SimMapNet: a Bayesian framework for gene regulatory network inference using gene ontology similarities as external hint Authors: Maryam Shahdoust, Rosa Aghdam, Mehdi Sadeghi Published: 2026 ABSTRACT AbstractGene regulatory network (GRN) reconstruction is a fundamental challenge in computational biology, and is crucial for understanding gene interactions. In this study, we aim to incorporate Gene Ontology (GO) similarities into the construction of GRNs. Our key assumption is that genes with higher similarity in Molecular Function, Biological Process, or Cellular Component categories are more likely to be functionally related and, therefore, more likely to be connected in the network. We introduce SimMapNet, a Bayesian framework that estimates the precision matrix, which serves as the adjacency matrix in a Gaussian Graphical Model for undirected GRN inference. SimMapNet enhances network inference by integrating GO similarities, which inform the hyperparameters of the prior distribution through a kernel function, incorporating biological prior knowledge in a principled manner. We evaluate SimMapNet on three datasets: two datasets from the SOS DNA-repair response pathway in Escherichia coli and one dataset from Drosophila melanogaster. The results demonstrate the algorithm’s superior performance compared to state-of-the-art methods such as GLASSO, GENIE3, and KBOOST in terms of F1-score. SimMapNet has low time complexity, making it suitable for constructing large networks. Our simulation results confirm that SimMapNet is particularly well-suited for scenarios with limited sample sizes, where traditional methods often struggle.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s12859-026-06542-9.Keywords: Gene regulatory networks, Bayesian inference, Gene ontology similarities, Gaussian graphical model Introduction Developing an integrative framework for Gene Regulatory Network (GRN) that combines multiple data sources, machine learning algorithms, and network-based computational strategies remains a crucial challenge in the field, requiring continuous innovation and interdisciplinary collaboration [1]. Decades of biomedical research have led to the accumulation of extensive biological knowledge, including pathway information, transcription factor-target interactions, gene ontology (GO) annotations and other functional information, which are now available in public databases [2, 3]. Leveraging prior biological data can significantly enhance the statistical power and interpretability of GRN reconstruction, particularly in relation to complex phenotypes. However, traditional network inference methods rely solely on gene expression data [4–9], often resulting in high false positive and false negative rates in the inferred network due to noise and limited coverage [10]. Studies have demonstrated that incorporating prior biological knowledge can refine network structure, reduce uncertainty, and improve predictive power [11–14]. For instance, the Prior Lasso (pLasso) method partitions gene interactions based on pathway knowledge [14] and PriorPC that uses soft priors that assign to edges a probability of existence [12]. More recent approaches, such as GRNPT, integrate transformer-based embeddings from large-scale biological data to capture regulatory patterns from single-cell RNA sequencing [15]. In this study, we incorporate GO similarities into the construction of undirected GRNs. Our key assumption is that genes with higher similarity in the Molecular Function (MF), Biological Process (BP) categories are more likely to be functionally related and, therefore, more likely to be connected in the network. This assumption is biologically motivated, as genes involved in similar biological processes or sharing molecular functions often participate in the same regulatory pathways [16–18]. Additionally, we also utilize Cellular Component (CC) similarities to assess the impact of different types of external information in improving the accuracy and robustness of GRN construction. We introduce SimMapNet, a Bayesian framework that estimates the precision matrix, which serves as the adjacency matrix in a Gaussian graphical model (GGM) for GRN inference. In this approach, the precision matrix,i.e., the inverse of the covariance matrix, represents the network structure, where non-zero entries indicate regulatory relationships between genes [19]. To estimate the precision matrix, we adopt a Bayesian approach, incorporating GO similarities into the estimation of the hyperparameters of the prior distribution. Specifically, GO similarities help to define the prior covariance structure, which guides the Bayesian inference process. Previously, we developed F-MAP [20], which combined Gaussian graphical models with a Bayesian framework to infer the GRN of one species while incorporating gene expression data from related species. In F-MAP, external information was integrated by estimating hyperparameters in the prior distribution through factor analysis of the covariance matrix of related species. Following a similar conceptual framework, SimMapNet introduces a distinct methodology by incorporating GO-based functional relationships through kernel functions. Unlike F-MAP, which relies on cross-species covariance structures, SimMapNet directly integrates functional similarity measures into the prior distribution, enabling GO similarities to systematically refine the inferred network structure. We evaluate SimMapNet using three different microarray gene expression datasets. Two datasets belong to the SOS DNA-repair response pathway in Escherichia coli, one with 9 samples and the other with 466 samples, representing different sample sizes of a small-scale GRN [21–23]. The third dataset, from Drosophila melanogaster [24], is considered due to its high dimensionality, which allows us to assess SimMapNet’s performance on large-scale data. The results demonstrate SimMapNet’s ability to more accurately identify functionally related genes. This novel approach results in a more robust and biologically interpretable network by incorporating functional relationships. Compared to established methods, including Graphical Lasso (GLASSO) [25], KBOOST [26], and GENIE3 [27],SimMapNet exhibits superior performance in ensuring biologically relevant connections