interpretable [57], or by using post-hoc methods, a separate group of techniques that provide explanation independently from the inner works of the original model, such as SHAP and LIME [58–61]. Besides, Jiang et al. [62] have distinguished between approaches that generate post-hoc explanations and approaches that are inherently interpretable. Whereas, Yuan et al. [53] have distinguished between inherent interpretable models and explainable models that could be understood by post hoc explanation techniques. Yet, the term “explainability” denotes the rationale behind the model’s outcomes, and “why” the model generate such predictions [19, 20, 61, 63]. In this way, the AI-based model tend to generate explanations after the training process of the model [64]. This is achieved through using the so-called post-hoc techniques [53]. Hence, by doing so, an explainable model is still black box (nontransparent), but its predictions are human-understandable [65, 66]. Typically, in the context of KG for SL prediction, post-hoc techniques involve assessing the importance of each component in the graph [67]. To achieve this, several post-hoc explainability techniques, such as GNNExplainer, gradient-based XAI, Sensitivity analysis, perturbation are employed, to uncover the most significant components of the input graph (nodes, edges, or subgraphs) that contribute the most to a given prediction [68, 69]. Besides, the term “interpretability” denotes the “ how” AI-based model function [63]. Specifically, it refers to methods that make the inner working of an AI-based model transparent [19, 20, 61, 66]. In this way, the model, is no longer a black box (transparent), at least to some extent [53, 65, 66]. Based on this, self-interpretable methods tend to generate interpretability during the training process of the model [20, 64, 69]. To highlight the importance of each component within the model, most of techniques are built upon attention mechanisms that have the ability to focus on the most pertinent components (nodes or edges) [67–69]. Thereafter, based on this categorization, we divide our reviewed models into inherently interpretable (ante-hoc) and post-hoc approaches. Review methodology The present review was conducted following a structured literature selection in order to identify studies relevant to GNN- and KG-based SL prediction. We performed a systematic search across Google Scholar, Pubmed, IEEE Xplore, and Springer Nature databases, using the following key-words: “Synthetic lethality,” “GNN,” “Knowledge graph,” “XAI,” “Interpretability,” “omics,” “multi-omics,” and cancer. The search was limited to articles published between 2021 and 2026 translating the and fast expansion of DL-based SL prediction methods in this period. We included studies that (i) proposed or evaluated a computational model for SL interaction prediction, (ii) incorporated GNN architectures and/or KG representations as a core methodological component, (iii) reported quantitative performance metrics (AUROC, AUPR, or F1-score), (iv) addressed model interpretability or explainability (v) integrated heterogeneous/omics data types. We excluded studies that: (i) were limited to wet-lab experimental SL screens without a computational prediction component, (ii) relied exclusively on nongraph ML approaches, (iii) did not report reproducible performance evaluations, or (iv) were review articles or conference abstracts. After deduplication and screening of titles and abstracts, 29 studies were selected for full-text review. The resulting corpus was curated to capture the breadth of methodological approaches, spanning diverse graph-based architectures, heterogeneous data modalities, and frameworks for interpretability. This allowed us to tackle the explainability aspect of SL prediction, and to highlight the contribution of KG and omics to the SL models performance and interpretability. Tables 2– 6 summarize the selected approaches detailing the name, publication year, used data, model, and interpretability of these methods. Tables 2– 6 summarize each selected study, detailing the model name (where applicable), publication year, data types integrated, architecture employed, and interpretability approach adopted. This structured selection ensures that the comparative analysis presented herein is grounded in methodologically coherent and reproducible evidence. Table 2A summarized overview of SL prediction models.ApproachYearKGMulti-omicsCell-specificGenerali zationModelInsightsLimitations Ante-hoc Post-hoc DatasetSLMGAE2021 × × × × Multiview GAE + Attention− Minimization of graph reconstruction loss via attentive merging.− Random negative sampling. − Limited Knowledge sources − Lack interpretability and generalization × × SynLethDBGCATSL2021 × × × × GAT−Dual attention mechanism−Limited knowledge sources − Computational complexity − Lack interpretability × × SynLethDBKG4SL2021✓ × × × KGNN− First known model to integrate a heterogeneous KG.− Random negative sampling − Lack of interpretability × × SynLethDB SynLethKGPiLSL2022✓✓ × ✓R-GCN + Attention− Generalizabilty − Explanation of SL mechanism via the weighted paths in enclosing graphs by the attention mechanism− Random negative sampling − Absence of cell line context✓SynLethKG[16]2023✓ × × × R-GCN + Attention− Self-attention mechanism to prune out the KG.− Random negative Sampling − Overlooking other discriminative data for training✓SynLethDBKR4SL2023✓ × ✓✓RED-GNN + Recurrent Neural Network (GRU)− Introduction of relational digraph reasoning − Fused textual-structural message propagation − Interpretability− Dependence on KG completeness − Absence of cell line context − Random negative sampling✓SynLethKGSLGNN2023✓ × × × KGNN + GAT− Modeling genes preferences for different SL-related factors in the KG. − Considering the importance of factors.− Gene representations are interconnected, which decreases the model’s accuracy. − Random negative sampling✓SynLethKG A summarized overview of SL prediction models. Table 6A summarized overview of SL prediction models.ApproachYearKGMulti-omicsCell-specificGenerali -zationModelInsightsLimitations Ante-hoc post-hoc DatasetStruct2SL2025 × × × × Multi-layer perceptron (MLP)− The potential of integrating protein structural features, sequence data, and PPI networks features on the efficacy of the Struct2SL modeL− Random negative Sampling − Limited Interpretability × × SynLethDBEFOL-SL2025✓ × × ✓GAT + Trans-former− Interpretable complex knowledge multi-hop reasoning model based on First-Order Logical query for SL prediction. − Type-Aware Negative Sampling to prevent false-negative classes, and class imbalance issue− Scalability of the model − Sensitivity to Natural Language understanding − Data sparsity✓SynLethDBSL-MERK2025✓ × × ✓LLMs+ Explainable GNN− Combining the explanations generated by RAG (Retrieval-Augmented Generation) and KG for SL mechanisms. − LLMs model fine tuning− Sensitivity to Natural Language understanding✓NexLeth SynLethKGSynLeaf2026✓✓ × ✓RGCN + VAE− UMT/UME dual training with adaptive selection − Per-cancer type SL prediction− Lack of interpretability − Irrealistic 1:1 class balancing × × SynLethKGMGANSL2026 × × × × GAN− Adversial alignment-orthogonality constraint to prevent information leakage− Random negative sampling − Absence of semantic Knowledge − Lack of interpretability × × SynLethDB A summarized overview of SL prediction models. Table 3A summarized overview of SL prediction models.ApproachYearKGMulti-omicsCell-specificGenerali-zationModelInsightsLimitations Ante-hoc Post-hoc DatasetSLwise2023 × ✓✓✓Graph-SAGE+ Trans-former− Prediction of cell-specific SL interactions. − Ability to generalize on new cell lines.− High context dependency − Computational complexity✓SynLethDBMSGT-SL2023 × ✓✓ × Multi-view GNN + Trans-former− multi-view GNN to learn nonlinear relationships among genes − omics data for enhancing generalizability.− Lack of interpretability − Computational complexity × × SLAD-CE2023 × × × ✓-− Analyzing SL interactions in specific cellular contexts − Identifying the associations between cellular damage and candidate SL genes abnormalities.− Limited availability of positive SL labels for various cell lines − Lack of gene regulation, cell–cell interactions and protein–protein interaction (PPI) networks✓SynLethDBHLASL2024✓ × × × Hyperbolic GNN+ Lorenz Graph Attention− Mapping information in hyperbolic space reduces distortion in data. − Lorentzian Attention mechanism to capture the hierarchical structure of data.− Computational complexity − Limited interpretability. × × SynLethKGMLEC-iSL2024 × ✓✓✓GCN+ Graph Trans-former− Cell-specific multi-omics data − Generalizability.− Computational complexity.