FungiGuard: identification of plant antifungal peptides with artificial intelligence

Abstract

Antifungal peptides (AFPs) are crucial for plant defense against biotic stress. Yet, no artificial intelligence tool specifically classifies plant AFPs. To fill this gap, we develop FungiGuard, which integrates Random Forest, Long Short-Term Memory, and attention mechanisms to identify AFPs using functionally annotated plant small peptides. FungiGuard outperforms existing generalized AFP model in classifying plant AFPs, and detects candidate AFPs in Arabidopsis, wheat, rice, and maize. It also discovers novel AFPs through randomly generated sequences. Experimental validation confirms the antifungal activity of candidate AFP against Botrytis cinerea. This tool deepens plant AFP understanding and facilitates novel AFP discovery.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13059-026-03983-6.

Background

Small peptides, typically consisting of 2 to 100 amino acid residues, are widely recognized as key regulators of numerous physiological processes in plants [1, 2]. The completion of plant reference genomes has made it more convenient for us to explore small peptides. The application of ribosome sequencing (Ribo-seq) and protein mass spectrometry in plants enables a deeper investigation of non-conventional peptides (NCPs), which are translated from regions of the genome that were once annotated as non-coding [3, 4]. These NCPs include those encoded by upstream open reading frames (uORFs), downstream open reading frames (dORFs), and non-coding open reading frames (ncORFs) [4]. Another type of plant small peptides originates from a long precursor peptide, with post-translational processing generating mature functional peptides [5]. These small peptides warrant further investigation due to their potential functional significance.

Plant pathogenic fungi reduce crop yield and quality, leading to significant losses in agricultural production [6–8]. Plant antifungal peptides (AFPs) serve as important components of the plant innate immune system and play a central role in defense against fungal pathogens. AFPs are categorized into several major families, including thionins, lipid transfer proteins, hevein-like peptides, snakins, defensins, cyclotides, knottin-like peptides and others, each with distinct sequence features and antifungal mechanism [9, 10]. Functionally, AFPs act through diverse modes, such as disrupting fungal cell membranes, inhibiting cell wall synthesis, generating reactive oxygen species, or interfering with intracellular signaling pathways [11].

For example, cyclotides and defensins are cysteine-rich peptides with N-terminal signal peptides that target them to the apoplast to combat fungi; their disulfide bonds stabilize their structure [12, 13]. Representative plant AFPs include RsAFP2 from Raphanus sativus and MsDef1 from Medicago sativa, both members of the plant defensin family (PDF) [14–16]. These cysteine-rich peptides exhibit broad-spectrum antifungal activity mainly by disrupting fungal membrane [17, 18]. In Arabidopsis thaliana, multiple PDF genes (e.g., PDF1.1–PDF1.5, PDF2.1) are induced by fungal infection, pathogen-associated molecular patterns (PAMPs), and hormone signaling [19–21]. These AFP classes illustrate the molecular diversity of plant antifungal responses. However, some peptides annotated as AFPs are neither cysteine-rich nor possess N-terminal signal peptides, such as those generated through post-translational processing or certain structurally simple small AFPs [5, 22]. Specifically, HR2-7, which consists of 24 amino acids and adopts a simple α-helical structure, exhibits potent and broad-spectrum antimicrobial activity, demonstrating strong inhibitory effects against four phytopathogenic fungi [22].

In addition to conventional AFPs, NCPs have recently been identified as important antifungal agents with regulatory and defense functions [23]. Research has demonstrated the antifungal activities of 25 identified NCPs, which show broad-spectrum antifungal efficacy with varying activity against different fungal species [24]. Given the rising prevalence of fungicide resistance in plant pathogens, AFPs represent promising alternatives for developing novel plant disease control strategies with enhanced sustainability and reduced environmental impact [1, 22, 25].

