ProjFusNet: deep neural network for peptide precursor prediction using projection-fused protein language model and structural features

Abstract

Peptide precursors, as the source molecules of bioactive peptides, play essential roles in neuroregulation, immune defense, and drug development. Their accurate identification is crucial for elucidating mechanisms of life regulation and developing novel therapeutics. However, the complexity and diversity of peptide precursor sequences pose significant challenges to prediction tasks. Existing methods predominantly rely on sequence features or structural features, hindering the full exploitation of complementary information between modalities and consequently limiting prediction performance. We introduce ProjFusNet, a deep learning framework that integrates evolutionary-scale protein sequence representations from ESM-2 with structural features via a projected multimodal fusion strategy. A bidirectional LSTM is further employed to model the complex interactions between sequence and structure. In rigorous five-fold cross-validation, ProjFusNet demonstrates improved performance across key metrics, including ACC, SN, AUC, SP, and MCC, compared to single-feature models.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13321-025-01117-x.

Scientific contribution

This study introduces ProjFusNet, a novel deep learning framework to integrate evolutionary-scale sequence representations (ESM-2) and protein structural features via a projected multimodal fusion strategy for peptide precursor prediction. This approach overcomes the limitations of existing methods that rely solely on single modalities, enabling the effective exploitation of complementary information and achieving higher accuracy and robustness across key metrics. By providing a powerful tool for the efficient and accurate identification of peptide precursors, ProjFusNet has the potential to contribute to neuropharmacology research and peptide-based drug development. Future work should focus on further validating the model’s generalizability and exploring its application in broader contexts.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13321-025-01117-x.

Introduction

Bioactive peptides are short-chain polymers of amino acids that play key roles in regulating fundamental physiological processes across all domains of life [1–5]. Derived from precursor molecules, these peptides serve as critical mediators in various biological functions, including neuromodulation [6], immune defense [7, 8], and pharmacological applications [9]. They exert their effects by binding to specific target receptors. For example, peptides derived from proenkephalin activate μ-opioid receptors to modulate neural signaling [10, 11]. Cathelicidin precursors are processed to yield LL-37, which neutralizes endotoxins and maintains immune homeostasis [12, 13]. Proglucagon processing products help regulate metabolic balance [14, 15]. The functional maturation of bioactive peptides critically depends on the precise proteolysis of their precursor molecules [16–19]. Consequently, identifying peptide precursors is a fundamental prerequisite for unraveling regulatory mechanisms and developing targeted therapeutic strategies.

Traditional biochemical methods for detecting peptide precursors, including bioactivity assays [20], biochemical characterization [21], genomics-based comparisons [22, 23], and proteomics-driven mass spectrometry [24, 25], continue to play an essential role in experimental validation. However, these methods are characterized by high costs (over $500 per sample) and are also time-consuming (requiring 2–3 weeks per analysis) [26]. These efficiency limitations have driven the development and adoption of artificial intelligence (AI) methods, particularly based on sequence-based machine learning approaches [27–33]. Pioneering tools like NeuroPID (2014) [34, 35] were followed by advancements such as NeuroPP (2019) [36] and NeuroRipp (2019) [37]. These methods have progressively optimized feature engineering to better capture complex, nonlinear patterns within precursor sequences.

Despite these advancements, sequence-based approaches have several limitations. First, reliance on handcrafted sequence features inherently restricts their ability to model essential high-order structural properties critical for precursor function. Second, their predictive sensitivity markedly decreases for precursors with low sequence homology, defined as sequence similarity below 30% [38]. Most critically, existing methods largely neglect the integration of protein structural information [39–42]. However, integrating evolutionary semantics derived from sequences with structural information offers significant potential for substantially improving peptide precursor identification [43]. Therefore, developing an integrated framework that effectively harmonizes sequence, structural, and evolutionary information represents a pivotal strategy for overcoming current limitations and advancing peptide precursor identification.

To address the limitations of traditional peptide precursor identification methods, we present ProjFusNet, a novel deep learning framework designed for accurate peptide precursor prediction. ProjFusNet integrates evolutionary-scale sequence representations derived from ESM-2 (evolutionary scale modeling) [44] with key structural features, such as solvent accessibility and local structural motifs, through a novel projection-based multimodal fusion strategy. This strategy projects heterogeneous feature representations into a unified latent space, enabling the effective fusion of sequence and structural information. A bidirectional LSTM (BiLSTM) network [45, 46] is then employed to capture complex, nonlinear interdependencies between the fused sequence-structure representations. The BiLSTM excels at modeling long-range dependencies, critical for capturing spatially distant yet functionally linked sequence-structure patterns [47, 48]. Through rigorous evaluation, ProjFusNet outperforms single-feature models across multiple metrics, demonstrating superior generalization capability and robustness. ProjFusNet serves as a powerful high-throughput screening tool for identifying cryptic precursor genes, thereby accelerating the discovery of novel bioactive peptides. Additionally, it can facilitate the de novo design of engineered precursors featuring optimized cleavage efficiency for next-generation peptide therapeutics [49]. By enabling accurate and efficient large-scale identification of peptide precursors, ProjFusNet offers significant promise for accelerating research in neuropharmacology and advancing the development of peptide-based drugs.

