DeepPlantAllergy: deep learning for explainable prediction of allergenicity in plant proteins

Abstract

Allergy is an immune response triggered by specific peptides recognized by immune system effectors. While several bioinformatics tools have been developed to predict protein allergenicity, most rely on hand-selected features and lack interpretability. Improved predictive and explainable models are needed, especially for under-studied plant allergens. We present DeepPlantAllergy, a deep learning model that combines Convolutional Neural Networks (CNNs), Bidirectional Long Short-Term Memory (BiLSTM) networks, and Multi-Head Self-Attention (MHSA) to capture both local patterns and long-range dependencies within protein sequences. We evaluated four embedding techniques—including one-hot encoding, SeqVec, ProtBert, and ESM-1B—and employed Integrated gradients to identify residues contributing to allergenicity. Predictive performance was similar for ESM-1B and ProtBert embeddings, with no statistically significant difference, with an F1 score of 93.9% and 93.6% and AUC of 97.74% and 97.8%, respectively. Motif extraction revealed complementary strengths: ProtBert highlighted regions similar to OneHot patterns, while ESM captured distinct segments, and SeqVec identified additional regions overlapping with experimentally validated epitopes. Notably, molecular docking confirmed the biological plausibility of a predicted epitope, supporting the utility of residue-level predictions. DeepPlantAllergy thus offers both high predictive accuracy and interpretable insights, facilitating the discovery of allergenic motifs in under-characterized plant proteins. The source code, datasets used for training and evaluation, trained models, and the full pipeline for prediction and motif identification are available at the GitHub Repository: https://github.com/Lilly-dh/DeepPlantAllergy.

Graphical Abstract

Graphical Abstract.

Introduction

The rising prevalence of food allergies has driven the need for innovative tools to predict and pre-emptively identify novel allergens. Allergy has been defined as an adverse reaction to small proteins, known as allergens, encountered through the skin, ingestion, or inhalation. Though usually innocuous, allergens can elicit a reproducible immune response that can be potentially life-threatening [1]. Allergens are found across a wide range of organisms, including plants, animals, fungi, and bacteria. Among food allergens, eight major groups are recognized: milk, eggs, fish, shellfish, tree nuts, peanuts, wheat, and soy—four of which are plant-derived allergens [2]. Recent studies have shown that fruit allergy is increasingly prevalent in both children and adults [3]. In addition, allergies to peanuts and tree nuts—both plant-derived—often persist for life, unlike others that may be outgrown in adulthood [4]. This underscores the need to improve allergenicity prediction, especially for plant proteins, which are a major source of persistent food allergies.

Diagnosing food allergies remains challenging due to their high cost, limited specificity in identifying the precise allergen, and complexity of current tests—especially procedures like the Oral Food Challenge [4]. As a result, strict avoidance is often the only way to avoid allergy-related adverse, which significantly limits an individual’s dietary choices. Yet determining what to avoid is complicated by cross-reactivity: IgE antibodies sensitized to a primary allergen (e.g. birch pollen) can also react to structurally similar proteins in foods like almonds, hazelnuts, or walnuts, even without prior exposure [5].

To address these challenges, various in silico approaches have emerged. In 2001, the Food and Agriculture Organization (FAO) and the World Health Organization (WHO) defined potential allergens as proteins with either 6–8 contiguous amino acids or 35% identity over an 80-amino-acid window [2]. AllerMatch [6] implemented this rule via a sliding window and word-match strategy, although such homology-based tools often overlooked biochemical complexity. At a sequence level, allergens tend to cluster into a few protein families with shared biochemical functions, often characterized by specific linear or conformational motifs that contribute to their allergenic potential [4, 7]. Motivated by these insights, AlgPred [8] combined SVMs trained on peptide and dipeptide composition and Motif-based prediction (MEME/MAST), to predict allergenic proteins and map IgE epitopes. Dimitrov et al. [9] introduced five E-descriptors capturing key physicochemical properties, which were subsequently used to develop two alignment-free allergen prediction tools AllerTOP, based on k-nearest neighbors (kNN) and AllergenFP, a fingerprint-based method [10]. AllerCatPro2 [11] follows a structured decision pipeline: it first compares input proteins to known allergenic 3D structures, then applies the FAO/WHO sequence criteria, and includes a gluten-specific Q-repeat motif detection module, achieving high accuracy on different datasets; ProAll-D [12] trained on E-descriptors using an LSTM, achieved 91.5% accuracy. ALLERDET [13] employed Restricted Boltzmann Machines (RBM), and trained them on features extracted from pairwise alignment and achieved an accuracy of 97.26%.

Although current tools have shown promise in allergenicity prediction, they typically rely on handcrafted features like amino acid composition, motif searches, or physicochemical properties. This limits their generalizability to novel allergens and provides limited biological insight. Recently, Protein Language Models (pLMs), inspired by Natural Language Processing, have advanced biological applications such as drug discovery and function prediction [14]. These models, based on Long Short Term Memory (LSTM) or Bidirectional Encoder Representations from Transformers (BERT) architectures are trained in an unsupervised manner on large datasets, and are able to encode raw sequences into fixed-size vectors capturing rich representations of the composition and order of amino acids [14]. SeqVec, an LSTM-based model derived from ELMo and trained on unlabeled sequence data (UniRef50), captures context-specific biophysical features by processing protein sequences sequentially, making it well suited for modelling local dependencies [15]. It has been applied in Major Histocompatibility Complex class I (MHC I class) ligand prediction and protein function prediction [16]. Transformer-based models like ProtBert and ESM-1B learn broader sequence properties: ProtBert identifies spatially interacting residues and has been applied to diverse tasks such as protein structure refinement and Gene Ontology prediction, while ESM-1B encodes biochemical properties such as hydrophobicity and polarity, biophysical properties including molecular weight and charge differences, but was also able to cluster orthologous genes without evolutionary information provided [16–18].

