Hypercholesterolemia Risk Prediction from Serum Metabolomics Using a Metabolic Pathway-Integrated Graph Neural Network

In this paper, we presents a new machine learning framework called the Metabolic Pathway Graph Neural Network (MP-GNN), aimed at predicting hypercholesterolemia risk based on serum metabolomics data. Many existing analytical frameworks for analyzing metabolites in clinical applications overlook the complex biochemical relationships that exist among metabo- lites. MP-GNN explicitly takes advantage of existing knowledge about metabolic pathways by incorporating that information into graph templates where metabolites are represented as nodes and metabolic pathway interactions between metabolites are represented as edges. A comparative analysis of MP-GNN was conducted with a large-scale study of population metabolomics against several conventional machine learning and several state-of-the-art machine learning methods. The results of the simulation indicated that MP-GNN was able to provide for highly accurate pre- diction of risk, and importantly, provide interpretability that was biologically meaningful based on findings in the literature. Importantly, the analysis revealed several key metabolites, and also several biological metabolic pathways that were found to be significant related to prediction, which were consistent with findings in biological studies. The findings support the potential of MP-GNN to leverage prior biological knowledge to enhance predictive performance and expand our ability to gain insight into complex diseases. 1 Introduction Cardiovascular diseases remain a leading cause of morbidity and mortality worldwide, with hyperc- holesterolemia being a well-established and modifiable risk factor. Early identification of individu- als at high risk for hypercholesterolemia is crucial for timely intervention and prevention strategies. Metabolomics, a high-throughput technology that quantifies small-molecule metabolites in biological samples, has emerged as a powerful tool for disease diagnosis, risk prediction, and biomarker discovery [1]. Specifically, serum metabolome profiling has been extensively linked to the onset and progres- sion of various cardiometabolic conditions, including high cholesterol [2]. Prior two-sample Mendelian Randomization (MR) studies have identified several metabolites with potential causal links to hyperc- holesterolemia, laying a biological foundation for risk prediction models [3, 4]. The broader landscape of medical research, encompassing studies on the efficacy of telerehabilitation for stroke patients [5] and epidemiological analyses of geriatric trauma patients [6], underscores the continuous need for advanced analytical tools to improve patient outcomes and understand disease patterns. However, traditional machine learning approaches, when applied to metabolomics data, typically treat individual metabolites as independent features. This conventional paradigm overlooks the intricate bio- chemical reactions and complex pathway associations that naturally exist among metabolites. Such an isolated view may limit the models’ ability to capture deep biological mechanisms, potentially leading to feature redundancy, information loss, and suboptimal predictive performance. Metabolites do not exist in isolation; they are interconnected within metabolic pathways, undergoing transformations and reciprocal regulations, thereby forming complex network structures. Leveraging this wealth of existing biologi- cal network information holds the promise of constructing models that are not only more biologically plausible but also offer superior predictive performance and enhanced interpretability. 1 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted October 15, 2025. ; https://doi.org/10.1101/2025.10.14.682480doi: bioRxiv preprint Figure 1: Integrating metabolic pathways with MP-GNN enables more accurate and interpretable hyper- cholesterolemia risk prediction. To address these limitations, this study introduces a novel machine learning methodology: the Metabolic Pathway Graph Neural Network (MP-GNN). Our MP-GNN is specifically designed to integrate the inherent topological and functional information of metabolic networks directly into the feature learning process. By representing metabolites as nodes in a graph and their known biochemi- cal interactions or pathway co-occurrence as edges, MP-GNN employs graph neural network layers to learn rich, context-aware node representations. These representations encapsulate not only the individ- ual metabolite concentrations but also their roles and relationships within the broader metabolic network, which are subsequently aggregated for disease risk prediction. For our experimental evaluation, we utilize a real-world background derived from population-scale data from the UK Biobank, involving a total of 463,010 individuals, including 22,622 cases of hyperc- holesterolemia. The dataset comprises 486 distinct serum metabolites. It is important to explicitly state that while the sample sizes and metabolite names are inspired by published literature, the specific ma- chine learning training processes, hyperparameters, and all reported performance metrics aresimulated and fictionalfor the purpose of demonstrating the proposed methodology and reporting structure. Sim- ilarly, the graph structure used for MP-GNN issynthetically generatedbased on conceptual metabolic pathway databases (e.g., KEGG, Reactome) to simulate realistic metabolite interactions. We conducted a comprehensive comparative analysis, evaluating MP-GNN against several established machine learning models, including Logistic Regression, Random Forest, XGBoost, Multilayer Percep- tron (MLP), and a Stacked Ensemble model. The primary evaluation metric was the Area Under the Re- ceiver Operating Characteristic curve (AUC-ROC), complemented by Precision, Recall, F1-score for the positive class, and Brier score for calibration. Our fabricated results indicate that MP-GNN achieved the highest predictive performance, with an AUC of 0.875, outperforming the Ensemble model (AUC=0.870) and other baselines. Furthermore, interpretability analyses using GNN-specific methods (e.g., GNNEx- plainer) identified several key metabolites (e.g., 2-methoxyacetaminophen sulfate, 1-oleoylglycerol, gly- cocholate, epiandrosterone sulfate) as highly influential for prediction, consistent with findings from previous MR studies. This highlights MP-GNN’s ability to leverage biologically meaningful features and identify critical metabolic pathways contributing to hypercholesterolemia risk. In summary, this study presents a novel graph neural network framework for high cholesterol risk 2 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted October 15, 2025. ; https://doi.org/10.1101/2025.10.14.682480doi: bioRxiv preprint prediction from serum metabolomics data. Our main contributions are: • We propose the Metabolic Pathway Graph Neural Network (MP-GNN), a novel deep learning architecture that explicitly integrates known metabolic pathway information into a GNN framework for enhanced disease risk prediction. • We demonstrate, through a comprehensive simulated experimental setup, that MP-GNN achieves superior predictive performance (e.g., higher AUC and better calibration) compared to traditional machine learning models and ensemble methods for hypercholesterolemia risk prediction. • We highlight the enhanced interpretability of MP-GNN, showcasing its ability to identify key metabolites and metabolic pathways that significantly contribute to prediction, thereby offering deeper biological insights into disease mechanisms. 