among genes. Notably, to ensure a fair and direct comparison with our undirected graphical framework, the inherently directional importance scores of KBOOST and GENIE3 were symmetrically transformed into undirected edge weights. Methods SimMapNet constructs the GRN within the GGM framework [19, 28], assuming gene relationships follow a multivariate normal distribution, and estimates the precision matrix. The algorithm integrates Bayesian inference and kernel methods to estimate the precision matrix, enforce sparsity and then build adjacency matrices representing regulatory relationships. The Bayesian estimation, GO similarities, and the stepwise implementation of SimMapNet are detailed in the following subsections, with additional information on the Bayesian framework and GGM in the supplementary file. Bayesian inference of precision matrix Let be an n × p matrix, where each , for , is an independent and identically distributed multivariate normal observation, . Here, is an unknown p × p covariance matrix, and Θ is the corresponding precision matrix, which is assumed to be positive definite. Wishart prior distribution: The Inverse Wishart distribution is a commonly used prior for [29, 30]. Based on the relationship between the Wishart and inverse Wishart distributions, the prior distribution on Θ is the Wishart distribution: , where ν is the degrees of freedom, which must be greater than (p-1), and Ω is a p × p positive definite matrix [30]. The Wishart distribution is the conjugate prior for the population precision matrix of a multivariate normal distribution. Therefore, the posterior distribution of Θ follows the Wishart distribution, , where: The mode of the posterior distribution, known as the maximum a posteriori (MAP) estimate, serves as an estimator for Θ : Hyperparameter estimation: The parameter ν is empirically determined and can, in principle, be learned from the data, starting from an initial value greater than p - 1, where p denotes the number of variables (genes) [30, 31]. Following the recommendations of Zhang et al. [32], we adopt a stronger prior specification by setting . In this study, we fix ν = 2p to ensure a balanced contribution of the prior information within the Bayesian framework and to maintain consistency across different prior components. To estimate the hyperparameter Ω , GO similarities are transformed into a covariance structure using kernel functions, incorporating gene relationship data. GO similarities range from 0 to 1, and we calculate distances between genes as , which are then input into the kernel function. Finally, the hyperparameter Ω is estimated as: where, is a p × p diagonal matrix, whose diagonal entries represent the standard deviations of the corresponding genes. The p × p matrix K encodes prior information about gene correlations obtained from the kernel function. The matrix is a diagonal matrix, where is a positive parameter that ensures the estimated Ω is positive definite and possesses desirable algebraic properties. In this study, we focus on two stationary isotropic kernels, which depend only on the distance d(x, x') between inputs [33, 34]. Squared exponential (SE) kernel [13, 33, 35] (also known as Gaussian or RBF) and Ornstein-Uhlenbeck (OU) kernel, part of the Matérn group [35]. The formulas and parameter definitions for both kernels are provided in the supplementary file (Section 2, subsection “Kernel functions”).. Gene ontology (GO) similarities GO classifies gene functions into Biological Process (BP), Molecular Function (MF), and Cellular Component (CC) [16, 36]. GO similarities are computed using semantic similarity measures such as Resnik’s [37] and Wang’s similarity [38] (See the supplementary file, Section 3, for more details.). These similarities are mapped to kernel functions to define distances between genes. SimMapNet steps The steps of the algorithm are presented below. The algorithm is implemented as an R package, available at GitHub. Figure 1 displays a graphical abstract of the SimMapNet algorithm. Graphical abstract of the SimMapNet algorithm. The “ths” in the Sparsification section denotes the threshold applied to induce sparsity in the MAP estimation of the precision matrix Step 1: data pre-processing: The input to the algorithm consists of gene expression data which is an n × p gene expression matrix, where n and p represent the number of samples and the number of genes, respectively. This matrix can be obtained from various transcriptomic platforms. To ensure the data conforms to a multivariate normal distribution, a normalization transformation is applied. One common transformation is log-transformation, where the of the data is computed [39]. The gene expression vectors for each sample are then centered so that each gene has zero mean across samples. Step 2: similarities calculation and Kernel computation: Calculate the similarities (or differences) between variables. We calculate the GO similarities between genes using the GOSemSim R package [36]. These similarities are then converted into distances and passed to a kernel function to construct the kernel matrix (K). Step 3: Bayesian inference of the precision matrix: This step includes multiple stages: 3.1Set a Wishart prior for the precision matrix: . Set a Wishart prior for the precision matrix: . 3.2Choose a value equal to (or greater than) 2p as the prior degree of freedom ν . Choose a value equal to (or greater than) 2p as the prior degree of freedom ν . 3.3Estimate the hyperparameter Ω using Equation (3). Estimate the hyperparameter Ω using Equation (3). 3.4Estimate the posterior distribution: using Equation (1). Estimate the posterior distribution: using Equation (1). 