Protocol to explain graph neural network predictions using an edge-centric Shapley value-based approach.
Ontology highlight
ABSTRACT:
SUBMITTER: Mastropietro A
PROVIDER: S-EPMC9700376 | biostudies-literature | 2022 Dec
REPOSITORIES: biostudies-literature
Publications
Protocol to explain graph neural network predictions using an edge-centric Shapley value-based approach.
Mastropietro Andrea A Pasculli Giuseppe G Bajorath Jürgen J
STAR protocols 20221124 4
Here we present EdgeSHAPer, a workflow for explaining graph neural networks by approximating Shapley values using Monte Carlo sampling. In this protocol, we describe steps to execute Python scripts for a chemical dataset from the original publication; however, this approach is also applicable to any user-provided dataset. We also detail steps encompassing neural network training, an explanation phase, and analysis via feature mapping. For complete details on the use and execution of this protoco ...[more]
Similar Datasets
Project description:Graph neural networks (GNNs) recursively propagate signals along the edges of an input graph, integrate node feature information with graph structure, and learn object representations. Like other deep neural network models, GNNs have notorious black box character. For GNNs, only few approaches are available to rationalize model decisions. We introduce EdgeSHAPer, a generally applicable method for explaining GNN-based models. The approach is devised to assess edge importance for predictions. Therefore, EdgeSHAPer makes use of the Shapley value concept from game theory. For proof-of-concept, EdgeSHAPer is applied to compound activity prediction, a central task in drug discovery. EdgeSHAPer's edge centricity is relevant for molecular graphs where edges represent chemical bonds. Combined with feature mapping, EdgeSHAPer produces intuitive explanations for compound activity predictions. Compared to a popular node-centric and another edge-centric GNN explanation method, EdgeSHAPer reveals higher resolution in differentiating features determining predictions and identifies minimal pertinent positive feature sets.
| S-EPMC9483788 | biostudies-literature
Project description:Transportation networks play a crucial role in society by enabling the smooth movement of people and goods during regular times and acting as arteries for evacuations during catastrophes and natural disasters. Identifying the critical road segments in a large and complex network is essential for planners and emergency managers to enhance the network's efficiency, robustness, and resilience to such stressors. We propose a novel approach to rapidly identify critical and vital network components (road segments in a transportation network) for resilience improvement or post-disaster recovery. We pose the transportation network as a graph with roads as edges and intersections as nodes and deploy a Graph Neural Network (GNN) trained on a broad range of network parameter changes and disruption events to rank the importance of road segments. The trained GNN model can rapidly estimate the criticality rank of individual road segments in the modified network resulting from an interruption. We address two main limitations in the existing literature that can arise in capital planning or during emergencies: ranking a complete network after changes to components and addressing situations in post-disaster recovery sequencing where some critical segments cannot be recovered. Importantly, our approach overcomes the computational overhead associated with the repeated calculation of network performance metrics, which can limit its use in large networks. To highlight scenarios where our method can prove beneficial, we present examples of synthetic graphs and two real-world transportation networks. Through these examples, we show how our method can support planners and emergency managers in undertaking rapid decisions for planning infrastructure hardening measures in large networks or during emergencies, which otherwise would require repeated ranking calculations for the entire network.
| S-EPMC10734987 | biostudies-literature
Project description:Biomarkers are indispensable for precision medicine. However, focused single-biomarker development using human tissue has been complicated by sample spatial heterogeneity. To address this challenge, we tested a representation of primary tumor that synergistically integrated multiple in situ biomarkers of extracellular matrix from multiple sampling regions into an intratumor graph neural network. Surprisingly, the differential prognostic value of this computational model over its conventional non-graph counterpart approximated that of combined routine prognostic biomarkers (tumor size, nodal status, histologic grade, molecular subtype, etc.) for 995 breast cancer patients under a retrospective study. This large prognostic value, originated from implicit but interpretable regional interactions among the graphically integrated in situ biomarkers, would otherwise be lost if they were separately developed into single conventional (spatially homogenized) biomarkers. Our study demonstrates an alternative route to cancer prognosis by taping the regional interactions among existing biomarkers rather than developing novel biomarkers. Cancer prognosis using multiregion sampling is costly and not completely reliable due to the required biomarker homogenisation step. Here, the authors develop an intratumor graph neural network for prognosis in multiregion cancer samples based on in situ biomarkers and gene expression that does not need homogenisation.
| S-EPMC9307796 | biostudies-literature
Project description:Machine learning and deep learning have facilitated various successful studies of molecular property predictions. The rapid development of natural language processing and graph neural network (GNN) further pushed the state-of-the-art prediction performance of molecular property to a new level. A geometric graph could describe a molecular structure with atoms as the nodes and bonds as the edges. Therefore, a graph neural network may be trained to better represent a molecular structure. The existing GNNs assumed homogeneous types of atoms and bonds, which may miss important information between different types of atoms or bonds. This study represented a molecule using a heterogeneous graph neural network (MolHGT), in which there were different types of nodes and different types of edges. A transformer reading function of virtual nodes was proposed to aggregate all the nodes, and a molecule graph may be represented from the hidden states of the virtual nodes. This proof-of-principle study demonstrated that the proposed MolHGT network improved the existing studies of molecular property predictions. The source code and the training/validation/test splitting details are available at https://github.com/zhangruochi/Mol-HGT.