✓SynLethDBMPASL2024✓ × × ✓KGNN+ Attention− Introducing a discrepancy contrastive layer for better gene representations.− Increased complexity of the model. − Limited scalability due to the use of large datasets. − Limited interpretability × × SynLethDB SynLethKG A summarized overview of SL prediction models. Table 4A summarized overview of SL prediction models.ApproachYearKGMulti-omicsCell-specificGenerali-zationModelInsightsLimitations Ante-hoc Post-hoc DatasetBioPath-Net2024✓ × × ✓Neural Bellman- Ford Network (NBFNet)-based GNN− Interpretability of the results. − Type-aware stringent negative sampling to increase the prediction accuracy. − Background regulatory graph (BRG) for improved message passing.− Potential biases in the training data. − Limited scalability due to the use of large datasets.✓SynLethDB SynLethKGSL-Interpreter2024✓ × × × RED-GNN +Frontend vizualization− Interpretability of the results. − Human-AI collaboration− The dataset has limited biological relationships resulting in instability of explanations✓SynLethDBKDDSL2024 × ✓✓ × feed-forward neural network (FNN)− Interpretability of the results.− Limited number of available training samples.✓SynLethDBDSESL2024✓ × × ✓Ensemble GCN+ AttentionKG-built submodels to learn gene embedding at multiple focus levels− Lack of interpretability − Random negative Sampling − Overlooking other discriminative aspects such as, pathway analysis, cellular context, multi-omics, etc. × × SynLethKG SLKB[33]2024✓ × × × LLMs−The first natural language dataset for SL explanations, namely “NexLeth dataset.”− Increased risk of generating mistaken explanations. − Fine tuning increased the complexity of the model.✓NexLethDepGPS2024 × ✓✓✓Graph- SAGE+ Trans-former− Based on the perturbation analysis, DepGPS demonstrated its effectiveness in explaining candidate SL genes.− Only a single gene is perturbed, highlighting the need for a deeper network level analyses✓SynLethDB A summarized overview of SL prediction models. Table 5A summarized overview of SL prediction models.ApproachYearKGMulti-omicsCell-specificGenerali-zationModelInsightsLimitations Ante-hoc Post-hoc DatasetDGIB4SL2025✓ × × ✓GIN + DGIB− DGIB4SL replaces attention mechanisms with graph information bottlenecks (GIB) to overcome the inherent instability of explainibility and high-frequency biases of attention mechanisms.− The fixed number K for generating explanations, which may result in overlapping or incomplete explanations.✓SynLethDBMiT4SL2025✓✓✓✓− HGT (hetero-geneous graph trans-former) − HGT+GCN− Transfer Learning strategy to deal with scarcity of SL labeled data across cell lines.− Lack of interpretability. × × SLKB PrimeKGLukePi2025✓ × × ✓− HGT (heterogeneous graph transformer) − HGT+GCN− A novel self-supervised pretraining framework that pretrains GNN-based model on unlabeled data to deal with the poor generalizability of these models and the scarcity of labeled SL data issue.− Lack of interpretability - Random negative Sampling -Overlooking other distinctive modalities, including multi-omics data (transcriptomics, proteomics, etc.) × × SynLethDB PrimeKGMGH-K4SL2025✓ × × ✓GNN+ Attention+ Generative adversarial network (GAN)− Higher-order relationship Mining − Class imbalance Mitigating− Overlooking other discriminative aspects, such as pathway analysis, cellular context, multi-omics, etc. × × SynLethDB SynLethKGKG-SLomics2025✓✓✓✓RGAT− An enriched KG by integrating cell line-specific multi-omics.− Increased complexity of the model. − Random negative Sampling✓SLKB SynLethDB A summarized overview of SL prediction models. Overview of existing methods The primary objective of this section is to explore GNN architectures for SL prediction and interpretation, spanning KG-based reasoning, attention-based, and Transformer-based models. Accordingly, we organize the methods into five architectural categories based on their primary graph learning paradigm as follows: Knowledge graph-based graph neural networks approaches KG4SL [70] is the first known model KG to predict SL interactions in human cancers. The model constructs a gene-specific subgraph from the KG for each SL-related gene. Each subgraph has a fixed number of sampled neighbors for each gene. Edge weights are assigned to the edges of the subgraph using a gene-specific relation scoring function, and neighbor representations are aggregated through weighted averaging with linear transformation and ReLU activation. The model computes the SL prediction score for a gene pair using the normalized inner product of the two genes’ final representation with sigmoid activation for binary SL classification. Building on the same KG subgraph reasoning principle but introducing pairwise enclosing subgraph extraction, PiLSL [71] relies on pairwise interactions between genes, extracting an enclosing subgraph from each gene pair in the KG to capture the intersection of local neighborhoods. The model computes relation-aware attention scores per edge via attentive embedding propagation, controlling message passing from neighbors. PiLSL updates nodes’ representations through combining self-representation and aggregated neighbor messages. It also encodes multi-omics data using an MLP to derive explicit features. These are fused with the latent subgraph features to SL interactions. Whereas PiLSL aggregates neighbors uniformly per relation, SLGNN [10] introduces factor-aware message aggregation by modeling SL-related factors as weighted sums of KG relationship embeddings to capture fine-grained interaction dynamics. Initial gene embeddings are derived through explicit multi-layer KG convolution using element-wise relationship-entity products. The model deploys parallel graph attention networks (GATs) with each GAT isolating and generating a factor-specific gene embedding, assigning attention scores to the messages depending on the importance of the related factors. These factor-specific representations are subsequently weighted via a gene–factor attention mechanism and aggregated into a unified representation. The final aggregated message is used to update the node’s representation, capturing both the influence of neighboring genes, and the most relevant biological factors for predicting SL interactions. Complementing this factor decomposition strategy with explicit subgraph pruning, Wu’s [16] approach extracts a local subgraph for each gene pair and augments node embeddings with positional vectors encoding shortest distances to both central genes. It applies a pruning mechanism via relationship-aware self-attention to score edges, with regards to neighboring nodes’ features and weight matrices. The model also integrates a R-GCN encoding module that performs message passing on the pruned subgraph using relation-specific weight matrices decomposed via basis functions. Final gene pair representations are scored via inner product followed by sigmoid activation for binary SL classification. Moving from single-hop subgraph extraction to multi-hop hierarchical propagation, MPASL [72] integrates KG into a multi-perspective learning using two complementary propagation strategies to predict SL. It recursively expands multi-hop entity neighborhoods from gene nodes via KG hierarchy propagation. Gene representations are refined by introducing fine-grained information from the KG, using Ripple propagation to connect genes with existing SL interaction preferences to candidate genes. Specifically, Ripple propagation initializes seed sets from existing SL interactions and propagates outward, computing weighted tail-entity responses via softmax-normalized similarity. A gene-entity module applies relation attention to score neighbors and enhance entities, while attention scores aggregate information contextually. Finally, an entity–entity module compares layer-wise embeddings through a discrepancy contrastive layer, with an attention aggregator producing final gene representations. Departing from Euclidean message passing, HLASL [73] models SL interactions by mapping node embeddings from Euclidean tangent space into Hyperbolic space using Lorentz manifolds. A Lorentzian attention mechanism then computes relational weights between head and tail entities in KG triplets, with softmax normalization applied across neighborhoods on the manifold. The message propagation is conducted by computing Lorentzian linear combinations of the genes’ neighboring nodes in hyperbolic space. Then relying on a hyperbolic Bi-interaction Aggregator, the model combines entity self-representations with aggregated neighborhood signals across both gene–gene interactions and gene-entity relationships. The prediction layer computes the embeddings’ inner product and outputs the probability score of