3.5Estimate the MAP of the precision matrix using Equation (2). Estimate the MAP of the precision matrix using Equation (2). Step 4: make sparse precision matrix: To induce sparsity in the MAP estimate, we apply a hard-thresholding method. In this process, different percentiles of the absolute values of the estimated partial correlations are used as thresholds.The sparsity threshold () can be selected either by maximizing a performance measure (e.g., F1-score) when a reference network is available, or by minimizing the EBIC(see the next subsection) when the true network is unknown. Step 5: make adjacency matrix: The final binary adjacency matrix is derived from the sparse precision matrix, where non-zero elements (set to one) indicate the presence of an edge between the corresponding gene pairs. Computational complexity and implementation details The computational framework of SimMapNet is implemented entirely within the R statistical environment. The primary theoretical computational bottleneck of the Gaussian Graphical Model optimization routine involves p × p matrix inversions, resulting in a worst-case time complexity of , where p denotes the number of genes. However, by deriving exact Maximum A Posteriori (MAP) closed-form optimization updates, SimMapNet functions as an efficient shrinkage estimation framework. It completely bypasses the computationally intensive, iterative stochastic search loops often required by alternative network learning algorithms. This structural formulation allows the model to execute highly accelerated structural learning by updating the precision matrix directly through fast matrix algebra. Empirical benchmarking profiles across varying sample dimensions demonstrate that network reconstruction scales efficiently, completing model convergence well within seconds on standard desktop hardware architectures (see Supplementary Section 9 for complete runtime comparisons). Reference-independent hyperparameter selection using EBIC In benchmark datasets with available reference networks, the hyperparameters of SimMapNet can be selected by maximizing network reconstruction accuracy. However, in practical applications, the true network structure is generally unknown and performance measures such as the F1-score cannot be computed. Therefore, we additionally employ the Extended Bayesian Information Criterion (EBIC) as a reference-independent model selection criterion. For a candidate sparse precision matrix Θ , the EBIC is computed as: where is the Gaussian graphical model log-likelihood evaluated at the sparse precision matrix Θ , |E| denotes the number of edges in the inferred network (i.e., the number of non-zero off-diagonal elements in the upper triangular part of Θ ), n is the number of samples, p is the number of genes, and is a tuning parameter controlling the complexity penalty. Following previous studies, we set γ = 0.5. For a given dataset, a grid of candidate hyperparameter values, including kernel hyperparameters, covariance regularization parameters, sparsity thresholds, and, when desired, the prior degree of freedom (ν ), is explored. The EBIC is calculated for each candidate model, and the parameter configuration yielding the minimum EBIC is selected as the optimal model. Configurations yielding undefined likelihoods, which typically arise from non-positive-definite sparse precision matrices, are excluded from consideration. Unlike F1-score-based selection, the EBIC does not require knowledge of the true network structure and is therefore applicable to real biological datasets where a reference network is unavailable. Configurations yielding undefined likelihoods, which typically arise from non-positive-definite sparse precision matrices, are excluded from consideration. Results The results of implementing SimMapNet on different datasets are evaluated by following performance metrics, such as True Positive rate (TPR), False Positive rate (FPR), precision (PPV), accuracy (ACC), and F1-score [40]. The mathematical definition of these metrics are provided in supplementary file Section 4. TPR and FPR are also used to plot the receiver operating characteristic (ROC) curve, and the area under the ROC curve (AUC) is calculated. Similarly, the Precision-Recall (PR) curve is plotted using PPV and TPR, and the area under this curve (PRAUC) is calculated [41, 42]. The performance of SimMapNet is benchmarked against three well-known methods: GENIE3 [27], KBOOST [26], and GLASSO [25]. In addition, to more precisely evaluate the role of GO similarities, Euclidean distances have been used as prior information for calculating the kernel functions, following the approach described in the SimMapNet methodology. The parameters of the algorithm and other methods are trained based on their corresponding diagnostic measures, specifically the F1-score. The optimal parameter set is selected by maximizing the F1-score. To ensure the reliability and robustness of SimMapNet, we systematically evaluated its performance using bootstrap resampling of the input samples and computed the 95% confidence interval (CI) for the F1-score obtained through bootstrap sampling [43]. The CIs are calculated according to the percentile method using percentiles of the bootstrap distribution; . These evaluations allowed us to assess the stability of SimMapNet in reconstructing GRNs and its sensitivity to variations in input data. The F1-score is used as the primary evaluation metric, and the bootstrap procedure reflects variability induced by sampling uncertainty in the observed data. To evaluate the robustness of SimMapNet, we performed bootstrap resampling over the input samples (observations). Specifically, bootstrap datasets were generated by sampling n observations with replacement from the original data matrix. For each bootstrap replicate, the complete SimMapNet pipeline was re-applied, including network estimation, sparsification, and performance evaluation against the reference network. We used 1 000 bootstrap replicates to estimate the variability of performance metrics. For each replicate, the F1-score was computed, and 95% confidence intervals were obtained using the percentile method based on the empirical bootstrap distribution. We additionally report the bootstrap mean and standard deviation across replicates. To assess statistical significance of performance differences, paired bootstrap tests were performed between the best-performing SimMapNet variant and competing methods. The full results of these analyses are provided in Supplementary Section 5. SOS DNA repair network dataset The SOS DNA repair network is a small-scale, experimentally validated gene expression dataset consisting of nine genes. There are two real microarray SOS datasets, varying in terms of their sample size (referred here as SOS1 and SOS2). SOS1 contains nine samples [21, 22], while SOS2 (version 4, build 6) contains 466 samples for the