Materials and methods

Dataset construction

We constructed a high-quality dataset of peptide precursors (positive samples) and non-precursor proteins (negative samples) protein sequences to support robust model training. The initial data were sourced from the UniProt database (Release 2025_03) [50]. For the positive samples, specific search criteria were applied to identify proteins annotated as precursors of bioactive peptides, including those containing precursor, signal peptide, and propeptide. Only manually reviewed (Swiss-Prot) entries were selected to ensure reliability. For negative samples, we employed a complementary search strategy. Specifically, we excluded entries annotated as precursor, signal peptide, or propeptide. And we only selected manually reviewed (Swiss-Prot) entries lacking these annotations to represent biologically relevant non-precursor proteins. Then, both positive and negative samples were subjected to the same filtering steps to improve quality and diversity: (1) removing sequences containing non-standard amino acids; (2) excluding sequences longer than 1000 residues; (3) applying CD-HIT (using an 85% sequence identity threshold) to reduce redundancy. Although CD-HIT effectively removed highly redundant sequences (> 85% identity), we further assessed potential moderate homology between the training and testing sets. A large-scale similarity analysis on 450,000 random sequence pairs revealed that 10.11% exhibited sequence similarity $\ge$ 30%, a threshold known to challenge traditional methods.

To avoid potential model bias arising from differences in sequence length, we performed length-stratified sampling on the negative set to match its length distribution to that of the positive set. (1) Analyzing the length distribution of the positive samples; (2) Randomly sampling negative sequences proportionally to align their length distribution with the positives. The resulting Kernel Density Estimation (KDE) plot (Fig. 1) demonstrates the high similarity in length distribution between the final positive and negative sets, with both sets showing primary peaks around 100–200 residues and secondary peaks near 400–500 residues, confirming effective balancing.

Fig. 1.

KDE (Kernel Density Estimation) Plot of Sequence Lengths for Positive and Negative Sets

The final curated dataset consisted of 7526 samples, comprising 3763 positive samples and 3763 negative samples. To rigorously assess model performance, we adopted a two-tier data splitting strategy. Initially, the dataset was randomly divided into a training set (80%, 6020 samples) and an independent testing set (20%, 1506 samples), ensuring an equal class distribution in both subsets. Subsequently, we applied fivefold cross-validation on the training set. In each fold, 80% of the training set was utilized for training, while the remaining 20% served as the internal validation set. For each fold, the best model was selected based on its performance on the internal validation set during training. The performance metrics reported in this study, including accuracy (ACC), area under the receiver operating characteristic curve (AUC), and Matthews correlation coefficient (MCC), represent the average results obtained from the best model in each fold on the independent testing set.

Comprehensive analysis of the dataset’s structural integrity was conducted across three dimensions [40] (sequence length distribution, amino acid proportional composition, and amino acid counts by sample class), with separate visualizations for training and validation sets (Fig. 2). The sequence length distribution demonstrated effective balancing between positive and negative samples in both data partitions, showing highly similar range characteristics and peak alignments. Analysis of amino acid proportional composition revealed that residue frequency rankings within the positive class were identical between training and validation sets, and similarly identical within the negative class, with specific hydrophobic residues consistently dominant and certain rare amino acids minimally represented across all sets.

Fig. 2.

Dataset Characterization: Sequence length distribution; Amino acid proportional composition; Amino acid counts by sample class. Left column: Training set (n = 6020). Right column: Validation set (n = 1506)

Model architecture

ProjFusNet employs a dual-branch architecture comprising four core components as illustrated in Fig. 3. The processing pipeline initiates with simultaneous feature extraction from both modalities: ESMFold processes input sequences to generate predicted tertiary structures while ESM-2 concurrently extracts evolutionary-scale contextual embeddings. These modality-specific representations are then independently projected into a unified latent space: structural descriptors through a fully connected layer with ReLU activation, batch normalization (BatchNorm), and Dropout regularization, while sequence embeddings are similarly projected to matching dimensionality. The fusion module concatenates these aligned structural and sequence representations. This unified feature vector is then processed by the classification head, where a bidirectional LSTM (BiLSTM) layer captures sequence dependency, followed by a fully connected layer with ReLU and Dropout, and finally passing through the output layer to achieve binary classification. The projection mechanism ensures complementary feature alignment while BatchNorm and Dropout enhance model stability for peptide precursor identification.

Fig. 3.

The workflow of ProjFusNet. Overview of the ProjFusNet model for identification of peptide precursors. The model process can be divided into three parts: (1) Feature Extraction: extract structural features and evolutionary-scale contextual embeddings; (2) Feature Projection and Fusion: project heterogeneous features into a unified dimensional space and concatenate aligned representations; (3) Classification: process the unified feature set and output classification results

Feature extraction module

The feature extraction module serves as the cornerstone of ProjFusNet, with its core objective being to extract rich and complementary feature representations from input protein sequences. This module employs a dual-branch architecture, leveraging advanced protein language models to capture sequential patterns and tertiary structural information, respectively, fully highlighting the importance of these two dimensions in protein functions and interactions.