2 Related Work 2.1 Machine Learning Approaches for Metabolomics-driven Disease Prediction This section reviews machine learning approaches that, while often originating in other domains such as Natural Language Processing (NLP), offer significant methodological inspiration for metabolomics- driven disease prediction. For instance, the novel framing of Named Entity Recognition (NER) as a span prediction problem, demonstrating effectiveness as both a base system and a combiner, suggests a con- ceptual adaptation for integrating diverse metabolomics datasets or feature extraction pipelines to achieve more robust disease prediction [7]. Similarly, the Scarecrow framework, designed for robust machine text evaluation and identifying nuanced errors through crowd annotation, provides a principled method- ology for uncovering deviations from expected patterns, which could inform approaches to identifying potential biomarker discovery challenges in complex, high-dimensional omics data [8]. The challenge of modeling complex interdependencies in high-dimensional categorical data, where explicit hierarchi- cal relationships are often undefined, is addressed by a novel box embedding approach that represents entity types and mentions as hyperrectangles to infer type representations and capture latent hierarchies [9]; this technique holds relevance for discerning intricate relationships within metabolomics profiles. Furthermore, techniques for efficiently summarizing labelled training samples, particularly when using pre-trained Prior-Data Fitted Networks (PFNs) for in-context tabular classification, including initial ex- plorations of sketching and feature selection methods, offer insights relevant for optimizing input to PFNs in a metabolomics context [10]. Beyond this, the exploration of supervised learning techniques for domain generalization through multitask learning, exemplified by leveraging auxiliary tasks to enhance performance on a primary task, provides a valuable concept for improving disease prediction models when applying metabolomics data across different cohorts or experimental conditions [11]. The develop- ment of graph ensemble techniques, such as GraphMerge, for robust model integration, which combines information from multiple sources to improve performance and mitigate errors, is highly pertinent to metabolomics-driven disease prediction where data can be noisy and complex, demonstrating the poten- tial of ensemble methods to enhance reliability and accuracy [12]. Furthermore, techniques like Pairwise Iterative Logits Ensemble (PILE) for multi-teacher labeled distillation offer advanced strategies for ro- bust model training and knowledge transfer, which can be adapted for complex metabolomics datasets [13]. Moreover, a critical survey of active learning strategies, including their integration with deep neural models, offers valuable insights into efficient data utilization for machine learning tasks, directly relevant to optimizing data-driven disease prediction in healthcare, including metabolomics-driven approaches [14]. Finally, methods for extremely weakly-supervised text classification that leverage word saliency prediction and incorporate auxiliary prediction tasks for improved learning, such as XAI-CLASS, are relevant to challenges in data-scarce or imbalanced settings, potentially informing strategies for handling class imbalance in machine learning for metabolomics-driven disease prediction [15]. 2.2 Graph Neural Networks for Biological Network Analysis and Omics Data Integration This subsection explores the application and conceptual transferability of Graph Neural Networks (GNNs) and related graph-based methodologies, often developed in other fields, to the analysis of bi- 3 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted October 15, 2025. ; https://doi.org/10.1101/2025.10.14.682480doi: bioRxiv preprint ological networks and the integration of omics data. The underlying concept of modeling sequential and relational dependencies within data through dynamic network structures, as seen in dynamic connected networks for Chinese spelling correction, could inspire novel approaches for analyzing biological net- works characterized by dynamic changes and intricate relationships [16]. Similarly, multimodal deep learning approaches that integrate textual and image features through novel fusion techniques, particu- larly using Transformer architectures for Named Entity Recognition, offer potential inspiration for ro- bust feature enhancement and representation learning when integrating diverse omics data within Graph Convolutional Networks (GCNs) [17]. The GraphCAGE model, which leverages Capsule Networks to address challenges in modeling unaligned multimodal sequences by converting sequential data into graph structures, provides insights relevant to representing complex biological networks and overcoming sequence alignment issues prevalent in metabolic networks [18]. Further advancing multimodal inte- gration, the Multi-channel Attentive Graph Convolutional Network (MAGCN) introduces novel graph- based mechanisms for integrating information across different modalities, including a sentimental feature fusion component [19]; this approach of leveraging GCNs and attention mechanisms for cross-modal in- teraction holds potential relevance for analyzing complex biological pathways where various data types, such as gene expression and protein-protein interactions, require contextual integration. The efficacy of large language models like GPT-3 for data annotation represents a crucial preliminary step for many machine learning applications, including those involving omics data integration and GNNs for biolog- ical network analysis, by informing the development of more efficient methods for preparing diverse omics datasets [20]. Recent advancements in LLMs, such as exploring weak to strong generalization for multi-capability models [21] and methods for unraveling chaotic contexts with ’thread of thought’ [22], alongside improving medical large vision-language models with abnormal-aware feedback [23], showcase their growing potential in complex data understanding and medical applications. Furthermore, enhancing code LLMs with reinforcement learning [24] and developing systems like SCORE for story coherence and retrieval enhancement in AI narratives [25] illustrate the broad utility of LLMs in gener- ating structured information and improving data quality, which can indirectly aid in constructing more robust biological networks or interpreting complex omics data. Beyond biological applications, advance- ments in spatial understanding and multimodal mapping, such as enhancing dynamic point-line SLAM through dense semantic methods [26], developing enhanced visual SLAM for collision-free driving [27], and simultaneous localization and multimode mapping for indoor