same nine genes [23]. The reference network includes 24 regulatory interactions (edges). GO similarities for MF and BP were mapped for eight genes. As a result, the edges associated with umuDC were removed from the reference network, yielding a final network with 22 edges. Figure 2 displays the SOS reference networks, labeled “Reference Network”, along with the reconstructed networks using GO similarities MF (MF-GO), GO similarities BP (BP-GO) for SOS1 and SOS2. Due to the absence of CC similarity data for some genes, the network construction based on CC similarity was performed separately. In the reference network, edges lacking CC information were excluded, resulting in the reference network (SimMapNet_CC), which contains only eight edges (see Figure S1 for the Reference Network). The constructed GRN using CC similarities are in Figure S1. In addition, the performance metrics for both datasets are in Table 1 and Table 2. Figure 3 displays box plots of the F1-scores achieved from 1000 bootstrap sampling for different methods. The results show that SimMapNet, particularly with molecular function (SimMapNet_MF) and biological process (SimMapNet_BP) GO similarities, achieves the highest F1-scores across both SOS1 and SOS2 datasets, outperforming GLASSO, GENIE3, and KBOOST. The boxplots of bootstrap sampling for both SOS1 and SOS2 confirm SimMapNet robustness, as it maintains a high median F1-score with reasonable variability. SOS Gene regulatory networks based on MF and BP similarities. The Reference Network represents the true network where each node shows a gene and each edge shows the true relationship between genes. True positive edges are in black and false positives are in red. MF-GO (SOS1) and BP-GO (SOS1) are the reconstructed networks for SOS1 using SimMapNet with MF and BP GO similarities, respectively. Similarly, MF-GO (SOS2) and BP-GO (SOS2) are for SOS2 Table 1Performance metrics of different methods for SOS1 constructed networksMethodTPTNFPFNTPRFPRPPVAccuracyF1_scoreAUCPRAUCBootstrap CI (95%)SimMapNet_MF178210.940.200.890.890.920.870.810.78 ± 0.090.65 – 0.97SimMapNet_BP168220.890.200.890.860.890.820.840.77 ± 0.070.61 – 0.89SimMapNet_CC76110.880.140.880.870.880.790.730.63 ± 0.120.38 – 0.88SimMapNet_dist166420.890.400.800.790.840.710.760.72 ± 0.070.58 – 0.84GENIE3152830.830.800.650.610.730.460.610.70 ± 0.040.63 – 0.78GLASSO122860.670.800.600.500.630.290.510.56 ± 0.130.29 – 0.76KBOOST172810.940.800.680.680.790.490.610.75 ± 0.040.70 – 0.84Performance of SimMapNet on SOS2 using different GO similarities (SimMapNet_MF, SimMapNet_BP, SimMapNet_CC) and Euclidean distances (SimMapNet_dist), alongside GENIE3, GLASSO, and KBOOST. TP, TN, FP, and FN represent True Positives, True Negatives, False Positives, and False Negatives. Bootstrap performance results. denotes the bootstrap mean and standard deviation of F1-score computed over 1000 bootstrap resamples. Bootstrap CI (95%) denotes the 95% confidence interval for the F1-score obtained using the percentile bootstrap method Performance metrics of different methods for SOS1 constructed networks Performance of SimMapNet on SOS2 using different GO similarities (SimMapNet_MF, SimMapNet_BP, SimMapNet_CC) and Euclidean distances (SimMapNet_dist), alongside GENIE3, GLASSO, and KBOOST. TP, TN, FP, and FN represent True Positives, True Negatives, False Positives, and False Negatives. Bootstrap performance results. denotes the bootstrap mean and standard deviation of F1-score computed over 1000 bootstrap resamples. Bootstrap CI (95%) denotes the 95% confidence interval for the F1-score obtained using the percentile bootstrap method Table 2Performance metrics of different methods for SOS2 constructed networksMethodTPTNFPFNTPRFPRPPVAccuracyF1_scoreAUCPRAUCBootstrap CI (95%)SimMapNet_MF164620.890.600.730.710.800.660.710.75 ± 0.040.70 – 0.80SimMapNet_BP173710.940.700.710.710.810.600.700.75 ± 0.040.71 – 0.81SimMapNet_CC73410.880.570.640.670.740.660.600.69 ± 0.070.53 – 0.84SimMapNet_dist172810.940.800.680.680.790.660.710.77 ± 0.030.72 – 0.84GENIE3117370.610.300.790.640.690.580.750.63 ± 0.040.56 – 0.69GLASSO138250.720.200.870.750.790.710.830.74 ± 0.030.66 – 0.79KBOOST163720.890.700.700.680.780.700.810.75 ± 0.030.68 – 0.78Performance of SimMapNet on SOS2 using different GO similarities (SimMapNet_MF, SimMapNet_BP, SimMapNet_CC) and Euclidean distances (SimMapNet_dist), alongside GENIE3, GLASSO, and KBOOST. TP, TN, FP, and FN represent True Positives, True Negatives, False Positives, and False Negatives. Bootstrap performance results. denotes the bootstrap mean and standard deviation of F1-score computed over 1000 bootstrap resamples. Bootstrap CI (95%) denotes the 95% confidence interval for the F1-score obtained using the percentile bootstrap method Performance metrics of different methods for SOS2 constructed networks Performance of SimMapNet on SOS2 using different GO similarities (SimMapNet_MF, SimMapNet_BP, SimMapNet_CC) and Euclidean distances (SimMapNet_dist), alongside GENIE3, GLASSO, and KBOOST. TP, TN, FP, and FN represent True Positives, True Negatives, False Positives, and False Negatives. Bootstrap performance results. denotes the bootstrap mean and standard deviation of F1-score computed over 1000 bootstrap resamples. Bootstrap CI (95%) denotes the 95% confidence interval for the F1-score obtained using the percentile bootstrap method The box plots show the F1-scores obtained from 1000 bootstrap samples for different methods, where (a) corresponds to SOS1 and (b) to SOS2. SimMapNet is evaluated using GO similarities in Molecular Function (MF), Biological Process (BP), and Cellular Component (CC), denoted as SimMapNet_MF, SimMapNet_BP, and SimMapNet_CC, respectively. SimMapNet_dist refers to the version of SimMapNet that incorporates Euclidean distances between gene expressions. For comparison, the results of three other methods–GLASSO, KBOOST, and GENIE3–are also included Comparing the results of SOS1 and SOS2 reveals that the number of false positive edges in SOS2 is higher than in SOS1, possibly due to greater expression variability resulting from its larger sample size. To assess the impact of sample size on the performance of the constructed networks, we conducted a benchmark on SOS2 using different sample sizes (20, 50, 100, and 200). The samples were selected