each gene pair reflecting a potential SL interaction. By extending KG-based relational reasoning to heterogeneous multi-omics contexts, KG-SLomics [74] constructs a large heterogeneous KG capturing biological entities including genes, compounds, diseases, and pathways with their relationships. Afterwards, pretraining it using RotatE and combining its embeddings with multi-omics features from cancer cell lines to represent both topological and genetic information; these features are neighbor filled for nongene entities. The R-GAT aggregates heterogeneous neighbors with relation-dependent weights via relation-aware multi-head attention, that also highlights relevant entities of the KG. Furthermore, the model uses the skip connections to mitigate over smoothing. Hence, KG-SLomics captures both global biological knowledge and local cancer-specific contexts identifying clinically relevant SL pairs. By adopting an ensemble strategy over multiple KG-derived submodels rather than a single unified GNN, DSESL [75] is a two-stage ensemble framework that initially combines four submodels to leverage multiple aspects of the complex biological logic behind SL interactions. The knowledge graph embedding (KGE) submodel embeds SynLethKG’s entities using a Translating Embeddings for Modeling Multi-relational Data (TransE) framework, and channel attention to weight neighboring entity types for each gene via the identification of genes with shared biological contexts across the KG’s different entities. The meta submodel extracts the features of indirect relational pathways between genes in the KG capturing higher-order biological relationships linked to KG. SL-graph submodel applies GCN to an SL subgraph of known SL pairs to exploit the modular organization of SL interactions and predict new SL genes within functional clusters. Subsequently, the pathway submodel utilizes the textual and positional reactome pathway graph, processed through GCN to identify SL interactions between genes showing functional redundancy in parallel pathways which link common cellular processes. Thereafter, the outputs of the four submodels are integrated subsequently through an MLP meta learner that learns nonlinear combinations of the base model predictions. Also grounded in KG-based GNN reasoning but distinguished by its information-theoretic training objective, DGIB4SL [76] is a KG-based GNN that integrates three key components. The GIB principle enables the model to learn compressed subgraphs by maximizing mutual information with SL labels while minimizing redundant information from the original input graph. The motif-based GNN encoder decomposes each IB-graph into 13 specialized adjacency matrices corresponding to different three-node motifs (chains, cycles, etc.); processes each through separate GNN networks, then concatenates their representations. The model uses variational inference with KL-divergence regularization for graph representation learning, and an MLP classifier to produce the final SL predictions. Knowledge graph-based path reasoning KR4SL [77] is based on the construction of relational directed graphs between a primary gene and all its potential partners using dynamic programming. The model combines KG’s structural features with textual semantics, extracted via a pretrained BERT, during message propagation. Moreover, the model integrates a GRU to effectively capture sequential path dependencies and semantic relationships, and uses attentive aggregation weights to measure the importance of messages from neighboring entities. Finally, a feedforward scoring decoder computes the probability of each candidate gene being a synthetic lethal partner of the primary gene and ranks them accordingly. Adopting a complementary path-based formulation grounded in iterative Bellman-Ford propagation rather than GRU-based sequential scoring, BioPathNet [78] is built upon NBFNet and performs path-based link prediction in biomedical KGs. It first initializes boundary conditions for KG source nodes, subsequently propagating relational path representations across successive iterations using message-passing GNNs. It then uses BRG to augment message passing without altering the final downstream loss computation. The model also employs entity-aware negative sampling to improve discriminative performance and tighten structural decision boundaries. The learned representations are then decoded by a sigmoid activated feedforward network to produce the final pairwise link prediction scores. By extending KR4SL’s path reasoning engine with an interactive human-in-the-loop refinement layer, SLInterpreter [62] uses KR4SL as its backend GNN engine, in addition to a frontend visual tool assisting domain experts in a deeper exploration of the underlying mechanisms through the predictions analysis and the paths interpretation. Owing to [KG Refinement], biologists refine the acquired KG using metapath strategies and building on [SL Interpretation], SLInterpreter further enables iterative refinement of AI predictions, supporting iterative human–AI collaboration. Graph attention-based approaches Within this category, SLMGAE [79] integrates an SL graph representing genes as nodes and edges as known SL relations, as a main view; and other data sources as support views including gene ontology (GO) and PPI. For each view, a two-layer GCN encoder with LeakyReLU activation propagates node through the normalized graph Laplacian to produce gene embeddings. These embeddings are subsequently utilized by a decoder that reconstructs the adjacency matrices of the input graphs. SLMGAE then implements an edge-level softmax attention mechanism to assign different weights to the reconstructed graphs from each view. The final SL prediction is achieved by combining the reconstructed main view graph with the weighted sum of the support view graphs. The entire model is trained by minimizing a composite loss function that balances reconstruction accuracy with predictive performance. By sharing the multi-view attention fusion principle but replacing the autoencoder objective with a discriminative dual-level GAT formulation, GCATSL [80] is the first adaptation of SL interactions’ prediction, adopting a dual-level attention. At the node level, standard GAT is adopted to distinguish global neighbors, identified via random walk with restart, from local neighbors. Multi-head attention is applied independently across both topological scales, using learned coefficients to aggregate localized neighborhood representations. Next, a bi-interaction MLP merges these direct, local, and global features into a unified, graph-specific embedding. To integrate multi-source data, a feature-level attention mechanism weights the relative contributions of multiple heterogeneous graphs (e.g. GO and PPI). Final gene representations are decoded via a sigmoid-activated reconstruction matrix to output pairwise SL probabilities. Graph-Transformers-based approaches MSGT-SL [17] combines multi-view GCN with Graph Transformer in order to capture both immediate and long-range dependencies gene interactions. The two-layer multi-view GCN leverages SL network alongside omics data views to aggregate localized neighborhood features independently within each data modality. The model then selects a constrained subset of primary SL genes and auxiliary omics-derived genes via cross-omics random walk sampling. Features from all views are concatenated and fed into a single standard multi-head self-attention layer to capture long-range gene dependencies. The resulting core gene representations are then paired and fed into a standard cross-entropy classifier for binary SL prediction. MLEC-iSL [81] is another graph-transformer-based predictive framework, which extends this GCN–Transformer hybrid design to a two-stage regression-then-classification pipeline. It first processes population-based and cell-specific multi-omics features through separate MLPs, and learns local biological topologies across integrated BioGRID, KEGG, and Reactome networks via multi-view GCN. The learned multi view embeddings are passed through a multi-head self-attention Transformer encoder that outputs continuous SL connectivity scores trained with a regression objective. In the second stage, these individual connectivity metrics serve as inputs for a downstream logistic regression classifier, which classifies pairwise binary SL interactions. Whereas MLEC-iSL applies the Transformer sequentially after GCN, SLWise [82] treats SL prediction as a multimodal graph alignment problem, integrating cell-specific views across four distinct biomedical modalities: exclusive mutation patterns, gene effect scores, paralog relationships, and