As shown in Fig. 4, the ESM-2 branch employs a deep Transformer architecture (specifically, the ESM-2 model with 650 million parameters) to process the raw amino acid sequence and generate a high-dimensional residue embedding (S_feat) [51]. The model is pre-trained on a large-scale protein sequence dataset and captures long-range dependencies and evolutionarily conserved patterns through a self-attention mechanism. The sequence features it outputs cover biochemical properties, functional semantics, and evolutionary relationships, providing rich sequence-level information for subsequent analysis.

Fig. 4.

Feature Extraction architecture of the ProjFusNet model. The figure shows the feature extraction module of the ProjFusNet model, which uses a dual-branch structure to process the input protein sequence. The right branch uses the ESM-2 model to generate sequence embeddings (S_feat), which are interpreted as representing information about evolutionary conservation, biochemical properties, and functional semantics. The left branch uses the ESMFold model to generate structural embeddings (Struc_feat), which are interpreted as representing information about spatial constraints, solvent accessibility, and topological features

Meanwhile, the ESMFold branch first predicts the tertiary structure of proteins, outputting atomic coordinates (PDB files). Structural features are then programmatically extracted from these predicted structures to form the structural embeddings (Struc_feat). These structural embeddings encode the spatial constraints between residues, the local structural environment (such as secondary structure propensity and solvent accessibility), and global topological features, thereby forming a geometric representation of protein conformation. The introduction of the ESMFold branch enables the model to have an in-depth understanding of proteins from the perspective of spatial conformation. To evaluate the confidence of ESMFold-predicted structures used in our study, we analyzed the per-residue predicted local distance difference test (plDDT) scores for all structures in both training and testing sets. The plDDT score, ranging from 0 to 100, provides a confidence estimate for each residue’s structural prediction, with higher scores indicating greater reliability. The overall confidence statistics for both datasets are summarized in Table 1. The analysis of these plDDT scores reveals a moderate to good level of confidence in the overall structural predictions, with mean scores of approximately 77 for both datasets. The distributions are consistent between the training and testing sets, indicating no significant data bias. We observe that 40% of the residue structural prediction confidences fell in the high range (plDDT $\ge$ 90), providing a reliable structural basis for feature extraction.

Table 1.

Confidence statistics of ESMFold-predicted structures based on plDDT scores

Metric	Training Set	Testing Set
plDDT (Mean $\pm$ SD)	76.86 $\pm$ 21.17	77.86 $\pm$ 20.33
Median plDDT	85.81	86.50
plDDT $\ge$ 90 (High Confidence)	40%	40.8%
70 $\le$ plDDT $\le$ 90 (Medium Confidence)	28.5%	29.5%
plDDT $<$ 70 (Low Confidence)	31.6%	29.8%

This dual-path independent processing mechanism fully utilizes the advantages of ESM-2 in sequence modeling and the expertise of ESMFold in structure prediction. In this way, the model constructs a dual perspective covering protein sequence evolution and spatial conformation, providing a solid foundation for subsequent feature fusion.

Training Configuration: The ESMFold and ESM-2 feature extraction modules were utilized in a frozen state, providing static embeddings for each input sequence. All subsequent modules (the projection layers, fusion mechanism, and the BiLSTM classifier) were trained on the dataset. This approach allows the model to learn how to best align and combine the generated representations from the pre-trained models for accurate peptide precursor prediction without modifying the foundational features themselves.

Feature projection and fusion module

The feature projection and fusion module is the core innovative component of ProjFusNet, which is responsible for the unified representation and deep integration of the sequence feature (S_feat) from the ESM-2 branch and the structural features (Struc_feat) from the ESMFold branch. As shown in Fig. 5, this module projects heterogeneous features into a unified dimensional space through an independent linear transformation layer. Specifically, the 1280-dimensional sequence features are compressed to 512 dimensions through a fully connected layer, and the 256-dimensional structural features are also mapped to 512 dimensions through similar processing, thereby eliminating the dimensional differences of the original features. The projection process uses the ReLU activation function to enhance the nonlinear representation capability, and introduces batch normalization and Dropout (0.2) regularization to improve the generalization of the model. The projection layer formula is as follows:

$$pro{j}_{\mathit{seq}}=Dropout\left(BatchNorm\left(ReLU\left(Linear\left(S_{\mathit{feat}}\right)\right)\right)\right)$$ 1

$$pro{j}_{\mathit{struct}}=Dropout\left(BatchNorm\left(ReLU\left(Linear\left(Stru{c}_{\mathit{feat}}\right)\right)\right)\right)$$ 2

where Linear denotes the fully connected layer, BatchNorm represents the batch normalization layer, ReLU represents the rectified linear unit activation function, and Dropout represents the Dropout regularization.

Fig. 5.

