KAN-PROSPECT: a Kolmogorov–Arnold Networks–integrated framework for predicting the effects and adverse reactions of natural products via transfer learning

Abstract

The current reliance on wet-lab experiments for evaluating the efficacy and adverse effects of natural products remains a major obstacle in new drug discovery. We propose KAN-PROSPECT, a novel deep learning framework that integrates transfer learning with Kolmogorov-Arnold Networks (KAN). This model can simultaneously predict the efficacy and adverse effects of natural products based solely on molecular SMILES representations, thereby addressing the limited generalizability of existing models in predicting these aspects for natural products. Beyond methodological advances, KAN-PROSPECT also contributes to more sustainable and resource-efficient drug discovery. Leveraging a cross-modal transfer learning strategy pretrained on approximately 3,800 drugs and fine-tuned on 400 natural products, KAN-PROSPECT consistently outperforms baseline models in dual-label prediction tasks. Notably, it demonstrates exceptional robustness in addressing data scarcity, excelling particularly in few-shot and zero-shot scenarios. Through the transfer learning strategy, the model partially alleviates the data scarcity issue commonly encountered in natural product prediction tasks. In addition, the incorporation of KAN layers enhances the ability to model complex nonlinear relationships between molecular structures and associated pharmacological or adverse reaction profiles, contributing to improved predictive performance. Furthermore, the framework demonstrates strong generalization ability, enabling high-accuracy predictions for the efficacy and adverse effects of entirely new natural products. KAN-PROSPECT was further applied to comprehensively predict natural products from the MEC and NPASS databases, with Icaritin from Epimedium used as a representative case study. Overall, KAN-PROSPECT is the first framework to unify transfer learning with the KAN architecture for dual-label prediction of natural products, showing great potential for large-scale bioactivity and toxicity prediction, new drug development, and drug repositioning of natural products.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1038/s41598-026-57761-2.

Introduction

Accurate prediction of the Anatomical Therapeutic Chemical (ATC) classification of natural products is considered essential for the elucidation of their pharmacological roles and is regarded as a critical component supporting basic research and drug development^1,2. In the risk–benefit assessment of drugs, precise estimation of adverse reaction frequency is regarded as a key component³. However, many adverse reactions are often identified only after market approval^4–7, which not only compromises patient safety but also results in substantial economic losses for pharmaceutical companies^8–10. Due to their chemical diversity and structural complexity, natural products and their derivatives have become an important source of antitumor and other therapeutic agents, which have been utilized as core templates for new drug design and have driven the development of derivatization and biomimetic synthesis strategies^11–13. Nevertheless, the systematic prediction of natural product ATC and adverse reactions (ADR) remains a major challenge. On one hand, most existing approaches focus primarily on structure discovery, bioactivity prediction, or target identification of natural products^14–18 and still show clear limitations in predicting effects and adverse reactions. On the other hand, current drug-related predictive models are generally restricted to single-task learning—predicting either ATC or ADR—without a unified framework capable of addressing both simultaneously¹⁹. Furthermore, fragmented natural product databases, insufficient molecular representation, and incomplete annotation quality have been recognized as major obstacles to comprehensive data acquisition and the development of high-performance predictive models for natural products^20–23.

In view of the aforementioned challenges—such as single-task limitations, scattered data sources, poor model generalization, and the absence of a unified prediction framework for ATC and ADR—a novel deep learning framework, termed KAN-PROSPECT, is proposed. Notably, a series of representative models have been established for ATC classification and ADR prediction. Classic ATC prediction baselines include ATC-CNN²⁴, ATC-NLSP²⁵, and CGATCPred²⁶, while widely used ADR prediction methods cover Galeano et al. model²⁷, MGPred²⁸, DSGAT²⁹, and SDPred³⁰. Most of these existing approaches adopt single-task architectures and are mainly validated on general drug datasets, with limited generalization capability for natural products. These methods are selected as comparative baselines in our experiments, with further technical details provided in the supplementary. The framework is driven by fusion-based transfer learning³¹ and utilizes drug and natural product data collected from public databases for model training and knowledge transfer, thereby enabling dual-task joint prediction based solely on molecular SMILES representations. Through this design, ATC and ADR profiles of natural products can be simultaneously generated without requiring additional target or biomarker inputs. Compared with previous models that focus on single tasks or rely heavily on complex feature engineering, KAN-PROSPECT demonstrates superior generalization and cross-domain transfer capability, particularly in few-shot and zero-shot scenarios. It can effectively identify potential pharmacological features of previously unseen natural products³². The introduction of KAN layers within the framework further improves nonlinear representation learning capability, facilitating modeling of complex relationships between molecular features and ATC/ADR-related outputs.

KAN-PROSPECT provides an efficient and unified framework for predicting the ATC and ADR profiles of natural products. The use of SMILES-only input design significantly reduces the cost and complexity of data acquisition^22,33, the term “SMILES-only” indicates that prediction can be achieved solely from molecular SMILES, without requiring additional physicochemical features, target information, or predefined ATC/ADR labels from the user. The ATC/ADR embeddings are generated internally by the model and concatenated with molecular embeddings during the prediction pipeline. By integrating transfer learning with the KAN architecture, high prediction accuracy and robustness are achieved, even under imbalanced data conditions. Moreover, the architecture enables integrated modeling of molecular representations together with ATC and ADR feature information, facilitating analysis of potential associations between molecular characteristics and pharmacological or adverse reaction profiles. This framework provides a unified strategy for dual-label prediction of natural products and may help address some limitations associated with conventional single-task drug prediction frameworks. KAN-PROSPECT may therefore support pharmacological research, natural product–based drug development, drug repositioning, and clinical risk assessment. Compared with traditional experimental screening approaches, artificial intelligence–based in silico models enable early-stage prediction of potential efficacy and safety signals, thereby helping reduce experimental burden and supporting a more efficient discovery workflow for natural product research.

Methods

A model construction and evaluation pipeline for natural products was developed based on transfer learning, as illustrated in Fig. 1. The model was constructed using the largest-scale dataset currently available. A dual-model framework integrating GAT and GCN layers was first constructed to extract graph-structured molecular features^34,35, while nonlinear KAN layers were introduced to enhance high-dimensional nonlinear feature modeling capability³⁶. Using large-scale annotated drug ATC and ADR datasets as the source domain, the two branches of the framework were independently trained for ATC and ADR prediction tasks. During source-domain training, the model learned both molecular topological features and task-specific representations associated with ATC or ADR annotations, thereby establishing transferable feature representations. The pretrained feature extraction modules and model parameters were subsequently transferred to the natural product domain, where fine-tuning was performed for downstream ATC and ADR prediction tasks. The resulting framework, termed KAN-PROSPECT, enables the simultaneous prediction of ATC classifications and ADR profiles for natural products using molecular SMILES as the primary input.