L1000 perturbation profiles. The model uses two-layer GraphSAGE encoders to aggregate the local structural neighbors within each graph view. Then a multi-head Transformer cross-attention module aligns and merges the representations from the first three omics views by computing cross-modal attention weights, and the L1000 perturbation features via direct vector concatenation. This comprehensive multi-omics representation is combined with baseline SL graph structural features and fed into an MLP for binary classification. SLWise’s interpretability was tested by SLAD-CE [21], an explanatory framework that generates testable biological hypotheses for SL prediction. Grounded in the premise that many SL interactions manifest as triple-gene relationships, the model decodes why a predicted gene pair is lethal in a given cellular context. It achieves this by intersecting two sets of significantly downregulated genes derived from the L1000 connectivity map, capturing downstream genes disrupted by the loss of either primary gene. These intersected targets are then filtered via DepMap scores to isolate essential or tumor-driver genes. And by performing gene set enrichment analysis on these key genes, SLAD-CE maps out the dysregulated pathways leading to cell death operating on a similar parallel GraphSAGE–Transformer architecture but oriented toward dependency score prediction rather than binary classification, DepGPS [83] projects multi-omics profiles (i.e. expression, CNV, and mutations) into a latent space enriched with gene2vec semantic vectors and PPI-derived structural embeddings. The core representation learning module consists of two stacked parallel layers, each containing a GraphSAGE module and a Performer linear Transformer operating simultaneously on the same gene embeddings; their outputs are summed and processed by an MLP. A final prediction module translates these consolidated features into continuous gene dependency scores. Ultimately, the framework employs in silico perturbation analysis, systematically altering model gene statuses to surface context-aware synthetic lethal candidates based on variations in these predicted dependency metrics. Shifting from task-specific training to pretrained heterogeneous graph transformer (HGT) representations, LukePi [18] bases its architecture on an HGT encoder trained on PrimeKG optimizing two joint tasks. The model applies topology-based node degree classification to infer degree labels from modified graphs and capture local structural patterns. Additionally, it leverages semantics-based edge recovery that evaluates the presence and the type of masked/pseudo-edges to map inter-relation semantics. The optimized HGT encoder is then fine-tuned for downstream SL and DTI prediction. Building directly on LukePi’s HGT pretraining strategy and extending it to a multimodal triplet framework, MiT4SL [84] constructs multimodal triplet representations combining three parallel encoders. HGT maps PrimeKG to generate BKG-based gene representations capturing semantic biological interactions, while ESM-2 protein language model extracts sequence-level structural and functional representations, and a DeeperGCN processes localized PPI networks to model cell-line-specific context. All resulting vectors are concatenated into unified triplet embeddings. These representations are fed to a downstream MLP predictor that optimizes them jointly through a multi-task loss function combining cross-entropy classification, contrastive learning, and decision-level regularization. Rounding out this category, EFOL-SL [85] has a modular pipeline that turns KGs into interpretable multi-hop logical reasoning, encoding SL queries as first-order logical graph queries and processing them in three stages. The model converts edges into relation nodes and unifies entities/relations into directed nodes using triplet transformation for consistent querying; it also masks intermediate entities during pretraining and inference so the model predicts masked nodes, enabling explicit reasoning. The model essentially leverages a Sparse Transformer Encoder to apply masked self-attention over the graph and learn global semantic embeddings while respecting KG sparsity. Additionally, it uses GAT to focus on local neighborhood aggregation and produce visualized multi hop reasoning chains. Training components include two stage masked pretraining (i.e. large dense subgraphs then small meta graphs) plus type aware negative sampling and supervised fine-tuning on SL labels. Similarly grounded in KG-based relational encoding but distinguished by its dual-stage knowledge distillation training strategy, SynLeaf [86] predicts SL across pan- and single-cancer contexts relying on a dual-stage multimodal fusion framework. It integrates an omics VAE cross-encoder with Product of Experts fusion and a KG encoder to aggregate relational neighborhoods per gene. A dual-stage training strategy then applies either feature-level knowledge distillation via Uni-Modal teacher or probability averaging via uni-modal ensemble (UME), with adaptive selection based on AUC validation. For a gene pair, SynLeaf concatenates learned embeddings and outputs a probability score via a shared classifier. Generative-artificial intelligence and large language models-based approaches Hypergraph modeling has emerged as a promising approach in graph-based analysis. It is increasingly applied to biomarker prediction, including novel cancer biomarkers [87, 88] and disease-related ceRNA biomarkers [89, 90]. Unlike conventional pairwise graphs, hypergraphs capture complex multi-way molecular interactions, thereby improving predictive performance. In the context of SL, MGHK4SL [8] represents SL graph as a Hypergraph to mitigate the complex genetic relations that bypass the common pairwise interactions representations in reality. This is done by leveraging heterogeneous GNN to extract hyperedges, and attention to weigh their importance combining both low- and high-order structural information. Then to frame the prediction within the biological context of SL, the model extracts semantic information from SynLethKG using CODER. Both structural and semantic embeddings are combined utilizing multimodal adaptive fusion based on learnable attention assigning dynamic weights to each modality. MGHK4SL also relies on GAN, with a Generator to produce negative samples and a Discriminator for the prediction module. Diversely, MGANSL [91], a dual-module GAN-based framework, integrates multi-network representation learning across PPI, GO, and SL. The network-common encoding module deploys a shared encoder acting as a generator, opposed by three pairwise discriminators trained adversarially to align cross-network representations into a unified common subspace. Simultaneously, the network-specific encoding module uses three independent private encoders to preserve intra-network features in separate subspaces. The model also imposes orthogonality constraints to ensure that these shared and private features remain structurally independent mitigating information leakage, while relying on a reconstruction loss to maintain fidelity to original network representations. The final gene-pair representation concatenates both subspaces and is classified by a two-layer MLP trained with binary cross-entropy. In regard to the use of the explanatory potential of LLMs, Zhang and his co-workers [33] built the NexLeth dataset in a first attempt to generate natural language explanations for SL mechanisms. They first used New Bing (GPT-4-powered) to automatically query literature and generate explanations for SL gene pairs from SynLethDB. These explanations were validated by domain experts, and key mechanistic phrases were annotated. The final dataset comprises 707 gene pairs, with 145 factual and 562 hypothetical explanations. The second part of the study consisted in fine-tuning four language models (BioGPT, BioGPT-Large, BioMedLM, and PMC-LLaMA) using parameter-efficient LoRA adaptation. Differently from other models relying on subgraph or feature importance scores requiring domain expertise to interpret their predictions, the authors combined instruction templates with KG-derived domain information to improve the models’ performance, resulting in BioMedLM achieving superior performance. This framework can also be integrated downstream of SL predictors to enhance biological interpretability and potentially improve accessibility of complex biological relationships for researchers and clinicians. Another LLM-based framework is SL-MERK [34] that generates natural language explanations of SL mechanisms by integrating GraphRAG with SynLethKG. LLMs serve three cascaded