Feature Projection and Fusion Mechanism. The module adopts a dual-path architecture: one path compresses the 1280-dimensional sequence features of ESM-2 to 512 dimensions through a projection layer (linear transform + ReLU + BatchNorm + Dropout), and another path performs the same projection on the 256-dimensional structural features of ESMFold. The bimodal features are connected to form a 1024-dimensional unified representation and finally input into the LSTM classifier to realize peptide precursor prediction

After feature projection, we concatenate the two features in the same dimensional space to achieve information and form a 1024-dimensional fusion feature vector (as shown in Formula 3). This fusion feature vector completely retains the evolutionary semantics of the sequence and the geometric constraint information of the structure, establishing a complete feature foundation for protein function prediction.

$$F_{\mathit{fused}}=\left[pro{j}_{\mathit{seq}}\oplus pro{j}_{\mathit{struct}}\right]$$ 3

where $pro{j}_{\mathit{seq}}$ and $pro{j}_{\mathit{struct}}$ are the sequence features and structural features after projection processing, respectively.

This projection-splicing mechanism effectively combines multimodal features from ESM-2 and ESMFold branches to produce a 1024-dimensional unified feature representation for the LSTM classifier. It automatically aligns sequence and structural features through implicit spatial mapping and learnable parameter matrices, which not only preserve key sequence evolution semantics and structural details but also enhance feature representation through nonlinear activation and splicing operations, thereby preventing information loss. This design streamlines feature processing, lowers computational costs, and boosts the accuracy of protein function prediction while increasing the model’s flexibility.

Classification module

The classification module uses a two-layer bidirectional long short-term memory (LSTM) network to process the 1024-dimensional fused feature vector generated by the feature projection layer. LSTM is a recurrent neural network architecture suitable for processing sequence data, which can capture complex temporal patterns and long-distance dependencies in high-dimensional feature representations while maintaining stable regularization.

Figure 6 shows the workflow of the classification module. After feature fusion, the input sequence is first encoded by a bidirectional LSTM with two hidden layers, each with a hidden dimension of 128. This process obtains a hidden state for each time step. Then, the hidden state of the last time step is selected as the global representation of the sequence, which integrates information in both the positive and negative directions, so its dimension is 256:

$$h_{\mathit{last}}=\left[{\begin{array}{c}\to \\ h\end{array}}_{\mathit{last}},{\begin{array}{c}\leftarrow \\ h\end{array}}_{\mathit{last}}\right]$$ 4

where ${\begin{array}{c}\to \\ h\end{array}}_{\mathit{last}}$ and ${\begin{array}{c}\leftarrow \\ h\end{array}}_{\mathit{last}}$ are the last hidden states of the forward and backward passes, respectively.

Fig. 6.

Classification module diagram

The representation passes through a fully connected layer, which includes a linear transformation, a ReLU activation function, and a Dropout regularization, with the Dropout probability set to 0.2. These operations further reduce the dimension of the feature to 128 dimensions to improve the model’s ability to handle nonlinear relationships and enhance its generalization performance to unseen data:

$$F_{\mathit{fc}}=ReLU\left(Linear\left(h_{\mathit{last}}\right)\right)$$ 5

$$F_{\mathit{dropout}}=Dropout\left(F_{\mathit{fc}},p=0.2\right)$$ 6

Finally, the output layer converts the 128-dimensional features into logits of two categories:

$$y_{\mathit{logits}}=W_{\mathit{out}}·F_{\mathit{dropout}}+b_{\mathit{out}}$$ 7

In the model training phase, the cross-entropy loss function is used to optimize the network parameters. This ensures that the model can learn to distinguish features of different categories:

$$\mathcal{L}=-\sum _i{y}_{i}log\left({\widehat{y}}_i\right)$$ 8

where $y_i$ is the true label, and ${\widehat{y}}_i$ is the predicted probability.

In the inference phase, the final category probability distribution is obtained through the Softmax function to achieve accurate classification prediction of the sample.

$${\widehat{y}}_i=\frac{\text{exp}\left(y_{logits,i}\right)}{{\sum }_j\text{exp}\left(y_{logits,j}\right)}$$ 9

To prevent overfitting, the model integrates a Dropout layer with a Dropout rate of 0.2 and an early stopping strategy with a patience value of 5 epochs. Through these methods, the risk of the model on the training set is effectively reduced, while improving the model’s generalization ability to new data.

Evaluation metrics

In this study, we have conducted an exhaustive performance evaluation of the ProjFusNet framework using a comprehensive set of five diverse metrics: Accuracy (ACC), Sensitivity (SN), Specificity (SP), the Matthews Correlation Coefficient (MCC), and the area under the Receiver Operating Characteristic curve (AUC) [52, 53]. These metrics offer a comprehensive assessment of peptide precursor identification, collectively evaluating various critical aspects of classification performance to ensure a robust understanding of the framework’s effectiveness. The formulas for calculating ACC, SN, SP, MCC are provided below [54–57]:

$$\begin{array}{c}ACC=\frac{TP+TN}{TP+TN+FP+FN}\end{array}$$ 10

$$\begin{array}{c}SN=\frac{\mathit{TP}}{TP+FN}\end{array}$$ 11

$$\begin{array}{c}SP=\frac{\mathit{TN}}{TN+FP}\end{array}$$ 12

$$\begin{array}{c}MCC=\frac{TP\times TN-FP\times FN}{\sqrt{\left(TP+FN\right)\left(TP+FP\right)\left(TN+FP\right)\left(TN+FN\right)}}\end{array}$$ 13

where true positive (TP) represents the number of correctly predicted positive samples (e.g. correctly identified peptide precursors), true negative (TN) represents the number of correctly predicted negative samples (e.g. correctly identified non-peptide precursors), false positive (FP) refers to the number of negative samples incorrectly predicted as positive, and false negative (FN) represents the number of positive samples incorrectly predicted as negative.