| S-EPMC8811943 | biostudies-literature
Project description:BackgroundLung cancer is a global disease with high lethality, with early screening being considerably helpful for improving the 5-year survival rate. Multimodality features in early screening imaging are an important part of the prediction for lung adenocarcinoma, and establishing a model for adenocarcinoma diagnosis based on multimodal features is an obvious clinical need. Through our practice and investigation, we found that graph neural networks (GNNs) are excellent platforms for multimodal feature fusion, and the data can be completed using the edge-generation network. Therefore, we propose a new lung adenocarcinoma multiclassification model based on multimodal features and an edge-generation network.MethodsAccording to a ratio of 80% to 20%, respectively, the dataset of 338 cases was divided into the training set and the test set through 5-fold cross-validation, and the distribution of the 2 sets was the same. First, the regions of interest (ROIs) cropped from computed tomography (CT) images were separately fed into convolutional neural networks (CNNs) and radiomics processing platforms. The results of the 2 parts were then input into a graph embedding representation network to obtain the fused feature vectors. Subsequently, a graph database based on the clinical and semantic features was established, and the data were supplemented by an edge-generation network, with the fused feature vectors being used as the input of the nodes. This enabled us to clearly understand where the information transmission of the GNN takes place and improves the interpretability of the model. Finally, the nodes were classified using GNNs.ResultsOn our dataset, the proposed method presented in this paper achieved superior results compared to traditional methods and showed some comparability with state-of-the-art methods for lung nodule classification. The results of our method are as follows: accuracy (ACC) =66.26% (±4.46%), area under the curve (AUC) =75.86% (±1.79%), F1-score =64.00% (±3.65%), and Matthews correlation coefficient (MCC) =48.40% (±5.07%). The model with the edge-generating network consistently outperformed the model without it in all aspects.ConclusionsThe experiments demonstrate that with appropriate data=construction methods GNNs can outperform traditional image processing methods in the field of CT-based medical image classification. Additionally, our model has higher interpretability, as it employs subjective clinical and semantic features as the data construction approach. This will help doctors better leverage human-computer interactions.
| S-EPMC10423350 | biostudies-literature
Project description:Prenatal opioid exposure has been linked to adverse effects spanning multiple neurodevelopmental domains, including cognition, motor development, attention, and vision. However, the neural basis of these abnormalities is largely unknown. A total of 49 infants, including 21 opioid-exposed and 28 controls, were enrolled and underwent MRI (43 ± 6 days old) after birth, including resting state functional MRI. Edge-centric functional networks based on dynamic functional connections were constructed, and machine-learning methods were employed to identify neural features distinguishing opioid-exposed infants from unexposed controls. An accuracy of 73.6% (sensitivity 76.25% and specificity 69.33%) was achieved using 10 times 10-fold cross-validation, which substantially outperformed those obtained using conventional static functional connections (accuracy 56.9%). More importantly, we identified that prenatal opioid exposure preferentially affects inter- rather than intra-network dynamic functional connections, particularly with the visual, subcortical, and default mode networks. Consistent results at the brain regional and connection levels were also observed, where the brain regions and connections associated with visual and higher order cognitive functions played pivotal roles in distinguishing opioid-exposed infants from controls. Our findings support the clinical phenotype of infants exposed to opioids in utero and may potentially explain the higher rates of visual and emotional problems observed in this population. Finally, our findings suggested that edge-centric networks could better capture the neural differences between opioid-exposed infants and controls by abstracting the intrinsic co-fluctuation along edges, which may provide a promising tool for future studies focusing on investigating the effects of prenatal opioid exposure on neurodevelopment.
| S-EPMC9117006 | biostudies-literature
Project description:The human connectome is modular with distinct brain regions clustering together to form large-scale communities, or networks. This concept has recently been leveraged in novel inferencing procedures by averaging the edge-level statistics within networks to induce more powerful inferencing at the network level. However, these networks are constructed based on the similarity between pairs of nodes. Emerging work has described novel edge-centric networks, which instead use the similarity between pairs of edges to construct networks. In this work, we use these edge-centric networks in a network-level inferencing procedure and compare this novel method to traditional inferential procedures and the network-level procedure using node-centric networks. We use data from the Human Connectome Project, the Healthy Brain Network, and the Philadelphia Neurodevelopmental Cohort and use a resampling technique with various sample sizes (n=40, 80, 120) to probe the power and specificity of each method. Across datasets and sample sizes, using the edge-centric networks outperforms using node-centric networks for inference as well as edge-level FDR correction and NBS. Additionally, the edge-centric networks were found to be more consistent in clustering effect sizes of similar values as compared to node-centric networks, although node-centric networks often had a lower average within-network effect size variability. Together, these findings suggest that using edge-centric networks for network-level inference can procure relatively powerful results while remaining similarly accurate to the underlying edge-level effects across networks, complementing previous inferential methods.