roles: entity extraction (Gemma-2-9b), mechanistic subgraph distillation, and explanation integration (GPT-4). This architecture yields measurable performance gains over GPT-4 alone demonstrating that grounding LLM outputs in structured biological knowledge improves mechanistic precision. Extending coverage beyond the curated literature base through intersection-based KG reasoning, and validated empirically on unseen gene pairs (RB1/CDK4/6; PIK3CA/PTEN) make the model generalizable. SL-MERK offer a post-hoc interpretability via natural language explanations that are generated as a decoupled step applied to opaque LLM computations, with mechanistic subgraphs providing only partial structural transparency at the input level. Other architectures KDDSL [54] is a hierarchical knowledge-embedded DNN that incorporates biological processes (BP) knowledge to reveal the mechanisms underlying SL interactions. Modeling each BP term as a subsystem composed of biologically meaningful neurons, the model hierarchically connects these subsystems based on GO relationships, enabling structured propagation of gene expression features through biologically relevant pathways. It processes transcriptomics from the LINCS L1000 project, propagating data through sequential gene and subsystem transformation layers via linear mapping, Tanh activation, and batch normalization, with each subsystem represented by two neurons. A cross-entropy-based FNN is attached to each subsystem to score its importance. Replacing GO-hierarchical propagation with molecular-level protein representations, Struct2SL [92] relies on three distinct protein-level feature modalities to predict SL. First, protein sequences are encoded into vector representations by averaging hidden states from an ELMo-based seq2vec model. Then to capture the proteins’ 3D structural attributes, the model leverages pairwise distance matrices from AlphaFold2 that are transformed into binary adjacency contact graphs and embedded via node2vec random walks. PPI network context is similarly extracted by applying node2vec to the STRING interaction network. These sequential, structural, and topological features are concatenated, normalized, and mapped to establish unified gene-level representations. Finally, an SL graph refines these embeddings, and a multi-layer regularized MLP utilizing ReLU activations and dropout executes link prediction. Discussion, challenges, and vision Performance comparison The compared methods considered the SL prediction either as a binary classification task, or a ranking task. To evaluate their models’ performance, authors mainly used AUC, AUPR, and F1 for binary classification, while the metrics used for raking evaluation consisted essentially in NDCG, Precision, and Recall for @ predicted SL partners. AUC, the area under the ROC curve is a threshold-independent measure of how well a model ranks positive cases above negative ones. While AUPR, the area under the precision–recall curve summarizes the trade-off between precision and recall and is especially useful for imbalanced data, along with F1, the harmonic mean of precision and recall, used to balance false positives and false negatives in a single score [93, 94]. NDCG@N measures how well relevant items are ordered near the top of a list, with lower-ranked items discounted logarithmically [95]; while Precision@N is the fraction of relevant documents among N retrieved items and t is used to evaluate ranking quality at a fixed cutoff [96]. To assess the models’ performance in more realistic and challenging sets, some of the approaches included additional validation scenarios. We describe these scenarios as follows: E1 is the transductive scenario. In this setting, data from the test set could be present in the training set [77]. E2 is a semi-inductive scenario, a more realistic scenario, includes one of the two SL partner genes in the training set and leaving the other one out. E3 is the inductive scenario, the most challenging case, indicates that neither of the SL genes present in the test set are included in the training set [8, 77]. The results are shown in Tables 7– 11. As can be seen from the results, in the E1 scenario, all KG-based approaches perform well, outperforming the others likely due to their effective co-occurrence patterns of relations shared between training and test splits. This confirms that for transductive reasoning, relying on known SL pairs, gene features (GO and PPI) or KG as inputs support new SL pairs prediction [77]. However, the performances of most of approaches, including KG-based models, start to struggle significantly under partial inductivity (E2), this is mainly due to relying on graph topology or gene embeddings likely not seen in the training data [77]. Critically, most of KG-only models fail to report E3 results entirely. Findings showcase the effectiveness of the multi-modal representation [91]. In fact, some models with KG-independent gene representations derived from omics or biological ontologies still sustain discrimination in inductive settings [82]. Moreover, by refining the KG with more informative sources [8, 86], or logic-based dependencies [85], models retain substantially more predictive capacity. Findings also demonstrate that multi-perspective learning KG enhances the model’s generalization capability in inductive scenarios [72, 75]. This confirms that by adopting a paradigm shift which enables the multimodal learning leads into maintaining the performance in inductive scenarios. Notably, the parallel integration of GraphSAGE with Transformer yielded competing performance [82, 83], even in inductive settings [82]. This combination, applied to DTI prediction produced strong results, which suggest its efficacity in biomedical link prediction tasks [97]. Such architecture alleviates over-smoothing and over-squashing through Transformer that applies message aggregation from distant nodes and captures global dependencies via self-attention [98]. Simultaneously, it preserves local connectivity with GraphSAGE aggregating features from neighboring nodes using graph’s edge structure, overcoming transformer’s inability to capture edge information [98, 99]. The findings also validates the distinctive outcomes of hypergraph modeling in dealing with unseen genes, while providing significant potential under inductive setting [8]. Table 7Performance of SL prediction under E1.ModelAUROCAUPRF1SLMGAE0.9170.9420.871GCATSL0.9370.9480.810KG4SL0.9470.9560.887PiLSL0.9530.959–KGsum0.9400.9430.883SLGNN0.9630.9710.908SLwise0.9260.9240.683MSGT-SL0.819–0.749MLEC-iSL0.7270.719–KDDSL0.9510.9490.796MPASL0.9650.979–DSESL0.9440.9510.882MiT4SL0.881––LukePi0.900.787–KG-SLomics0.920.910.86Struct2SL0.9810.9790.939EFOL-SL0.9690.933–SynLeaf0.9650.9660.909MGANSL0.9570.9610.886 Performance of SL prediction under E1. Table 11SL ranking approaches performance under E3.ModelNDCG@50Recall@50Precision@50KR4SL0.38970.5950.5958BioPathNet0.4630.6240.625MGHK4SL0.20330.3607—- SL ranking approaches performance under E3. Table 8Performance of SL prediction under E2.ModelAUROCAUPRF1SLMGAE0.7680.8150.714GCATSL0.8320.8670.8PiLSL0.7940.815–KGsum–––SLGNN–––SLwise0.9020.9020.700MSGT-SL–––MLEC-iSL0.7070.735–MPASL0.8760.894–KDDSL0.9400.8550.655DSESL0.8820.9010.811MiT4SL0.624––LukePi–0.7880.679KG-SLomics0.700.680.65EFOL-SL0.900.927–SynLeaf0.8620.8750.795 Performance of SL prediction under E2. Table 9Performance of SL prediction under E3.ModelAUROCAUPRF1SLMGAE0.5610.4970.730GCATSL0.7250.7140.689PiLSL0.6650.670–KGsum–––SLGNN–––SLwise0.8950.8960.653MSGT-SL0.591–0.559MLEC-iSL0.6770.679–KDDSL0.4870.70.34MPASL–––DSESL0.8240.8430.754MiT4SL0.575––LukePi0.7220.6790.651KG-SLomics–––EFOL-SL0.8400.814–SynLeaf0.7400.7610.715 Performance of SL prediction under E3. Table 10SL ranking approaches performance under E1.ModelNDCG@50Recall@50Precision@50KR4SL0.4860.61970.6243BioPathNet0.4510.6430.644MGHK4SL0.30680.5064– SL ranking approaches performance under E1. Multi-omics in synthetic lethality prediction The integration of multi-omics data in SL prediction offers several benefits (Table 12). Specifically, using expression and mutation-based codependencies helps identify functionally related gene pairs with a higher SL likelihood [71, 79]. Otherwise, combining mutation exclusivity with expression data provides functional similarity signals beyond graph structure, improving prediction accuracy and reliability across datasets [79]. Transcriptomic data deliver important cell-specific signals needed to predict SL interactions based on cellular environment and gene activity [84]. Including multiple omics perspectives allows models to effectively distinguish between cell line contexts. It also