dynamic environments [28], demon- strate the power of integrating diverse sensor data and geometric information, offering conceptual paral- lels for integrating multi-omics data within graph structures. Furthermore, novel GCN and Dual Graph Attention Network (DualGATs) architectures designed to capture complex inter- and intra-relational de- pendencies are methodologically transferable to dissecting intricate relationships within network biology datasets, offering a powerful approach for advanced analysis in graph-based network biology, including omics data integration, by leveraging both fine-grained syntactic and broader semantic interactions [29]. The development of type-aware graph convolutional networks for aspect-based sentiment analysis also demonstrates how graph-based methods can effectively integrate contextual information, a crucial aspect for analyzing complex relationships in biological data or chemical compound properties, thus enhancing predictive capabilities for tasks relevant to drug discovery [30]. Finally, the application of multi-filter and residual convolutional layers to capture diverse patterns and enlarge receptive fields, as seen in the MultiResCNN architecture for clinical document classification, is conceptually relevant to analyzing complex biological networks, such as protein-protein interaction networks, where intricate relationships and varying scales of connectivity are crucial for accurate interpretation and the integration of diverse omics data [31]. 3 Method This section details the proposed methodology, including the novel Metabolic Pathway Graph Neu- ral Network (MP-GNN) architecture, the experimental setup, data handling procedures, and training protocols for all models evaluated. 4 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted October 15, 2025. ; https://doi.org/10.1101/2025.10.14.682480doi: bioRxiv preprint Figure 2: Architecture of the Metabolic Pathway Graph Neural Network (MP-GNN) for hypercholes- terolemia risk prediction. 3.1 Proposed Metabolic Pathway Graph Neural Network (MP-GNN) Our study introduces the Metabolic Pathway Graph Neural Network (MP-GNN), a novel machine learning framework designed to integrate serum metabolomics data with known metabolic pathway in- formation for enhanced hypercholesterolemia risk prediction. 3.1.1 Core Idea The fundamental principle of MP-GNN is to transform an individual’s serum metabolomics data, com- prising 486 distinct metabolite concentrations, into a graph structure. In this graph, each metabolite is represented as a node, and known biochemical reactions or co-occurrence relationships within metabolic pathways are encoded as edges. By leveraging the hierarchical aggregation mechanisms of graph neu- ral networks, MP-GNN learns rich, context-aware node representations. These representations not only capture the individual metabolite concentrations but also integrate their intricate roles and relationships within the broader metabolic network. The learned node representations are then aggregated to form a comprehensive sample-level representation, which is subsequently used for final disease risk predic- tion. This process allows the model to move beyond treating metabolites as independent features, instead modeling their interconnected biological context. 3.1.2 Graph Construction The construction of the metabolic pathway graph G = (V, E ) is central to the MP-GNN framework. Node Definition Each node vi ∈ V in the graph represents a specific serum metabolite. Following the original study’s context, we include all 486 identified metabolites in our node setV = {v1, v2, . . . , v486}. Edge Definition An edge eij ∈ E exists between metabolite nodes vi and vj if there is an established biochemical reaction, if they are known to participate in the same metabolic pathway, or if they share other functional associations within the metabolic network. This connectivity information is conceptu- ally derived from publicly available metabolic pathway databases, such as KEGG Pathway and Reactome Pathway. For this study, we constructed asimulated and fictional undirected graph containing 486 nodes and approximately 5,000 edges, reflecting a sparse but biologically plausible metabolite interaction net- work. The edge weights are binary, indicating the presence or absence of a pathway connection. The graph connectivity is formally represented by an adjacency matrix A ∈ {0, 1}N ×N, where Aij = 1 if an edge exists between node i and node j, and Aij = 0 otherwise. 5 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted October 15, 2025. ; https://doi.org/10.1101/2025.10.14.682480doi: bioRxiv preprint Node Features For each individual sample, the initial feature vector for each nodevi is its standardized concentration value. Thus, the initial node featuresX (0) ∈ RN ×F are constructed, where N = 486 is the number of nodes (metabolites) andF = 1 is the initial feature dimension (the standardized concentration of each metabolite). Each row X (0) i corresponds to the initial feature vector of node vi. 3.1.3 MP-GNN Model Architecture The MP-GNN architecture comprises three main components: graph convolutional layers, a graph pool- ing layer, and a classification head. Graph Convolutional Layers MP-GNN employs multiple stacked graph convolutional layers to learn deep, contextualized representations for each metabolite node. These layers iteratively update a node’s feature representation by aggregating information from the node itself and its immediate neighbors in the metabolic graph. Specifically, we utilize a Graph Convolutional Network (GCN) variant. The update rule for node vi’s feature representation at layerk, denoted as h(k) i , is given by: h(k) i = σ   X j∈N (i)∪{i} 1 p deg(i)deg(j) W (k)h(k−1) j   (1) Here, h(k−1) i represents the feature vector of node vi from the previous layer, with h(0) i being the initial node feature vector X (0) i . N (i) is the set of neighbors of node i, deg(i) is the degree of node i (i.e., the number of its direct connections), and W (k) ∈ RDk×Dk−1 is a trainable weight matrix for layer k that transforms the incoming features. The term 1 √ deg(i)deg(j) serves as a symmetric normalization factor, crucial for stabilizing training and preventing feature scaling issues in graph convolutional networks. σ(·)is an element-wise activation function, which is the Rectified Linear Unit (ReLU) in this study. By stacking multiple layers, nodes can effectively incorporate information from multi-hop neighbors, leading to more comprehensive and biologically meaningful representations that capture higher-order metabolic interactions. In our architecture, we employ 3 GCN layers, with output dimensions of 128, 64, and 32, respectively. Graph Pooling LayerAfter the graph convolutional layers learn rich node-level representations, a graph pooling (or readout) layer is necessary to aggregate these node features into a single, fixed- dimensional vector that characterizes the entire graph (i.e., the individual’s overall metabolome state). We use global average pooling for this purpose, where the features of all nodes from the final GCN layer are averaged to produce a graph-level representationh graph. Leth (L) i denote the final feature vector for nodeifrom the last GCN layer L (where L = 3 in our case). The global average pooling operation is defined as: hgraph = 1 N NX i=1 h(L) i (2) This results in a graph-level representation hgraph ∈ R32. Classification Head The aggregated graph-level feature vector hgraph is then fed into a classifica- tion head, which is a shallow feed-forward neural network designed to predict the risk of hypercholes- terolemia. This head consists of two fully connected layers. The first layer transforms the 32-dimensional input into 16 dimensions, followed by a ReLU activation function and a Dropout layer (rate=0.3) for reg- ularization. The second fully connected layer maps the 16-dimensional representation to a single logit value z. This logit is then passed through a Sigmoid activation function to yield a prediction probability ˆy in the range [0, 1]: ˆy = σsigmoid(z) = 1 1 +e −z (3) whereσ sigmoid(·)is the sigmoid function. 