based on their variability in gene expression. Specifically, we sorted the samples according to their standard deviation across all genes and selected the top n samples with the highest standard deviations. Figure 4 displays the changes in F1-score and AUC for different sample sizes across various methods. The complete results are in Table S1. a The trend of changes in F1-score of the SOS network across different sample sizes for various methods. b The trend of changes in AUC of the SOS network across different sample sizes for various methods. SimMapNet using GO similarities MF, BP, and CC are represented as SimMapNet_MF, SimMapNet_BP, and SimMapNet_CC, respectively. SimMapNet_dist represents the SimMapNet method using Euclidean distances between gene expressions. Other methods are indicated by their names: GLASSO, KBOOST, and GENIE3. A sample size of 9 corresponds to the results from SOS1 data, while the other sample sizes reflect results from sample selections in SOS2. The sample size of 466 represents the results of the main SOS2 dataset (whole information) Drosophila melanogaster dataset To assess SimMapNet’s performance on high-dimensional datasets, we applied it to gene expression data from D. melanogaster (amel), obtained from Kalinka et al. [24] (ArrayExpress, E-MTAB-404). The dataset includes over 3,000 genes, but we focus on those in the reference network, which involves 12 transcription factors (TFs) and their target genes. The reference network is derived from ChIP-chip data from MacArthur et al. [44], which includes relationships between 12 TFs and their 2,049 target genes. In other hand, since GO similarities are unavailable for all genes, we limited the study to 1,441 genes with available all GO similarity information. Genes without GO similarity were excluded from both the study and the reference network. The best performances for the Drosophila melanogaster datasets are shown in Table S2. Networks constructed using SimMapNet with GO similarities consistently achieved higher F1-scores than those built by other methods. Each network’s parameters were selected based on the configuration yielding the maximum F1-score, resulting in networks with varying levels of sparsity. To ensure a fair comparison, we searched each method’s results to find the performance corresponding to networks with a similar number of edges to the reference network (about 4,967 edges). Table 3 presents these results. Even under matched conditions, SimMapNet using GO similarities clearly outperforms the competing methods. Table 3Performance metrics of different methods forDrosophila melanogaster data constructed networksMethodTPTNFPFNTPRFPRPPVAccuracyF1_scoreSimMapNet_MF19929290297329590.400.240.400.660.40SimMapNet_BP15328824343934190.310.280.310.600.31SimMapNet_CC20769366289728750.420.240.420.660.42SimMapNet_dist14368732353135150.290.290.290.590.29GENIE314318798346535200.290.280.290.590.29KBOOST14608717354634910.290.290.290.590.29GLASSO821207418948690.020.010.300.710.03Performance of SimMapNet on Drosophila melanogaster data using different GO similarities (SimMapNet_MF, SimMapNet_BP, SimMapNet_CC) and Euclidean distances (SimMapNet_dist), alongside GENIE3, GLASSO, and KBOOST. TP, TN, FP, and FN represent True Positives, True Negatives, False Positives, and False Negatives Performance metrics of different methods forDrosophila melanogaster data constructed networks Performance of SimMapNet on Drosophila melanogaster data using different GO similarities (SimMapNet_MF, SimMapNet_BP, SimMapNet_CC) and Euclidean distances (SimMapNet_dist), alongside GENIE3, GLASSO, and KBOOST. TP, TN, FP, and FN represent True Positives, True Negatives, False Positives, and False Negatives To investigate whether combining networks enhances performance, we integrated the networks constructed using different GO similarity measures by identifying common edges among them (Fig. 5, see also Table S3). The bar plot illustrates performance across various GO combinations. The MF&CC network achieved the highest TPR, while MF&BP showed superior Precision, and BP&CC offered a balanced trade-off between the two. Combining all three GO types (MF&BP&CC) resulted in a lower FPR and improvements in both Precision and Accuracy, but with a reduction in TPR. This indicates a trade-off: enhancing precision may come at the expense of detecting fewer true positive edges. Performance comparison of gene regulatory networks constructed using SimMapNet with different GO similarity measures in Drosophila melanogaster. The bar plot presents True Positive Rate (TPR), False Positive Rate (FPR), Precision, Accuracy, and F-score for networks based on pairwise similarities (MF&BP, MF&CC, BP&CC) and the combined network (MF&BP&CC). The MF&BP&CC network, which integrates all three GO similarities, achieves the highest accuracy but a lower TPR compared to pairwise networks Robustness analysis of hyperparameters and prior specificity To evaluate the stability and structural dependency of SimMapNet, we conducted a multi-tiered sensitivity analysis across the model’s key hyperparameters (α , , ν , and the sparsification threshold ) using the SOS1 dataset. The parameter acts as an essential regularization term that guarantees the positive definiteness of the estimated scale matrix Ω '(equation 3). Detailed parameter configurations, evaluation ranges, and comprehensive multi-metric performance plots are available in Section 7 of the Supplementary Material. Our sensitivity analysis demonstrated that the network reconstruction accuracy of SimMapNet is highly stable and largely insensitive to changes in the Bayesian hyperparameter settings (α , , and ν ) within the evaluated ranges. In contrast, the threshold controls network density systematically, capturing the classic trade-off between edge sparsity and biological signal retention. Contribution of GO-based prior information To assess the contribution of GO-based prior information within the SimMapNet framework, we performed a controlled perturbation analysis of the GO similarity matrix. Specifically, we evaluated four prior configurations: (i) the original GO similarity matrix, (ii) an uninformative prior without GO information, (iii) a permuted GO matrix that preserves the overall similarity distribution while disrupting gene-specific