MGANSL: multi-network representation generating with generative adversarial network for synthetic lethality prediction Authors: Jinxin Li, Xinguo Lu, Zihao Li, Xing Liu, Hongrui Liu, Jingjing Ruan Published: 2025 ABSTRACT AbstractBackground:Cancer is a complex disease that arises from the simultaneous mutations of multiple biological molecules. An effective therapeutic strategy is to exploit synthetic lethality (SL) by targeting the SL partner of cancer driver genes. Computational approaches have emerged as efficient complements to traditional methods. Although some methods integrate heterogeneous sources to learn multi-network representations, they often neglect consistent information shared across different networks and specific characteristic specific to individual network. Therefore, a comprehensive representation learning framework for capturing both multi-network consistency and network-specific information of gene pair is needed.Results:We proposed a novel approach capturing Multi-network consistent and specific representation with Generative Adversarial Network for Synthetic Lethality prediction (MGANSL). MGANSL employs network-aligned and network-specific encoding modules to cooperatively learn comprehensive multi-network representations of gene pair. In particular, network-aligned encoding module can capture cross-modal consistent information via cross-network adversarial generation, and network-specific encoding module can capture single network specific information via intra-network adversarial generation.Conclusions:Comprehensive experiments conducted on two human synthetic lethality datasets demonstrate the superiority of proposed method in SL prediction. Moreover, the novel predicted SL associations could aid in designing anti-cancer drugs and providing potential drug targets.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12859-025-06345-4.Keywords: Representation learning, Generative adversarial network, Synthetic lethality, Anti-cancer drug repositioning, Multiple biological network. Introduction Cancer arises from the co-mutations of diverse biological molecules which lead to aberrant cell proliferation and eventual metastasis. Elucidating these complex molecular interactions is essential for understanding the mechanism behind cancer and developing innovative therapeutic strategies. Traditional targeted cancer therapies focus on gene products that are mutated or overexpressed to a particular cancer type [1]. However, it is well known that targeting loss-of-function mutations and some instances of gene overexpression faces with huge challenges, which presents obstacles in drug development. Encouragingly, the concept of synthetic lethality (SL) offers a promising alternative for addressing these by indirectly targeting intractable genes [2]. Specifically, synthetic lethality depicts a type of genetic relationship wherein the simultaneous inhibition of two genes results in lethality, but the inhibition of each gene individually being viable. It enables the targeting SL partner of cancer driver, which effectively mitigates the challenges posed by direct targeting of mutated or overexpressed genes [3]. The clinical approval and successful application of the PARP inhibitor olaparib further highlight the therapeutic potential of SL-based strategies [4]. Consequently, synthetic lethality has attracted much attention from the area of anti-cancer drug development, which prompt a series of works dedicated to the exploration and identification of SL gene pairs. In recent decades, numerous wet lab-based and statistics-based works have been presented to infer novel SL interactions. For example, Jerby-Arnon et al. propose a data-driven approach (DAISY) to infer SL gene pairs by integrating cancer patients, somatic mutation, and shRNA data [5]. However, these methods are costly, time-consuming and off-target effects. The rapid advancements of high-throughput technologies have produced abundant biological heterogeneous data, such as protein-protein interactions [6], gene functional annotations [7] and so on. Therefore, computational models for SL prediction with these heterogeneous sources have been effectively complementary to traditional approaches. Biological networks depict complex interaction between different biological molecules (gene, protein) and genetic interactions are affected by complex network regulation and other biological interaction. Mining the structural and functional properties embedded within these networks enables researchers to uncover potential mechanisms underlying synthetic lethality (SL) and identify candidate SL gene pairs. Early works in synthetic lethality (SL) prediction primarily utilized single network approaches, focusing on individual biological network such as protein-protein interaction (PPI) network [8], gene signaling network [9] or SL interactive network [10]. These methods typically analyze topological features or perturbation-induced network alterations. The representative approach is DDGCN [11] that learn gene representation in SL interactive network with graph convolutional network to explore SL network topological structure information. However, single network-based methods cannot model complex interaction between genes and neglect other biological networks which provides more comprehensive insights and information to mechanism hidden in SL. As research progressed, the field has