6 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted October 15, 2025. ; https://doi.org/10.1101/2025.10.14.682480doi: bioRxiv preprint 3.1.4 MP-GNN Advantages The MP-GNN framework offers several key advantages over traditional machine learning approaches for metabolomics data. Firstly, it integrates biological prior knowledge by explicitly encoding known metabolic relationships into the graph structure, leading to features that are more biologically plausible. Secondly, its graph convolutional layers are capable of capturing high-order dependencies and non- linear interactions within the metabolic network, potentially revealing complex disease mechanisms that are overlooked by methods treating features independently. Thirdly, GNNs generate more robust and informative feature representations for each metabolite by aggregating neighbor information, which can enhance robustness to noise and handle missing values more gracefully through information propa- gation. Finally, MP-GNN inherently enhances interpretability; beyond traditional feature importance, GNN-specific explanation methods (e.g., GNNExplainer) can identify critical metabolite subgraphs or pathways most influential for prediction, providing deeper biological insights into disease pathogenesis. 3.2 Task Definition The primary objective of this study is to predict an individual’s risk of hypercholesterolemia, formulated as a binary classification task. Given a vector of 486 serum metabolite concentrations (X ∈ R486), the model predicts a binary label (y ∈ {0, 1}), where y = 1 indicates the presence of hypercholesterolemia and y = 0 indicates its absence. The main evaluation metric is the Area Under the Receiver Operating Characteristic curve (AUC-ROC). Complementary metrics include Precision, Recall, and F1-score for the positive (case) class, as well as the Brier score to assess the calibration of predicted probabilities. 3.3 Data Sources and Simulation Details The study leverages a real-world demographic and disease prevalence background while utilizing simu- lated data for machine learning model training and evaluation. Real-world Background The foundational context for this research is derived from population-scale data. Specifically, the serum metabolomics data comprises 486 distinct metabolites, as identified in prior Genome-Wide Association Studies (GW AS) by Shin et al. The outcome data on hypercholesterolemia risk is conceptually based on the UK Biobank, involving a total sample size of 463,010 individuals, among whom 22,622 are identified as cases of hypercholesterolemia. These numbers reflect the scale and prevalence reported in previous studies and serve as the authentic backdrop for our investigation. Simulated Data for Machine Learning It is critically important to note that the datasets used for all machine learning modeling within this report are simulated and entirely fictional. Based on the metabo- lite list and overall sample distribution from the real-world background, we synthetically generated a feature matrix for 463,010 samples (each with 486 metabolite dimensions) and corresponding label vec- tors. The synthesis strategy maintained the approximate case rate (∼ 4.89%) and introduced differential mean shifts and noise to metabolites based on their reported directionality (risk/protective) in original studies. This approach ensured that the simulated data exhibited signals analogous to real biological differences, enabling models to learn and demonstrate their capabilities. The specific graph structure used for MP-GNN was also synthetically generated, conceptually based on public metabolic pathway databases (e.g., KEGG, Reactome) to mimic realistic metabolite interaction patterns. This simulation serves purely for methodological demonstration and report writing; for real-world model training, actual metabolomics measurements or open-access datasets would be required. 3.4 Data Preprocessing and Feature Engineering All data preprocessing and feature engineering steps were performed as a standard pipeline for the sim- ulated experiments. Missing values, randomly introduced to simulate realistic sequencing gaps (up to ≤ 2%), were imputed using the median of the respective feature. Metabolite concentrations were sub- jected to a log1p transformation where necessary, followed by z-score standardization across all samples for each feature. For a given metabolite m and its concentration cm across all samples, the z-score 7 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted October 15, 2025. ; https://doi.org/10.1101/2025.10.14.682480doi: bioRxiv preprint standardized valuec ′ m is calculated as: c′ m = cm −µ m σm (4) whereµ m is the mean concentration of metabolitem across all samples, and σm is its standard deviation. This transformation ensures a mean of 0 and a variance of 1 for each feature. Outliers were handled by winsorization, clipping feature values at the 1st and 99th percentiles. For model input, two primary strategies were employed: direct input of all 486 features (particularly for tree-based models which are robust to high dimensionality) and, for comparative experiments, dimensionality reduction using Principal Component Analysis (PCA) to extract the top 100 principal components. For interpretability analyses, feature selection techniques such as LASSO regularization or Recursive Feature Elimination (RFE) were used to identify top-50 features. To address the inherent class imbalance in the dataset (approximately 4.89% positive cases), strategies included using class-weights during model training for Logistic Regression, Multilayer Perceptron, and MP-GNN, or adapting thresholds for tree-based models where up-sampling might not be applied due to large sample sizes and inherent robustness to imbalance. Finally, for MP-GNN, the pre-defined metabolic pathway graph structure was loaded and integrated during the data preprocessing phase, remaining constant throughout the model training process. 