Results and discussion

In this study, we conducted a comprehensive performance evaluation of the proposed ProjFusNet model, verifying its reliability and generalization capabilities through a series of experiments. To ensure consistent performance across different data subsets, we employed a rigorous fivefold cross-validation approach. The experimental results show that ProjFusNet demonstrates superior performance across multiple key metrics, outperforming existing baseline models. Specifically, ProjFusNet achieved an average ACC of 95.43%, an AUC of 0.9819, and an MCC of 0.9087. Furthermore, the model achieved an SP of 95.19% and SN of 95.67%, indicating its well-rounded capacity to accurately identify both positive and negative instances.

Ablation studies on feature sources and fusion strategies

In this section, we conducted ablation experiments to clarify the contribution of each feature source and fusion strategy to the overall performance of the ProjFusNet model. These experiments employed a unified training and verification process to systematically evaluate the impact of different feature sets and fusion strategies.

Multimodal feature ablation

To evaluate the contributions of each feature modality and the impact of multimodal fusion on ProjFusNet’s performance in peptide precursor identification, we conducted systematic feature ablation experiments using fivefold cross-validation. These experiments employed the same dataset, model architecture, training and validation procedures, and evaluation metrics to ensure fair comparisons. We assessed the contributions of structural features, ESM features, sequence features, and the multimodal fusion features of ProjFusNet to the model’s performance.

Table 2 presents the key performance metrics of each configuration on the validation set. The fusion of ESM-2 embeddings with structural features achieves the best performance, significantly surpassing single modality. When using only ESM embeddings, the ACC is only 65.07% and the SN is only 56.31%. This shows that when only using the language model, it is easy to miss difficult-to-judge samples. Structural features alone achieved near-random performance (ACC = 53.92%, MCC = 0.0787), indicating weak standalone discriminative ability. Multimodal fusion improved ACC by 30.36 percentage points and MCC by 0.6025 over ESM-only baselines. These results demonstrate that the sequence information captured by the language model is complementary to the structural features, and false positives are distinguished by structural features. Multimodal fusion effectively leverages this complementarity to significantly expand the model’s decision boundary.

Table 2.

Performance Evaluation of Multimodal Feature Ablation

Feature Type	ACC (%)	AUC	SN (%)	MCC	SP (%)
ESM + Structure	95.43	0.9819	95.67	0.9087	95.19
ESM	65.07	0.6407	56.31	0.3062	73.84
Structure Only	53.92	0.5608	49.00	0.0787	58.83
Sequence Only	61.29	0.6679	60.03	0.2258	62.55

Figure 7 visualizes the learned feature representations of the validation set samples using the Uniform Manifold Approximation and Projection (UMAP) dimensionality reduction technique. Each sub-figure corresponds to a distinct input feature configuration evaluated in the ablation study. The visualizations provide compelling support for the quantitative findings in Table 2. When only sequence features or only structural features are used, samples exhibit high dispersion with substantial overlap between positive and negative classes. This blurred decision boundary directly correlates with the high misclassification rates observed quantitatively. Although only using ESM features offers improved cluster coherence compared to using only sequence or structural, the centers of the positive and negative clusters remain relatively close. Samples near the decision boundary are susceptible to misclassification. In contrast, the fusion of ESM-2 and structural features yields the most distinct separation. Positive and negative samples form well-defined, separated clusters in the embedding space. This visual evidence directly substantiates the boundary expansion brought by complementarity, explaining the significant performance improvement.

Fig. 7.

UMAP visualization of features on the validation set

These results conclusively demonstrate the effectiveness of multimodal feature fusion in ProjFusNet. Furthermore, they reveal the complementary relationship between different modal features. ESM features perform well in capturing deep sequence semantic information, while structural features provide critical conformation-specific discriminative power and have an advantage in distinguishing false positives. The combination of these modalities not only improves the overall performance of the model but also enhances the model’s ability to identify complex or challenging peptide precursors.

Ablation study on fusion mechanisms

Building upon the validated efficacy of multimodal feature fusion, we further investigated how different feature integration strategies impact model performance. An ablation study was conducted comparing three fusion approaches under identical experimental conditions: Projection Fusion (ProjFusNet’s core methodology), Cross-Attention Fusion, and Feature Concatenation. All strategies were evaluated using consistent training/validation procedures, hyperparameters, and assessment metrics to ensure comparability.