addresses labeling issues across cell lines and improves the ability to generalize to new genes and cellular environments [17, 75, 84]. In addition, integrating omics data help models focus on relevant biological signals and represent dynamic gene relationships in specific cellular contexts [82, 83]. Nonetheless, omics data still present some pertinent challenges to address regarding their multi-dimensionality, heterogeneity and incomplete modalities [100, 101]. In fact, multi-omics data are large and complex while SL labeled samples are limited; this might increase models’ instability and overfitting. Moreover, different omics layers have distinct scales, sparsity, and noise patterns which makes their integration in the same feature space challenging. Besides, poorly matched omics can degrade performance, implying modality specific mismatches as indicated in [17]. Also, omics-derived features can be dominated by a few over studied genes, harming generalization to unseen genes and cancers [102]. Another unresolved issue arises from the scarcity of context-specific omics data, particularly in rarer cancers where tissue- or cancer-type–specific profiles are limited [103]. Additionally, uneven coverage across widely used sources introduces additional bias, since mutation, copy number variation, and expression panels are not uniformly represented across all cell lines and genes [17, 74, 103]. These limitations underscore the need for more comprehensive and standardized omics resources to ensure robust and generalizable SL predictions. Table 12Multi-omics data type.ApproachYearGenomicsTranscriptomicsProteomicsReactomicsEpigenomicsPiLSL2022✓✓–––SLwise2023✓✓–✓–MSGT-SL2023–✓–––KDDSL2024–✓–––MLEC-iSL2024✓––✓–DepGPS2024✓✓–––MiT4SL2025–✓✓––KG-SLomics2025✓✓✓––SynLeaf2026✓✓––✓ Multi-omics data type. Generalizability Generalization in SL prediction models is mainly evaluated by inductive learning scenarios. While these models perform well on partially or totally unseen data subsets, they still operate on a closed world presuming that the testing and training distributions are identical. Consequently, these systems exhibit severe vulnerability when encountering out-of-distribution inputs, such as novel biological entities, unmapped genetic feature combinations, or entirely unseen cell lines [18]. This vulnerability is a well-documented bottleneck in GNNs and KG completion frameworks [23, 104], a challenge that directly undermines the reliability of current computational SL predictions when applied to real-world clinical data. In addition, the scarcity of data per cell line makes it challenging for the models to generalize to new cell lines. Conversely, adopting a pan-cancer framework allows models to extract broad and generalizable molecular relationships that transcend tissue boundaries, focusing on universal SL interactions rather than cell-line-specific anomalies [92]. Furthermore, some recent pretraining efforts for SL prediction demonstrate that integrating diverse biological data into transferable node embeddings increases generalization. Long et al. [80] employ an unsupervised graph reconstruction model combining protein sequences, PPI networks, and gene ontology graphs to generate high-quality protein embeddings for downstream link prediction. Whereas Tao et al. [18] adopt a self-supervised strategy on PrimeKG, using custom pretext tasks (i.e. node degree and edge recovery) without task-specific labels. Their model performs well across multiple tasks, especially under low-data and distribution-shift conditions, as shown in SL and DTI prediction. Pretraining is therefore a promising direction in SL prediction to generate robust and transferable embeddings and generalize across new biological distributions. While multiple methods rely on inductive learning sets to ensure their model’s generalizability, others adopt transferability among cell lines and different datasets to evaluate the model’s performance under challenging conditions (Table 13). Pretraining in SL prediction is a recent practice that was mainly adopted in studies like [80] and [18]. The former is an unsupervised graph reconstruction model of protein and drug networks [80]. For SL prediction task, the authors integrate heterogeneous data sources combining protein sequences encoded into a feature matrix along with PPI and GO graphs, modeling the interactions between proteins and semantic information related to them, respectively, in order to produce high-quality protein embeddings. These node representations encapsulate knowledge from nodes’ attributes and the relational context in the used networks; they serve to reconstruct original PPI and GO graphs, and later as fixed input features for downstream link prediction models. On the other hand, Tao et al. [84] pretrain on PrimeKG following a self-supervised learning trend with custom pretext tasks, namely node degree and edge recovery, without any task-specific labels. The model transfers the general understanding of biomedical entity relationships encoded into the GNN’s parameters and rich node embedding. This makes it a universal model with the ability to perform well on multiple downstream tasks particularly in challenging low-data and distribution-shift scenarios in the instance of SL and DTI prediction. These examples prove that fusing diverse biological data into robust, transferable node embeddings boosts accuracy, speed, and generalization for specific link prediction problems. Table 13Generalizability analysis.ApproachYearGeneralizability mechanismGCATSL2021Inductive learningKR4SL2023Inductive learningPiLSL2022Inductive learningSLwise2023Transfer learningMPASL2024Inductive learningMLEC-iSL2024Inductive learningBioPathNet2024Inductive learningDSESL2024Inductive learningMiT4SL2025Transfer learningLukePi2025Self-supervised pretraining frameworkMGHK4SL2025Inductive learningKG-SLomics2025Transfer/inductive learningSynLeaf2026Inductive learning Generalizability analysis. Interpretability/explainability concerns Regardless of the interpretability/ explainability, Table 14 presents the most promising inherently interpretable/post-hoc methods that we have identified. Within this context, attention mechanisms have been widely adopted in dealing with interpretability/explainability concerns. In fact, attention can reveal domain-specific perceptions that help discern the focus and reasoning of GNN models [55, 78]. Specifically, they showcase which nodes, edges, or features [15, 16, 74, 80] are emphasized during message passing which ensures traceability. Hence, attentive subgraph extraction over KG helps leverage the enclosed semantics in entities and edges related to SL genes [10, 71, 72], while self-attention pruning removes irrelevant gene entity edges and highlights important ones [16]. In the same line, multi-hop path reasoning over relational subgraphs in KG [77, 78, 82], or via masked querying [85], discloses critical paths to SL interactions making the decision process transparent and interpretable. However, while attention mechanism seems to be a meaningful tool that could be adopted by SL researchers, it is argued that attention weights, although plausible, might not be faithful to the model’s causal decision process [20]. Although with external evaluation [71, 74, 77, 105], attention can be trusted as an explanation. But some issues remain related to attention’s variation across random seeds, heads, or small input perturbations and could undermine reproducibility as an explanation [76, 106]. Nevertheless, the integration of both high-order and low-order structural information could alleviate some of these challenges as it provides additional semantic information that makes attention more valuable, context-aware and informative [8, 76]. Chen et al. [76] address unstable attention mechanisms with a diverse GIB using a discriminative process to generate multiple diverse subgraph explanations per gene pair. This ensures faithful, diverse, and biologically meaningful explanations for gene interactions. Table 14Interpretability/explainability analysis.ApproachYearModel Ante-hoc Post-hoc MechanismPiLSL2022R-GCN + Attention✓–Path-based reasoning through attention mechanism[16]2023R-GCN + Attention✓–Self-attention mechanism to prune the KG and capture high-importance relationshipsKR4SL2023RED-GNN+ Recurrent Neural Network✓–Path-based reasoning through attention mechanismSLGNN2023KGNN + GAT✓–KG message aggregation via GATsSLWise + SLAD-CE2023GraphSAGE + Transformer–✓SLAD-CEMLEC-iSL2024GCN + Graph Transformer✓–Genes connectivity learning through transformer + GraphBioPathNet2024Neural Bellman- Ford Network (NBFNet) based GNN✓–NBFNet, a message-passing neural network to