annotation mapping, and (iv) a random similarity matrix generated from independent uniform random values. The perturbation analysis was conducted across the two SOS datasets and Drosophila dataset using all three GO ontologies (BP, MF, and CC). The complete results are presented in Supplementary Figures S3 –. Across datasets, biologically meaningful GO information consistently improved network reconstruction performance. In sample-limited settings, such as the Drosophila dataset and SOS1, substantial reductions in performance were observed when GO information was permuted or randomized. Importantly, the permuted GO control preserves the overall distribution of similarity values while removing gene-specific biological correspondence. Therefore, the observed degradation in performance indicates that SimMapNet benefits from biologically meaningful gene-level GO information rather than generic similarity structure alone. Moderate performance was still obtained under non-informative or perturbed priors, reflecting the fact that gene expression data itself provides a substantial baseline signal for network reconstruction. Consequently, the improvement achieved using the original GO prior represents additional predictive information beyond expression-based inference. For the SOS2 dataset, the influence of GO perturbation was less pronounced. This behavior is expected because the large sample size provides sufficient information for network estimation from expression data alone, thereby reducing sensitivity to prior perturbations. Such behavior is consistent with Bayesian inference principles, where the relative influence of prior information decreases as the amount of observed data increases. Overall, these results demonstrate that SimMapNet effectively exploits biologically meaningful GO information when it is informative, while remaining robust when such information is unavailable or degraded. The perturbation analyses also mitigate concerns regarding potential circularity between GO annotations and the reference networks, since disruption of biologically meaningful GO structure consistently reduces reconstruction performance. Discussion This study introduces SimMapNet, an undirected network reconstruction approach that leverages GO similarity to improve gene relationship inference. Three datasets were used to evaluate SimMapNet: two small SOS networks [21, 23] and a large-scale Drosophila melanogaster dataset [24]. We found that GO similarities greatly increase network inference accuracy. Further, the results of bootstrap sampling confirm that SimMapNet balances both high accuracy and stability, making it a more reliable approach for network inference. Specifically, an expanded 1,000-replicate paired bootstrap analysis demonstrates that SimMapNet’s performance advantages over purely data-driven baselines are highly reproducible, yielding empirical 95% confidence intervals that strictly exclude zero against GLASSO, GENIE3, and KBOOST. While the optimal prior configuration shifts depending on sample scale–with the Molecular Function (MF) and Biological Process (BP) variants exhibiting statistically comparable performance tiers as sample sizes expand–the framework consistently remains robust against sampling artifacts (see Supplementary Section 7). The inferred networks were validated against reference networks and benchmarked against GENIE3, KBOOST, and GLASSO. SimMapNet typically outperformed various algorithms, with improvements varying by similarity metric. GO-based similarities were more accurate than Euclidean distances, showing the value of biological context over statistics. Our key assumption is that genes sharing functional similarity are likely to participate in related biological processes and be regulated by similar mechanisms. This often results in potential interactions within a gene regulatory network, where one gene can directly influence the expression of another, enabling coordinated control over a common function [16]. Incorporating such similarities allows the algorithm to better capture the underlying biological structure and mitigate noise in gene expression data, particularly in cases with limited sample sizes. We also incorporate CC similarities, as they may provide complementary insights into gene interactions, despite their implications differing from those of functional similarity. Our results suggest that functional similarity, particularly utilizing MF and BP, generally improves network inference performance. In the construction of SOS GRNs using both SOS1 and SOS2 datasets, networks built with BP and MF similarities outperform those based on CC similarity. In the Drosophila melanogaster dataset, the differences are less pronounced, although the constructed network utilizing MF and BP still achieves slightly higher AUC scores than CC (Fig. 2 and Figure S1). To ensure a fair comparison, we selected parameter sets that generated a comparable number of network edges, approximately matching the reference network’s edge count (Table 3). Under these conditions, SimMapNet using MF remains the strongest predictor, while CC outperforms BP. These findings suggest that while functional similarity is generally more informative for GRN reconstruction, CC similarity may still provide valuable insights depending on the dataset and network structure. Further investigation is needed to fully understand the implications of this result. Assessment of SimMapNet’s performance across varying sample sizes for the SOS dataset (Table S1) reveals distinct performance patterns across GO categories. SimMapNet using MF similarities demonstrates the most stable performance, particularly excelling in small sample sizes where functional constraints are more informative. The GRN constructed by SimMapNet applying BP performs comparably but shows slight variations due to the complexity of process-level interactions. SimMapNet using CC exhibits robust performance but is more sensitive to increasing sample sizes, likely due to structural dependencies captured in cellular localization data. Despite these variations, all GO-constrained versions of SimMapNet outperform traditional methods across the tested sample sizes. As the sample sizes increase, we observe that SimMapNet, like other GRN inference methods, produces slightly