As shown in Table 3, projection fusion demonstrated superiority across all evaluation metrics. It achieved excellent performance by effectively modeling complex cross-modal interactions. In contrast, feature concatenation showed predictive ability close to chance, reflecting severe model bias. Cross-attention fusion showed intermediate competency, but still not as good as projection fusion. The dominance of projection fusion originates from its high-dimensional alignment mechanism, which preserves modality-specific semantics while eliminating feature redundancy. In addition, parametric projection matrices maximize complementary information exchange between ESM embeddings and structural characterization. This dual-alignment mechanism enhances robustness and superiority in peptide precursor identification. These findings suggest that the feature integration strategy plays a critical role in the ProjFusNet framework.

Table 3.

Performance evaluation metrics of different fusion strategies

Strategy	ACC (%)	AUC	SN (%)	MCC	SP (%)
Projection Fusion (ProjFusNet)	95.43	0.9819	95.67	0.9087	95.19
Cross-Attention Fusion	86.61	0.9264	84.99	0.7340	88.23
Concatenation	50.17	0.5025	79.84	0.0135	20.50

Interpretable analysis of multimodal feature contributions

To elucidate the decision logic of ProjFusNet and understand the relative contributions of structural and sequential features, we employed SHAP (Shapely Additive exPlanations) value analysis based on cooperative game theory. This approach quantifies feature importance through a unified global and local interpretability metric, revealing how multimodal features collectively drive predictions.

The global feature importance ranking is shown in Fig. 8, which lists the top 20 features by mean SHAP value. Notably, ESM-derived features (e.g., Esm_f234 and esm_f1160) dominate the highest ranks, indicating that evolutionary sequence information plays a decisive role in driving model predictions. These features capture conserved biochemical and functional semantics across homologous sequences, providing strong discriminative signals for precursor identification. Crucially, structural features also contribute significantly, with struct_f41 appearing among the top 20 features. This underscores the complementary role of structural information, which appears to act as a spatial filter to reduce false positives caused by sequence-based ambiguity by incorporating structural constraints inaccessible to sequence-only models.

Fig. 8.

Global Feature Importance Ranking

The SHAP summary plot (Fig. 9) provides a detailed view of how individual features influence the model’s predictions. The top ESM-derived sequence features (e.g., esm_f234, esm_f1160) show the broadest distribution of impact, with their points spanning both highly positive and highly negative SHAP values. This bidirectional impact highlights their role in capturing complex, context-dependent evolutionary semantics that can significantly support classification, depending on the specific sequence context. In comparison, the structural feature struct_f41 exhibits a different and more directional pattern. An observable trend indicates that higher values of struct_f41 tend to be associated with negative SHAP values. In comparison, its lower values show a slight tendency to be associated with neutral to positive SHAP values. This moderate negative trend suggests that struct_f41 often functions as a contributory factor against a precursor classification. The model may have learned that elevated values for this structural descriptor are often indicative of a structural context that is less commonly associated with peptide precursors.

Fig. 9.

SHAP Summary Plot: Top 20 Features for ProjFusNet

Moreover, the feature synergy map (Fig. 10) reveals strong correlations between key ESM features (e.g., esm_f234 and esm_f1160, r = 0.69), and moderate correlations between ESM and structural features (e.g., esm_f234 and struct_f41, r = 0.21). This indicates that the model leverages both independent and synergistic information from both modalities: ESM embeddings provide the foundational evolutionary context, while structural features refine these predictions by incorporating spatial and conformational constraints.

Fig. 10.

Multimodal Feature Synergy Map

In summary, ProjFusNet fully leverages a division of labor between its feature modalities: ESM-2 sequence embeddings provide the primary signal for precursor identification, while structural features act primarily as a filter to exclude non-precursor proteins. This multimodal fusion strategy effectively combines these complementary approaches to achieve high accuracy and low false rates.

Comparative evaluation with benchmark methods

To comprehensively evaluate the generalization ability and advantages of ProjFusNet, we conducted comparative analyses from multiple perspectives. Specifically, we compared it with state-of-the-art methods, common deep learning architectures, and classical machine learning and deep learning models using fixed feature sets. To ensure robustness and comparability of the results, all experiments employed fivefold cross-validation.

Comparison with state-of-the-art methods

ProjFusNet was evaluated against the leading peptide precursor identification tools: NeuRipp [37] and NeuroPP [36]. NeuRipp is a Paralle-CNN-LSTM hybrid model. NeuroPP is an SVM [58, 59] model based on handcrafted amino acid composition (AAC), dipeptide composition (DPC), and tripeptide composition (TPC) features.

For a fair comparison with state-of-the-art methods, each benchmark model was evaluated using its most appropriate and author-recommended approach. In the case of NeuRipp, whose official source code was available, we rigorously retrained the model from scratch on our training set using the author’s provided code and default configurations before evaluating the resulting model on our independent testing set. In contrast, for NeuroPP, where no training source code was provided and only a prediction tool was available, we adhered to the methodology outlined in the original publication by utilizing the authors’ official prediction tool to generate predictions for sequences in our independent testing set. These strategies ensured that each model was evaluated under its intended and optimal conditions, thereby providing an equitable and realistic comparison.