Techniques such as position-weight matrices, hidden Markov models, traditional machine learning models, as well as artificial neural networks and other deep learning approaches, have been widely applied in protein classification [26–29]. For example, AmPEP applied Random Forest for Antimicrobial Peptides (AMPs) classification, AntiBP2 utilize Long Short-Term Memory (LSTM) as the core for identifying AMPs [28, 30], and iAMP-CA2L combined Convolutional Neural Network (CNN), Bi-directional LSTM (BiLSTM) and Support Vector Machines for detecting AMPs [27]. When data is sufficient, utilizing models such as Transformer and Bidirectional Encoder Representations from Transformers (BERT) can lead to better classification results for AMPs [26]. Furthermore, deep learning techniques have been applied to predict AFPs. AFPDeep employs a CNN-LSTM hybrid architecture combined with character embedding to identify AFPs from sequences [31]. Deep-AFPpred utilizes transfer learning, implemented via pretrained embeddings from seq2vec, alongside a CNN-BiLSTM hybrid model for prediction [32]. Among these, AFP-MFL demonstrates outstanding performance in AFP prediction. It leverages pre-trained protein language models and integrates multiple feature sources through a co-attention mechanism to enhance predictive accuracy [33]. However, these advanced technologies have not been applied to facilitate AFP discovery in plants.

The vast potential of AFPs within plant genomes could serve as a substantial defense against plant pathogenic fungi. However, the discovery of AFPs remains predominantly driven by experimental approaches [24]. In contrast to the well-established datasets for AMPs, datasets for identified plant AFPs remain relatively limited [34]. Moreover, due to the relatively low sequence similarity and short length of AFPs, traditional bioinformatics methods still present significant challenges. Therefore, we aim to use artificial intelligence (AI) solutions to autonomously learn sequence features, potentially identifying candidate AFPs by recognizing features from genomic sequences, thereby reducing the workload of traditional experimental screening.

In this study, we developed FungiGuard, an AI hybrid model for identifying plant AFPs training from a comprehensive dataset with functionally annotated plant small peptides. Combining with random forest, and LSTM with attention, FungiGuard displayed high performance on classifying plant AFPs. Our model predicted hundreds of AFPs in plants and identified key sites through saturation mutagenesis. The antifungal activity of one candidate AtcAFP5 against Botrytis cinerea, a pathogen inducing gray mold were experimentally confirmed. Additionally, we discovered novel AFPs through random generation, thereby enhancing our understanding and potential for discovering new AFPs.

Results

Overview of the plant-based small peptide resource for training plant AFP classifier

The PlantPepDB dataset manually collected 3848 plant-derived peptides [34]. Among these, 541 peptides were annotated as plant AFPs, and 506 of them were shorter than 100 amino acids, representing 93.53% of the total AFPs. Among the 5643 non-AFP peptides, 2638 with functional annotation have lengths below 100 amino acids, representing 46.75% of the total non-AFPs. The small peptides with lengths below 100 were selected for further analysis. These 506 plant AFPs were derived from 176 distinct species, with Arabidopsis thaliana contributing the highest number, followed by Viola odorata (Additional file 1: Fig. S1A) (Additional file 2: Table S1).

To further investigate the distinguishing characteristics of AFPs and non-AFPs, we compared their sequence lengths and cysteine content. Interestingly, AFPs displayed longer sequences than non-AFPs after the initial length-based filtering (Additional file 1: Fig. S1B). As expected, AFPs showed a markedly higher proportion of cysteine residues compared to non-AFPs (Additional file 1: Fig. S1C), suggesting a potential role for disulfide bonds in structural stabilization and antifungal activity. However, we also observed that the proportion of cysteine residues varies among different AFP families. Among the 506 known AFPs from PlantPepDB [34], we selected the top 10 families by peptide count and calculated their cysteine content. The Cyclotide family exhibited the highest average cysteine proportion, reaching approximately 20%. This was followed by Hevein, Snakin, Thionin, Knottin and Defensin families (Additional file 1: Fig. S1D). In contrast, families such as Thaumatin and Urease showed relatively low cysteine content, with average proportions below 5%. These findings suggest that some AFPs are not cysteine-rich and may stabilize their mature protein structures through alternative mechanisms (Additional file 1: Fig. S1D). Taken together, 506 high-confidence plant AFPs and 2638 plant non-AFPs peptides from this functionally annotated dataset were used for training plant AFPs predictors.