DeepAlgPro [19] used one-hot encoded sequences as input to a CNN-MHSA (Multi-Head Self-Attention) and achieved an accuracy of 91.62% while also improving interpretability through identification of potential epitopes. AllerTrans [20] employed a straightforward neural network with three nonlinear hidden layers, reaching 94.44% accuracy by combining ESM-2 and ProtT5 embeddings. Alg-MFDL [21] integrated CNNs, transformers, and bidirectional gated recurrent units (BiGRUs) to process ESM-2, ProtT5, and amino acid composition features, achieving an accuracy of 97.3%. SEP-AlgPro [22] showed the superior discriminative power of pLM-derived features over traditional descriptors by selecting top 10 feature sets and 7 classifiers to build a stacking ensemble, achieving 95.7% accuracy on the DeepAlgPro dataset. Despite their high performance, most of these models assessed allergenicity only at the whole-protein level, without identifying specific regions or epitopes.

In this study, we sought to bridge the gap between computational prediction and biological insight. We developed a deep learning framework that integrates sequence-based embeddings—including SeqVec, ProtBert, and ESM-1B—with CNNs, BiLSTM networks, and MHSA mechanisms. This architecture not only achieves high predictive performance (Accuracy, F1 Score, MCC, Precision) but also enables the extraction of biologically meaningful motifs potentially linked to allergenicity.

Materials and Methods

Data preparation

To construct a high-quality dataset of plant-derived allergens and non-allergens, we adapted the methodology of DeepAlgPro [19] with modifications to improve sensitivity and reproducibility. MMseqs2 [23] was used instead of CD-HIT [24], and the identity threshold was lowered to 20% to enable the assessment of advanced embeddings (ProtBert, SeqVec, ESM) in capturing similarities below the twilight zone. The final dataset comprised 1228 proteins.

Positive set: Allergens sequences (n = 955) were collected from Allerbase [25] (http://bioinfo.unipune.ac.in/AllerBase/Home.html, Last Access 21-Sep-2021). To remove highly similar proteins, sequences were clustered with MMseqs2 (version 16.747c6) using at identity 85%, coverage 0.8, coverage mode 0, and sensitivity 7.0 resulting in 614 clusters. This ensured that highly similar proteins were removed while preserving biological diversity. The positive set consisted exclusively of plant allergens.

Negative set: Nonallergen candidates (n = 1817) were retrieved from UniProt (Last Access 07-Sep-2024) using the following query: NOT(taxonomy_id:9606) NOT (keyword:KW-0020) AND (reviewed:true) AND (Protein Existence: Evidence at protein level) NOT (cc_allergen:allergic) NOT “antigen” NOT “Cancer” to avoid including immunogenic features that could bias the model’s learning, with additional filters for sequence length (10–850 amino acids), annotation score (>3), and exclusion of taxonomy_id:10239 (Virus) and taxonomy_id:33154 (Homo sapiens). This yielded 1817 sequences, 87% of which belonged to Viridiplantae. Redundancy was further reduced with MMseqs2 clustering at 20% identity (same parameters as above), leaving 798 representatives. To reduce potential bias from similarity with the positive set, sequences sharing >20% identity with allergens were removed, resulting in 769 nonallergens. Approximately 20% of the final negative set consisted of nonplant proteins (bacteria, protozoa, fungi, algae), included to mitigate bias toward plant-specific features.

Balancing and splitting: The initial dataset was slightly imbalanced, containing 614 positives and 769 negatives. Both positive and negative sets were clustered at 20% sequence identity, and all sequences within the same cluster were assigned to the same split. The positive set was divided into training (70%, 431 sequences), validation (20%, 122 sequences), and test (10%, 61 sequences). The negative set was downsampled to match these sizes, maintaining a 1:1 class ratio, while longer negative sequences and any remaining sequences were set aside to ensure that the main datasets remained balanced and unbiased (Figure 1a: Dataset Preparation).

Figure 1.

Protein allergenicity prediction workflow: a schematic representation of the complete workflow, from dataset collection and preprocessing to embedding extraction, model training, final allergenicity prediction, and motif extraction.

Protein representation and feature extraction

We next explored the influence of embedding strategies on allergenicity prediction; we used four types of embeddings as input: One-Hot Encoding, SeqVec, ESM-1B, and ProtBert. The embeddings were generated using the Bio-Embeddings [26] and Bio-Transformers modules from the DeepChain library. One-hot encoding was considered as baseline, providing a simple positional encoding of sequences with no biological context. SeqVec embeddings [15] are based on LSTM with a 1024-dimension vector extracted from the last layer. Two transformer-based Models ESM-1B [17] and ProtBert [18] were also employed to encode the sequences, respectively, to 1280 and 1024 vectors. Each raw protein sequence in the dataset was passed through the selected model to obtain per-residue embeddings, where each residue is represented by a high-dimensional vector. After processing, the output vectors were padded to 1000 residues to obtain a fixed-size input vector (embedding size, length = 1000, number of sequences).