learn GNNs path representationsSLInterpreter2024RED-GNN + Frontend vizualization–✓Metapath strategiesKDDSL2024FNN✓–FNNNexLeth2024LLMs–✓LLMsDepGPS2024GraphSAGE+ Transformer–✓Perturbation-based methodsDGIB4SL2025GIN + DGIB✓–information bottleneckKG-SLomics2025RGAT✓–Features importance visualization via GATsEFOL-SL2025GAT + Transformer✓–multi-hop reasoning model based on the first-order logical querySL-MERK2025LLMs+ Explainable GNN–✓KG retrieval-augmented LLM for explainability Interpretability/explainability analysis. On the other hand, post-hoc approaches used counterfactual perturbation [83] that allows to probe how changes in a gene’s state affect the other genes revealing context-specific gene dependencies. Other methods extract metapath explanations that are evaluated by experts to refine the KG and retrains the model iteratively so explanations align with domain knowledge, in an AI-Human collaborative frame [62]. Whereas Xing et al. [21] explore associations between cellular damage and candidate SL genes abnormalities, and adopts enrichment analysis to explain SL underlying mechanisms biology grounded explanatory paradigm. LLMs shift SL interpretability from graph-structural outputs towards mechanistic narratives achieved either through fine-tuning biomedical language models on KG-augmented personalized prompts [33], or by combining GraphRAG’s multi-level community analysis with KG subgraph extraction [34]. By grounding explanations in established BPs (i.e. DNA repair, cell cycle regulation, etc.) both frameworks contextualize SL predictions within authoritative literature, enabling hypothesis generation for wet-lab validation of computationally identified SL interactions. Knowledge graph for synthetic lethality prediction KGs have emerged as a powerful tool for representing the real-world knowledge [107]. KGs have shifted SL prediction by overcoming the major limitations of homogeneous graph approaches. Prior models to KG considered gene pairs as isolated samples, missing how genes interact with broader biological entities like pathways, diseases, and chemical compounds [70]. By contrast, heterogeneous KGs map these diverse entities seamlessly which shifts how models process data. Typically, these graphs are utilized both in the premodel phase for feature extraction and post-modeling for causal reasoning, directly supplying the rich context needed for explainable outputs [69, 108]. This was leveraged in recent frameworks generating multi-hop logical paths for cell-aware predictions, merging expanded graphs with multi-omics data for cross-cancer generalization, or leveraging protein sequence triplets to bypass label scarcity. However, in the spite of these innovations, several limitations persist. Specifically, real-world noise and the scarcity of known interactions creates systemic inconsistencies that downstream models struggle to resolve [32, 109]. Moreover, large-scale KG training imposes substantial computational costs, imposing a heavy resource burden that remains a major hurdle for widespread deployment [110]. KG Completion is another issue to be addressed [107]. To address this challenge, KG reasoning approaches are among solutions that have caught great attention among academia [77, 111]. Notably, KGs are used either for entity/relation extraction, aiming to improve prediction accuracy by adding internal or external knowledge to datasets, or by increasing the explainability over a KG as an evidence-based approach that is more trustable [112]. Indeed, finding reasoning paths between source and target entities in KG proved to make the link prediction for unseen entities interpretable as it provides support evidence for the inferred link [111, 113]. Besides, recent researches position KG-based methods as a structurally distinctive form of input-based interpretable AI; by grounding model predictions in explicit entity-relation representations rather than latent feature spaces [114, 115]. Hence, KG embedding, which aims to embed the entities and relations of a KG into low-dimensional vector spaces while maximally preserving its topological properties and leverages these representations for link prediction, is an emerging computational approach for biomedical discovery [111, 116, 117]. Hence, using such methods that make KGs more accessible helps leverage KG’s enclosed semantics, and thus contributes in improving the explainability of KG models. In other respects, reasoning over the KG offers traceability and transparency which attenuate the blackbox nature of GNNs, affording biologists and clinicians meaningful pathways for result validation [114]. Very recently, integrating KG with RAG has emerged as a promising direction to improve the causal reasoning in graph-augmented LLMs [34, 111]. Cell-line-specific synthetic lethality prediction Cell-line-specific SL prediction data are a challenging aspect of cancer therapeutics, as most SL interactions remain highly context-dependent [118, 119]. Table 15 summarizes the different cell-line-specific data type implicated in SL prediction. Studies using combinatorial CRISPR screening revealed significant heterogeneity, with some studies finding no overlapping synthetic lethal gene pairs among different cell lines. As mentioned in [119], notable examples include [120] that screened 2628 gene pairs across three cell lines (A549, HELA, and 293T) and identified no synthetic lethal interactions common to all three lines, and Horlbeck et al. [121], who reported only 128 shared synthetic lethal pairs out of 222 784 gene pairs screened in K562 and Jurkat leukemia cells. In addition, Pu et al. [82] who predicted SL interactions in A549, A375, and HT29, found no shared pairs among the three lines. This context-dependency is generally driven by tumor cell heterogeneity, genetic background variation, and microenvironmental factors [119, 122] representing a major translational obstacle limiting the clinical utility of preclinical SL findings. The distinction of cancer cell lines is particularly crucial for developing effective targeted therapies for cancers with recurrent genetic alterations that currently lack treatment options, as robust SL interactions offer greater promise for clinical translation than those confined to specific cellular contexts [2]. Integrating cell line genomics, expression, proteomics, and drug response measurements is also crucial to mapping perturbation outcomes onto mechanistic features that inform therapeutic targeting [123]. Simultaneously, recent approaches span semisupervised networks to graph-based deep models, multi-view GCNs, transformer GNN encoders, and forest ensembles combining sequence and omics features [81, 123, 124]. These advances in explainable feature learning and cell line adapted transfer learning are therefore poised to yield more reliable, interpretable SL predictions. However, some challenges still remain, including label scarcity, selection bias across tested gene sets, and cross line generalization [84]. To mitigate these issues, recent works propose explicit cell line embeddings, triplet representation transfer, and patient/drug screen integration [82, 103]. Table 15Cell-line-specific data type.ApproachYearA375A549HT29K562JurkatMewo22Rv11MIA Paca-2DMS53SLwise2023✓✓✓––––––MSGT-SL2023–––✓✓–––SLAD-CE2023✓✓✓✓–––––MLEC-iSL2024–––✓✓––––KDDSL2024–✓–✓–––✓–DepGPS2024✓–––––––✓MiT4SL2025–––––✓✓–KG-SLomics2025✓✓–✓✓✓–– Cell-line-specific data type. Although SL predictive models have known important advances in recent years, through enhancing their interpretability/explainability, as mentioned above, they are still faced with some noticeable challenges. Firstly, most models use exclusively either KG or multi-omics, working on combining both sources could yield better results as proved in [18, 71]. Cancer gene regulatory networks are also under-exploited; they are generally built on gene expression data but they capture different heterogeneity [125]. Therefore, including them adds a new layer of information to other omics data types, which helps harness the complexity of SL interactions. Besides, the imbalanced proportions of SL and non-SL interactions and the scarcity of high-quality negative samples could affect the models’ performance and evaluation [126]. Limitations and future perspectives for synthetic lethality prediction Although SL predictive models have known important advances in recent years, they still face various challenges. Class imbalance and scarcity of data Class imbalance between positive and negative samples, specifically the scarcity of negative samples is still a challenging issue for SL [32, 74]. In order to address this issue, by ensuring a balance between positive and