Single model using functionally annotated small peptides for plant AFP identification

Using the annotated plant small peptides as training set, we applied deep learning and traditional machine learning techniques to construct AFP prediction models, including four deep learning models from two categories and traditional machine learning models. Our base model utilized a neural network model with LSTM as the core layer, which has been demonstrated to be effective in AMP identification [27, 35]. To address LSTM's limitation in long sequence memory, we incorporated an attention mechanism [36]. Moreover, to further enhance feature representation, we incorporated Graph-based Learned Unsupervised Embedding (GLUE) into selected models to integrate contextual information from sequences and improve the modeling of structural features. During training, the loss curves of the deep learning models showed overall good convergence, indicating stable fitting without obvious overfitting (Additional file 1: Fig. S2). Due to the relatively small dataset size, we also applied the support vector machine classifier (SVC) and random forest classifier (RFC) to plant AFP classification. Hyperparameters of both deep learning and traditional machine learning models were optimized.

Using a holdout validation approach (10% of data) with balanced sampling during training, we systematically evaluated six machine learning models for plant AFP prediction. Initial assessment via Receiver Operating Characteristic (ROC) analysis revealed distinct performance tiers among the models (Fig. 1A). The RFC emerged as the top performer with an Area Under the Curve (AUC) of 0.93, significantly outperforming other models (AUC range: 0.86–0.89) including various LSTM architectures (standard, bidirectional, and attention-enhanced variants) and SVC. Comprehensive metric evaluation and confusion matrix analysis demonstrated RFC's balanced performance across all measures (accuracy = 90.46%, precision = 79.76%, Recall = 60.91% and F1 = 69.07%), while SVC showed particularly poor recall (20%) and F1 score (42.46%), leading to its exclusion from subsequent analyses (Fig. 1B-C and Additional file 1: Fig. S3). The RFC and biLSTM with attention models achieved the highest accuracy, approximately 90.46% and 88.08%, respectively (Fig. 1B). Probability distribution analysis (Fig. 1D) of the five retained models revealed that RFC maintained the clearest separation between AFPs (high probability concentration near 1.0) and non-AFPs (low probability near 0.0), while attention mechanisms in LSTM models introduced some high-confidence false negatives. Overall, we obtained 5 base models that can be further combined for identifying plant AFPs.

Fig. 1.

Performance evaluation of models for predicting plant AFPs using database of functionally annotated plant small peptides. A The combined Receiver Operating Characteristic (ROC) curves of the six models, along with their corresponding Area Under the Curve (AUC) values, illustrate their ability to discriminate between AFPs and non-AFPs. B The distribution of true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN) for each model, with precision values displayed above the bars. C A comparative summary of performance metrics—Accuracy, Precision, Recall, and F1 Score—across all models. D Probability distributions of prediction accuracy for true AFPs (top) and non-AFPs (bottom) generated by the five models

The architecture and performance of FungiGuard for plant AFP identification

Integrating five machine learning models, including LSTM, biLSTM, their attention-enhanced variants, and the top-performing RFC, we developed FungiGuard, an ensemble framework, through a majority voting scheme to improve the accuracy and reliability of plant AFP identification (Fig. 2A). Performance evaluation revealed that prediction confidence increased with the number of agreeing models, with unanimous consensus (5/5 votes) achieving the highest precision (89.58%) (Fig. 2B). Confusion matrix analysis demonstrated that samples classified as AFPs by all five models exhibited a highest true positive proportion (~ 89.6%), significantly outperforming predictions with partial model agreement (Fig. 2C and Additional file 1: Fig. S4). While full consensus resulted in a modest decrease in recall and F1 score (Fig. 2D), the substantially improved precision was prioritized for selecting high-confidence AFP candidates for experimental validation. Taken together, these results established the 5-model agreement as FungiGuard’s integration criterion, ensuring high-confidence AFP predictions for downstream experimental validation. By leveraging the complementary strengths of diverse architectures, the framework enhances the identification of functional antifungal peptides while mitigating individual model limitations, providing a robust tool for plant small peptide characterization.