Fig. 1.

Overview of the KAN-PROSPECT framework. The workflow begins with drug data preprocessing, followed by model construction, and pretraining. Subsequently, natural product data are processed and model parameters are fine-tuned, resulting in a predictive model capable of identifying both effects and adverse reactions of natural products, the KAN-PROSPECT training process follows the sequence shown in the figure. The framework further includes model performance evaluation, generalization assessment, embedding space analysis, and prediction of novel effects, adverse reactions of natural products.

To evaluate the transferability of the framework, model performance before and after transfer learning was systematically compared through both internal validation and independent testing, demonstrating improved predictive performance and enhanced generalization capability following adaptation to natural product data. In addition, the embeddings learned during training were further analyzed to investigate the molecular representations captured by the model. The framework was also applied to predict previously uncharacterized ATC and ADR associations of natural products, providing potential insights for downstream feature interpretation and hypothesis generation.

Dataset

Pretraining and transfer datasets

The drug–effect dataset proposed by Chen *et al.³⁷ and the drug–adverse reaction datasets employed by Galeano et al.²⁷ and Zhao et al.²⁸ were adopted as benchmark datasets. To extend their coverage, complementary information from DrugBank³⁸, KEGG³⁹, and ADReCS⁴⁰ was integrated to fill missing drug entries, and a unified dataset linking pharmacological effects with adverse reactions was established. Compared with the datasets from Chen et al., Galeano et al., and Zhao et al., our integrated dataset shows marked improvements in scale and coverage and represents one of the most comprehensive resources currently available for related predictive studies. This expanded dataset provides sufficient training data and enables transfer learning from approved drugs to natural products, thereby supporting robust and generalisable model performance.

To construct the natural product dataset (Transfer dataset), natural product drugs were extracted from the NPASS²¹ and HERB⁴¹ databases and matched to the updated drug entries by SMILES. Successfully matched entries were designated as the natural product transfer dataset, consisting of molecules directly or indirectly derived from natural sources. Unmatched entries were assigned to the drug pretraining dataset, representing synthetic compounds. Based on the benchmark dataset, this strategy established a clear distinction between drug and natural product data, providing a foundation for the subsequent transfer learning framework. The final dataset includes 8,040 drug–effect pairs and 39,794 drug–ADR pairs, as well as 1,289 natural product–effect pairs and 6,910 natural product–ADR pairs(Table 1). Compared with drug data, natural product data remain relatively sparse and incomplete, further highlighting the necessity of transfer learning in this setting. The similarity comparison between the pretraining dataset and the transfer learning dataset, along with its detailed explanation, is provided in the supplementary.

Table 1.

Data statistics of drug and natural product.

Dataset	Dataset-type	Drug-number	NP-number	Relation-number
Benchmark dataset	Drug-ATC	3,785	–	5,432
Benchmark dataset	Drug-ADR	759	–	37,441
Pretrain dataset	Drug-ATC	3,951	–	8,040
Pretrain dataset	Drug-ADR	1,522		39,794
Transfer dataset	NP-ATC	–	375	1,289
Transfer dataset	NP-ADR	–	195	6,910

To sum up Benchmark Dataset: Publicly available standard drug ATC/ADR datasets for baseline comparison and representation analysis. Pretrain Dataset: Extended from the benchmark dataset by integrating DrugBank, KEGG, and ADReCS, used for source-domain pre-training of the model. Transfer Dataset: Natural product data collected from the NPASS and HERB databases, used for model fine-tuning, performance assessment, and external validation. In Table S1 for the source, composition and specific role of each data set.

In the drug–adverse reaction relationship, the datasets used by Galeano et al. and Zhao et al. include frequency grades for adverse reactions. For instance, common adverse reactions such as headache exhibit an average reported clinical frequency of 3.52, representing the majority of entries in the matrix^27,42–44. To prevent the dominance of common adverse reactions and to enhance the model’s ability to predict rare events, frequency grading was not incorporated. Instead, the presence or absence of each drug–adverse reaction relationship was encoded in a binary form (0/1). This design mitigates class imbalance arising from frequency stratification and allows the model to learn from and predict both frequent and rare adverse reactions. Since experimentally validated true negative associations are largely unavailable in current biomedical databases, unobserved drug–effect and drug–adverse reaction pairs were treated as candidate negative samples rather than biologically confirmed true negatives^45,46. The construction of negative samples is detailed in the supplementary.

Structural classification datasets

Structurally classified natural product data were collected from the MedChemExpress (MCE) database, including categories such as ketones, alkaloids, and phenolics^47,48. In total, 16 structural classes comprising 14,232 natural products were identified, screened, and curated for subsequent analysis. These structurally categorized natural products were further used for effect and adverse reaction prediction within their respective classes, enabling systematic assessment of potential prediction bias and model performance differences across distinct structural categories.

Structural similarity datasets

Natural product compounds with structural similarity information were collected from the NPASS database. Structural similarity annotations were obtained from the MCE database, in which each compound contains up to 200 structurally similar reference compounds derived from public chemical databases. Molecular structural similarity was quantified using the Tanimoto coefficient based on molecular fingerprint similarity calculations, which is a widely used metric in cheminformatics studies.

Based on the reported similarity scores, 63,629 natural products were screened and stratified into four structural similarity intervals (≥ 0.85, 0.75–0.85, 0.60–0.75, and < 0.60)^18,49. Pharmacological effect and adverse reaction prediction tasks were subsequently performed within each similarity interval to systematically evaluate the robustness and generalization capability of KAN-PROSPECT across compounds with different levels of structural similarity to the transfer learning dataset.

This analysis was designed to evaluate whether the model maintained stable predictive behavior across distinct chemical similarity levels, thereby reducing potential overreliance on structurally similar compounds present in the training data.