denser networks, which may introduce additional false positives. This effect is common in network reconstruction due to the increasing number of detectable correlations in large datasets. However, SimMapNet’s Bayesian framework mitigates this issue better than alternative approaches by integrating biological priors. The results indicate that SimMapNet does not require a large sample size to perform well, which makes it particularly effective when data are limited. Assessment of SimMapNet’s performance across varying sample sizes for the SOS dataset (Table S1) reveals distinct performance patterns across GO categories. SimMapNet using MF similarities demonstrates the most stable performance, particularly excelling in small sample sizes where functional constraints are more informative. The GRN constructed by SimMapNet applying BP performs comparably but shows slight variations due to the complexity of process-level interactions. SimMapNet using CC exhibits robust performance but is more sensitive to increasing sample sizes, likely due to structural dependencies captured in cellular localization data. Despite these variations, all GO-constrained versions of SimMapNet outperform traditional methods across the tested sample sizes. As the sample sizes increase, we observe that SimMapNet, like other GRN inference methods, produces slightly denser networks, which may introduce additional false positives. This effect is common in network reconstruction due to the increasing number of detectable correlations in large datasets. However, SimMapNet’s Bayesian framework mitigates this issue better than alternative approaches by integrating biological priors. The results indicate that SimMapNet does not require a large sample size to perform well, which makes it particularly effective when data are limited. This interpretation is also supported by the GO prior perturbation analysis. In sample-limited settings, substantial reductions in performance were observed when biologically meaningful GO similarities were replaced by permuted or random priors, demonstrating that informative GO structure contributes directly to network reconstruction accuracy. In contrast, the effect of prior perturbation was less pronounced in larger datasets, where the abundance of expression data provides stronger information for network estimation. This behavior is consistent with Bayesian inference principles, whereby the influence of prior information is greatest when data are limited and gradually decreases as sample size increases [45]. GENIE3 performs well at smaller sample sizes but fails to show consistent improvement as sample sizes increase, suggesting that its tree-based structure may be more sensitive to data variability. While GLASSO maintains relatively stable, though suboptimal, performance across larger sample sizes, it is worth mentioning that GLASSO often struggles with overfitting or unstable covariance estimates in low-data scenarios [46, 47]. However, SimMapNet outperforms KBOOST not only when applied to SOS data with varying sample sizes but also in constructing gene regulatory networks for Drosophila flies, particularly benefiting from GO-based constraints that improve stability when data is limited. SimMapNet employs a MAP shrinkage estimation framework to infer the precision matrix of gene expression data, which serves as the foundation for constructing a GGM of the underlying undirected gene association network. While precision matrix estimation for multivariate normal data using regularized or Bayesian shrinkage is a well-established technique [15, 48–50], most existing approaches are not specifically tailored for biological network construction and lack the capacity to seamlessly incorporate external continuous prior matrices. To demonstrate this limitation, Section 6 and Table S4 of the Supplementary Material present the benchmarking results of implementing a widely utilized shrinkage method,Ledoit and Wolf [48],on the SOS and Drosophila melanogaster datasets. By contrast, SimMapNet explicitly integrates functional similarity information into the regularized framework. Injecting this structural knowledge not only introduces an adaptive biological penalty mechanism that enhances the interpretability and topological accuracy of the inferred networks, but it mathematically simplifies the optimization objective. As a result, SimMapNet achieves remarkably short computation times, making it a highly efficient tool for undirected gene network inference. By deriving direct closed-form updates, it functions as an accelerated shrinkage estimation framework that completely bypasses the computationally intensive iterative search loops typical of alternative structural learning algorithms. This enables it to reconstruct networks well within a minute on standard computing architectures, provided that gene functional similarity matrices are available (see Supplementary Section 9 for complete empirical runtime benchmarks). Future studies could explore hybrid approaches that integrate full Bayesian sampling strategies with advanced biological or topological constraints to further enhance undirected network reconstruction accuracy. To incorporate GO similarities, we employ a kernel function that transforms similarity values into a structured prior covariance matrix. This kernel-based approach ensures that prior biological knowledge is treated as a probabilistic influence rather than a strict constraint, making the framework more flexible and biologically realistic. We utilize two isotropic kernel functions: the squared exponential (SE) kernel, which allows smooth transitions in similarity contributions and demonstrated robust performance [13, 51], and the Ornstein-Uhlenbeck (OU) kernel, which enforces a stronger locality constraint to reduce false positive edges [33]. The choice of the kernel function is left to the user, who can select the one that yields the best-performing network. In this study, we evaluate both kernels and determine which one provides the most biologically meaningful results. The other parameters of SimMapNet, such as the degree of freedom of the prior distribution (ν ), must be specified by the user. Here, we set ν = 2p since varying it did not strongly affect the results. In this paper, our primary goal is to