Fig. 2.

Architecture and performance of FungiGuard for plant AFP classification. A The architecture of FungiGuard using a voting ensemble of five machine learning models: LSTM, biLSTM, LSTM with attention, biLSTM with attention and RFC. FC: fully connected layers, FV: feature vector, and GeLU: Gaussian error linear unit (activation functions). B Prediction performance of the ensemble model using different numbers of voting classifiers (1 to 5), showing the percentage of true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN). Precision values are displayed for each configuration. C Column-normalized confusion matrices (columns sum to 100%) illustrating the distribution of actual AFPs and non-AFPs predicted by FungiGuard. Top-left and bottom-left cells: TN and FP proportions among negative samples. Bottom-right and top-right cells: TP and FN proportions among positive samples. D Performance metrics (Accuracy, Precision, Recall, and F1 Score) for ensemble models with increasing numbers of voting classifiers (1 to 5)

FungiGuard outperforms current tools on the task of plant AFP identification

The deep learning model AFP-MFL integrates features extracted from a pre-trained protein language model and utilizes a co-attention mechanism, outperforming existing state-of-the-art methods across four independent test datasets [33]. As AFP-MFL was trained on AFPs from diverse sources [37–40], its predictive accuracy for plant-derived AFPs may be suboptimal. To test this hypothesis, we compared the predictive performance of AFP-MFL with the core models of FungiGuard for plant AFP identification. Firstly, we applied AFP-MFL to predict all plant small peptides included in both the training and test sets of FungiGuard dataset from PlantPepDB, achieving an AUC of 0.78 (Fig. 3A). This result indicates moderate discrimination between AFPs and non-AFPs, but its confusion matrix revealed a high false positive rate of 66.5% (Fig. 3B), with an overall accuracy of 74.96%, precision of 33.47%, and F1 score of 44.71%. We hypothesized that the phylogenetic relationship between plant AFPs and animal AFPs affects the predictive accuracy of AFP-MFL. Therefore, we performed BLAST searches to identify homologous proteins of all AFPs obtained from PlantPepDB within the AFP-MFL dataset and calculated their sequence identities. Our analysis revealed that plant AFPs that were predicted as true positives (TP) by AFP-MFL exhibit higher sequence similarity to AFPs from PlantPepDB, whereas those predicted as false negatives (FN) show lower similarity (Fig. 3C). These results suggest that the limited representation of plant AFPs in the AFP-MFL dataset, likely contributes to the lower predictive performance for plant-derived peptides.

Fig. 3.

Performance comparison between the AFP-MFL model and FungiGuard. A Receiver Operating Characteristic (ROC) curve of the AFP-MFL model, with its corresponding Area Under the Curve (AUC) value, showing its classification performance across all plant small peptides in PlantPepDB. B Confusion matrix of AFP-MFL predictions, with actual labels on the vertical axis and predicted labels on the horizontal axis. The matrix displays samples classified as AFPs by AFP-MFL. Top-left and bottom-left cells: True negative (TN) and false positive (FP) proportions among negative samples. Bottom-right and top-right cells: True positive (TP) and false negative (FN) proportions among positive samples. C The density plot showing the protein sequence similarity (Identity %) of predicted small peptides with proteins in AFP-MFL dataset, with the legend indicating the category corresponding to each curve. Statistical differences between groups were assessed using the Mann–Whitney U test. D, E ROC curve (D) and confusion matrix (E) showing AFP-MFL's performance when evaluated on the 10% test set. F Comparison of performance metrics (Accuracy, Precision, Recall, and F1 Score) between AFP-MFL and FungiGuard on the 10% test set