increasingly shifted toward multi-network integration to leverage diverse biological contexts, including PPI networks [12], Gene Ontology (GO) functional networks, and so on. To address these challenges, recent efforts have focused on integrating multiple biological networks that encode different interaction patterns [13, 14]. For example, to integrate different gene feature network, Long et al. [15] proposed a graph attention network method GCATSL to learn gene multi-network features with graph attention network in which two-level attention mechanism are utilized to aggregate different networks. Similarly, Hao et al. proposed a graph auto-encoder method SLMGAE [16] to learn gene cross-view representation from different gene networks, in which attention mechanism are utilized to aggregate different network representation. Furthermore, Wang et al. proposed a method KG4SL to integrate different kinds of entity relationships for learning gene interpretable representation from different gene knowledge networks. However, current multi-network methods concentrate on learning consistent gene representations from various biological networks, they implicitly assume that different network properties have same distributions. Different networks depict different types of genetic interactions and thus gene characteristics in diverse networks have different distribution [17]. Furthermore, each network usually contains private information which describes specific biological interactive pattern [18]. Existing methods often overlook these network-specific features, resulting in incomplete gene-pair representations [19]. Recently, Generative Adversarial Networks (GANs) have gained substantial traction in the deep learning community and have been successfully applied to tasks such as image synthesis, information retrieval, and graph representation learning [20]. A GAN model typically model antagonistic game between generators and discriminators in which generators aim to produce a fake image or representation, and discriminators aim to distinguish the image or representation produced by generators [21]. For example, Zhang et al. proposed a Self-Attention Generative Adversarial Network (SAGAN) for image generation tasks with attention-driven and long-range dependency [22]. Furthermore, GAN have been extended in multi-modal data for learning cross-modal representation [23] from different data sources. For example, Hu et al. [24] proposed a multi-modal adversarial network to retrieve samples from different modalities by learning comprehensive representation, in which two encoder module are utilized to learn modal-common and modal-specific during adversarial game. Vukotic et al. verify the effectiveness of GANS to obtain multi-modal representation in retrieving video segments task [25]. These successes highlight the strong potential of GANs for handling heterogeneous biological networks with diverse distributions and interaction patterns. Inspired by these, we present a novel approach to learn multi-network representation of gene pair with Generative Adversarial Networks for SL prediction (MGANSL). MGANSL comprised of two encoding modules to cooperatively learn comprehensive multi-network representation from diverse network properties. In network-common encoding module, MGANSL utilize encoder based on generative adversarial network to learn cross-network consistent representation of gene pair by projecting different network features into a common subspace. In network-private encoding module, we adopt network-specific encoders to individually extract network-specific representation by projecting different network features into network -private subspaces. Inter-network consistent representation and intra-network specific representation are concatenated to learna multilayer perceptron (MLP) classifier trained in a supervised manner. Comprehensive experiments performed on benchmark SL dataset verify that our approach is superior to baseline approaches. Additionally, ablation study demonstrates that cross-network consistent representation and network-specific representation provide complementary view for SL prediction. Four kinds of case study further illustrate the potential in anti-cancer drug discovery of SL pairs predicted by our approach. Method Data process We evaluated MGANSL on two human synthetic lethal datasets, including SynLethDB and SynLethDB2.0. SynLethDB [26] contains 19667 SL associations between 6375 genes, while SynLethDB2.0 [27] contains 35943 SL associations between 9856 genes respectively. For each dataset, an equal number of unlabeled gene pairs (i.e., gene pairs without known SL associations) were randomly selected as negative samples (19667 for SynLethDB and 35943 for SynLethDB 2.0). Therefore, we select 19667 non-SL samples in SynLethDB and 35943 non-SL samples in SynLethDB2.0. For each sample, we utilize three network representations: PPI structure network, GO functional network, and SL interactive network. For PPI structure network, we download human protein-protein interactions from HPRD database and calculated the gene topological similarity network in PPI network with FSweight [28]. For GO functional network, we download the Gene Ontology annotation data from Gene Ontology dataset [7] and calculate the gene functional similarity network following the method described in [10]. For SL interactive network, we calculate the gaussian