negative sample pairs [85, 92], random selection of gene pairs that are not identified as negative samples is commonly adopted [18, 32]. However, this approach is not representative of real-world scenarios, and the selection of the reliable negative samples, make the process more challenging [75]. To mitigate this issue, self-supervised learning and transfer learning strategies have exhibited promising achievements in dealing with the scarcity issue of SL negative samples [84]. Very recently, more sophisticated approaches based on GANs have been used to mitigate the challenges of negative sample scarcity in SL prediction [8]. Multimodal data integration: knowledge graph, epegenetics, and multi-omics Besides, with the main aim of improving the interpretability and SL prediction, fusing data from multiple sources, including multi-omics data and KG has emerged as a trend and hotspot area of research that has gathered great interest very recently [86]. Likewise, while still under-exploited in the context of SL prediction, recent research has drawn great attention towards the implication of epigenetics data, especially in the DNA damage response pathways. In fact, as epigenetic abnormalities appear in the early stage of carcinogenesis and premalignant status, they might be studied further to develop strategies for cancer prevention [127]. Hence, built on gene expression data, epigenetic data integration may capture different heterogeneity that can help building new mechanisms for SL prediction [125]. Yet, while combining multi-source information could yield better results [18, 81]. However, most of approaches still struggle with the effective fusion of high-dimensional and heterogeneity of data [8, 86]. To overcome this issue, recent researches have been focused on developing effective multimodal learning frameworks. Within this context, an adaptive multimodal fusion module has been designed to dynamically learn the heterogeneous contributions of each modality [8]. Similarly, another promising solution to tackle this issue is owing to knowledge distillation mechanism [86]. Other avenues of development Yet, it is worth mentioning that all developed models for SL prediction have oriented their findings toward cancer therapeutics. The majority of SL-related studies have the same objective of curing cancer through the discovery of new potential SL pairs [2, 71, 128, 129]. However, SL is implicated in many more diseases like neurological disorders and infectious illness, namely Huntington’s [130], Tuberculosis [131, 132], and Sars-cov-2, and other viruses [133]. SL is also implicated in Ferroptosis [134] that is becoming a hotspot for the treatment and improvement of the prognosis of related systemic diseases, such as inflammatory diseases, pulmonary fibrosis, hepatitis, inflammatory bowel disease, neurodegenerative diseases, etc. [135]. Therefore, identifying SL interactions involved in these diseases could give good results, providing a promising new path to serve the field of healthcare. Moreover, recent studies reported the existence of SL triplets, which represent SL interaction among three genes [136–138]. Although SL triplets have not been totally validated [136], they are an example of biological interaction complexity. This discovery could lead to new cancer therapeutic directions. Architectural aspects of synthetic lethality prediction Regarding architectural aspects of SL prediction, several emerging paradigms offer significant potential for clinical translation. Structurally transparent mixture-of-experts frameworks [139] can segment predictions by assigning distinct expert modules to specific biological pathways or cancer-type-specific vulnerabilities. Similarly, we believe that fuzzy inference systems [140] that rely on human readable rule bases offer a promising yet still underexplored way to embed existing biological knowledge (i.e. gene essentiality) directly into a model’s architecture. LLMs are another under-explored aspect in SL research; although NexLeth, built on pretrained language models, has established the first attempt to generate explanations for SL interactions, it does not predict them, raising the need for frameworks to achieve both prediction and explainability [33]. Notably, we also encourage AI-Human collaborative work benefiting from the validation of experts to guide models make accurate and practical decisions [62]. Another promising direction is by leveraging information bottleneck-constrained GNNs to isolate invariant causal subgraphs [141], which holds strong potential for uncovering SL interaction subnetworks by rethinking specific biological contexts that are invariant generalize reliably across diverse cancer types. Furthermore, to advance this field, future methodologies should prioritize structural robustness and interpretable design. Given the growing evidence that attention mechanisms fail to offer true interpretability [142], the field should pivot toward counterfactual reasoning frameworks beyond attention-based path explanations. By determining the exact data perturbations required to alter a prediction [143], these frameworks yield causally grounded, contrastive explanations for predicted gene pairs. Also, explanation stability remains largely unquantified in SL model evaluations, and must be incorporated as a mandatory metric alongside traditional performance indicators like AUROC and AUPR to ensure consistent explanations for highly similar inputs [144]. Besides, to avoid the well documented pitfalls of computational biology models’ interpretability, namely the over dependance XAI method and the gap between model predictions and practical biological interpretation [145], we recommend future studies to employ multiple, complementary XAI techniques and systematically validate their outputs against established pathway databases and functional annotations. We trust that the future of computational SL relies essentially on transitioning from opaque and attention-based models to architecturally transparent, causally grounded, and rigorously validated frameworks that ensure clinical utility across different cancer types. Key Points We introduce existing synthetic lethality (SL) knowledge graph (KG) databases. We introduce existing synthetic lethality (SL) knowledge graph (KG) databases. We analyze the interpretability/ explainability aspects of SL predictive models. We analyze the interpretability/ explainability aspects of SL predictive models. We highlight the implication of KG, omics and cell-line-specific data as a recent shift in SL research. We highlight the implication of KG, omics and cell-line-specific data as a recent shift in SL research. We shed light on the importance of large language models and Generative artificial intelligence as promising solutions to boost the findings. We shed light on the importance of large language models and Generative artificial intelligence as promising solutions to boost the findings. We emphasizes the limitations of the reviewed approaches and suggest future possible directives to improve SL research. We emphasizes the limitations of the reviewed approaches and suggest future possible directives to improve SL research. Contributor Information Salsabil Felloussi, Department of Computer Science, L3IA Laboratory, Faculty of Sciences Dhar El Mahraz, University Sidi Mohamed Ben Abdellah, B.P. 1796 - Atlas, 30003, Fez, Morocco. Fatima El Haoussi, Department of Computer Science, L3IA Laboratory, Faculty of Sciences Dhar El Mahraz, University Sidi Mohamed Ben Abdellah, B.P. 1796 - Atlas, 30003, Fez, Morocco. Hamid Tairi, Department of Computer Science, L3IA Laboratory, Faculty of Sciences Dhar El Mahraz, University Sidi Mohamed Ben Abdellah, B.P. 1796 - Atlas, 30003, Fez, Morocco. Noura Aherrahrou, Department of Computer Science, L3IA Laboratory, Faculty of Sciences Dhar El Mahraz, University Sidi Mohamed Ben Abdellah, B.P. 1796 - Atlas, 30003, Fez, Morocco. Conflicts of interest None declared. Funding None declared. Data availability The data (the links to the processed datasets) and source codes (the links to the codes) for the reviewed studies from their original publications and repositories are publicly available in the GitHub repository. https://github.com/Felloussi-Aherrahrou-et-al/Awesome-Synthetic-Lethality-Prediction-Approaches. References Associated Data Data Availability Statement The data (the links to the processed datasets) and source codes (the links to the codes) for the reviewed studies from their original publications and repositories are publicly available in the GitHub repository. https://github.com/Felloussi-Aherrahrou-et-al/Awesome-Synthetic-Lethality-Prediction-Approaches.