When tested on the same holdout of 10% test set, FungiGuard exhibited superior performance reducing false positives by nearly half (33.3% vs. AFP-MFL’s 61.2%) (Fig. 3D, E). Most notably, FungiGuard achieved a precision of 89.58%—more than triple that of AFP-MFL (28.96%)—while maintaining robust recall (Fig. 3F). This substantial improvement highlights FungiGuard’s specialization for plant AFP identification, providing higher-confidence predictions for downstream experimental validation. The comparative analysis underscores that domain-specific optimization, as implemented in FungiGuard, is critical for accurate plant peptide characterization. These findings indicate FungiGuard’s advantage in enhancing predictive precision and minimizing false positives in the context of plant AFP identification.

Identification of candidate AFPs from short ORFs in plants using FungiGuard

To identify novel AFPs in plants, we employed FungiGuard to screen small open reading frames from A. thaliana, wheat, rice, and maize for potential antifungal activity. Specifically, we first extracted peptides encoded by short ORFs (sORFs) shorter than 100 amino acids from these plant species, with wheat yielding the highest number of sequences at 6,392 and A. thaliana containing the fewest with 2,829 sequences. These sequences were subjected to AFP prediction using FungiGuard. Our analysis revealed distinct patterns of AFP distribution among these plant species (Fig. 4A) (Additional file 3: Table S2). Among the predicted AFPs, A. thaliana exhibited the highest proportion (1.24%), with 35 peptides identified, representing the largest percentage relative to its total sORF-derived peptides. In contrast, maize showed the lowest proportion (0.22%), with only 7 predicted AFPs. This disparity is related to the species distribution within the training dataset used by FungiGuard, potentially influencing prediction sensitivity across different plant species (Additional file 1: Fig. S1A).

Fig. 4.

Identification of AFPs from short ORFs in four plant species by FungiGuard. A Distribution of AFPs and non-AFPs encoded by sORFs in Arabidopsis, rice, wheat, and maize, as identified by FungiGuard. B Bubble plot displaying the enriched GO terms of AFPs in Arabidopsis. Bubble size reflects gene ratio; color indicates adjusted p-value (p.adjust); x-axis shows GO terms. C Heatmap displaying z-score normalized FPKM values for genes encoding candidate AFPs (y-axis) under various biotic stresses (x-axis). The number of transcriptome datasets per stress condition is indicated. Genes and conditions are clustered by expression patterns. Gene names highlighted in red represent those associated with the GO term “defense response to fungus.” D Schematic representation of the LCR76 protein domain architecture. Signal peptides are highlighted in purple, while cysteine residues are marked in yellow

To explore the function of AFPs, we performed a detailed analysis of the predicted AFPs in A. thaliana. First, Gene Ontology (GO) enrichment analysis revealed that these AFPs were significantly enriched in two GO terms: “killing of cells of other organism” and “defense response to fungus” (Fig. 4B). This enrichment indicates that the predicted AFPs are functionally associated with antimicrobial activity and the plant’s innate immune response against fungal pathogens, supporting their potential role in plant defense mechanisms. Subsequently, we analyzed the expression profiles of the predicted AFPs in A. thaliana using transcriptomic data. Under various biotic stress conditions, genes encoding antifungal peptides exhibited differential expression patterns [41]. Several of these genes, highlighted in red, are annotated with the Gene Ontology term “defense response to fungus” (Fig. 4C). AT1G55010, AT2G26020, and AT5G44430 were significantly upregulated during Sclerotinia sclerotiorum infection (Fig. 4C), suggesting their potential involvement in the plant’s defense response. Among them, AT1G55010 and AT2G26020 are known members of the PDF family and were included in the original training dataset (Additional file 3: Table S2). In contrast, AT5G44430, although showing strong induction upon infection, is not annotated as a PDF family member and was not part of the training set. Similarly, AT2G31953 (LCR76), which also lies outside the training set and PDF annotation, encodes a small, secreted, cysteine-rich protein with sequence similarity to the PCP (pollen coat protein) family (Fig. 4D).