kernel similarity of gene with known SL associations. Finally, we adopt Principal Component Analysis (PCA) to extract more discriminative network representations from three gene similarity network and concatenate the two gene features as representation of gene pair. Overview of MGANSL The workflow of MGANSL is illustrated in Fig. 1. The cross-network representation learning module contains a network-common encoder and three discriminators for extracting cross-network consistent information. Network-common encoder aims to align multiple network representations with various distributions into a common subspace with same distribution, and three discriminators are regarded as opponent to align distributions during adversarial learning. Meanwhile, the intra-network feature representation learning module contains three network-specific encoders and three evaluators for extracting network-private information, which is complementary to cross-network consistent information. Finally, by incorporating network-common and network-specific information, synthetic lethal prediction module is adopted to predict novel human synthetic lethal interactions. The overall framework of MGANSL. a The workflow of MGANSL. b Network-common feature learning. c Network-private feature learning Cross-network representation learning To align gene-pair representations across multiple biological networks, MGANSL employs a generative adversarial learning framework, where the network-common encoder acts as a generator and multiple discriminators guide the cross-network alignment process. Formally, given three representations of gene pair from three network: PPI structural network , GO annotation network and SL interactive network respectively, MGANSL learn network consistent representation of gene pair with network-common encoder , which can be calculated as : Where represents the representing function cross all networks, is the trainable parameters of network-common encoder, (P, O, S) represent the PPI structure network, GO functional network and SL interactive network, and , and denotes the feature dimension from three networks respectively. To facilitate the integration of multi-network information, we set the feature dimensions of the three networks to be identical,i.e., ===, where denotes the input size of model. By utilizing Eq. 1, the features of different networks can be mapped into a common subspace, and the diversities of different networks can be significantly diminished by optimizing parameter [29]. To obtain above objective, we adopt adversarial learning strategy to train generating network. Specially, we employ three discriminators , and as adversaries to prompt network-common encoder for better capturing network-common representation during the training process. For discriminator , it aims to classify representation from network P as true but representation from network O as false. In reverse, network-common encoder attempts to confuse three discriminators, rendering them unable to distinguish the different network representations. Finally, the differences of representation from different network will be vanished when three discriminators cannot distinguish generated representations. The corresponding loss function of discriminator to distinguish the representation from network P and O can be formulated as follows: Where represents the results predicted by discriminators. The total discriminative loss of three discriminators can be calculated as follows: Discriminators try to predict representation of one network as 1 and predict the representation of other network as 0, thus trained by minimizing Eq. 3. Three discriminators are learned with discriminative loss to accurately distinguish representations from different networks. On the other hand, network-common encoder is trained by generative loss and try to produce cross-network aligned representations that will confuse discriminators unable to differentiate them. And the generative loss of network-common encoder can be formulated as follows: To train the cross-network generative network, during the training phase, the encoder and discriminators are optimized by minimizing Eqs. 4 and 3 respectively. Then, to fully fuse of representations from three networks for extracting cross-network consistent feature of gene pair, we concatenate three network representations and adopt fusion network to extract high-level consistent representation of gene pair. The can be obtained by following: Where w is the weight and b is the bias of fusion network. Intra-network representation learning In addition to cross-network consistent information, intra-network specific information is also an effective complementary to network-common information. To learn network-specific representations for capturing intra-network information, we utilize intra-network representation learning module to learn network-specific representations in which three encoders and three evaluators are adopted in a cooperative way. Specifically, the specific representation of network k can be calculated: Where represents the private encoder of network k and is learnable parameter of network k encoder. To assess the difference of representations in network-specific subspace, we introduce three evaluators to differentiate network-specific representation. Specially, we employ three evaluators , and to distinguish the representation of one network from the