Feature extraction and fusion

Molecular structure features

Molecular structures of natural products and drugs were represented by their SMILES strings obtained from the DrugBank database. To enhance the robustness of the model to different SMILES representations and ensure that it learns chemical structure rather than string order, a SMILES randomization strategy was adopted during data preprocessing (see supplementary for details). For each natural product, a canonical SMILES was retained as the original representation and generated additional random SMILES for non-unique molecules using the open-source RDKit chemistry toolkit^50–53. This procedure expanded data diversity and reduced the model’s dependence on any fixed SMILES representation. For molecular feature extraction, RDKit was used to compute atom-level descriptors including atom count, element symbol, atomic degree, explicit, and implicit hydrogens, valence electrons, and aromaticity^54–56. These attributes were one-hot encoded, while chemical bond connections were extracted to construct molecular graphs, yielding both feature vectors, and graph vectors for each compound^57,58 (see supplementary for details).

ATC/ADR coding features

ATC and ADR coding features were incorporated to capture hierarchical pharmacological and phenotypic information. Both ATC and ADR codes possess structured hierarchies, where different levels carry specific semantic meanings^59–61. We first split each code into hierarchical fragments according to its official classification system, then performed standardization and one-hot encoding. The hierarchical one-hot vectors were concatenated and flattened into a fixed-length 1D tensor, which was fed into Conv1d layers for feature extraction. The detailed transformation instances and implementation pipeline are provided in supplementary.

Feature fusion strategy

a feature fusion strategy was applied to integrate molecular embeddings with ATC/ADR embeddings. Specifically, the feature embedding of a natural product was concatenated with the corresponding ATCs or ADRs embedding to form a composite vector.

1

represents the embedding of natural product features in the model, where represents the number of natural products and each natural product has features. : represents the embedding of ATC or ADR in the model, formed through one-dimensional convolution. It has n ATCs or ADRs, with each ATC or ADR having features. Again, n equals the batch size. The symbol concatenates two one-dimensional vectors end to end, forming a longer one-dimensional vector. : is the concatenated tensor, which we refer to as the “feature embedding”. Through KAN, we train the natural product features and their corresponding ATC or ADR features as a sum of univariate functions and predict the binding probability(see supplementary for details).

KAN-PROSPECT architecture

KAN-PROSPECT is a transfer learning framework designed for predicting the ATC classes and ADRs of natural products. The framework consists of two independent models: KAN-PROSPECT-ATC for ATC classification and KAN-PROSPECT-ADR for ADR prediction. Although the two models share a similar architecture, their parameters are trained separately. The model architecture contains three stages: molecular representation learning, feature fusion, and prediction. First, molecular structures represented by SMILES are converted into graph representations and processed using GAT and GCN layers to capture both local and global structural information. The resulting molecular features are further refined through KAN layers to improve nonlinear feature representation.

In parallel, ATC or ADR codes are encoded as hierarchical feature vectors and processed using Conv1d layers followed by KAN layers to capture semantic relationships within the coding systems. The generated molecular embeddings and ATC/ADR embeddings are then concatenated and passed through additional KAN layers for feature fusion and classification.

Compared with conventional MLP-based architectures, the introduction of KAN layers improves nonlinear modeling capability while maintaining a more structured representation of feature interactions. To address the limited availability of labeled natural product data, KAN-PROSPECT adopts a transfer learning strategy in which the model is first pretrained on large-scale drug datasets and subsequently fine-tuned on natural product datasets. A schematic overview of the framework is shown in Fig. 2. Detailed network configurations, layer parameters, and embedding structures are provided in supplementary.

Fig. 2.

The framework of the KAN-PROSPECT model. The frameworks for predicting the ATCs and ADRs of natural products are both combinations of the models shown in the figure (see supplementary for details).

Model performance evaluation

The model was trained and evaluated using five-fold cross-validation to construct predictive frameworks for the pharmacological effects and adverse reactions of natural products^62,63. Since BCEWithLogitsLoss was adopted during model training, the model output corresponded to raw logits rather than sigmoid-normalized probabilities. Therefore, binary labels were assigned using a logit threshold of 0, where outputs greater than 0 were classified as positive samples (label = 1), and outputs less than or equal to 0 were classified as negative samples (label = 0). This strategy is mathematically equivalent to applying a sigmoid activation followed by a probability threshold of 0.5^64,65. The corresponding evaluation metrics and formulas are provided in supplementary. To further evaluate predictive performance, KAN-PROSPECT was compared with a series of baseline models using the transfer dataset. The machine learning baselines included SVM, KNN, Logistic- regression, Naive Bayes, Decision Tree, and Random Forest^66–68. The deep learning baselines comprised graph-based architectures (GCN, GAT, GIN, and GraphConv)^34,69 and non-graph architectures (KAN, MLP, CNN, RNN, and Transformer)⁷⁰.

Existing models for predicting drug actions and adverse reactions generally achieve strong performance on benchmark datasets. Representative models for drug action prediction include ATC-CNN, ATC-NLSP and CGATCPred, while those for adverse reaction prediction include Galeano et al.’s model, MGPred, DSGAT and SDPred (see supplementary for details). However, in real-world applications, the associations between drugs and their ATC or ADR labels follow a pronounced long-tail distribution: most drugs are linked to only a few ATC or ADR categories, whereas a minority are associated with many. This imbalance is particularly evident for drugs and natural products, substantially constraining model generalizability. As a result, conventional evaluation strategies based solely on random five-fold cross-validation may obscure performance deficiencies under realistic conditions, leading to biased predictions.

To more accurately assess model robustness, the benchmark dataset was re-organized into three evaluation scenarios, Majority, Few-shot, and Zero-shot to replace the traditional five-fold cross-validation approach (see supplementary for details)⁷¹. Within each scenario, the dataset was evenly partitioned into five subsets; four subsets were used for model training and one for validation. The process was repeated across five iterations so that every subset served once as the test set, enabling a comprehensive assessment across varying data-scarcity conditions.

Feature learning and spatial analysis of KAN-PROSPECT

From the perspectives of both drugs and natural products, spatial features capture the model’s ability to integrate molecular characteristics during representation learning. By examining the distribution patterns of these high-dimensional feature vectors within the embedding space, as well as performing similarity and distribution analyses of the model-generated representations, to evaluate whether the model effectively organizes molecular features. Specifically, KAN-PROSPECT was evaluated whether structurally or functionally similar molecules cluster together in the latent space, while molecules of different classes remain well separated^72–75.