The presence of the signal peptide and cysteine-rich composition are key features of antifungal peptides [42]. Among the predicted AFPs in A. thaliana, six peptides were found to contain signal peptides predicted by SignalP 6.0 software [43], accounting for approximately 17% of the total (Additional file 3: Table S2). In other plant species, we also identified several potential AFPs, highlighting the applicability of our model in non-model crops. Only two of these AFPs were annotated as secreted proteins (Additional file 3: Table S2), which may be attributed to limited homology of signal peptide-containing AFPs across species in the training dataset. Overall, these results indicate that our model can recognize known AFPs from the dataset and may also help identify novel candidate peptides with potential antifungal activity.

Identification of candidate AFPs from non-canonical ORFs in Arabidopsis

Current studies have revealed that the non-canonical ORFs in 5’ UTR and 3’ UTR of mRNAs and long non-coding RNA regions can encode small peptides, known as NCPs [24]. We employed FungiGuard to identify AFPs in Arabidopsis, analyzing 2868 peptides coded by non-canonical ORFs. These included 2113 uORFs, 209 dORFs, 546 ncORFs identified via a super-resolution Ribo-seq dataset [4] and with peptide lengths under 100 amino acids.

Our analysis revealed that 163 peptides encoded by uORFs, 15 by dORFs, and 13 by ncORFs were predicted to have antifungal potential (Fig. 5A). Among these, 39 candidate AFPs encoded by uORFs contained predicted signal peptides (Additional file 3: Table S2) whereas the remaining candidates lacked typical secretion signals. We focused on these candidate AFPs and analyzed transcriptome data from A. thaliana infected with Botrytis cinerea, a destructive fungal pathogen responsible for gray mold disease. Among the genes encoding candidate AFPs, 18 showed differential expression in response to B. cinerea infection, with 10 uORF-encoded genes and 1 dORF-encoded gene being significantly upregulated. These 11 AFPs were designated AtcAFP1 through AtcAFP11 (Additional file 1: Fig. 5B). These results suggest that a subset of non-canonical ORFs in Arabidopsis may play previously unrecognized roles in fungal defense responses.

Fig. 5.

Identification and functional validation of Arabidopsis non-conventional peptides (NCPs) with antifungal activity against Botrytis cinerea. A Classification of Arabidopsis proteins encoded by different NCP types (uORFs, dORFs and ncROFs) using FungiGuard, showing their distribution of potential AFPs. B The heatmap with z-score normalized FPKM values showing the up-regulated AFP genes during Botrytis cinerea infection. Genes originating from uORFs are shown in green, while those from dORFs are highlighted in red. The number after gene ID indicate start and end position of main ORF. C Dose–response curve of the growth inhibition of B. cinerea by AtcAFP in vitro. The data represent the mean OD values measured after 48 h of incubation at each protein concentration. Means and standard errors (SEs) were calculated from three biological replicates (*P < 0.05; **P < 0.01; Student’s t-test). Asterisks indicate the minimum inhibitory concentration (MIC). D Infiltration of N. benthamiana leaves with 100 μmol/L AtcAFP5 caused necrosis, whereas water-infiltrated leaves showed no symptoms. Photos were taken 24 h after infiltration. E Disease symptoms on N. benthamiana leaves sprayed with AtcAFP5 protein at three concentrations (32, 16, and 8 μmol/L) or with water were photographed 3 days after inoculation with B. cinerea (n = 3). F Disease symptoms on Arabidopsis leaves sprayed with 16 μmol/L AtcAFP5 were photographed under the same conditions (n = 5). E and F Bars = 2.5 cm. Relative diseased areas were quantified, and means ± SEs were calculated from the respective biological replicates (*P < 0.05; **P < 0.01; Student’s t-test). G Mean OD values 48 h after treatment with 16 μmol/L AtcAFP or AtcAFP5^T19A,T20A (n = 3, means ± SEs; *P < 0.05; **P < 0.01; Student’s t-test). H Disease symptoms on N. benthamiana leaves sprayed with AtcAFP, AtcAFP5^T19A,T20A (16 μmol/L), or water were photographed 3 days after inoculation with B. cinerea (n = 3, means ± SEs; *P < 0.05; **P < 0.01; Student’s t-test). Bar = 2.5 cm. Relative diseased areas were quantified, and means ± SEs were calculated from three biological replicates. I Confocal microscopy of B. cinerea mycelia treated with AtcAFP5 or AtcAFP5^T19A,T20A for 3 h. Bars = 500 μm. J Effect of AtcAFP5 or AtcAFP5.^T19A,T20A concentration on nucleic acid leakage (n = 3; means ± SEs; *P < 0.05, **P < 0.01, Student’s t-test)