representation of another network. For evaluator , it aims to assess specific representation of P network as true but specific representation of O network as false. Cooperatively, network-specific encoders try to learn network-private subspace by the assessed results of evaluators. This approach enables our model to learn network-specific representations that includes unique information from different networks, and the corresponding loss function of evaluator can be formulated as follows: Where represent the results assessed by evaluator respectively. And the total evaluator loss of three evaluators can be calculated as follows: Network-specific encoder is trained by generative loss and try to produce intra-network specific representations that can be effectively assessed by evaluators. The generative loss of encoder of network P can be formulated as follows respectively: And the total generative loss of intra-network feature generating network can be calculated as follows: To maintain the original information from initial representation, we employ network-common representation and network-specific representation to reconstruct the initial representation of gene pairs. We implement a reconstruction module consisted of two layers of neural network to recover the raw representation of different network. Given specific representation of network k and cross-network consistent representation , the initial representation of network k can be reconstructed by: Where represents the reconstructed representation of network k and represents the decoder of reconstruction. In train reconstruction module, we adopt mean square error to calculate the reconstruction loss, which is formulated as follows: Where is the square of Frobenius norm. In addition to reconstruction loss, to encourage network-common and network-specific encoders for learning disparate representations of gene pairs for each network, we impose orthogonal constraint in network-common feature subspace and network-specific feature subspace. Detailly, the representation in network-common and network-specific subspaces are enforced by orthogonality, and corresponding loss function can be computed as follows: Where is the square of F-norm, represents the feature matrix in network-common subspace and represents the feature matrix in network-specific subspace. SL prediction To accurately describe multi-network information of gene pair, MGANSL first constructs representations by concatenating network-common and network-specific features of gene pair, and the multi-network representation of gene pair is formulated as follows: Since deep neural networks (DNN) can capture complex non-linear relationships, thus we adopt DNN as SL decoder to predict the categories of gene pair . Given gene pair and its multi-network representation , we can predict its category as follows: Where represents the SL prediction module which consist of two layers neural network represents the probability of gene pair being SL pair. We adopt cross entropy to construct the classification loss of SL decoder, which is formulated as follows: Where represents the real label of gene pair . Overall loss and optimization In training, the learning process includes two steps to alternatively learning proposed model. One step is to train two representation learning module and SL prediction module during the generating phase, the other step is to train discriminator and evaluator during adversarial phase. Training two representation learning module and SL prediction module includes the generative loss of network-common encoder and network-specific encoders, the reconstruction loss of input decoder, orthogonality loss and SL classification loss, and the total generating loss function can be calculated as the linear addition of above loss terms: Where and are weight coefficient for controlling the effects of reconstruction loss and orthogonality loss. Training two discriminator and three evaluator includes discriminative loss and evaluator loss , and the total adversarial loss can be calculated as follows: We optimize with the loss during the generating training phase, and we optimize MGANSL with the loss during the adversarial training phase. We employ Adam optimizer to train our model. After optimization, all gene pairs are ranked to determine the most likely SL candidates. The code of MGANSL is implemented in Python 3.6 using Pytorch framework, and its open access is available in github (https://github.com/lijinxinchina/MGANSL/). Results Experimental setups Experimental setting and measuring metrics To evaluate the performance of different methods in discovering novel SL pairs and identifying SL partners for genes lacking known SL associations, we implement two kinds of data partition strategies to conduct cross validation: (1)CV1: 5-fold cross validation based on SL pairs. (2)CV2: 5-fold cross validation based on genes. Given the extremely large number of unlabeled gene pairs, we randomly sample an equal number of unknown pairs as negative samples. For CV1, all positive and negative samples are randomly divided into five subsets, with four subsets used for training and one for testing. For CV2, all genes are randomly divided into five subsets, and samples involving four subsets are used as the training set, while those involving the remaining subset form the testing set. To ensure the robustness and mitigate potential sampling