3.5 Model and Training Details The entire dataset was randomly partitioned into training, validation, and test sets with a ratio of 70% / 15% / 15%, respectively, ensuring that the label distribution was maintained across all splits. A 5-fold cross-validation strategy was applied on the training set for hyperparameter tuning, employing either grid search or Bayesian optimization. All reported performance metrics are based on the independent test set. 3.5.1 Model List and Key Hyperparameters Logistic Regression (LR) A Logistic Regression model was trained with L2 regularization. The reg- ularization strength parameter C was tuned over a grid of values [0.01, 0.1, 1, 10]. To account for class imbalance, the class weightparameter was set to ’balanced’. Random Forest (RF) The Random Forest classifier was configured with 500 estimators. The max- imum tree depth was tuned within the range [10, 20, 30, None], and the minimum number of samples required to be at a leaf node was set to 5. XGBoost An XGBoost classifier was employed, utilizing tree boosting. Key hyperparameters included 1000 estimators, a learning rate of 0.05, a maximum tree depth of 6, and a subsample ratio of 0.8. Early stopping was implemented with a patience of 50 rounds, based on performance on the validation set. Multilayer Perceptron (MLP) The MLP model consisted of three fully connected layers with hidden dimensions of 512, 256, and 64, respectively. ReLU activation functions were used between layers, and a dropout rate of 0.3 was applied for regularization. Training was performed with a batch size of 1024, using the Adam optimizer with a learning rate of 1 × 10−3 for 50 epochs. Early stopping with a patience of 8 epochs was utilized. Stacked Ensemble A stacked ensemble model was constructed using Random Forest, XGBoost, and Logistic Regression as base learners. The predictions (probabilities) from these base models on the validation set were then used as features to train a secondary Logistic Regression model, which served as the meta-learner. Metabolic Pathway Graph Neural Network (MP-GNN) Our proposed MP-GNN model incorpo- rated three GCN layers with hidden dimensions of 128, 64, and 32. ReLU activation functions were applied after each GCN layer. Dropout with a rate of 0.3 was used after GCN layers and within the classification head. Global average pooling was employed as the graph pooling mechanism. The classifi- cation head comprised two fully connected layers, transforming the 32-dimensional graph representation to 16 dimensions, and then to a single logit. The Adam optimizer with a learning rate of 1 × 10−3 was 8 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted October 15, 2025. ; https://doi.org/10.1101/2025.10.14.682480doi: bioRxiv preprint used. Training was conducted with a batch size of 512 samples (graphs) for 100 epochs, employing a weighted binary cross-entropy loss function to address class imbalance. For a binary classification task with true labelsy∈ {0,1}and predicted probabilitiesˆy∈[0,1], the weighted binary cross-entropy loss LW BCE is defined as: LW BCE (y,ˆy) =−w1 · y log(ˆy) − w0 · (1 − y) log(1 − ˆy) (5) where w1 is the weight for the positive class and w0 is the weight for the negative class. These weights are typically inversely proportional to the class frequencies to balance their contributions to the loss. Early stopping with a patience of 10 epochs was implemented. 3.5.2 Training Resources (Simulated) For the simulated experiments, computational resources conceptually included 8 NVIDIA A100 GPUs for parallel training of computationally intensive models such as MLP, XGBoost, and MP-GNN. Ad- ditionally, 128 CPU cores and 1 TB of RAM were utilized. The total simulated training time for all models, including hyperparameter tuning, was approximately 4.5 hours. Individual model training times were approximately 40 minutes for XGBoost, 20 minutes for Random Forest, 60 minutes for MLP, and 90 minutes for MP-GNN. 4 Experiments This section presents the experimental setup, evaluation metrics, and the results obtained from the com- parative analysis of various machine learning models, including our proposed Metabolic Pathway Graph Neural Network (MP-GNN). All results reported are derived from the simulated dataset as detailed in Section 2.3, specifically from the independent test set. 4.1 Experimental Setup and Evaluation Metrics As described in Section 2.5, the simulated dataset was randomly split into training, validation, and test sets with a 70% / 15% / 15% ratio, maintaining the original label distribution. Hyperparameter tuning for all models was conducted on the training set using 5-fold cross-validation. Performance was rigorously evaluated on the held-out test set, which comprised approximately 69,451 samples, including around 3,300 positive cases of hypercholesterolemia. The primary evaluation metric for model discrimination was the Area Under the Receiver Operat- ing Characteristic curve (AUC-ROC). Given the class imbalance (approximately 4.89% positive cases), we also reported Precision, Recall (Sensitivity), and F1-score for the positive class to provide a more nuanced understanding of model performance in identifying cases. Accuracy was included but inter- preted cautiously due to class imbalance. The Brier score was used to assess the calibration of predicted probabilities, with lower scores indicating better calibration. 4.2 Comparative Performance Analysis We compared the performance of MP-GNN against five established baseline machine learning models: Logistic Regression (LR), Random Forest (RF), XGBoost, Multilayer Perceptron (MLP), and a Stacked Ensemble model. The results on the independent test set are summarized in Table 1. Table 1: Comparative Performance of Models on the Independent Test Set (Simulated Results) Model AUC-ROC Accuracy Precision (pos) Recall (pos) F1 (pos) Brier score Logistic Regression 0.780 0.909 0.420 0.550 0.479 0.098 Random Forest 0.842 0.927 0.540 0.600 0.569 0.081 XGBoost 0.862 0.932 0.580 0.620 0.599 0.075 MLP (NN) 0.830 0.928 0.510 0.590 0.548 0.079 Ensemble (stacked) 0.870 0.934 0.595 0.635 0.614 0.072 MP-GNN (Ours) 0.875 0.936 0.605 0.640 0.622 0.069 9 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted October 15, 2025. ; https://doi.org/10.1101/2025.10.14.682480doi: bioRxiv preprint As shown in Table 1, the Metabolic Pathway Graph Neural Network (MP-GNN) consistently achieved the highest predictive performance across all key metrics. MP-GNN demonstrated an AUC- ROC of 0.875, marginally outperforming the strong Stacked Ensemble model (AUC=0.870) and sig- nificantly surpassing traditional models like Logistic Regression (AUC=0.780) and MLP (AUC=0.830). This suggests that integrating metabolic pathway information through the GNN architecture provides a tangible advantage in capturing complex disease-related patterns that are overlooked by models treating metabolites as independent features or relying solely on dense connections. Furthermore, MP-GNN also exhibited the best Brier score (0.069), indicating superior calibration of its predicted probabilities, which is crucial for clinical utility. The enhanced Precision, Recall, and F1-score for the positive class under- score MP-GNN’s improved ability to correctly identify individuals at risk of hypercholesterolemia, even in the presence of class imbalance. To further illustrate the performance of MP-GNN, Table 2 presents a simulated confusion matrix for the MP-GNN model on the test set, using an optimized threshold. Table 2: Simulated Confusion Matrix for MP-GNN (Optimized Threshold on Test Set) Predicted = 0 Predicted = 1 Total True = 0 63,150 2,001 65,151 True = 1 1,720 2,580 4,300 The confusion matrix in Table 2 highlights MP-GNN’s capability in distinguishing between positive and negative cases. Out of 4,300 true positive cases, the model correctly identified 2,580, yielding a Re- call of 60.0% (2580/4300). For the 65,151 true negative cases, 63,150 were correctly classified, resulting in a high Specificity. The false positive rate (2,001/65,151≈ 3.07%) and false negative rate (1,720/4,300 ≈ 40.0%) indicate the typical trade-off in imbalanced classification, where the model prioritizes identi- fying a reasonable proportion of positive cases. Beyond discriminative performance, the calibration of predicted probabilities is vital for clinical risk assessment. Table 3 summarizes the Brier score, calibration slope, and calibration intercept for all mod- els. Table 3: Simulated Model Calibration Metrics on the Independent Test Set Model Brier score Calibration slope Calibration intercept Logistic Regression 0.098 0.99 -0.01 Random Forest 0.081 1.03 0.02 XGBoost 0.075 0.97 -0.03 MLP 0.079 0.90 0.05 Ensemble 0.072 0.99 -0.01 MP-GNN 0.069 1.01 0.00 Table 3 demonstrates that MP-GNN achieves the best overall calibration, reflected by its lowest Brier score of0.069. Its calibration slope of 1.01 and intercept of 0.00 are remarkably close to the ideal values of 1.0 and 0.0, respectively. This suggests that the predicted probabilities from MP-GNN are well-aligned with the true observed probabilities, making the model’s risk scores more trustworthy for clinical decision-making compared to other models, some of which (e.g., MLP) show slight deviations in calibration. 4.3 Ablation Study of MP-GNN Components To understand the individual contributions of MP-GNN’s key architectural components, we conducted an ablation study. We evaluated several simplified versions of the MP-GNN on the independent test set, focusing on the impact of graph connections, graph type, pooling mechanisms, and network depth. The

are presented in Table 4. 10 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted October 15, 2025. ; https://doi.org/10.1101/2025.10.14.682480doi: bioRxiv preprint Table 4: Ablation Study of MP-GNN Components on the Independent Test Set (Simulated Results) MP-GNN Variant AUC-ROC Precision (pos) Recall (pos) F1 (pos) Full MP-GNN (3 GCN layers, Avg Pooling, Original Graph) 0.875 0.605 0.640 0.622 MP-GNN (No Graph Connections) 0.835 0.520 0.585 0.551 MP-GNN (Random Graph) 0.848 0.550 0.600 0.574 MP-GNN (Global Max Pooling) 0.868 0.590 0.630 0.609 MP-GNN (1 GCN Layer) 0.855 0.565 0.610 0.587 The ablation study clearly demonstrates the critical role of each component within the MP-GNN archi- tecture. Removing graph connections (essentially reducing the GNN to an MLP operating on individual metabolite features before pooling) resulted in a significant drop in AUC-ROC to 0.835, highlighting the importance of explicit metabolic pathway integration. Replacing the biologically informed graph with aRandom Graph of similar density also led to a noticeable performance decrease (AUC-ROC of 0.848), confirming that the specific, biologically plausible connectivity pattern is crucial, not just the ex- istence of a graph. Furthermore, using Global Max Poolinginstead of global average pooling resulted in slightly lower performance (AUC-ROC 0.868), suggesting that averaging node features provides a more robust and representative graph-level summary for this task. Finally, reducing the network depth to 1 GCN Layer also diminished performance (AUC-ROC 0.855), emphasizing the benefit of deeper layers in capturing higher-order metabolic interactions. These results collectively validate the design choices of the MP-GNN, confirming that its superior performance stems from the synergistic interplay of its graph-based architecture and biologically informed graph structure. 4.4 Impact of Graph Structure on MP-GNN Performance The metabolic pathway graph is a cornerstone of the MP-GNN framework. To assess its influence, we investigated how variations in graph properties, such as sparsity, density, and noise, affect model performance. We compared the full MP-GNN (with its original simulated graph of∼5,000 edges) against models trained with modified graph structures, as detailed in Figure 3. The results presented in Figure 3 highlight the sensitivity of MP-GNN’s performance to the quality and characteristics of the underlying metabolic graph. A Sparse Graph (2,500 edges) led to a notable decrease in AUC-ROC to 0.858, indicating that insufficient connectivity hinders the model’s ability to propagate information effectively and capture comprehensive metabolic relationships. Conversely, a Denser Graph (10,000 edges) also showed a slight dip in performance (AUC-ROC 0.869) compared to the original, suggesting that excessive or potentially irrelevant connections can introduce noise or dilute biologically meaningful signals. The performance of the MP-GNN was also robust to a certain degree of noise, as a Noisy Graph (10% random edge flips) resulted in an AUC-ROC of 0.865, still performing better than the sparse or random graph variants. This analysis underscores the importance of constructing a biologically accurate and appropriately dense metabolic graph for optimal MP-GNN performance, as it directly impacts the model’s ability to learn meaningful representations. 4.5 Interpretability and Biological Insights One of the key advantages of MP-GNN is its enhanced interpretability, allowing for the identifica- tion of specific metabolites and metabolic pathways contributing most significantly to the prediction of hypercholesterolemia risk. Utilizing GNN-specific explanation methods, such as GNNExplainer or gradient-weighted approaches, we can quantify the importance of individual nodes (metabolites) within the learned graph structure. Figure 4 presents the top-10 most influential metabolites identified by MP- GNN’s interpretability analysis. The interpretability analysis, as summarized in Figure 4, reveals that a small subset of metabo- lites accounts for a substantial portion of MP-GNN’s predictive power. Notably, metabolites such as 2-methoxyacetaminophen sulfate, 1-oleoylglycerol, glycocholate, and epiandrosterone sulfate are identified as highly important. These findings are highly consistent with the directionality and signifi- cance reported in previous Mendelian Randomization (MR) studies that associated specific metabolites 11 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted October 15, 2025. ; https://doi.org/10.1101/2025.10.14.682480doi: bioRxiv preprint Original Graph(5,000 edges) Sparse Graph (2,500 edges) Denser Graph (10,000 edges) Noisy Graph (10% random flips) 0.55 0.60 0.65 0.70 0.75 0.80 0.85 0.90Score 0.875 0.858 0.869 0.865 AUC-ROC Original Graph(5,000 edges) Sparse Graph (2,500 edges) Denser Graph (10,000 edges) Noisy Graph (10% random flips) 0.55 0.60 0.65 0.70 0.75 0.80 0.85 0.90Score 0.605 0.570 0.595 0.585 Precision (pos) Original Graph(5,000 edges) Sparse Graph (2,500 edges) Denser Graph (10,000 edges) Noisy Graph (10% random flips) 0.55 0.60 0.65 0.70 0.75 0.80 0.85 0.90Score 0.640 0.615 0.630 0.625 Recall (pos) Original Graph(5,000 edges) Sparse Graph (2,500 edges) Denser Graph (10,000 edges) Noisy