Based on the benchmark datasets and transfer datasets, molecular embeddings for drugs and natural products were extracted during the training of KAN-PROSPECT-ATC and KAN-PROSPECT-ADR, respectively. Clustering analysis was first performed on the drug embeddings, and their spatial distributions within the embedding space were visualized. Subsequently, both intra-class and inter-class similarities were quantified to evaluate the structural consistency and discriminability of the learned representations. Scatter plots and box plots are then used to illustrate the distribution characteristics, a similar analysis was also conducted for natural products to examine their embedding distributions. Each similarity value is standardized and the is defined as follows:

2

Where and represent the embedding vectors of two drugs or natural products, denotes their cosine similarity value, and and correspond to the minimum and maximum cosine similarity values of the drug and natural product embeddings within the current or comparative model, respectively. The normalized similarity distribution is subsequently visualized using scatter plots and box plots to demonstrate the clustering characteristics and discriminative capacity of the embedding representations. Further explanation can be provided from the model architecture, where the combined advantages of GAT, GCN, and KAN can be illustrated.

External sample prediction and biological analysis

Ten natural products with well-documented effects and adverse reactions were randomly selected from the transfer dataset as external validation samples, which were not included in the model training process. After each round of model training, these samples were used for prediction to obtain corresponding outputs. This external validation was designed to assess the model’s ability to accurately predict the effects and adverse reactions of unseen natural product data.

In parallel, an equal number of negative samples were randomly selected from the negative subset of the transfer dataset. These negative data were used to evaluate whether the model could capture potential ATC or ADR signals and to explore the underlying mechanisms through pathway enrichment analysis. During prediction, KAN-PROSPECT was observed that several natural products displayed novel predicted associations (true label = 0, predicted label = 1), suggesting that KAN-PROSPECT may identify previously unrecognized relationships between compounds and their biological effects. To further examine the biological plausibility of these predictions, two representative compounds were selected for in-depth analysis: nifedipine, a synthetically modified derivative of a natural product; lovastatin, a naturally derived drug. Known molecular targets of these compounds were obtained from the KEGG database, and pathway enrichment analysis was performed using the DAVID and BioRender tools^76,77 to evaluate whether the predicted effects or adverse reactions exhibited potential biological consistency with previously reported pathways and target information.

Prediction and mechanistic analysis of natural products effects and adverse reaction

Based on the KAN-PROSPECT-ATC and KAN-PROSPECT-ADR models, this study predicted the effects and adverse reactions of natural products obtained from the MCE and NPASS databases. Using all ATC and ADR codes from the transfer dataset, corresponding ATC and ADR prediction labels were generated for each natural product. To optimize the utilization of natural product data and comprehensively evaluate the model’s predictive performance, two analytical strategies were implemented⁷⁸. (1) For natural products with available structural classification information, compounds were categorized according to their structural types to assess the comprehensiveness of model predictions across different molecular classes. The results indicated that the predicted effects and adverse reactions were evenly distributed among various structural categories, suggesting broad coverage, and balanced predictive ability. (2) For natural products derived from the NPASS database, prediction results were further grouped and analyzed according to structural similarity categories. This approach was designed to evaluate whether the model could maintain consistent predictive performance under conditions of low structural similarity, thereby assessing its robustness and generalization capability beyond structurally similar compounds⁷⁹.

Based on these two strategies, this study aimed to delineate the application scope of the KAN-PROSPECT model for predicting the effects and adverse reactions of natural products. KAN-PROSPECT was then applied to perform comprehensive predictions, providing systematic ATC and ADR labels for the natural products under investigation.

To further assess the model’s capability to predict the effects and adverse reactions of novel natural products, icaritin, a approved prenylated flavonoid, was selected as a case study^80–82. These emerging pharmacological activities were used as targets to evaluate the predictive potential of KAN-PROSPECT. By integrating all ATC and ADR codes from the benchmark, pretraining, and transfer datasets, KAN-PROSPECT-ATC and KAN-PROSPECT-ADR were applied to predict icaritin’s effects and adverse reactions. The predicted outcomes were then compared with published pharmacological data and clinical adverse reaction reports to assess prediction accuracy and biological plausibility.

Results

To address key challenges in natural product research—including the absence of dedicated prediction models for effects and adverse reactions, limited training data and the poor generalization of drug-based models to natural products—KAN-PROSPECT, a transfer learning–based framework, was developed. First, a cross-modal transfer learning strategy was implemented, in which approximately 3,800 drugs were used for pretraining to capture deep feature associations between molecular structures and their effects or adverse reactions. The pretrained model was then transferred to the natural product domain and fine-tuned on a dataset of approximately 400 natural products, effectively mitigating the limitations imposed by scarce training data. Second, the model incorporates a multi-mechanism feature extraction and modeling strategy by integrating GAT, GCN, and KAN modules. While GAT and GCN capture molecular structural features, KAN enhances the modeling of complex nonlinear relationships between molecular features and biological outcomes. Finally, a unified dual-label prediction approach was implemented, enabling the simultaneous modeling of effects and adverse reactions of natural products. This overcomes the single-label restriction of conventional models and provides a scalable tool for joint evaluation of natural product effect and adverse reaction. The following sections present the model performance, experimental results and case analyses.

Comparison of baseline model information

Existing models for predicting the effects and adverse reactions of natural products were systematically reviewed and analyzed. To date, no study has addressed the joint prediction of both effects and adverse reactions for natural products. Current approaches primarily focus on single-task prediction in the drug domain. For drug effect prediction, models such as ATC-CNN, ATC-NLSP, and CGATCPred infer ATC categories using substructure or multimodal features. For drug adverse reaction prediction, methods including DSGAT, MGPred, SDPred, and Galeano et al. model capture potential drug–ADR associations using graph neural networks, attention mechanisms or embedding-based frameworks. However, these models do not cover natural product applications and lack a unified framework for simultaneous prediction of both effects and adverse reactions. These models were further compared in terms of methodological characteristics and input data sources, as summarized in Fig. 3. Notably, KAN-PROSPECT achieves the simplest feature requirements for input data while supporting dual-label prediction (Table 1).

Fig. 3.

Comparison of the proposed KAN-PROSPECT framework with existing models for drug effect and adverse reaction prediction. Each row represents a model, and each column indicates a type of information or feature. Colored markers denote whether the model uses or provides the corresponding information.

Performance evaluation of the KAN-PROSPECT model

The KAN-PROSPECT model was successfully constructed and trained using a two-stage training strategy. In the first stage, the model was pretrained on a large-scale drug dataset to capture general pharmacological patterns; in the second stage, it was fine-tuned using natural product data to adapt the learned representations to naturally derived compounds. All experiments were repeated using five-fold cross-validation, and the mean ± standard deviation of each evaluation metric was reported to assess model robustness and stability across different data splits.