bias, this 5-fold cross validation was independently repeated 10 times, and the final performance metrics are the averaged values over these runs. For assessing the effectiveness of model, we employ AUC, AUPR, F1 and Matthews Correlation Coefficient (MCC) as measurement indicators. Baseline methods We compare our approach with art-of-state methods to comprehensively assess the performance in SL prediction, and these baseline methods are listed as follow: Mashup [30] integrates different types of gene networks to learn gene consistent embedding for SL prediction. Mashup [30] integrates different types of gene networks to learn gene consistent embedding for SL prediction. DDGCN [11] learns gene interpretable representation in SL network by capturing neighbor node information for SL prediction. DDGCN [11] learns gene interpretable representation in SL network by capturing neighbor node information for SL prediction. GCATSL [15] fuses different feature graphs to construct gene features and predict SL associations with learned factor matrix. GCATSL [15] fuses different feature graphs to construct gene features and predict SL associations with learned factor matrix. SL2MF [10] learns probabilistic latent representations of genes to infer SL matrix by incorporating GO and PPI information. SL2MF [10] learns probabilistic latent representations of genes to infer SL matrix by incorporating GO and PPI information. GRSMF [31] learns gene representations and correlation matrix between genes from known SL matrix. GRSMF [31] learns gene representations and correlation matrix between genes from known SL matrix. CMF [32] fuses multiple biological sources to obtain gene latent representation with collective matrix factorization (CMF). CMF [32] fuses multiple biological sources to obtain gene latent representation with collective matrix factorization (CMF). TARSL [33] adopt a cross-network representation learning to learn gene representation for inferring novel SL associations. TARSL [33] adopt a cross-network representation learning to learn gene representation for inferring novel SL associations. Performance comparison Comparison of SL prediction for new association To assess the ability to predict SL associations of proposed method, We first compare our method with baseline methods in SLDB dataset under CV1 setting, and the results are illustrated in Tables 1 and 2 respectively. On SLDB dataset, MGANSL achieves best performance with AUC=0.9715, AUPR=0.9737, F-score=0.9051 and MCC=0.8127, which is higher 4.75%, 2.88%, 4.53% and 7.65% than the second-best method. Similarly, on SLDB2.0 dataset, MGANSL again achieves best results with AUC=0.9574, AUPR=0.9612, F-score=0.8869 and MCC=0.7754, outperforming the second-best method by 3.95%, 1.69%, 2.03% and 3.53% respectively. These results illustrate that MGANSL can be used in identifying novel SL associations effectively. Table 1Performance evaluation of different methods on SLDB dataset under CV1 settingMethodsAUCAUPRF-scoreMCCDDGCN0.8809±0.00490.9170±0.00360.8358±0.00840.7303±0.0117GCATSL0.9248±0.00400.9386±0.00350.8616±0.00340.7331±0.0072CMF0.8693±0.00300.8761±0.00330.8076±0.00370.6244±0.0125SL2MF0.8507±0.00650.9050±0.00460.8151±0.00700.7075±0.0082TARSLmashup0.9125±0.00260.9289±0.00200.8409±0.00420.6939±0.0072GRSMF0.9215±0.01030.9374±0.01030.8493±0.01030.7193±0.0258MGANSL0.9715± 0.00160.9737±0.00150.9051±0.00240.8127±0.0033The boldface denotes the best result and the underline denotes the second best Performance evaluation of different methods on SLDB dataset under CV1 setting The boldface denotes the best result and the underline denotes the second best Table 2Performance evaluation of different methods on SLDB2.0 dataset under CV1 settingMethodsAUCAUPRF-scoreMCCDDGCN0.8398±0.01290.8974±0.00450.8252±0.00370.7074±0.0063GCATSL0.9109±0.00410.9151±0.00300.8612±0.00380.7182±0.0077CMF0.8931±0.00480.8916±0.00370.8304±0.00640.6597±0.0141SL2MF0.9173±0.0016TARSL0.9421±0.00190.8607±0.00310.7470±0.0078mashup0.8926±0.00410.9078±0.00480.8224±0.00620.6466±0.0169GRSMF0.8993±0.00440.9245±0.00290.8394±0.00580.7032±0.0093MGANSL0.9574±0.00180.9612±0.00200.8869±0.00240.7754±0.0046The boldface denotes the best result and the underline denotes the second best Performance evaluation of different methods on SLDB2.0 dataset under CV1 setting The boldface denotes the best result and the underline denotes the second best Comparison of SL prediction for new genes To assess the ability of our method for predicting SL partner for genes lacking known SL associations, we compare MGANSL with baseline methods under CV2 setting. The performance results on the SynLethDB and SynLethDB2.0 datasets are illustrated in Tables 3 and 4 respectively. On the SLDB dataset, MGANSL achieves best results, with an average AUC of 0.9403, AUPR of 0.9438, F-score of 0.8610 and MCC of 0.7197, higher 10.32%, 7.44%, 11.60% and 8.11% than the second. On SynLethDB2.0 dataset, MGANSL also significantly outperforms all baseline methods, obtaining an average AUC of 0.9368, AUPR of 0.9391, F-score of 0.8645 and MCC of 0.7308, surpassing the second-best method by 14.6%, 11.93%, 15.77% and 30.45% respectively. In this challenging scenario, mashup also obtain global second performance in all methods, and the possible reason is that mashup integrated embedding captures generalizable biological information applicable even to unseen genes. These metric indicators