High-quality training data is crucial for deep learning models, but the availability of known PPI data between plants and pathogens is very limited. Therefore, training models on known protein interaction datasets to predict PPIs in new host-pathogen systems becomes particularly important. In this study, to improve the accuracy of the model while preserving the biological significance to the greatest extent, protein-related data of pine nematode and host pine as well as PPI data verified by biological experiments, were selected to build a dataset suitable for training a pine nematode-pine PPIs prediction model.

First, the raw protein sequence data for both the pine wood nematode and pine tree, which served as the basis for subsequent construction of protein 3D structures and extraction of sequence features, was obtained. The protein sequence data for the pine wood nematode was gained from the NBIC database (Taxonomy ID: 6326), totaling 53,412 sequences. The protein sequence data for the pine tree was obtained from UNIPROT(https://www.uniprot.org), totaling 200,806 sequences. The genome annotation project database for pine trees was TreeGenes (https://www.treegenesdb.org). To minimize potential errors from raw protein sequence data, sequences containing short protein sequences (e.g., lengths less than 50 bp) and homologous sequences were removed.

The experimentally validated protein-protein interaction data between the pine wood nematode and its host pine tree was gained from the PHI-base (https://www.phi-base.org) database.

PPI data for both the pine wood nematode and pine tree were queried from the STRING database. Specifically, there were 6,231 interaction pairs for the pine wood nematode and 2,009 interaction pairs for the pine tree. All interaction data selected here were physical interactions, excluding weak and transient interactions. The interaction data between the two were analyzed and compared, and the overlapping data was regarded as interspecific interaction. Finally, combining data obtained from PHI-base, 8,259 protein-protein interaction pairs were constructed using 4,792 proteins, including 5,258 positive reference datasets. It is worth noting that protein interactions can take various forms, only direct physical interactions were considered as positive data. Non-interacting proteins cannot be directly obtained from PPI databases, so proteins with no interactions in the PPI network were marked as negative samples not recorded in the PPI dataset.

AlphaFold can obtain the representation information of 3D structure through protein sequences. Therefore, AlphaFold DB would be used to view the structural information of protein interactions generated based on the above rules, and structure prediction would be made for protein sequences lacking structural information to generate PDB files to establish a protein structure database for PPI prediction of pine wood nematode disease.

According to the structural graph of protein network interactions, an interaction between two proteins was marked as 1, otherwise as 0. There were 8,259 positive and negative samples in total, which were divided into training, testing sets in an 8:2. The validation set was used to determine the optimal parameters for the model, and the trained model with the determined optimal parameters was used for training. To improve model performance, different proportions of positive and negative samples were set, and 1/2 of the positive (negative) samples from the training set were mixed with the negative (positive) samples from the training set to enhance the model’s generalization capability under imbalanced sample conditions.

An end-to-end deep learning framework MFGA-PPI was introduced for identifying plant-pathogen PPIs. Here, we take the sequence and structure information of the two proteins as input and define it as a binary classification problem, with the final output being a set of 0 or 1 predictions as to whether they interact or not. The overall architecture of the MFGAC-PPI model is shown in Figure 1 . This model architecture consisted of three parts: the feature extraction module, the feature aggregation module, and the prediction module. In the first part, to better represent protein structures, both tertiary and primary structure information and design feature extraction modules were used for each. The second part used a linear interpolation method to effectively combine the two feature vectors, achieving a multidimensional protein representation. The third part inputed the resulting protein pairs into an attention network, calculated attention scores, and used them for PPI prediction.

In this study, Intel(R) Xeon(R) Gold 6354 CPU @ 3.00GHz×2, NVIDIA GeForce RTX 3090×4 graphics processor, 256.0GB memory, 19TB MR9364-8i storage were used, and operating system was Ubuntu 18.04.4 LTS.

In the process of model training, 5-fold cross-validation was adopted. For each fold training set, 1000 samples were randomly selected for training. For each selected sample, the characteristics and labels of the sample were obtained. In this experiment, the Adam optimizer was used to optimize the model. The initial value of the learning rate was 1 × 10⁻³, and the learning rate attenuated to half of the original value after every 10 rounds of training to prevent the model from jumping out of the optimal solution. In this model, three layers of GCN neural networks are used, and the embedding dimension of GCN in each layer was 20, which reduces the computational overhead on the premise of ensuring sufficient feature information capture. The feature dimension of the protein sequence captured by ESM was 1280 dimensions. Binary cross-entropy is used as the loss function of the model in this paper, and the formula is as follows:

where, y is the true label and p is the probability that the model predicts 1. y log(p) represents the loss when the true label is 1 and (1 − y) log(1 − p) represents the loss when the true label is 0.