Model design and training

To address plant protein allergenicity prediction, we propose a hybrid deep learning architecture integrating CNNs, BiLSTM, and MHSA. The model processes the padded per-residue embeddings derived from each of the four representation strategies and outputs a prediction score for each sequence and an importance score for each residue, enabling both classification and interpretability. The model’s parameters were optimized via Optuna [27], with a function to reduce the gap between validation loss and training loss. The best architecture was selected based on the best performance on validation set and computational cost. Additional information is also provided in Supplementary Material 1, Section 2.

The model begins by passing the input vector through a 1D convolutional layer with 64 filters and a kernel size of 5 for short-range motifs, followed by batch normalization and 3 stacked BiLSTM layers with 128 hidden units each. At this point, the encoded protein sequence has its contextual information and local patterns extracted and is ready for the MHSA block, which weighs the importance of different residues. The MHSA block uses a single layer with eight heads (d_model = BiLSTM hidden ×2); computes scaled dot-product attention on queries, keys, and values; and concatenates the head outputs through a linear layer. The resulting sequence output is then pooled using an Adaptive Average Pooling Layer to a fixed-size vector, which is then processed through three fully connected layers. The Rectified Linear Unit (ReLU) activation function introduces nonlinearity; a 0.5 dropout is applied for regularization. To assess robustness while avoiding data leakage, each model was trained and evaluated on an independent dataset using five different random seeds. Early stopping was implemented to halt the training when the validation loss does not improve for five successive epochs. Model training employed the Binary Cross-Entropy with Logits Loss (BCEWithLogitsLoss), which integrates a sigmoid activation with binary cross-entropy and is suitable for binary classification tasks. The Adam optimizer was used with an initial learning rate of 10⁻⁵ and no weight decay. A ReduceLROnPlateau scheduler was implemented to reduce the learning rate by a factor of 0.5 after five epochs without improvement in validation loss. Proteins with prediction scores >0.5 were considered candidate allergens. The model was implemented in PyTorch 2.4.0 + cu121 and trained on an NVIDIA RTX 4060 GPU (Figure 1b: Training and Interpretation).

Architecture comparison

To understand how each part of the model contributes to its overall performance, we evaluated alternative architectures of varying complexity. Each trial involved systematically toggling off a specific module while keeping all other parameters and training conditions constant. Three trials were performed: NoCNN (without the CNN layer), NoBiLSTM (without the BiLSTM layers), and NoAttention (without the multi-head self-attention mechanism), and compared to the Full model (CNN-BiLSTM-MHSA model). This comparison allowed to assess how each component contributes to predictive accuracy and motif identification.

Performance comparison

In addition to evaluating performance on the independent test set, the model’s performance was also assessed on the DeepAlgPro [19] dataset, which is widely used in recent allergen prediction studies [12, 21, 22]. To avoid bias, the dataset was filtered for homology relative to training set at 20% using MMseqs2, retaining 1104 sequences. Our model was compared to state-of-the-art allergen prediction tools—AlgPred2.0 [28], DeepAlgPro [19], AllerCatPro 2.0 [11], and SEP-AlgPro [22]—to evaluate its predictive efficiency. Model performance was assessed using standard classification metrics: Accuracy, Precision, Recall, F1 score, Matthews correlation coefficient (MCC), and Area Under the Curve (AUC) when available. These were computed using the functions accuracy_score, precision_score, recall_score, f1_score, matthews_corrcoef, and roc_auc_score from the sklearn.metrics module.

Biological interpretation

Motif extraction and processing

To better understand the importance of each residue in the model’s final decision, Integrated Gradients [29] was applied to the model predictions. Attribution scores generated by the prediction model were analysed to identify contiguous regions of high importance (motifs). First, raw attribution scores were summed across the embedding dimension. Then, attribution values for each residue were smoothed using a moving average filter with different window sizes (5, 7, 9, 11, and 13 residues). For each smoothed attribution profile, a threshold was computed to identify significant regions. Two strategies were implemented: (i) a percentile-based cut-off, where residues above the 80th percentile of the smoothed distribution were considered important, and (ii) a midpoint cut-off, defined as the average of the minimum and maximum smoothed values. Residues exceeding the chosen threshold were grouped into contiguous stretches, and motifs were defined as regions of at least three consecutive residues. We analysed the resulting motifs to compare the effect of the smoothing window size, the contribution of embeddings, and model type.

Molecular docking analysis

Peptide–MHC docking

To structurally validate the sequence-based prediction of immunogenic peptides, candidate peptides were evaluated for their binding to HLA-DR1 molecules. Five high-resolution crystal structures of HLA-DR1 were selected from the Protein Data Bank (PDB IDs: 1KLU, 3C5J, 3PDO, 5NI9, and 6ATF). All structures were prepared and minimized in MOE (Chemical Computing Group, Montréal, Canada). Peptides were built in extended conformation and docked into the canonical peptide-binding groove using the induced fit docking protocol with 50 initial placements followed by refinement of the top five poses. For each complex, docking scores (kcal/mol), interacting residues, and binding energies after energy minimization were recorded.