demonstrate that MGANSL could effectively discover SL associations for novel genes, showing its robust performance in this challenging scenario. Table 3Performance evaluation of different methods on SLDB dataset under CV2 settingMethodsAUCAUPRF-scoreMCCGCATSL0.7494±0.05050.7968±0.04520.7008±0.02270.3263±0.1795CMF0.6989±0.04950.7227±0.04430.6825±0.01820.2061±0.1228SL2MF0.8727±0.02200.9100±0.01640.8101±0.0276 TARSL0.8108±0.02360.8478±0.02110.7419±0.02790.4731±0.0968Mashup 0.5532±0.0726GRSMF0.8027±0.07930.8219±0.07890.7484±0.03400.4302±0.1131MGANSL 0.9403±0.0204 0.9438±0.0162 0.8610±0.0311 0.7197±0.0565 The boldface denotes the best result and the underline denotes the second best Performance evaluation of different methods on SLDB dataset under CV2 setting The boldface denotes the best result and the underline denotes the second best Table 4Performance evaluation of different methods on SLDB2.0 dataset under CV2 settingMethodsAUCAUPRF-scoreMCCGCATSL0.7050±0.04200.7547±0.03850.6499±0.04030.3065±0.0715CMF0.7256±0.04130.7367±0.03810.6922±0.02520.2574±0.1884SL2MF0.7499±0.03020.8357±0.02110.7235±0.0289 TARSL 0.8352±0.0184 0.4953±0.0469Mashup0.8153±0.0299 0.7444±0.02600.4653±0.0620GRSMF0.7512±0.03030.7982±0.03210.6922±0.00970.2667±0.1153MGANSL 0.9368±0.0145 0.9391±0.0157 0.8645±0.0183 0.7308±0.0402 The boldface denotes the best result and the underline denotes the second best Performance evaluation of different methods on SLDB2.0 dataset under CV2 setting The boldface denotes the best result and the underline denotes the second best Model analysis Parameter analysis There are many important parameters affecting the predictive performances, including input size of feature , final embedding dimension , the coefficient of reconstruction loss and the coefficient of orthogonality loss . In order to determine optimal parameters for SL prediction, we conduct parameter analysis based on SLDB datasets. For simplicity, we search optimal value for each parameter by fixing other parameters. For and , we select it from following range: . For and , we select if from following range:. As illustrated in Fig. 2(a), the performance of the model gradually improves as the input feature dimension increases, reaching best result when . This is because a smaller size of feature may lead to the loss of critical information about gene pairs, making it difficult to fully represent complex information across multiple networks. The effects of are shown in Fig. 2(b), the performance of MGANSL improves significantly with increasing , reaching its optimum when . The possible reason is that a too-small embedding dimension cannot capture sufficient information, while an excessively large dimension may introduce redundant or noisy features. Fig. 2(c) and Fig. 2(d) illustrate the effects of the and respectively. The results show that appropriate regularization coefficients can effectively improve model performance, while excessively small or large values can lead to suboptimal results. In summary, according to the performance under different parameter settings, the optimal parameters of MGANSL are determined as ,, , and respectively. The parameter analysis of MGANSL on SLDB dataset under CV1 setting Architecture analysis To investigate the effect of adversarial generation in capturing network-common and network-specific information, we analyze the architecture of network-common encoding (NCE) and the network-specific encoding (NSE) module, and assess the predictive performance under different architecture. Considering the number of encoding layers in network-common and network-specific generating module, we design different GAN structures for NCE module and NSE module respectively. For NCE module, the layer of generating network changes in 1,2,3,4,5,6 and the details structure are listed below: NCE1:256,64; NCE2:256,128,64; NCE3:256,256,128,64; NCE4:256,256,128,128,64; NCE5:256,256,128,128,64,64; NCE6:256,256,128,128,64,64,64; For NSE module, the layer of generating network changes in 1,2,3,4,5,6 and the details structure are listed below: NSE1: 256,64; NSE2: 256,128,64; NSE3: 256,128,128,64; NSE4: 256,128,128,64,64; NSE5: 256, 256, 128,128,64,64; NSE6: 256,256,128,128, 128,64,64; The performance of different network structure are illustrated in Fig. 3. The performance comparison on different GAN structure for adversarial learning As shown in Fig. 3(a), these results illustrate that the performance of MGANSL significantly improves with the increasing of layers of network-common encoding module and obtain the best result (AUC=0.9698, AUPR=0.9711) when employing the NCE6 structure. This suggests that cross-network shared information exhibits high nonlinearity, requiring a deeper generative model to fully capture such patterns. For subsequent evaluation to the effect of NSE module, we changed the NSE architectures based on NCE6 architecture. The effect of MGANSL under different NSE structure are illustrated in Fig. 3(b). We can discover that the MGANSL is relatively stable to the layers of network-specific encoding module, and the performance of MGANSL performs best when utilizing the NSE1 structure. The possible reason is that the network-specific information is uncomplicated and linear. Thereby, a simple NSE architecture can more effectively capture these patterns and avoid overfitting. Representations visualization To qualitatively interpret the ability of learned multi-network