Performance comparison before and after transfer learning

KAN-PROSPECT achieved significant improvements across multiple evaluation metrics—including Accuracy, Precision, Recall, AUC, AUPR, F1-score, and MCC, as shown in Table 2. Specifically, KAN-PROSPECT-ATC achieved an accuracy of 0.8811, precision of 0.8728, and AUC of 0.9496, outperforming the baseline Model-ATC (accuracy 0.8710, precision 0.8808, AUC 0.9364). Similarly, KAN-PROSPECT-ADR attained an accuracy of 0.8718 and AUC of 0.9553, showing clear superiority over Model-ADR (accuracy 0.8025, AUC 0.8759). KAN-PROSPECT-ATC and KAN-PROSPECT-ADR were first pre-trained on large-scale drug data and then fine-tuned on natural product data. By comparing the performance before transfer (trained only on drugs) and after transfer (drug pre-training + natural product fine-tuning), we clearly demonstrate that transfer learning significantly boosts prediction performance on natural products and better adapts the model to natural product-specific tasks.

Table 2.

Transfer learning–driven performance improvement of KAN-PROSPECT. Significant values are in bold.

Model	Acc	Pre	Rec	AUC	AUPR	F1-score	Mcc
KAN-PROSPECT-ATC	0.8811 ± 0.0142	0.8728 ± 0.0161	0.9416 ± 0.0107	0.9496 ± 0.0028	0.9604 ± 0.0025	0.9059 ± 0.0044	0.7491 ± 0.0618
Model-ATC	0.8710	0.8808	0.8934	0.9364	0.9415	0.8870	0.7369
KAN-PROSPECT-ADR	0.8718 ± 0.0151	0.8245 ± 0.0073	0.9419 ± 0.0035	0.9553 ± 0.0029	0.9586 ± 0.0024	0.8793 ± 0.0048	0.7514 ± 0.0732
Model-ADR	0.8025	0.7998	0.8020	0.8759	0.8949	0.8009	0.6051

Bold values indicate the highest value for the corresponding evaluation metric.

It is worth noting that KAN-PROSPECT-ATC contains approximately 10.6 million parameters, whereas KAN-PROSPECT-ADR contains approximately 31.8 million parameters, nearly three times larger in scale. This difference is primarily attributable to the substantially greater size and complexity of the ADR dataset compared with the ATC dataset, which increases the difficulty of feature learning and model optimization. Despite this discrepancy, both models remain within a medium-to-large parameter scale, providing sufficient representation capacity while maintaining a reasonable balance between predictive performance and generalization ability. Overall, these results demonstrate that KAN-PROSPECT exhibits robust predictive capability and considerable potential for practical applications in the evaluation of natural product–associated ATC classifications and ADR profiles.

Comparison with machine and deep learning models

The KAN-PROSPECT model was systematically compared with traditional machine learning and deep learning methods using the Transfer Dataset as a benchmark, including both graph-based and non-graph-based approaches, with detailed results presented in supplementary.

At the traditional machine learning level, as shown in Fig. 4A,B, KAN-PROSPECT markedly outperforms conventional models on the transfer dataset. Although the Decision Tree and Random Forest models achieved relatively high MCC values (> 0.90) under small-sample conditions, their performance declined substantially when evaluated on the larger pretraining dataset (Decision Tree* and Random Forest* in Fig. 4B), with MCC values decreasing to 0.44 and 0.49, respectively. This indicates that the generalization ability of traditional machine learning algorithms is severely limited for large-scale data. In contrast, KAN-PROSPECT-ATC and KAN-PROSPECT-ADR achieve MCC values of 0.75, along with superior accuracy and AUC, demonstrating stable performance even with expanded data. These results highlight the necessity of deep learning approaches for predictive tasks involving large-scale natural product datasets.

Fig. 4.

Comparison of KAN-PROSPECT with machine learning and deep learning models. (A) and (B) present the performance comparison of KAN-PROSPECT with traditional machine learning methods in predicting the effects and adverse reactions of natural products, respectively. (C–F) illustrate the comparative results between KAN-PROSPECT and both graph-based deep learning models and non-graph deep learning models for the same prediction tasks. The y-axis represents the normalized performance scores of different evaluation metrics, including Accuracy (Acc), Precision (Pre), Recall (Rec), AUC, AUPR, F1-score, and MCC, with values ranging from 0 to 1.

At the deep learning level, as shown in Fig. 4C–F, KAN-PROSPECT demonstrates clear advantages in overall performance. While maintaining high accuracy, its recall exceeds 0.94, reflecting highly sensitive identification of positive samples. The MCC values of KAN-PROSPECT (0.7491 for ATC/0.7514 for ADR) are substantially higher than those of GAT (0.5313) or GCN (0.5490) alone, indicating that the integrated GAT–GCN architecture synergistically captures molecular topology and chemical features more effectively than individual models.

At the model architecture level, further analysis showed that simply integrating KAN into GIN led to performance degradation (MCC decreased from 0.49/0.56 to 0.26/0.42), indicating that KAN is more suitable for continuous feature space modeling than for discrete topological aggregation. This is consistent with the design principle that GIN primarily relies on graph adjacency for structural representation, where excessive nonlinear transformation may weaken topological information. In contrast, graph neural networks inherently preserve atom-level and neighborhood information, enabling effective modeling of molecular structure and connectivity patterns.

By combining GAT, GCN, and KAN, KAN-PROSPECT integrates structural representation learning with nonlinear function approximation, enabling simultaneous modeling of molecular topology and complex feature–bioactivity relationships. This integration improves predictive performance and enhances representation capacity, particularly for natural products characterized by structural diversity and sparse feature distributions.

Generalizability evaluation

The distribution of drug–ATC and drug–ADR associations exhibited a pronounced long-tail pattern, as illustrated in Fig. 5A–F. Conventional evaluation settings may not adequately reflect model behavior under such highly imbalanced data distributions. Therefore, KAN-PROSPECT was further evaluated under three scenarios—majority, few-shot, and zero-shot—to assess predictive performance across different label-frequency conditions⁴², and the results were compared with those of existing models see supplementary.

Fig. 5.