IgE-Pap s 3 docking

To assess the structural accessibility of conformational epitopes, the full protein sequence was modeled using AlphaFold2. The IgE Fab fragment in complex with bovine β-lactoglobulin (PDB ID: 2R56) was used as a structural template for allergen recognition. The allergen was removed, and both Fab and allergen protein were prepared in MOE. Protein–protein docking employed the hydrophobic patch analysis protocol with automatic detection of complementarity-determining regions (CDRs). Rigid-body preplacement was performed with 100 000 poses, placement with 10 000 poses, and refinement with 100 poses, sampling the entire molecular surface. Protein–ligand interaction fingerprints (PLIFs) were used to identify key interface residues. To further refine predictions, a second docking round was carried out by restricting IgE interface residues to those experimentally validated in the 2R56 structure, with enhanced sampling parameters (preplacement 1 000 000; placement 100 000; refinement 1000).

Results

Overview of model architecture

DeepPlantAllergy is a modular framework, designed as a hybrid deep learning architecture combining convolutional layers, known for detecting local sequence motifs, with BiLSTM units that capture long-range dependencies across the protein sequence. The output of these successive layers is then processed by an attention mechanism to highlight residue-level signals critical for prediction. The CNN-BiLSTM-MHSA architecture is inspired by previous works on antimicrobial peptide prediction [30] and gene transcription enhancer identification [31], both of which have demonstrated strong performance in their respective prediction tasks. It also aligns with other studies where a CNN coupled with Attention achieved superior accuracy in peptide toxicity prediction [32], and provided valuable insight into lung cancer decision-making when coupled with Neural Networks [33].

Model performance across embeddings

Comparing embeddings and architecture on the independent test set

Model robustness was evaluated on the independent test set using five random seeds to vary initialization while keeping the split fixed. Trained on the Full model (CNN-BiLSTM-MHSA), ESM-1B and ProtBert achieved the highest predictive performance (F1 ≈ 93.9% and 93.6%; MCC ≈ 0.879 and 0.873; Receiver Operating Characteristic – Area Under the Curve (ROC-AUC) ≈ 97.7%–97.8%), SeqVec showed moderate stability (F1 ≈ 91.6%; MCC ≈ 0.834; ROC-AUC 95.2%), and One-hot encodings performed lowest (F1 ≈ 80.8%; MCC ≈ 0.617; ROC-AUC 86.2%) (Fig. 2, Supplementary Table S-1c).

Figure 2.

Bar plots showing the effect of the different embeddings and architectures on six main metrics: Accuracy, F1 Score, MCC, Recall, Specificity, and Precision.

Effect of architecture: The incremental contribution of each component (CNN, BiLSTM, and MHSA) is most evident in models trained on One-hot encodings, which exhibited pronounced drops in F1 score and Recall, with significant differences (P < .05) observed for Full versus No CNN, Full versus No LSTM, and No CNN versus No Attention. SeqVec-based models were also significantly affected in F1 score, Recall, and Specificity when the CNN was removed (P = .025, .025, and .004, respectively). This architectural change had a smaller effect on ESM-1B embeddings (P = .07) and an even smaller effect on ProtBert embeddings (P = .22). Overall, ESM-1B and ProtBert maintained stable performance across architectural variants, with no significant differences reported (P > .05). Notably, the removal of MHSA in ProtBert models highlighted the importance of CNN and BiLSTM components for specificity, with a significant drop from 0.9245 to 0.8983 and 0.9081, respectively (P = .03).

Effect of embeddings: We assessed the effect of embedding choice on model performance across metrics using paired t-tests, aligning predictions by seed. ESMb1 and Protbert embeddings exhibited no significant difference in performance metrics in Full, NoCNN, and NoLSTM models (P-value > > .05). Only when attention was removed, did ESMb1 perform significantly better than Protbert (F1 score = 93.93% versus 91.96%, P-value = .004, MCC = 0.8788 versus 0.8395, P-value = .004). While the ROC-AUC variation remained nonsignificant when attention was removed (P-value = .92), eliminating CNN and BiLSTM shifted the balance, giving advantage to ProtBert over ESM-1B embeddings, resulting in a borderline significance for NoLSTM (97.86% versus 97.28%, P = 0.049), and a pronounced significance for NoCNN (97.40% versus 96.63%, P-value = .007). Models trained on Seqvec embedding had borderline contribution to ESM in the Full model (P-value = .052), while when compared to Protbert, the removal of MHSA, led to a significant drop in performance of Seqvec. This demonstrates that Seqvec took advantage of both CNN and MHSA to score comparably to advanced embeddings like ESMb1 and Protbert. Onehot encodings performance was significantly lower than all embeddings, across different architecture variants, reflecting the limited complexity of their representation.

Across architectures, transformer-based embeddings (ESM-1B and ProtBert) consistently achieved the highest and most stable performance. SeqVec exhibited moderate stability, benefiting from the presence of CNN and MHSA, while One-hot encodings performed substantially lower across all metrics. Detailed parameter counts provided in (Supplementary Table S-1e) show that One-hot embeddings gained from the additional parameters in the Full model, compensating for their lack of contextual information. In contrast, advanced embeddings (SeqVec, ProtBert, ESM) were less affected by architectural changes. Their large dimensionality (1024–1280) caused a substantial increase in total parameters when the CNN was removed, leading to a significant decrease in F1 score for ESM and SeqVec, while ProtBert showed only a numerical, nonsignificant decrease. Removal of the BiLSTM reduced parameters by ~75%, yet none of the advanced embeddings experienced significant performance drops (P > .28). Removing the MHSA block had a minor effect on total parameters, consistent with negligible changes in statistical metrics.