To rapidly validate our findings, we assessed the antifungal activity of six candidate AFPs (AtcAFP1 to AtcAFP6) by Agrobacterium-mediated infiltration of leaves of Nicotiana benthamiana expressing either GFP or the candidate AFPs. The results showed that AtcAFP3 (AT3G14270.1_355_421) and AtcAFP5 (AT1G17050.1_84_153) were successfully expressed (Additional file 1: Fig. S5B) and significantly inhibited Botrytis cinerea infection (Additional file 1: Fig. S5A and C). The remaining peptides did not exhibit detectable activity under the current experimental conditions, which may reflect a lack of inherent antifungal activity, limitations of the crude experimental approach, or potential specificity towards other fungal species rather than B. cinerea.

We then assessed the expression of defense-related genes in N. benthamiana and found that AtcAFP3 did not significantly induce immunity-associated genes. In contrast, AtcAFP5 strongly upregulated marker genes for immune responses [19] (Additional file 1: Fig. S5D), suggesting that, in addition to its direct antifungal activity, AtcAFP5 may also exert immunomodulatory effects, enhancing plant immune signaling [20]. This dual activity is consistent with previously reported antimicrobial peptides such as stable antimicrobial peptides (SAMPs), which similarly combine direct antimicrobial effects with the ability to modulate host immunity [44].

Using AlphaFold3 [45], we observed that AtcAFP3 and AtcAFP5 adopt stable α-helical structures (Additional file 1: Fig. S5E), which may contribute to their antifungal activity. Similarly, Plantaricin J, a member of Class II AMPs, is predominantly composed of α-helices throughout its sequence [46]. HR2-7, a recently characterized broad-spectrum antifungal peptide, adopts a conformation highly similar to that of AtcAFP5 [22].

AtcAFP5 suppresses Botrytis cinerea infection

To further validate the antifungal activity of the candidate AFPs against B. cinerea, we performed systematic functional assays using chemically synthesized peptides. Due to the absence of hydrophilic residues, AtcAFP3 could not be successfully synthesized, and subsequent experiments focused on AtcAFP5. In vitro antifungal assays demonstrated that AtcAFP5 inhibited B. cinerea growth in a dose-dependent manner, exhibiting significant inhibition at 16 μmol/L, corresponding to the minimum inhibitory concentration (MIC) (Fig. 5C).

Next, in planta functional validation was conducted. Infiltration of 100 μmol/L AtcAFP5 into N. benthamiana leaves caused visible necrosis within 24 h, whereas water-infiltrated leaves remained symptomless (Fig. 5D), suggesting that high concentrations of AtcAFP5 may exert phytotoxic effects or trigger programmed cell death (PCD) in plant tissues [47]. When leaves were sprayed with AtcAFP5 prior to fungal inoculation, disease symptoms were significantly reduced in N. benthamiana, showing a clear dose-dependent effect (Fig. 5E). Quantification of relative diseased areas confirmed that higher peptide concentrations substantially suppressed B. cinerea infection. Similar treatments in Arabidopsis leaves also resulted in significant inhibition of fungal infection (Fig. 5F). Together, these results indicate that chemically synthesized AtcAFP5 effectively suppresses B. cinerea infection.