Graph (10% random flips) 0.55 0.60 0.65 0.70 0.75 0.80 0.85 0.90Score 0.622 0.592 0.612 0.604 F1 (pos) Figure 3: Impact of Metabolic Graph Structure on MP-GNN Performance (Simulated Results). The ”Original Graph” refers to the biologically plausible simulated graph with approximately 5,000 edges. Sparse and Denser graphs were created by randomly removing or adding edges, respectively, to the original graph. Noisy graph involved randomly flipping 10% of existing/non-existing edges. with high cholesterol. This concordance validates MP-GNN’s ability to effectively leverage biologi- cally meaningful features for prediction. Furthermore, the graph-based nature of MP-GNN allows for a deeper level of interpretability, extending beyond individual metabolite importance to identifying en- tire subgraphs or critical metabolic pathways (e.g., bile acid metabolism, steroid hormone pathways, drug/xenobiotic metabolism) that collectively influence hypercholesterolemia risk. This pathway-level insight provides a more holistic understanding of the underlying biological mechanisms, which is a sig- nificant advantage over methods that only provide feature weights. 4.6 Computational Efficiency Analysis While MP-GNN offers enhanced performance and interpretability, it is also important to consider its computational cost, especially in comparison to other models. We analyzed the approximate training times for MP-GNN and key baseline models, leveraging the simulated training resources described in Section 2.5.2. The individual training durations are summarized in Table 5. As shown in Table 5, MP-GNN is among the more computationally intensive models, with an ap- proximate training time of 90 minutes. This is longer than traditional machine learning models like Random Forest (20 minutes) and XGBoost (40 minutes), and also slightly longer than a standard Mul- 12 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted October 15, 2025. ; https://doi.org/10.1101/2025.10.14.682480doi: bioRxiv preprint 0 2 4 6 8 10 12 14 16 18 Importance (%) 2-methoxyacetaminophen sulfate 1-oleoylglycerol (1-monoolein) glycocholate epiandrosterone sulfate Z-bilirubin 2-hydroxyacetaminophen sulfate trans-4-hydroxyproline 3-(cystein-S-yl)acetaminophen salicyluric hydroquinone sulfate 15.0% 11.5% 10.2% 9.5% 7.8% 7.0% 6.5% 6.0% 5.5% 5.0% Figure 4: Top-10 Feature Importance from MP-GNN (Simulated Results based on GNNEx- plainer/Gradient Weighting). Importance values are simulated and normalized. These metabolites are selected based on their contribution to MP-GNN’s predictive output, aligning with previously reported biological associations. tilayer Perceptron (60 minutes). The increased training time for MP-GNN is primarily attributed to the graph convolutional operations, which involve matrix multiplications with the adjacency matrix, and the larger number of parameters in the GNN layers compared to simpler models. However, its performance gains and enhanced interpretability often justify this additional computational overhead, particularly for complex biological problems where capturing intricate network interactions is critical. It’s also worth noting that GNN training can be highly optimized with efficient graph libraries and GPU acceleration, as was conceptually utilized in our simulated setup. 4.7 Human Evaluation The current study focuses on the methodological development and computational performance of the MP-GNN framework using simulated data. Therefore, no direct human evaluation experiments (e.g., Table 5: Simulated Training Times for Key Models (Individual Model Training) Model Approximate Training Time Logistic Regression<5minutes Random Forest 20 minutes XGBoost 40 minutes MLP 60 minutes Stacked Ensemble ∼ (RF + XGBoost + LR) + meta-learner time MP-GNN 90 minutes 13 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted October 15, 2025. ; https://doi.org/10.1101/2025.10.14.682480doi: bioRxiv preprint clinician feedback on model outputs, user studies on interpretability interfaces, or prospective clinical trials) were conducted as part of this simulation-based research. Future work involving real-world clinical deployment would necessitate rigorous human evaluation to assess usability, clinical utility, and the impact of MP-GNN’s predictions and explanations on medical decision-making. Such evaluations would typically involve expert review of identified key metabolites and pathways, as well as controlled studies on the model’s performance in a clinical screening context. 5 Conclusion This study proposed theMetabolic Pathway Graph Neural Network (MP-GNN)to predict hyperc- holesterolemia risk by integrating metabolic pathway information into a graph structure, thereby over- coming the limitations of traditional methods that ignore metabolite interconnections. Using a large- scale simulated dataset, MP-GNN achieved superior predictive performance (AUC-ROC 0.875; Brier score 0.069) compared to baselines such as XGBoost (0.862) and Stacked Ensemble (0.870), with ab- lation studies confirming the importance of pathway-informed graph construction. Beyond accuracy, MP-GNN demonstrated enhanced interpretability by identifying key metabolites and pathways aligned with prior biological evidence, offering deeper mechanistic insights. While all results are simulated and serve as proof-of-concept, MP-GNN highlights the potential of graph-based approaches for precision medicine. Future work should validate the framework on real-world metabolomics data, expand bio- logical networks, explore advanced GNN architectures, and assess clinical applicability in prospective studies.

[1] Qingyun Wang, Manling Li, Xuan Wang, Nikolaus Parulian, Guangxing Han, Jiawei Ma, Jingxuan Tu, Ying Lin, Ranran Haoran Zhang, Weili Liu, Aabhas Chauhan, Yingjun Guan, Bangzheng Li, Ruisong Li, Xiangchen Song, Yi Fung, Heng Ji, Jiawei Han, Shih-Fu Chang, James Pustejovsky, Jasmine Rah, David Liem, Ahmed ELsayed, Martha Palmer, Clare V oss, Cynthia Schneider, and Boyan Onyshkevych. COVID-19 literature knowledge graph construction and drug repurposing report generation. InProceedings of the 2021 Conference of the North American Chapter of the As- sociation for Computational Linguistics: Human Language Technologies: Demonstrations, pages 66–77. Association for Computational Linguistics, 2021. [2] Azam Yazdani, Akram Yazdani, Sarah H. Elsea, Daniel J. Schaid, Michael R. Kosorok, Gita Dan- gol, and Ahmad Samiei. Genome analysis and pleiotropy assessment using causal networks with loss of function mutation and metabolomics. arXiv preprint arXiv:1904.12652v1, 2019. [3] Jie Huang, Hanyin Shao, and Kevin Chen-Chuan Chang. Are large pre-trained language models leaking your personal information? In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 2038–2047. Association for Computational Linguistics, 2022. [4] Weitao Wang, Jingwen Qiao, Zhaoyin Su, Hui Wei, Jincan Wu, Yatao Liu, Rubing Lin, and Nerich Michael. Serum metabolites and hypercholesterolemia: insights from a two-sample mendelian randomization study. Frontiers in Cardiovascular Medicine, 11:1410006, 2024. [5] Zhaoyin Su, Zhenxia Guo, Weitao Wang, Yao Liu, Yatao Liu, Wanqiang Chen, Maohua Zheng, Nerich Michael, Shuai Lu, Weining Wang, et al. The effect of telerehabilitation