Computational footprint

Training time, CPU usage, and GPU peak memory were evaluated across five seeds (Fig. 3). GPU usage was mainly driven by architectural components, with BiLSTM and MHSA contributing most; Full and NoCNN models used the most GPU memory. CPU memory depended primarily on the embeddings: ESM ≈ 6500 MB, ProtBert and SeqVec ≈ 5400 MB, and One-hot ≈ 1200 MB.

Figure 3.

Bar plots showing the effect of the different embeddings and architectures on training computational cost.

Removing attention reduced training time, while removing CNN increased it and introduced seed-to-seed variability, highlighting the efficiency contribution of convolutions. One-hot embeddings trained fastest, SeqVec slowest, particularly in Full and NoCNN models. ESM and ProtBert trained faster in Full and NoLSTM architectures than in NoCNN models.

Overall, transformer embeddings deliver high predictive performance with moderate computational cost, whereas simpler embeddings are lighter but less accurate. In some cases, simpler models required longer training than the Full model, making them less efficient in practice.

Embeddings visualization

To explore how feature representations evolve across the model’s layers, we applied Uniform Manifold Approximation and Projection (UMAP) to reduce dimensionality and visualize the input and intermediate feature spaces—specifically the outputs of the CNN, BiLSTM, MHSA, fully connected layers, and the final classification layer, as shown in Supplementary Material 1, Section 4.

In the initial embedding space, allergens and non-allergens show considerable overlap, reflecting the limited separability of raw features. This pattern was consistent across all embedding types, except for ESM-1B, which displayed a more structured distribution. After the convolutional layer, ESM-1B embeddings begin to show partial class separation, suggesting that local patterns relevant to allergenicity are being captured, while other embeddings remain dispersed. The BiLSTM layer further refines clustering, particularly for ESM-1B, ProtBert, and SeqVec, highlighting its ability to model sequential dependencies. In contrast, OneHot encodings continue to exhibit scattered distributions, underscoring their limited representational power. At the attention layer, separation between allergens and non-allergens becomes apparent for all embeddings, though to varying degrees, indicating that the models identified key sequence regions important for allergenicity. Clustering improves further through the fully connected layers, resulting in clear separation at the final output, with ESM-1B achieving the most distinct class distinction despite minor overlaps.

Model performance compared to state-of-the art tools

On independent test set

We evaluated state-of-the-art tools on our independent test set under the following settings (AlgPred 2 in Hybrid mode at a 0.5 threshold; AllerCatPro with only “Strong evidence” predictions considered positive; DeepAlgPro locally; SEP-AlgPro via the online server). Results are summarized in Table 1. SEP-AlgPro achieved the highest accuracy (94.21%), followed by AllerCatPro (90.91%) and AlgPred 2 (89.26%), and lastly by DeepAlgPro (85.95%). Our embedding-based models achieved comparable or superior performance: ESM-1b reached 93.88% accuracy, 0.8778 MCC, and 97.14% AUC, Protbert 93.55% accuracy with 94.33% recall, and SeqVec 91.57% accuracy. McNemar’s test showed no significant difference between SEP-AlgPro and either ProtBERT or ESM-1b (P = 1), while SeqVec, ProtBERT, and ESM-1b all significantly outperformed AlgPred 2 (P < .001; Supplementary Table 2). Unlike AllerCatPro, whose strong performance depends on structural information, and SEP-AlgPro, which leverages a complex ensemble of classifiers, our framework achieves competitive results using only sequence embeddings combined with deep learning architectures.

Table 1.

Comparison of the predictive performance of different embedding-based models on the independent test set^* (N = 121)

Embedding type	Acc (%)	F1 Score (%)	Pre (%)	Rec (%)	MCC	AUC (%)
Onehot	80.66 ± 2.65	80.53 ± 2.71	80.45 ± 3.1	80.67 ± 3.46	0.6137 ± 0.05	86.01 ± 2.73
SeqVec	91.57 ± 1.36	91.56 ± 1.21	91.19 ± 3.04	92 ± 1.39	0.8321 ± 0.03	95.19 ± 0.23
ESM-1B	93.88 ± 1.71	93.8 ± 1.73	94.29 ± 1.88	93.33 ± 2.04	0.8778 ± 0.03	97.74 ± 0.75
ProtBert	93.55 ± 1.48	93.55 ± 1.49	92.86 ± 2.4	94.33 ± 2.79	0.8719 ± 0.03	97.79 ± 0.42
Deep Alg Pro	85.95	86.61	82.09	91.67	0.7241	93.88
AllerCatPro	90.91	90.43	94.55	86.67	0.8209	NA
AlgPred	89.26	89.43	87.3	91.67	0.7862	NA
SEP-Alg Pro	94.21	94.31	92.06	96.67	0.8854	98.03

Note: Precision, Recall, and F1 Score are reported for Positive class, AUC was calculated when the model predictions included probabilities, and marked as NA when the model did not provide detailed scores. Our models’ performance is reported on the average across five seeds of Independent Test Set—N = 121) where one sequence was filtered out due to nonstandard amino acids. Maximum values are highlighted in bold.