As shown in Fig. 11, the experimental results on the unified validation set demonstrate that ProjFusNet significantly outperforms all network variants of NeuRipp with an accuracy of 95.43%. The best variant, LSTM, only reaches 92.30%, and the traditional method, NeuroPP, is only 57.55%. The performance gap is due to NeuRipp’s static feature aggregation not being able to model long-range residue dependencies, while NeuroPP’s artificial features lack cross-domain generalization abilities. ProjFusNet’s multimodal projection fusion dynamically aligns ESM embeddings with structural representations, effectively overcoming the fusion bottlenecks of heterogeneous features.

Fig. 11.

Accuracy Comparison with State-of-the-Art Methods

General deep learning benchmarking

We conducted a detailed comparison of ProjFusNet with several common deep learning architectures, including DNN, LSTM-Attention, BiLSTM, and Transformer.

Although these architectures are widely used in various deep learning tasks, they exhibit different performance profiles when applied to peptide precursor classification. As shown in Fig. 12, ProjFusNet outperforms the others. It shows especially strong results in ACC, reaching 95.43%, which is 4.63 percentage points higher than the second-best model, BiLSTM (90.80%). Additionally, ProjFusNet outperforms in other key metrics: SN, SP, AUC, and MCC, with improvements of 7.97 percentage points, 1.39 percentage points, 0.0289, and 0.0907, respectively. Its ability to combine local structural motifs and overall sequence semantics via multimodal fusion explains its advantage over sequence-only models.

Fig. 12.

Comparison of ProjFusNet and Baseline Deep Learning Models

Performance comparison under fixed feature set conditions

To evaluate the architectural advantages of ProjFusNet, we compared ProjFusNet with 10 benchmark models using ESM + structural features. The benchmark models include four deep learning models (Transformer, GRU, BiLSTM_Attention, MLP) and six classic machine learning models (AdaBoost, KNN, Logistic Regression, Gradient Boosting, Random Forest, SVM). By controlling the variables, we evaluated the classification performance of different algorithms under the same dataset features to accurately measure the advantages and disadvantages of various models. As shown in Table 4, ProjFusNet outperforms other models in all five key performance indicators.

Table 4.

Performance under Fixed Feature Set

Base Model	ACC (%)	SN (%)	SP (%)	AUC	MCC (%)	Precision (%)
ProjFusNet	95.43	95.67	95.19	0.9819	90.87	95.22
AdaBoost	86.26	85.61	71.50	0.9360	72.60	86.80
KNN	88.47	89.63	77.20	0.9506	76.70	87.61
LR	91.15	91.23	82.10	0.9626	82.10	91.09
GBM	90.80	89.20	79.80	0.9655	81.70	92.15
RF	91.56	88.80	82.50	0.9701	83.20	94.01
SVM	92.77	90.50	85.0	0.9747	84.60	94.82
Transformer	92.72	92.79	92.66	0.9704	85.54	92.77
GRU	93.60	93.69	93.52	0.9763	87.22	93.54
BiLSTM_Attn	93.62	93.75	93.49	0.9734	87.28	93.53
MLP	93.92	93.59	94.25	0.9779	87.87	94.24

As shown in Fig. 13, compared with other deep learning (DL) models, ProjFusNet achieves the highest accuracy (95.43%) among all deep learning models. It improves ACC by 1.51 percentage points and the AUC by 0.004 compared with the suboptimal MLP model. The difference between recall (SN) and precision for ProjFusNet is only 0.45 percentage points, which is much smaller than the 0.65 percentage points observed in the MLP model. Although the Transformer model is designed for temporal modeling, its ACC (92.72%) is still lower than MLP, highlighting the advantage of ProjFusNet’s multimodal feature fusion.

Fig. 13.

Deep Learning Model Comparison under Fixed Feature Set

Figure 14 shows that ProjFusNet outperforms all traditional machine learning (ML) models when using the same feature set. Although SVM’s ACC (92.77%) approaches that of some deep learning models, its SP is 10.19 percentage points lower than ProjFusNet’s. This could result in many more false positives in large-scale screening. Traditional models like SVM and Random Forest also have larger gaps between recall and precision, while ProjFusNet achieves a better balance between sensitivity and specificity through multimodal feature fusion.

Fig. 14.

Machine Learning Model Comparison under Fixed Feature Set

To quantitatively assess the statistical significance of the performance advantage demonstrated by ProjFusNet, we conducted rigorous hypothesis testing based on the results from fivefold stratified cross-validation. We compared ProjFusNet against the best-performing deep learning (MLP) and machine learning (SVM) benchmark models across all key performance metrics. The performance distribution across folds for each model and metric is visually summarized in Fig. 15 (see Supplementary Materials). Statistical significance was evaluated using independent samples t-tests for normally distributed metrics (ACC, F1) and the non-parametric Mann–Whitney U test for non-normally distributed metrics (AUC), as determined by Shapiro–Wilk normality tests. Effect sizes were quantified using Cohen’s d to assess the practical significance of observed differences.

Fig. 15.