Data statistics and generalization evaluation. (A–C) show the data statistics used by KAN-PROSPECT-ATC in the Benchmark dataset, Pretrain dataset, and Transfer Dataset; D–F) show the data statistics used by the KAN-PROSPECT-ADR in the Benchmark dataset, Pretrain dataset, and Transfer Dataset; (G–I) show the AUPR values of KAN-PROSPECT-ATC and the corresponding baseline models under the majority, few-shot, and zero-shot scenarios. (J–L) present the AUPR values of KAN-PROSPECT-ADR and the corresponding baseline models under the same scenarios. In Fig. 5(G–L), the y-axis represents the AUPR (Area Under the Precision–Recall Curve) values of different models under the majority, few-shot, and zero-shot evaluation settings.

For the ATC prediction task, models such as ATC-CNN, ATC-NLSP, and CGATCPred achieved competitive performance on the majority subset but showed noticeable decreases under few-shot and zero-shot conditions. By comparison, KAN-PROSPECT-ATC maintained comparatively stable performance across different evaluation settings and consistently achieved competitive AUPRC values relative to benchmark models. In several zero-shot tasks, certain evaluation metrics remained comparable to, or slightly higher than, those observed under conventional settings. One possible explanation is that the combined GAT, GCN, and KAN architecture facilitates learning of generalized molecular feature representations, while the transfer learning framework initialized from large-scale drug datasets improves representation robustness and reduces overfitting to specific label distributions^83,84. However, these observations should be interpreted cautiously, as the current analyses do not fully exclude the potential influence of latent structural similarity, dataset composition bias, or distribution-related effects between different evaluation settings, as shown in Fig. 5G–I.

As Fig. 5J–L, KAN-PROSPECT-ADR was further compared with DSGAT, MGPred, SDPred, and the model proposed by Galeano et al. The results indicate that KAN-PROSPECT-ADR maintained stable predictive performance across the three evaluation scenarios, with AUPRC values remaining consistently high relative to the baseline models. In contrast, several comparative models showed larger performance fluctuations when evaluated under few-shot and zero-shot conditions, suggesting that KAN-PROSPECT-ADR may possess improved adaptability for imbalanced prediction tasks involving sparse or previously unseen labels.

Spatial distribution of natural product embedding analysis

Before incorporating natural product data into the training process, based on the Benchmark Dataset, embeddings were extracted from the training processes of the drug ATC prediction model and the drug ADR prediction model and compared them with the benchmark models that demonstrated the strongest generalization performance, ATC-CNN and DSGAT.

For the ATC prediction model, using ATC-CNN as the control, drugs were grouped according to their ATC classification codes and the top three categories with the largest sample sizes were selected as representative examples (see supplementary for details). As shown in Fig. 6A–I, the embeddings generated by KAN-PROSPECT-ATC generally exhibited higher intra-class similarity than inter-class similarity, together with relatively clearer category clustering trends in the embedding space. Although partial overlap between similarity distributions remained, statistical analyses demonstrated significant differences between intra-class and inter-class cosine similarity distributions (Supplementary Tables S11–S12), suggesting that KAN-PROSPECT captured category-related representation patterns more effectively than the baseline models. In contrast, the category boundaries produced by ATC-CNN were relatively indistinct. These findings suggest that KAN-PROSPECT-ATC tends to preserve category-related structural representations more consistently than the comparison models, rather than indicating completely separable embedding distributions. A similar embedding analysis was performed for the ADR prediction model, using DSGAT as the control. As shown in Fig. 6J–R, the embeddings generated by KAN-PROSPECT-ADR displayed a tendency toward stronger intra-class aggregation and relatively improved inter-class separation compared with DSGAT, confirming the superior representational performance of our model. The corresponding P-values and T-statistics for cosine similarity between drugs are provided in supplementary.

Fig. 6.

Feature space distribution and cosine similarity analysis of representative ATC and ADR categories. (A–C) shows the scatter plots of drugs in the top three categories (Anti-infectives for systemic use, Antineoplastic and immunomodulating agents and Nervous system) based on the number of drugs for KAN-PROSPECT-ATC, with the target category represented in red and the other categories in pink. (D–F) shows the scatter plots of drugs in the top three categories based on the number of drugs for ATC-CNN, with the target category represented in blue and the other categories in grey. (G–I) shows the visualization of drug similarity between KAN-PROSPECT-ATC and ATC-CNN in the corresponding drug categories, with the red and pink box plots representing intra-class similarity and inter-class similarity for KAN-PROSPECT-ATC and blue representing intra-class similarity for ATC-CNN. (J–L) is the same as (A–C), showing the scatter plots of drugs in the top three categories (Vaginal discharge, Hypothyroidism, and Cardiac failure congestive) based on the number of drugs for KAN-PROSPECT-ADR, with the target category represented in red and the other categories in pink. (M–O) is the same as (D–F), showing the scatter plots of drugs in the top three categories based on the number of drugs for DSGAT, with the target category represented in blue and the other categories in grey. (P–R) is the same as (G–I), showing the visualization of drug similarity between KAN-PROSPECT-ADR and DSGAT in the corresponding drug categories, with the red and pink box plots representing intra-class similarity and inter-class similarity for KAN-PROSPECT-ADR and blue representing intra-class similarity for DSGAT (see supplementary for details).

After incorporating natural product data into training, embedding quality was further evaluated on the Transfer Dataset. Compounds were categorized according to ATC and ADR coding systems, and the top three largest classes were analyzed. Cosine similarity-based clustering showed that KAN-PROSPECT maintained higher intra-class similarity than inter-class similarity, with clear but not strictly separable category structures (Supplementary Information).

Overall, these results suggest that KAN-PROSPECT learns more consistent category-related representations in embedding space. However, embeddings are not fully separable across categories, which is expected given structural overlap among drugs and natural products. The integration of GAT, GCN, and KAN modules, together with transfer learning, further enhances representation learning and supports cross-domain generalization between drugs and natural products. Although median differences in cosine similarity appear moderate in boxplots, statistical tests still reveal significant distributional shifts, likely because large-scale pairwise comparisons capture subtle but robust structural patterns not fully visible in low-dimensional visualizations.