On DeepAlgPro filtered dataset

The effectiveness of Transformer-based embeddings was further demonstrated when our model was evaluated on the filtered DeepAlgPro test set, originally consisting of 1420 sequences, downsized to 1104 sequences after a 20% homology filtering. Although trained exclusively on plant proteins, our model achieved 85.6% accuracy and 80.54% F1-score with ESM-1b embeddings, and 82.43% accuracy and 78.4% F1-score with ProtBERT embeddings, comparable to DeepAlgPro (85.87% accuracy, 82.39% F1-score). Results are summarized in Table 2. McNemar’s test was used to compare prediction errors between our models and existing tools (Supplementary Table 5f). While all embeddings diverged significantly from SEP-Alg-Pro, both ESM-1b and ProtBERT were nonsignificant compared to DeepAlgPro (P-value = .88 and .73) and AlgPred (P-value > .05). Additionally, ProtBERT showed comparable error rates to AllerCatPro (P-value = .19), highlighting its competitive performance. Seqvec and Onehot performance was significantly lower than all benchmark tools (P-value << .05).

Table 2.

Comparison of the predictive performance of different embedding-based models on the DeepAlgPro filtered test set (N = 1104)

Embedding type	Acc (%)	Pre (%)	Rec (%)	F1 Score (%)	MCC	AUC
Onehot	71.47	60.08	69.59	64.49	0.4116	71.08
SeqVec	81.07	72.54	79.08	75.67	0.6037	80.66
ESM-1B	85.6	81.03	80.05	80.54	0.6911	84.47
ProtBert	82.43	72.28	85.64	78.4	0.6442	83.08
Deep Alg Pro	85.87	76.84	88.81	82.39	0.7121	0.9323
AllerCatPro	93.93	97.25	86.13	91.35	0.8709	NA
AlgPred	88.41	80.69	90.51	85.32	0.7613	NA
SEP-Alg Pro	95.38	96.15	91.24	93.63	0.9009	96.64

Interestingly, SEP-AlgPro achieved the best overall performance, with 95.38% accuracy, 91.24% recall, and an MCC of 0.9009. AllerCatPro attained the highest precision (97.25%) to the detriment of a lower recall (86.13%), while AlgPred2 achieved a relatively high recall of 90.51%. The diversity of the DeepAlgPro dataset posed a challenge for our model, which was specifically designed for plant allergen prediction, unlike the models it was compared against. To further understand the performance gap, we analysed the misclassified sequences by our embeddings-based models, compared to other tools, ESM1b, Protbert, and Seqvec misclassified a few plant allergens (17, 17, and 28, respectively). Mostly, the models missed sequences from “Animal” and “Microorganism” classes, which were not included in our training dataset (Figure 4a and b).

Figure 4.

(a, b) Distribution of misclassified allergen sequences by class and model. Y-axis indicates the classification group and the x-axis for tool names. (a) shows red heatmaps for the misclassified allergens, while (b) shows blue heatmaps for misclassified nonallergens.

Identification of novel allergens and biological interpretation

We evaluated our embedding-based models on 20 novel plant allergens approved by the International Union of Immunological Societies (IUIS) between 2022 and 2025, with lengths ranging from 82 to 846. These proteins were neither in our initial dataset, nor shared >20% homology to our training set. All four embedding-based models, with and without attention, and external tools were applied to this set. Results are provided in Supplementary table 4.

The role of MHSA in prediction

Global metrics (accuracy, F1, MCC) were similar for models with and without attention (Comparing Embeddings and Architecture  on the Independent Test Set section), but per-sequence analysis revealed embedding-specific effects. For ProtBert, attention slightly reduced per-sequence scores, though recall remained high (95%), suggesting the model already captures sufficient contextual information. ESM performed better without attention, with recall rising from 0.80 to 0.95, indicating that extra attention can sometimes confuse predictions. SeqVec showed a modest recall increase from 70% to 80% when attention was removed (P = .061), reflecting mixed effects across sequences, particularly in the AF197 family. OneHot embeddings consistently benefited from attention, with higher per-sequence scores (P < .02) and stable recall (75%). External tools (SEP-Alg-Pro, AllerCatPro, DeepAlgPro, AlgPred) achieved 90% recall. Overall, attention effects were embedding-dependent: strong embeddings like ProtBert and ESM showed mixed or slightly negative effects, simpler embeddings like OneHot improved consistently, and SeqVec displayed intermediate behavior.

At the family level, prediction performance reflected sequence conservation and allergen prevalence. Six allergens, mostly from well-characterized families (superoxide dismutases, polcalcin, gibberellin-regulated proteins, 2S albumins, defensin-like proteins; length 82–153), were consistently identified by all models. ESM with attention showed a significant family effect (Kruskal–Wallis P = .044), while ProtBert and SeqVec were borderline (P ≈ .07–.09), indicating some families are easier to classify. OneHot models were less influenced by family but strongly biased by sequence length (Pearson P = .021, .007). SeqVec struggled with the AF197 family when attention was applied but recovered predictions after its removal, while it consistently predicted enolases as nonallergens. Rare allergens like Pap s 3 and Pla a 5 remained challenging across embeddings, reflecting difficulties in generalizing to scarcely represented or structurally divergent families: for instance, late embryogenesis abundant (LEA) was represented by only three allergens in IUIS [Cic a 1, Gly m 7 (also included in the Train set), and Pap s 3], and adenosylhomocysteinase (Pla a 5) is the first reported member.