Comparison of Model Performance Distribution. Box plots show the distribution of ACC, F1, and AUC for the proposed ProjFusNet model and the best-performing benchmark model (MLP and SVM) across the five independent test folds. The central line in each box represents the median, and the box extends from the first to the third quartile (IQR), and the whiskers show the range of the data. The mean value for each distribution is indicated by a black dot

As shown in Tables 5 and 6, ProjFusNet demonstrated statistically significant performance improvements over both machine learning and deep learning benchmarks. It significantly outperformed the best ML model (SVM) across all metrics, with highly significant improvements in ACC and F1 (p < 0.001) and significant improvement in AUC (p = 0.008). While the AUC improvement over MLP demonstrated a large effect size (d = 1.66), it did not reach conventional statistical significance (p = 0.151), which may be attributed to the higher variance observed in MLP’s AUC scores across folds. The consistently large Cohen’s d values (all > > 0.8) provide strong evidence that ProjFusNet’s performance advantages represent both statistical and practical significance.

Table 5.

Comparison between ProjFusNet and the best ML model (SVM)

Metric	ProjFusNet (Mean $\pm$ SD)	SVM (Mean $\pm$ SD)	p-value	Cohen’s d
ACC (%)	95.43% ± 0.18%	92.77% ± 1.00%	< 0.001	3.29
F1	0.9544 ± 0.0017	0.9260 ± 0.0108	< 0.001	3.28
AUC	0.9819 ± 0.0005	0.9747 ± 0.0032	0.008	2.86

Table 6.

Comparison between ProjFusNet and the best DL model (MLP)

Metric	ProjFusNet (Mean $\pm$ SD)	MLP (Mean $\pm$ SD)	p-value	Cohen’s d
ACC (%)	95.43% ± 0.18%	93.92% ± 0.86%	0.009	2.18
F1	0.9544 ± 0.0017	0.9390 ± 0.0088	0.009	2.19
AUC	0.9819 ± 0.0005	0.9779 ± 0.0030	0.151	1.66

Limitations and future work

ProjFusNet represents an advance in peptide precursor prediction through its multimodal fusion approach, yet several limitations warrant consideration to properly contextualize these findings and guide future research. The main limitations concern dataset composition and methodological comparisons. Although the CD-HIT filtering at 85% identity effectively reduced high sequence redundancy, approximately 10.11% of sequence pairs between the training and testing sets still share $\ge$ 30% similarity. This residual homology suggests that while our results demonstrate substantial improvements, they may not fully represent the model’s performance on entirely novel, low-homology sequences. Additionally, the current dataset, while high-quality, could benefit from expanded diversity across peptide precursor families to enhance broader applicability. Another consideration arises from the ablation study on fusion strategies, where differing parameter counts between approaches mean that model capacity differences, rather than fusion mechanisms alone, could partially account for the performance outcomes.

To address these limitations, there are several directions for future work. First, we will create a more rigorous benchmark using advanced tools like MMseqs2 to enforce a strict sub-30% similarity threshold between training and testing sequences, enabling a more accurate assessment of generalization capability. We will also expand the diversity of training data to improve model robustness across biological contexts. To ensure fair comparisons between fusion strategies, we will implement controlled experiments with matched parameter capacity across all models, isolating the effect of the fusion mechanism itself. Furthermore, we will validate structural features by comparing the ESMFold predictions with experimentally determined structures to confirm conformational reliability. Finally, to enhance practical utility, we will develop methods for residue-level interpretability to connect predictions with biological mechanisms and design computationally efficient architectures suitable for large-scale screening applications.

Conclusions

This study presents ProjFusNet, a novel deep learning framework for accurate peptide precursor prediction by integrating ESM-2 and structural features through a projection-based multimodal fusion strategy. ProjFusNet achieves competitive performance (ACC 95.43%, AUC 0.9819) under rigorous fivefold cross-validation.

Feature ablation studies systematically confirmed the importance of multimodal synergy, combining structural features with evolutionary information. Fusion strategy ablation experiments demonstrated that the projection fusion strategy effectively improved model performance. The bidirectional LSTM architecture effectively captured the complex interdependencies between sequence and structure, outperforming traditional machine learning and other deep learning models.

ProjFusNet shows superior performance in peptide precursor prediction by effectively integrating multimodal features and using advanced fusion strategies. It provides a powerful tool for peptide precursor identification and is expected to promote neuropharmacology research and peptide drug development. Future work will extend the fusion architecture to functional peptide families.

Author contributions

Jinjin Li proposed the basic idea and designed the framework. Jinjin Li and Fang Fang performed the experiments. Hua Shi and Feifei Cui wrote the manuscript. Leyi Wei supervised the project and revised the manuscript. Zilong Zhang and Changhang Lin performed the visualization of the experimental results and analysis.

Funding

The work was supported by the Internal Research Grants of Macao Polytechnic University (No. RP/CAI-02/2023 and RP/CAI-01/2025) and the Science and Technology Development Fund of Macao (No. 0177/2023/RIA3).

Data availability

The datasets and source code for ProjFusNet are publicly available at the GitHub repository: [ProjFusNet] (https:/github.com/JinArry/ProjFusNet/tree/main). This repository includes the curated dataset in FASTA format, Python scripts for structure prediction, feature extraction, model training, and a README file.

Declarations

Competing interests

The authors declare no competing interests.