present SimMapNet to introduce an approach that can benefit from incorporating external biological information. Therefore, we optimize the parameters to achieve the best performance, particularly focusing on the F1-score using the available reference network. We select the level of sparsity that maximizes network accuracy rather than using predefined statistical criteria. However, users can apply alternative sparsity selection methods, such as the Bayesian Information Criterion (BIC) [52] or Extended BIC (EBIC) [49, 53], depending on their specific dataset and application. To address the practical scenario in which a reference network is unavailable, we additionally investigated hyperparameter selection using EBIC. While the EBIC-selected models did not always coincide with those obtained through F1-score optimization, the overall ranking of prior-information sources remained broadly consistent. In particular, GO-informed priors, especially BP and MF, continued to achieve the most favorable balance between model fit and network complexity. These findings suggest that the benefits of incorporating GO similarities are not solely dependent on reference-network-guided parameter tuning. The primary analyses in this study should be interpreted as oracle benchmarking experiments designed to assess the maximum achievable reconstruction performance when a reference network is available. In contrast, the EBIC-based procedure provides a practical and reference-independent alternative that can be applied to real biological datasets where the underlying network structure is unknown. Additionally, we are working on deriving a closed-form solution for certain parameters, particularly leveraging the algebraic properties of the estimations. This effort aims to reduce user dependency on manual parameter selection, thereby improving the usability and robustness of the framework. Conclusion In this study, we presented SimMapNet, a novel Bayesian framework for GRN inference that incorporates GO similarities as biologically meaningful prior knowledge. By embedding functional gene similarities into the prior distribution using kernel functions, SimMapNet enhances both the accuracy and interpretability of reconstructed networks, particularly in scenarios where sample sizes are limited—a common challenge in biological data analysis. Comprehensive evaluations on three benchmark datasets, including two SOS DNA repair networks and a high-dimensional Drosophila melanogaster dataset, demonstrate that SimMapNet consistently outperforms widely used methods such as GENIE3, KBOOST, and GLASSO. Furthermore, simulation results confirm that SimMapNet not only excels in data-scarce conditions but also offers scalability and computational efficiency, making it a robust and versatile tool for both small-scale and large-scale GRN reconstruction tasks. While this study focuses on GRN inference, the underlying framework of SimMapNet is highly adaptable to other biological datasets, such as microbiome networks. In addition, by leveraging different types of domain-specific similarities, SimMapNet has the potential to improve network inference across diverse biological systems. Supplementary Information Supplementary file 1 (pdf 978 KB) Acknowledgements This research was supported by the Iran National Science Foundation (INSF) under project number 4027812. Abbreviations Gene regulatory network Gene ontology Gaussian graphical model Molecular function Biological process Cellular component Graphical Lasso True positive rate False positive rate Precision Accuracy Receiver operating characteristic Area under the ROC curve Area under this curve Author contributions MSH and MS developed the idea. MSH collected the data. MSH and RA led all statistical analyses from data preprocessing to fitting models, as well as summarizing the results by the creation of figures. MSH wrote the initial complete draft of the manuscript. MSH and RA and MS contributed in interpretations, editing, and revision of the manuscript. All authors read and approved the final manuscript. Funding Not applicable. Data availability The first dataset (SOS DNArepair network in Escherichia coli with nine samples) is from the study ”Inferring gene regulatory networks from gene expression data by path consistency algorithm based on conditional mutual information” by Zhang et al. (2012), available at: http://csb.shu. edu.cn/subweb/grn.htm. The second dataset (SOS DNA repair network in Escherichia coli with nine samples with 466 samples) is from the study ”A probabilistic graphical model for system-wide analysis of gene regulatory networks” by Kotiang et al. (2020) available at GitHub repository: https://github. com/kotiang54/PgmGRNs.The Drosophila dataset is part of the data available in ArrayExpress (accession code E-MTAB-404): http://www.ebi.ac.uk/arrayexpress/experiments/E-MTAB-404/. All processed datasets used in this study, as well as the R package SimMapNet, are available at: https: //github.com/maryam-shahdoust/SimMapNet. Declarations Conflict of interest The authors declare no conflict of interest. Ethical approval Not applicable. Consent for publication Not applicable. Footnotes Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. M. Shahdoust and R. Aghdam contributed equally to this work. References Associated Data Supplementary Materials Supplementary file 1 (pdf 978 KB) Data Availability Statement The first dataset (SOS DNArepair network in Escherichia coli with nine samples) is from the study ”Inferring gene regulatory networks from gene expression data by path consistency algorithm based on conditional mutual information” by Zhang et al. (2012), available at: http://csb.shu. edu.cn/subweb/grn.htm. The second dataset (SOS DNA repair network in Escherichia coli with nine samples with 466 samples) is from the study ”A probabilistic graphical model for system-wide analysis of gene regulatory networks” by Kotiang et al. (2020) available at GitHub repository: https://github. com/kotiang54/PgmGRNs.The Drosophila dataset is part of the data available in ArrayExpress (accession code E-MTAB-404): http://www.ebi.ac.uk/arrayexpress/experiments/E-MTAB-404/. All processed datasets used in this study, as well as the R package SimMapNet, are available at: https: //github.com/maryam-shahdoust/SimMapNet.