Case studies: Established (2S Albumin) and emerging (Pap s 3) plant allergens

To further illustrate the biological relevance of our model, we analyzed its motif predictions. We compared the motifs extracted by each of our embeddings, with and without attention, across different smoothing windows (5, 7, 9, 11, and 13) for one established allergen Hel a 17 and one edge case Pap s 3. Motifs were extracted using a percentile threshold of 80th percentile of the attribution scores to focus on the most influential residues. Predicted T-cell epitopes were generated with NetMHCIIPan 4.1 [34] to identify potential HLA binders. Results are provided in Supplementary Table 5.

2S Albumin: Hel a 17

Hel a 17 is a recently identified allergen from sunflower seeds [35]. All embedding variants (OneHot, SeqVec, ESM, ProtBert), with or without attention, correctly classified it as an allergen (prediction scores: 0.61–0.92). The motifs extracted by our models generally aligned across smoothing windows, Protbert correctly identified regions at [23–34] and [73–85], and weaker binders (not included in Fig. 5) at [10–17] and [97–101], which were highlighted without MHSA addition. These predicted motifs overlapped with T-cell epitopes from NetMHCIIpan, as shown in Fig. 5a and b. Due to limiting the predicted epitopes to those under the 10th percentile, many highlighted hotspots did not align with these epitopes but appear in very weak binder such as [61–75] consistently highlighted by ESM embeddings. Overall, these results demonstrate that our embedding-based approach can identify biologically relevant immunogenic regions in novel allergens.

Figure 5.

Overlay of motifs extracted from attribution scores of Hel a 17 across different smoothing windows and embedding models, with predicted HLA-binding epitopes from NetMHCIIpan highlighted in green. (a) shows motifs predicted without MHSA, whereas (b) shows motifs predicted with MHSA.

Edge cases: Pap s 3

Pap s 3 is an LEA protein experimentally identified as an IgE-binding allergen with low diagnostic potential [36]. All embedding variants (OneHot, SeqVec, ESM, ProtBert), with or without attention, correctly classified it as an allergen, except SeqVec without attention and the full ESM model, while benchmark tools consistently predicted it as a non-allergen. NetMHCIIpan predicted several potential T-cell epitope regions, notably residues 14–22, 29–37, 50–58, 55–63, 70–83, 90–98, and 100–108. Predicted motif positions were generally stable across smoothing windows, as shown in Fig. 6a and b, though new hotspots emerged depending on window size and MHSA. OneHot highlighted 16–25 and 66–75, ESM without attention highlighted 70–93 and 99–106, SeqVec without attention highlighted 79–91, and ProtBert without attention highlighted 92–104 and 109–114, with partial overlaps to NetMHCIIpan regions. The 50–58 region, corresponding to weaker HLA binding, was mainly recognized by SeqVec trained on the Full model (with attention). Overall, differences between embeddings suggest a potential benefit of combining predictions. These identified motifs were then further evaluated through docking to assess their structural plausibility.

Figure 6.

Overlay of motifs extracted from attribution scores of Pap s 3 across different smoothing windows and embedding models, with predicted HLA-binding epitopes from NetMHCIIpan highlighted in green. (a) shows motifs predicted without MHSA, whereas (b) shows motifs predicted with MHSA.

Peptide binding to HLA-DR1

To structurally validate the sequence-based prediction of immunogenic peptides, three nonameric peptides (65 TRKEQIGTEG 74, which was predicted by Onehot and Protbert (Full model (Attention = Yes)), 75 YQEMGKKGG 83 ESM1b (Attention = Yes), and 84 LSTMEQSGG 92 ESM1b (Attention = No)) were evaluated for their binding to HLA-DR1 molecules. All three peptides showed consistently favorable docking scores, ranging from −9.3 to −13.2 kcal/mol across the five HLA-DR1 structures. YQEMGKKGG achieved the most favorable binding score (−13.2 kcal/mol), followed by TRKEQIGTEG (−11.5 kcal/mol) and LSTMEQSGG (−11.4 kcal/mol). Despite minor differences (<2 kcal/mol), all peptides demonstrated strong accommodation within the HLA-DR1 groove. Interaction analysis revealed that each peptide contacted between 8 and 15 residues of the binding pocket, consistent with the canonical MHC class II binding footprint. Energy minimization confirmed complex stability, with TRKEQIGTEG and YQEMGKKGG reaching highly favorable binding energies (−77.8 and − 73.6 kcal/mol, respectively), while LSTMEQSGG showed lower average values (−37.6 kcal/mol) but still maintained stable interactions across multiple receptor conformations.

IgE recognition of Pap s 3

Docking of the full Pap s 3 model against the IgE Fab fragment identified two main interaction regions. The dominant site spanned residues Met87-Gln97, with Glu88 contributing in 37 of the top 100 docking poses. A secondary region between Met1-Glu17 was also detected. Cluster analysis revealed five interaction clusters, with one dominant cluster representing 67% of the poses (mean score −9.7 kcal/mol). This cluster involved residues D5, E8, R12, Q15, E17, K80, E88, E93, E96, and Q97, with K80, E88, and E93 most consistently enriched.