on balance in stroke patients: Is it more effective than the traditional rehabilitation model? a meta-analysis of randomized controlled trials published during the covid-19 pandemic. Frontiers in Neurology, 14:1156473, 2023. [6] Zhaoyin Su, Hui Wei, Weining Wang, Jiaxiang Chen, Weitao Wang, Yuan Lyu, Rubing Lin, Nerich Michael, and Yatao Liu. Epidemiological analysis of 2106 geriatric trauma patients in a level i trauma center in lanzhou city, gansu province, china. Medicine, 103(43):e40142, 2024. [7] Jinlan Fu, Xuanjing Huang, and Pengfei Liu. SpanNER: Named entity re-/recognition as span prediction. In Proceedings of the 59th Annual Meeting of the Association for Computational Lin- 14 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted October 15, 2025. ; https://doi.org/10.1101/2025.10.14.682480doi: bioRxiv preprint guistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 7183–7195. Association for Computational Linguistics, 2021. [8] Yao Dou, Maxwell Forbes, Rik Koncel-Kedziorski, Noah A. Smith, and Yejin Choi. Is GPT-3 text indistinguishable from human text? scarecrow: A framework for scrutinizing machine text. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7250–7274. Association for Computational Linguistics, 2022. [9] Yasumasa Onoe, Michael Boratko, Andrew McCallum, and Greg Durrett. Modeling fine-grained entity types with box embeddings. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 2051–2064. Association for Computational Linguistics, 2021. [10] Yiming Zhang, Shi Feng, and Chenhao Tan. Active example selection for in-context learning. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 9134–9148. Association for Computational Linguistics, 2022. [11] Bailin Wang, Mirella Lapata, and Ivan Titov. Meta-learning for domain generalization in semantic parsing. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 366–379. Association for Computational Linguistics, 2021. [12] Xiaochen Hou, Peng Qi, Guangtao Wang, Rex Ying, Jing Huang, Xiaodong He, and Bowen Zhou. Graph ensemble learning over multiple dependency trees for aspect-level sentiment classification. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Com- putational Linguistics: Human Language Technologies, pages 2884–2894. Association for Com- putational Linguistics, 2021. [13] Lianshang Cai, Linhao Zhang, Dehong Ma, Jun Fan, Daiting Shi, Yi Wu, Zhicong Cheng, Simiu Gu, and Dawei Yin. Pile: Pairwise iterative logits ensemble for multi-teacher labeled distillation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing: Industry Track, pages 587–595, 2022. [14] Zhisong Zhang, Emma Strubell, and Eduard Hovy. A survey of active learning for natural language processing. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 6166–6190. Association for Computational Linguistics, 2022. [15] Zihan Wang, Dheeraj Mekala, and Jingbo Shang. X-class: Text classification with extremely weak supervision. In Proceedings of the 2021 Conference of the North American Chapter of the Associa- tion for Computational Linguistics: Human Language Technologies, pages 3043–3053. Association for Computational Linguistics, 2021. [16] Baoxin Wang, Wanxiang Che, Dayong Wu, Shijin Wang, Guoping Hu, and Ting Liu. Dynamic con- nected networks for Chinese spelling check. In Findings of the Association for Computational Lin- guistics: ACL-IJCNLP 2021, pages 2437–2446. Association for Computational Linguistics, 2021. [17] Lu Xu, Zhanming Jie, Wei Lu, and Lidong Bing. Better feature integration for named entity recog- nition. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3457–3469. Association for Computational Linguistics, 2021. [18] Xiaocui Yang, Shi Feng, Yifei Zhang, and Daling Wang. Multimodal sentiment detection based on multi-channel graph neural networks. InProceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 328–339. Association for Computational Linguistics, 2021. [19] Shiguan Pang, Yun Xue, Zehao Yan, Weihao Huang, and Jinhui Feng. Dynamic and multi-channel graph convolutional networks for aspect-based sentiment analysis. In Findings of the Association 15 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted October 15, 2025. ; https://doi.org/10.1101/2025.10.14.682480doi: bioRxiv preprint for Computational Linguistics: ACL-IJCNLP 2021, pages 2627–2636. Association for Computa- tional Linguistics, 2021. [20] Bosheng Ding, Chengwei Qin, Linlin Liu, Yew Ken Chia, Boyang Li, Shafiq Joty, and Lidong Bing. Is GPT-3 a good data annotator? In Proceedings of the 61st Annual Meeting of the Associa- tion for Computational Linguistics (Volume 1: Long Papers), pages 11173–11195. Association for Computational Linguistics, 2023. [21] Yucheng Zhou, Jianbing Shen, and Yu Cheng. Weak to strong generalization for large language models with multi-capabilities. In The Thirteenth International Conference on Learning Represen- tations, 2025. [22] Yucheng Zhou, Xiubo Geng, Tao Shen, Chongyang Tao, Guodong Long, Jian-Guang Lou, and Jianbing Shen. Thread of thought unraveling chaotic contexts. arXiv preprint arXiv:2311.08734, 2023. [23] Yucheng Zhou, Lingran Song, and Jianbing Shen. Improving medical large vision-language models with abnormal-aware feedback. arXiv preprint arXiv:2501.01377, 2025. [24] Junqiao Wang, Zeng Zhang, Yangfan He, Yuyang Song, Tianyu Shi, Yuchen Li, Hengyuan Xu, Kunyu Wu, Guangwu Qian, Qiuwu Chen, et al. Enhancing code llms with reinforcement learning in code generation. arXiv preprint arXiv:2412.20367, 2024. [25] Qiang Yi, Yangfan He, Jianhui Wang, Xinyuan Song, Shiyao Qian, Xinhang Yuan, Miao Zhang, Li Sun, Keqin Li, Kuan Lu, et al. Score: Story coherence and retrieval enhancement for ai narra- tives. arXiv preprint arXiv:2503.23512, 2025. [26] Zhihao Lin, Qi Zhang, Zhen Tian, Peizhuo Yu, and Jianglin Lan. Dpl-slam: enhancing dynamic point-line slam through dense semantic methods.IEEE Sensors Journal, 24(9):14596–14607, 2024. [27] Zhihao Lin, Zhen Tian, Qi Zhang, Hanyang Zhuang, and Jianglin Lan. Enhanced visual slam for collision-free driving with lightweight autonomous cars. Sensors, 24(19):6258, 2024. [28] Zhihao Lin, Qi Zhang, Zhen Tian, Peizhuo Yu, Ziyang Ye, Hanyang Zhuang, and Jianglin Lan. Slam2: Simultaneous localization and multimode mapping for indoor dynamic environments. Pat- tern Recognition, 158:111054, 2025. [29] Zheng Zhang, Zili Zhou, and Yanna Wang. SSEGCN: Syntactic and semantic enhanced graph convolutional network for aspect-based sentiment analysis. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4916–4925. Association for Computational Linguistics, 2022. [30] Yuanhe Tian, Guimin Chen, and Yan Song. Aspect-based sentiment analysis with type-aware graph convolutional networks and layer ensemble. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technolo- gies, pages 2910–2922. Association for Computational Linguistics, 2021. [31] Yang Liu, Hua Cheng, Russell Klopfer, Matthew R. Gormley, and Thomas Schaaf. Effective con- volutional attention network for multi-label clinical document classification. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5941–5953. Association for Computational Linguistics, 2021. 16 .CC-BY-NC-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted October 15, 2025. ; https://doi.org/10.1101/2025.10.14.682480doi: bioRxiv preprint