External test and prediction of natural products effects and adverse reactions

External test

Ten natural products were randomly selected from the Transfer Dataset as external test samples, including compounds with known ATC and ADR annotations, as shown in Fig. 7A. The ATC prediction set comprised Spectinomycin, Cannabidiol, Mitomycin, Melatonin, Acetylcholine, Vinblastine, and Papaverine hydrochloride, while the ADR prediction set included Spectinomycin, Cannabidiol, Mitomycin, Melatonin, Norepinephrine, Dronabinol, and Dactinomycin. As shown in Fig. 7B,C, KAN-PROSPECT achieved strong predictive performance, with AUC and AUPR values exceeding 0.90 for ATC prediction and reaching or exceeding 0.89 for ADR prediction. Notably, four compounds (Spectinomycin, Cannabidiol, Mitomycin, and Melatonin) shared both ATC and ADR annotations, and their predictions remained consistently accurate, indicating stable dual-task performance. The detailed prediction accuracy proportions for each natural product are provided in supplementary.

Fig. 7.

External dataset prediction and biological mechanism analysis of nifedipine and lovastatin. (A) shows the names of natural products in 10 external test datasets, corresponding to the names of natural products with both ATCs and ADRs; (B) shows the AUC values for predicting the ATCs and ADRs of natural products in the external test datasets and (C) shows the Aupr values for predicting the ATCs and ADRs of natural products in the external test datasets; (D, E) show the false positive prediction results that exist in the negative dataset during the prediction of external test datasets, and analyze the new effect prediction of artificial and synthetic drug nifedipine, with (D, F) being the pathway enrichment analysis and molecular mechanism analysis of nifedipine, (F) created with BioRender, respectively; (E, G) show the new adverse reaction analysis of natural product drug lovastatin, with (E, G) being the pathway enrichment analysis and molecular mechanism analysis of lovastatin, (G) created with BioRender.

During prediction, several samples with true negative labels were predicted as positive by the model. Given the incompleteness of current biomedical databases, these cases may represent either false positives or previously unreported biological associations. To further explore potential biological relevance, exploratory analyses were conducted on nifedipine and lovastatin as representative compounds. These analyses were intended for hypothesis generation rather than validation.

For nifedipine, the model predicted a novel correspondence with the ATC code B01AA03, with its targets significantly enriched in multiple coagulation and immune regulation⁸⁵, related pathways in Fig. 7D. This observation shows partial functional consistency with pathways related to B01-class drugs, although no direct experimental evidence was provided in this study to confirm the predicted association. Moreover, its known target NR1I2 overlaps with that of warfarin⁸⁶, suggesting potential functional convergence at the receptor level in Fig. 7F.

For lovastatin, the model predicted a potential adverse reaction corresponding to ADR code 15.05.05.001, which involves muscle weakness, pain, and atrophy. Mechanistically, lovastatin inhibits HMG-CoA reductase and the mevalonate (MVA) pathway, thereby reducing cholesterol biosynthesis⁸⁷ in Fig. 7G. Further molecular analysis revealed significant target enrichment in immune regulatory and chromatin modification pathways in Fig. 7E. These observations are partially consistent with previous reports describing statin-associated myopathy or rhabdomyolysis⁸⁸, which are typically reversible upon drug withdrawal. Experimental evidence also supports that statins can act on related pathways through DNA methylation and gene expression regulation.

Overall, these case studies suggest that KAN-PROSPECT can generate biologically plausible hypotheses for both pharmacological effects and adverse reactions of natural products and related compounds. However, these predictions should be interpreted as exploratory computational inferences rather than experimentally validated conclusions, and require further experimental and clinical verification.

Effects and adverse reaction prediction based on structural classification and similarity

The natural products obtained from the MCE database were categorized into 16 structural classes, encompassing a total of 14,232 compounds (Fig. 8A; database link provided in the Supplementary Information). Using KAN-PROSPECT, systematic predictions of both ATC classifications and ADR profiles were performed. As shown in Fig. 8E,F, the average positive prediction ratios were approximately 0.60 for ATC effects and 0.47 for adverse reactions, indicating balanced predictive outputs across different structural classes.

Fig. 8.

Large-scale prediction of natural product effects and adverse reactions based on structural classification and similarity. (A) shows the structural classification categories of natural products. (B) shows the molecular structure of icaritin. (C, D) display the proportions of natural products from the NPASS database grouped according to their structural similarity values. (E, F) present the proportions of positive predictions for natural products classified by structure using the KAN-PROSPECT model, while (G, H) show the proportions of positive predictions for natural products grouped by structural similarity using the KAN-PROSPECT model.

Based on the NPASS database, 63,629 natural products with varying structural similarity to compounds in the training and transfer learning datasets were further analyzed. The distribution of similarity scores across predefined intervals is shown in Fig. 8C,D. Prediction results across these intervals Fig. 8G,H showed that ATC-related positive ratios remained stable at 0.600–0.612, while ADR-related ratios ranged from 0.472 to 0.480. Since large-scale natural product datasets generally lack experimentally validated labels, these values represent the proportion of predicted positive associations rather than conventional accuracy metrics, and are used here to assess the stability of model outputs across different similarity levels.

Overall, KAN-PROSPECT produced consistent prediction distributions across both structural categories and similarity intervals, suggesting robust behavior across diverse chemical spaces and indicating that its performance is not overly dependent on compounds highly similar to those in the training set.

Prediction and analysis of icaritin effects and adverse reactions

Icaritin is an active constituent derived from Epimedium. It is documented in traditional Chinese medicine literature to have recorded connotations of tonifying the kidney and Yang, strengthening muscles and bones, and eliminating wind-dampness^89–92. Modern pharmacological investigations have further confirmed its anti-tumor, anti-inflammatory, and immunoregulatory bioactivities^93–96. Nevertheless, many newly discovered pharmacological actions and potential adverse reactions of icaritin have not yet been formally annotated with standardized ATC or ADR classification codes⁹⁶. The molecular structure of icaritin is presented in Fig. 8B.

Based on the model output distribution, threshold values of 0.6 for effect prediction and 0.4 for adverse reaction prediction were applied, as shown in Fig. 9A,B. The results indicate that the predicted ranges of icaritin’s effects and adverse reactions met these criteria. KAN-PROSPECT not only accurately identified icaritin’s traditional pharmacological effects, but also revealed novel predicted associations in Fig. 9A, including anti-tumor(ATC: L01), anti-inflammatory(A07E), and immune-modulatory(L04A) pathways⁹⁴. Further mechanistic analysis suggested that icaritin may exert its effects through the suppression of inflammatory mediators (IL-6, TNF-α signaling) and remodeling of the tumor immune microenvironment, particularly via PD-L1 regulation⁹⁷. Several predicted associations were partially consistent with findings reported in recent experimental and clinical studies on icaritin, including studies related to antitumor activity, immune regulation, and pathway modulation, thereby providing preliminary support for the biological plausibility of the predicted results. However, these findings should still be interpreted as exploratory computational inferences rather than direct experimental validation^98–100.