| S-EPMC9838718 | biostudies-literature
Project description:Accurately predicting activation energies is crucial for understanding chemical reactions and modeling complex reaction systems. However, the high computational cost of quantum chemistry methods often limits the feasibility of large-scale studies, leading to a scarcity of high-quality activation energy data. In this work, we explore and compare three innovative approaches (transfer learning, delta learning, and feature engineering) to enhance the accuracy of activation energy predictions using graph neural networks, specifically focusing on methods that incorporate low-cost, low-level computational data. Using the Chemprop model, we systematically evaluated how these methods leverage data from semiempirical quantum mechanics (SQM) calculations to improve predictions. Delta learning, which adjusts low-level SQM activation energies to align with high-level CCSD(T)-F12a targets, emerged as the most effective method, achieving high accuracy with substantially reduced data requirements. Notably, delta learning trained with just 20-30% of high-level data matched or exceeded the performance of other methods trained with full data sets, making it advantageous in data-scarce scenarios. However, its reliance on transition state searches imposes significant computational demands during model application. Transfer learning, which pretrains models on large data sets of low-level data, provided mixed results, particularly when there was a mismatch in the reaction distributions between the training and target data sets. Feature engineering, which involves adding computed molecular properties as input features, showed modest gains, particularly in thermodynamic properties. Our study highlights the trade-offs between accuracy and computational demand in selecting the best approach for enhancing activation energy predictions. These insights provide valuable guidelines for researchers aiming to apply machine learning in chemical reaction engineering, helping to balance accuracy with resource constraints.
| S-EPMC11815826 | biostudies-literature
Project description:We developed a method to apply artificial neural networks (ANNs) for predicting time-series pharmacokinetics (PKs), and an interpretable the ANN-PK model, which can explain the evidence of prediction by applying Shapley additive explanations (SHAP). A previous population PK (PopPK) model of cyclosporin A was used as the comparison model. The patients' data were used for the ANN-PK model input, and the output by ANN was the clearance (CL). The estimated CL value from the ANN were substituted into the one-compartment with one-order absorption model, the concentrations were calculated, and the parameters of ANN were updated by the back-propagation method. Kernel SHAP was applied to the trained model and the SHAP value of each input was calculated. The root mean squared error for the PopPK model and the ANN-PK model were 41.1 and 31.0 ng/ml, respectively. The goodness of fit plots for the ANN-PK model represented more convergence to y = x compared with that for the PopPK model, with good model performance for the ANN-PK model. The most influential factors on CL output were age and body weight from the evaluation using Kernel SHAP, and these factors were incorporated into the PopPK model as the significant covariates of CL. The ANN-PK model could handle time-series data and showed higher prediction accuracy then the conventional PopPK model, and the scientific validity for the model could be evaluated by applying SHAP. Study Highlights WHAT IS THE CURRENT KNOWLEDGE ON THE TOPIC? A black-box property of an artificial neural network (ANN) decreases the scientific confidence of the model, and making it difficult to utilize the ANN in the medical field. Moreover, difficulty in handling the time-series data is a significant problem for applying the ANN for pharmacometrics study. WHAT QUESTION DID THIS STUDY ADDRESS? How can we apply the ANN for predicting the time-series pharmacokinetics (PKs) , and confirm the scientific validity of the ANN model? WHAT DOES THIS STUDY ADD TO OUR KNOWLEDGE? Using the ANN in combination with a conventional compartment (ANN-PK) model enabled to handle the time-series PK data, and the predicting performance of the model was higher than that of the population PK model. Furthermore, we could evaluate the scientific validity of the ANN model by applying the Shapley additive explanations. HOW MIGHT THIS CHANGE DRUG DISCOVERY, DEVELOPMENT, AND/OR THERAPEUTICS? We expect that our study will contribute to develop the interpretable ANN model, which can predict the time-series PKs, drug efficacies, and side effects with high prediction performance.
| S-EPMC8302242 | biostudies-literature
Project description:Optimization of metal-ligand asymmetric catalysts is usually done by empirical trials, where the ligand is arbitrarily modified, and the new catalyst is re-evaluated in the lab. This procedure is not efficient and alternative strategies are highly desirable. We propose the Homogeneous Catalyst Graph Neural Network (HCat-GNet), a machine learning model capable of aiding ligand optimization. This method trains models to predict the enantioselectivity of asymmetric reactions using only the SMILES representations of the participant molecules. HCat-GNet allows high interpretability indicating from which atoms the model gathers the most predictive information, thus showing which atoms within the ligand most affect the increase or decrease in the reaction's selectivity. The validation of the model's selectivity predictions is made using a new class of ligand for rhodium-catalyzed asymmetric 1,4-addition, demonstrating the ability of HCat-GNet to extrapolate into unknown chiral ligand space. Validation with other benchmark asymmetric reaction datasets demonstrates its generality when modeling different reactions.
| S-EPMC11889620 | biostudies-literature