Refined docking using the restricted IgE-binding interface confirmed the hotspot between Met87-Asp103. In this region, Glu88, Glu93, Glu96, Gln97, and Asp101 accounted for 239 of 997 contacts, highlighting their dominant role. Fingerprint analysis indicated that interactions were predominantly mediated by hydrogen bonding and ionic contacts, consistent with the electrostatic complementarity typical of IgE-allergen interfaces. Cluster analysis further supported this result, with the dominant cluster (398/493 poses, docking score −10.5 kcal/mol) mapping to residues M84-T97.

Overall, embedding-based models, particularly ProtBert and ESM, generalized effectively to this novel allergen and slightly outperformed benchmark tools. The recovered motifs largely matched predicted T-cell epitopes, suggesting reliance on biologically relevant features rather than dataset memorization. Nonetheless, edge cases like Pla a 5 underscore the need to improve coverage of rare allergen families. The MHSA block affects the regions emphasized by the model. While it sometimes highlights biologically meaningful motifs, in other cases it appears to shift attention away from expected regions, making its overall utility inconclusive.

Discussion

Accurate prediction of protein allergenicity remains challenging, as many tools perform well at classification but lack residue-level interpretability. To address this, we developed a modular deep learning framework combining CNN, BiLSTM, and MHSA layers, trained on a curated dataset of plant proteins. Building on DeepAlgPro, BiLSTM modules captured sequential dependencies, and advanced embeddings including SeqVec, ESM-1B, and ProtBert improved predictive performance, achieving accuracies of 91.6%–93.9%, comparable to state-of-the-art tools. Importantly, unlike existing tools that only provide protein-level classification, our framework can pinpoint specific allergenic regions, enabling assessment of surface exposure and accessibility to immune recognition. Such residue-level insights support downstream experimental validation, epitope mapping, or protein engineering.

Attribution analyses revealed embedding-dependent effects: ProtBert and ESM captured contextual information effectively, OneHot models benefited from attention, and SeqVec showed intermediate behavior. Predicted motifs depended on the presence of MHSA module, enabling to capture different regions, and on the window smoothing size used in motif construction. These motifs overlapped with conserved regions in major allergen families, such as 2S albumins, highlighting immunologically relevant regions. Rare allergens, such as Pap s 3, were also captured, whereas Pla a 5, from a biochemical family not previously recognized as allergenic, remained challenging, consistent with benchmark tools. Structural validation reinforced these predictions. Peptide–MHC docking confirmed that predicted Pap s 3-derived peptides bind stably and favorably to HLA-DR1, supporting effective antigen presentation. IgE docking identified a dominant conformational epitope spanning residues 87–97, overlapping two of the predicted peptides without the MHSA block, with interactions largely driven by electrostatic complementarity. These results indicate that Pap s 3 contains structurally accessible and energetically stable epitopes. The convergence of sequence-based and structural analyses highlights the reliability of our motif predictions and demonstrates the advantage of integrating multiple embeddings. While the MHSA block occasionally emphasizes biologically relevant motifs, its effect is inconsistent, preventing a definitive assessment of its contribution.

Our framework is highly flexible, allowing users to select model components and embeddings with a complete pipeline for predicting protein allergenicity and identifying potentially allergenic regions. The final output includes prediction scores for each protein and attribution scores for each residue. These can be further analyzed using the motif construction module to generate candidate motifs for downstream molecular docking.

Aggregating outputs from multiple embeddings, rather than considering them individually, may further improve predictive accuracy, particularly for challenging or rare allergens. Moreover, by providing residue-level insights, our approach enables the identification of candidate immunogenic regions that could guide experimental validation, epitope mapping, or protein engineering.

Several limitations should be acknowledged. First, the model was trained exclusively on plant proteins, which may limit generalizability to other allergen categories; future work will expand the dataset to assess cross-domain robustness. Second, the current framework relies primarily on sequence-based features and does not fully account for structural determinants of allergenicity; integration of structure-aware features using tools like AlphaFold2 is planned. Third, although advanced embeddings were used, newer protein language models (e.g. ESM-2, ProtT5) were not tested due to computational constraints. Finally, a user-friendly interface and web-based version are under development to improve accessibility. In summary, combining advanced embeddings, interpretable deep learning, and structural validation enables accurate allergenicity prediction and provides actionable insights into immunologically relevant motifs, offering a versatile tool for allergen screening in plant-derived foods.

Key Points

• DeepPlantAllergy represents an incremental advance in allergy prediction research, providing robust predictive metrics with a modular configuration.

• Its utility extends beyond classification by identifying motifs and epitope-like regions, thereby enhancing biological interpretability.

• Molecular docking results support its ability to pinpoint potential regions of allergy recognition.

Author contributions

Wala Dhouib (Data curation, Software, Formal analysis, Visualization, Writing—original draft), Fakher Frikha (Conceptualization, Methodology, Writing—review & editing), Najla Kharrat (Conceptualization, Methodology, Writing—review & editing), and Ahmed Rebai (Writing—review & editing). All authors had full access to all the data in the study and agreed to the published version of the manuscript.

Funding

None declared.

Data availability

The source codes and data are available at GitHub Repository: https://github.com/Lilly-dh/DeepPlantAllergy.

Ethics

There are no ethical issues involved.