representation to distinguish SL and non-SL samples, we visualize learned final multi-network representations embedding of SL and non-SL samples using t-SNE algorithm which widely used in the field of visualization. The results are shown in Fig. 4, and we can find a clear boundary between SL and non-SL samples, which indicate that learned multi-network representations are informative and discriminative. These results confirm that MGANSL effectively learns inter-network consistent information and intra-network specific information by constructing common and private subspaces, which significantly contribute to distinguishing SL and non-SL samples. The visualization for SL and non-SL sample with t-SNE Ablation study Our proposed method mainly includes two vital modules: (i) Network-shared encoding module for capturing cross-network consistent information, and (ii) Network-private encoding module for extracting intra-network specific information. To study their effects on SL prediction, we implement two variant models of MGANSL as follows: MGANSL-S: it only uses network-shared encoding module to capture cross-network consistent information for SL prediction. MGANSL-S: it only uses network-shared encoding module to capture cross-network consistent information for SL prediction. MGANSL-C: it only uses network-private encoding module to extract intra-network specific information for SL prediction. MGANSL-C: it only uses network-private encoding module to extract intra-network specific information for SL prediction. We conduct experiment on SLDB dataset with our method and two variants under CV1 setting, and the results are illustrated in Fig. 5. In these results, MGANSL-C acquires better performance than MGANSL-S, which indicates that network-specific information plays a more prominent role in SL prediction. Nevertheless, the full MGANSL model achieves the highest performance among all variants, demonstrating that shared and private information are complementary and that integrating both types of features yields a more comprehensive and robust representation for SL prediction. Comparison between MGANSL and its two variants on two dataset under CV1 setting Case study SL pairs verified by other works To further validate the biological relevance of the SL gene pairs predicted by MGANSL, we conducted a case study using the SynLethDB dataset. Specifically, the optimal cross-validated model was retrained on all known SL pairs, and all unknown sample are sorted by their prediction scores and are checked with SLDB2.0 dataset. The experimental illustrates that 61 gene pairs in top-1000 gene pairs were reported in SLDB2.0. Table 5 illustrates 10 selected SL pairs, and the rest gene pairs are provided in Supplementary Table S1. Table 5Top 10 predicted SL pairs with literature evidenceNumberGene1Gene2Pubmed IDNo. 1KRASMAPK126627737No. 2MAP2K7KRAS24104479No. 3KRASRNASEL27655641No. 4RETKRAS27655641No. 5KRASPAK127655641No. 6MAPKAP1KRAS27655641No. 7HCKMYC22623531No. 8KRASRAP1A24104479No. 9POLR2AKRAS27655641No. 10FHKRAS24104479 Top 10 predicted SL pairs with literature evidence As shown in Table 5, these gene pairs are validated by other methods and computational tools. For instance, No. 3 (KRAS and RNASEL) and No. 4 (RET and KRAS) is detected by RNAi Screening [34]. No. 1(KRAS and MAPK1) is detected by High Throughput screening method [35]. Although some pairs are only detected by text mining, some additional evidences also can support these pairs. For example, HCK regulates the expression of CDK6 through the ERK1/2-c-Myc axis, which contributes to the self-renewal and proliferation of LSCs [36]. These evidences verify that our proposed method can effectively discover novel SL associations, which provides strong candidate for biological clinical experiment. Clinical survival analysis To assess the clinical utility of the predicted SL pairs, we performed a survival analysis using the top 1000 predicted SL pairs on breast cancer patients. Clinical and copy number alteration data were obtained from cBioPortal. The patients are divided into two groups with the status of copy number for predicted SL pair : (i) Group1: both mutation of and . (ii) Group2: single mutation or no mutation of g1 and g2. Using the log-rank test, we found that 20 SL pairs exhibit significantly better prognosis (p-value0.05), and the results are provided in Supplementary Table S2. Kaplan–Meier survival curves for four representative pairs are shown in Fig. 6. The results represent that the group1 with g1 and g2 both mutated have longer survival time, with statistical significance (p-value0.05). These results show that predicted SL pairs may have potential clinical relevance and could inform targeted therapeutic strategies. The survival curves of 4 predicted SL pairs plotted by Kaplan-Meier. a CARS and NRAS, b CMPK1 and HRAS, c HRAS and ITGB3BP, d HRAS and PLK3 Drug repositioning analysis To explore the potential of predicted SL pairs in clinical anticancer drug discovering, we conduct drug repositioning analysis with top-1000 predicted SL pairs by retrieving whether one gene is cancer driver and another gene is targeted by known drugs. We download cancer drivers from cancer gene census list [37] and drug-targets associations from Drugbank [38] respectively. We obtain 311 gene pairs targeted by known drugs and the detail are provided in Supplementary Table S3. Fig. 7 illustrates 