Fig. 9.

Molecular structure of icaritin predicted pharmacological effects and adverse reactions by KAN-PROSPECT. (A, B) present the predicted pharmacological effects and adverse reactions of icaritin, respectively. In (A), Top 20, Top 50, and Top 100 indicate the ranking of predicted pharmacological effects for icaritin. Colored markers denote the predicted active ATC categories within each range, while blue squares within the Top 100 represent predictions that match the known therapeutic effects of icaritin. (B) displays the predicted adverse reactions of icaritin, where red indicates higher-ranked predicted adverse reactions compared with those shown in other colors.

In terms of adverse reaction prediction, the KAN-PROSPECT model identified multiple potential adverse events in Fig. 9B, including diarrhea, anorexia, fatigue, vomiting and elevated transaminase levels, which were largely consistent with those reported in clinical trials and post-marketing surveillance (coverage rate ≈ 86%). Notably, the model did not predict severe or fatal adverse events, such as cardiotoxicity or severe hepatotoxicity, aligning with existing pharmacological safety evaluations¹⁰¹, thereby demonstrating the model’s potential value for preclinical safety assessment¹⁰². In addition to recapitulating known adverse reactions, the model also predicted several previously unreported events, such as mild rash and neurofatigue syndrome, which may represent potential long-term or combination-related toxicities¹⁰³. These findings highlight the model’s sensitivity to subtle safety signals and suggest that such predictions could guide future experimental validation and pharmacovigilance efforts.

Discussion

Natural products play an important role in drug discovery and repositioning; however, their ATC classification and ADR prediction remain challenging due to limited data availability, severe class imbalance, missing labels, and the reliance of existing methods on task-specific or overly complex multi-source features. These issues reduce model generalization and limit applicability to natural products with sparse annotations. To address these challenges, we developed KAN-PROSPECT, a unified framework for simultaneous prediction of pharmacological effects (ATC) and adverse drug reactions (ADR), which also supports transferable pre-training for drug-related prediction tasks.

KAN-PROSPECT primarily relies on SMILES-based molecular representations, reducing dependence on multi-omics and network-based features. The integration of GAT, GCN, Conv1D, and KAN modules enables the model to capture structural, contextual, and nonlinear molecular relationships in a unified architecture. In addition, a transfer learning strategy is employed to leverage large-scale drug knowledge and adapt it to natural product prediction, thereby alleviating data scarcity in this domain.

It should also be noted that KAN-PROSPECT-ATC and KAN-PROSPECT-ADR contain approximately 10.6 million and 31.8 million parameters, respectively. Although these parameter scales are relatively large compared with the number of unique compounds, model training was performed on large-scale compound–ATC and compound–ADR association pairs, substantially increasing the effective number of supervised samples. Similar high-capacity architectures have been widely applied in recent deep learning studies for biomedical prediction tasks requiring complex nonlinear representation learning^69,104. Nevertheless, relatively large parameter scales may still increase the risk of overfitting, particularly under limited external validation settings and long-tail label distributions. Therefore, the current independent validation results should primarily be interpreted as proof-of-concept evidence rather than definitive confirmation of broad generalization capability. Future studies involving larger multi-source natural product datasets and broader external validation will be necessary to further assess the robustness and practical applicability of the framework.

Across external validation and large-scale similarity-based evaluations, the model exhibited stable predictive behavior. When applied to natural products from MEC and NPASS, KAN-PROSPECT produced consistent prediction trends across different structural categories and similarity partitions, indicating robustness across diverse chemical spaces. The observed prediction distributions (approximately 0.600 for ATC effects and 0.470 for ADR) provide a reference range for large-scale computational screening. In some cases, the model predicted positive associations for samples annotated as negative, which may reflect incomplete biomedical annotations where some unlabeled pairs correspond to unreported but biologically meaningful relationships.

Case studies on compounds such as icaritin further demonstrated that the model can recover known pharmacological effects and ADRs while also suggesting additional potential pathway-level associations, partially consistent with literature reports, supporting its utility for hypothesis generation in natural product research.

Despite these encouraging results, several limitations remain. First, negative samples were constructed from unobserved associations due to the lack of experimentally validated negative data, which may introduce false negatives and bias model training, although this is a common issue in association prediction tasks. Future work may benefit from more biologically informed strategies such as PU learning, semi-supervised learning, or confidence-aware negative sampling. Second, although transfer learning improves generalization, some zero-shot results may still arise from residual structural similarity between training and test compounds; thus, these findings should be interpreted as evidence of potential rather than universal zero-shot capability. Third, the current natural product dataset remains limited in size and annotation quality, and external validation was performed on a relatively small number of compounds, making the results primarily preliminary rather than fully generalizable. Fourth, the model does not incorporate patient-level clinical information, cross-species variation, or dynamic biological processes such as drug–drug interactions and temporal toxicity changes, which may limit its translational applicability.

In summary, KAN-PROSPECT provides a unified and relatively lightweight computational framework for ATC and ADR prediction of natural products, achieving competitive performance across benchmark comparisons, zero-shot evaluations, and exploratory case studies. The results suggest its potential utility in natural product pharmacology, drug repositioning, and early-stage safety screening. Future work will focus on expanding dataset scale, improving biological validation, integrating richer clinical and molecular context, and enhancing applicability in real-world drug discovery scenarios.

Supplementary Information

Below is the link to the electronic supplementary material.

Supplementary Material 1 ^{(1.1MB, docx)}

Author contributions

DNZ: Funding acquisition, methodology, project administration, supervision, writing—review and editing. ZSD: Data curation, software, validation, visualization, methodology, writing–original draft, writing—review and editing. ZJW, YC, BYL, XW, TYR, and HWC: Visualization, writing—original draft. LL and QJ: Writing—original draft. YLZ, YNZ and JGB: Writing—review and editing. XKR, XJC and HBX: Methodology, supervision, writing–review and editing.

Funding

This study was supported by the National Natural Science Foundation of China (62072144), the Heilongjiang Postdoctoral Initiation Grant (LBH-Q20159) and the Heilongjiang Provincial Education Science 14th Five-Year Plan Key Project (Grant No. GJB1425364).

Data availability

The datasets analyzed during the current study are publicly available. Additional processed data supporting the findings of this study are available from the corresponding author upon reasonable request.

Declarations

Competing interests

The authors declare no competing interests.