In the experiment, eight common indicators were used to evaluate the model: Precision, Recall, Accuracy, Specificity, F1-score, Matthews Correlation Coefficient (MCC), Area Under the P-R Curve(AUPRC) and Area Under the ROC Curve (AUROC). Their relevant definitions are as follows:

The basic framework of the MFGAC-PPI model mainly consisted of four submodules: ESM, improved GCN, Scaled dot-product attention, and FNN. According to the aforementioned hyperparameter settings, each submodule was treated as a different variable and the “control variable method” was used to investigate the influence and contribution of each submodule to the proposed model. Therefore, 5-fold cross-validation was used to conduct ablation studies on these four structures, six indicators were used for evaluation, and the maximum value of the results of each ablation was selected, as shown in Table 1 . It can be seen that, when any of the sub-modules was ablated, the overall prediction performance of the model was degraded, indicating that each submodule played a role, and the structural design was reasonable without structural redundancy. Similarly, it was found that, the model performance decreases the most among the evaluation metrics when the improved GCN module was ablated, especially specificity and MCC, which dropped by 1.73% and 2.19%, respectively. In contrast, when the FNN module was ablated, the model performance changed the least across all metrics, with precision only dropping by 0.01%, and the largest change being in MCC, which decreased by 1.39%.

From the results, it was evident that the improved GCN module contributed the most to the overall model and significantly affected its prediction performance. In contrast, the FNN had the least impact on the predictive ability of the model.

To assess the importance of each sub-module more intuitively and comprehensively, the results of the five-fold cross-validation in each fold of the six evaluation metrics are presented here using box plots, as shown in Figure 2 . It can be visually observed from the figure that when the improved GCN was ablated, the model showed the lowest average values and the lowest outliers across all metrics. The model without FNN showed the most stable performance in the five-fold cross-validation, with the highest average values across all metrics. However, in precision and specificity, it had nearly the same values as the model without the attention module, but the model without attention had lower dispersion points. In summary, each sub-module had an important contribution and influenced the model’s effectiveness to some extent. Among them, the improved GCN contributed the most to the model’s overall performance, followed by the ESM module, both of which reflected the importance of feature fusion to some extent. Then there was attention and FNN, with FNN having the smallest impact on the overall model.

The model constructed in this study differed from the GAT module by introducing a Scaled dot-product attention to capture subtle differences within proteins and their interaction interfaced while retaining more interpretable biological characteristics, which helped to identify amino acid residues with important functions, and had better robustness when dealing with proteins of different lengths. To validate that the introduction of the scaling factor improved the predictive ability of the model, as well as to explore the impact of different attention mechanisms on the model, three new models were added to the original structure: self-attention replacing scaled dot-product attention, mutual-attention replacing scaled dot-product attention, and multi-head-attention replacing scaled dot-product attention. Keeping other sub-modules and hyperparameters unchanged, these models were subjected to ablation experiments using 5-fold cross-validation, and evaluated using three comprehensive metrics: specificity, F1, and MCC for evaluation. Table 2 shows the maximum values in the cross-validation.

To clearly illustrate the variation of each fold value in cross-validation, radar charts of Specificity, F1, and MCC on the validation dataset were plotted as shown in Figure 3 . It can be seen that the model using mutual-attention had the smallest area in the metrics, while the MFGAC-PPI model using the scaled dot-product attention mechanism had the largest area in the three comprehensive metrics, indicating a significant reduction in false positive (FP) samples and an increase in true positive (TP) samples. Moreover, it was apparent from Figures 3A, C that the Specificity indicator’s area for the multi-head-attention model was much larger than that for the self-attention model, but the area in MCC was the opposite.

This may be due to the different focus of the two attention mechanisms when processing input data. Multi-head-attention captured different feature subspaces through multiple heads, performing better on negative sample features, while self-attention uses a single attention head to compute attention globally, focusing on optimizing overall performance, thus performing better in F1 and MCC metrics.

Combining Table 2 and Figure 3 , it was found that the applicability of the attention modules, from highest to lowest, was scaled dot-product attention > self-attention > multihead-attention > mutual-attention.

Additionally, different λ values were tested to evaluate the contribution of sequence and structural features to the model during feature fusion. During model training, the λ value for feature fusion in the constructed MFGAC-PPI model was adjusted while keeping other parameters constant, and the results are shown in Table 3 . It can be seen that when the λ values were set to 0.5 and 0.7, the performance of various evaluation metrics was the optimal, especially when λ=0.7, F1 and MCC reached the best.

The constructed MFGAC-PPI model were compared with five classic PPI prediction models: DeepPPI (Du et al., 2017), PIPR (Chen et al., 2019), Struct2Graph (Baranwal et al., 2022), AFTGAN (Kang et al., 2023), and TAGPPI (Song et al., 2022), and the results were shown in Table 4 . The proposed model exhibited the best performance across accuracy, precision, recall, specificity, and F1, although MCC was slightly lower than that of Struct2Graph. TAGPPI’s performance was close to the best across all metrics, particularly specificity, which was only 0.0017 lower than MFGAC-PPI. The sequence-based models PIPR and DeepPPI showed relatively low performance in MCC but maintained a relatively balanced performance across other metrics. Meanwhile, AFTGAN showed poor performance on the constructed dataset, significantly lagging behind other methods in key metrics such as F1 and recall, which were 0.1588 and 0.1555 lower than MFGAC-PPI, respectively, indicating a notable deficiency in recognizing positive samples.