10 predicted SL pairs related with specific cancer and targeted by known drug. For predicted SL pair KRAS and MAPK1, KRAS gain-of-function mutations are causal mechanism for hepatic vascular cavernomas [39]. As the SL partner of KRAS, MAPK1 signaling inhibition rescues hepatic vascular cavernoma formation in endothelial cell and MAPK1 has potential as a novel therapeutic target. Acetylsalicylic acid (ASA) is the active compound of aspirin, and it is the primary drug used to prevent colorectal cancer [40]. It indicates that drug Acetylsalicylic acid have promising to treat cancer mutated with KRAS by inhibiting MAPK1. These results verify that our predicted SL associations have potentials in clinical anticancer drug discovery. Drug repositioning analysis with predicted SL pairs Drug sensitivity analysis To evaluate the drug response of predicted SL pairs, we conduct drug sensitivity analysis with predicted SL pairs. The pharmacogenomics profiles are download from GDSC dataset, which comprised of IC50 values of 12 cell lines targeted by 251 drugs. In drug sensitivity analysis, gene pairs satisfying two conditions will be considered for following analysis: (i) One gene of SL pair is a druggable target in GDSC database. (ii) Another gene of SL pair both possesses the cases of mutation and non-mutation in different cancer cell lines. Then, we check the drug response of two cases with Wilcoxon rank sum test. In the predicted top 1000 SL pairs, 20 combinations of gene pair and drug show significantly high drug sensitivity. Table 6 shows 10 combinations of gene pairs and drug with significant drug sensitivity, and the detail are provided in Supplementary Table S4. From Table 6, we can see that gene EGEF is sensitivity in many combinations of gene and drug, which illustrates that EGEF has potential be treated as therapeutic target in clinical treatment. These results indicates that our predicted SL pairs are sensitivity to target drug in cancer cell lines, which provide potential drug candidate for therapeutic target. Table 6SL interactions related to drug sensitivity in cancer cell linesNo.Gene AGene BDrug targeted on Gene AMean of drug sensitivity (IC50)p-valueCell lines with Gene B mutationCell lines with Gene B wild type1ERBB2HRASAZD89311.6441.9680.0022ERBB2HRASAfatinib1.9172.2190.0143EGFRPDGFRAAZD89311.6601.8710.0274EGFRPDGFRACetuximab5.9176.0550.0055EGFRPDGFRALapatinib2.3802.6350.0096KITPAK1Dasatinib0.1110.5860.03357KITPDGFRADasatinib-0.0750.6230.0028EGFRPAK1AZD89311.6331.9480.0179EGFRPAK1Cetuximab5.8946.1110.01610EGFRPAK1Lapatinib2.3852.6860.037 SL interactions related to drug sensitivity in cancer cell lines Conclusion In this study, we proposed multi-network adversarial generating model for capturing intricate feature of gene pairs. Specifically, a novel computational model named MGANSL was proposed to predict novel SL associations. MGANSL integrates a network-common encoder and a network-specific encoder to jointly learn shared and private representations from heterogeneous biological networks. Comprehensive experiments demonstrate MGANSL can effectively predicting SL interactions by capturing comprehensive multi-network feature of gene pairs. However, there are some limitations that needed be addressed in the future works. First, incorporating additional biological data types (e.g., methylation and co-expression profiles) may further enrich the learned representations. Second, more advanced multi-network integration strategies could improve the robustness of feature fusion. Finally, experimental validation of predicted SL pairs will be essential for translating computational predictions into biological and clinical applications. Supplementary Information Supplementary Material 1. Author contributions XGL and JXL supervised and designed this work. JXL wrote the main manuscript text and implement the proposed model. ZHL downloaded and processed the used data. XL wrote the section 3.2 and section 3.3. HRL prepared figures 2-7. JJR wrote the section 3.1. HRL and JJR performed the experiments. All authors reviewed the manuscript. Funding The work was supported by the National Natural Science Foundation of China under Grant 62372160. Data availability All data used in this manuscript were downloaded from public repositories. The synthetic lethal interaction samples are download from SLDB (https://synlethdb.sist.shanghaitech.edu.cn/) and SLDB2.0 (https://synlethdb.sist.shanghaitech.edu.cn/v2). The pretein-protein interactive data are download from HPRD dataset (http://hprd.org/download). The gene functional annotations are obtained from Gene Ontology dataset (http://geneontology.org/docs/download-go-annotations/). Declarations Ethics approval and consent to participate Not applicable Consent for publication Not applicable Competing interests The authors declare no competing interests Footnotes Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1. Data Availability Statement All data used in this manuscript were downloaded from public repositories. The synthetic lethal interaction samples are download from SLDB (https://synlethdb.sist.shanghaitech.edu.cn/) and SLDB2.0 (https://synlethdb.sist.shanghaitech.edu.cn/v2). The pretein-protein interactive data are download from HPRD dataset (http://hprd.org/download). The gene functional annotations are obtained from Gene Ontology dataset (http://geneontology.org/docs/download-go-annotations/).