To further assess the robustness and generalization capability of the model, the proposed model was compared with five other competitive algorithms using the original dataset and an independent dataset, ara data (Zheng et al., 2023). The AUROC and P-R curves for these models on both datasets were calculated, as shown in Figures 4 , 5 .

As shown in Figure 4A , in the constructed original dataset, the model’s AUC area was the largest, reaching 0.95, which was 0.03, 0.09, 0.13, 0.17, and 0.19 higher than TAGPPI, Struct2Graph, PIPR, DeepPPI, and AFTGAN, respectively. Additionally, the analysis of the P-R curve in Figure 5A indicated that MFGAC-PPI and Struct2Graph performed the best. The results were consistent with the performance of various metrics in Table 4 , where the top three models in overall performance on the constructed dataset were MFGAC-PPI, TAGPPI, and Struct2Graph, respectively. As shown in Figure 4B , when training on the independent dataset, the AUC of MFGAC-PPI dropped by 0.03, whereas TAGPPI increased by 0.03, making TAGPPI to be the best performance in AUROC. The AUROC curves of the other models remained relatively stable, similar to the results on the original dataset, although Struct2Graph shows a slight decline. In addition, the analysis of Figure 5B showed that the performance of MFGAC-PPI on the P-R curve in the ara data dataset declined slightly, while Struct2Graph exhibited the best performance.

In general, MFGAC-PPI demonstrated the best performance in predictive ability and comprehensive metrics, but its performance declined when tested in different data sets, indicating that its generalization ability had room for improvement.

In actual biological systems, most proteins interact with specific proteins rather than random combinations. Therefore, only a small fraction of protein pair combinations has true interactions, especially when PPI-related data is scarce, and the imbalance between positive and negative samples in protein interactions can reach a ratio of 1:100 or more. To verify the superiority of the proposed model in the case of unbalanced samples, the MFGAC-PPI model was trained using the constructed dataset with balanced (1:1) to unbalanced (1:3, 1:5, 1:10, 1:20, 1:30) data, using precision, recall, and MCC as evaluation metrics to illustrate the advantages of the proposed model in imbalanced datasets.

The results are shown in Table 5 , the MFGAC-PPI model showed better performance in various indexes in unbalanced data sets. Although the performance of each index decreased with the increase of the degree of imbalance, the precision and MCC remained above 91%. At the same time, recall remained relatively stable, indicating that the model can correctly identify positive samples even under highly imbalanced conditions. Overall, the MFGA-PPI model was robust, but its overall predictive performance decreased with increasing unevenness, especially affecting precision and MCC.

The optimized plant-pathogen PPI prediction model was used to predict 20,765 sets of pine nematode and pine tree protein interaction data, and the full prediction results are shown in Supplementary Table 1 . In particular, protein pairs with interaction scores greater than 0.5 were judged to have interactions, and a total of 2,688 pine-pine wood nematode protein-protein interaction networks were generated, which contained 46 pine wood nematode proteins and 354 pine proteins, with the results shown in Supplementary Table 2 . Among the predicted pine-pine wood nematode PPIs results, 16 predicted PPIs were validated by previous biological experiments. Meanwhile, the comparison of different models revealed that about 19% or so of the PPIs could be derived from sequence-based prediction models, and there existed about 21% of the interaction pairs that could be predicted by structure-based methods, and there was a high degree of overlap between these two parts of the PPIs. These results indicated that a multi-dimensional feature fusion approach can effectively uncover new PPI data and was successfully applied in the study of the pine wilt disease system.

A topological analysis of the predicted PPI network revealed that pine-pine wood nematode PPIs exhibited scale-free properties similar to other biological networks. Notably, pine nematode proteins had more interaction links than pine proteins, with one pine nematode protein able to interact with an average of four pine proteins and at least 20 pine wood nematode proteins interact with over 10 pine proteins each. Here the nodes with higher degrees were analyzed for centrality, and the results are shown in Supplementary Figure 1 , and a PPI network diagram was drawn as shown in Figure 6A . From Supplementary Figure 1 and Figure 6A , it can be seen that the proteins with the most links are the effector protein BxSap1 of pine nematode as well as the autophagy gene BxATG16, which played crucial roles in the virulence of the pine wood nematode. This was followed by A0A1l7SCF8 BURXY, Bx tlp 1 of pine wood nematode and P41649.2 protein of pine. This result suggested that potentially pathogen-associated proteins were more involved than resistance-associated proteins in the pine nematode system. Surface representations of the 3D structures of the interacting proteins are shown in Figures 6B, C , demonstrating that MFGAC-PPI can efficiently predict regions of plant-pathogen protein interaction residues, illustrated by the surface interaction between effector proteins SapB3 (blue) and P41649.2 (green), with the interaction region in red. By revealing 3D structural analysis and protein surface interaction regions, it helped understand molecular communication between pine trees and pathogens, explaining how pathogens evaded or suppressed the host immune response and effectively invaded host cells.