Furthermore, we performed alanine scanning on AtcAFP5 using FungiGuard to identify potential functional residues. Mutations at positions 19 and 20 (AtcAFP5^{T19A, T20A}) resulted in a complete loss of antifungal activity. Interestingly, these mutations disrupted the α-helical structure of the peptide (Additional file 1: Fig. S5E). To investigate the role of these specific residues in planta, Nicotiana benthamiana expressing the mutant peptide exhibited markedly reduced antifungal activity compared with wild-type AtcAFP5 (Additional file 1: Fig. S5G and H). Analysis of defense-related gene expression in N. benthamiana revealed that, relative to AtcAFP5, AtcAFP5^{T19A, T20A} induced lower levels of key immune marker genes (Additional file 1: Fig. S5I).

Subsequently, we synthesized AtcAFP5^{T19A, T20A} and evaluated its antifungal activity in vitro, the mutant peptide displayed reduced inhibition of B. cinerea growth compared with wild-type AtcAFP5 (Fig. 5G and Additional file 1: Fig. S5F). Likewise, leaves sprayed with AtcAFP5^{T19A, T20A} developed more severe disease symptoms than those treated with wild-type AtcAFP5 (Fig. 5H). Confocal microscopy revealed that AtcAFP5 treatment caused pronounced structural alterations in B. cinerea hyphae, whereas the mutant peptide had a markedly weaker effect (Fig. 5I). Furthermore, nucleic acid leakage assays indicated that AtcAFP5 significantly disrupted fungal cell membranes, while the disruptive effect of AtcAFP5^{T19A, T20A} was attenuated (Fig. 5J). Taken together, these results demonstrate that AtcAFP5 effectively suppresses B. cinerea infection, and that residues T19 and T20 are critical for its antifungal activity and the activation of plant immune responses.

Plant AFP prediction from randomly generated small peptides

To broaden our understanding and discovery of plant AFPs, it is essential to explore AFPs derived from non-natural sources. In this study, we sought to generate novel AFPs through a targeted randomization strategy. To elucidate the sequence characteristics of AFPs and inform peptide design, we compiled a curated dataset of known AFPs. Analysis of amino acid composition indicated that cysteine is the most prevalent residue, followed by glycine (Fig. 6A). With respect to sequence length, the majority of AFPs consist of approximately 20 to 50 amino acids (Fig. 6B).

Fig. 6.

Features of predicted AFPs identified by FungiGuard and known AFPs in plant small peptides dataset. A Boxplot showing the percentage composition of the 20 standard amino acids in known AFPs; individual data points are represented by dots. B Histogram illustrating the frequency distribution of sequence lengths in known AFPs. C Distribution of AFPs and non-AFPs encoded by randomly generated sequences, as identified by FungiGuard. D to R. Violin plots illustrating the distribution of each feature across datasets, with blue lines representing mean values. D and E represent protein secondary structure features; F to K represent the chemical properties of amino acid residues; L to R correspond to the physicochemical properties of amino acid residues (*P < 0.05; **P < 0.01; ***P < 0.001; ns > 0.05; Mann–Whitney U test)

We generated 10,000 such peptides and analyzed them using FungiGuard, identifying 514 peptides classified as AFPs by all five models (Fig. 6C) (Additional file 3: Table S2). We integrated all predicted AFPs, including those encoded by various plant sORFs as well as NCPs from Arabidopsis. Amino acid composition analysis revealed that cysteine exhibited the highest content, followed by glycine, consistent with known AFP characteristics (Additional file 1: Fig. S7A). The sequence length distribution analysis showed that peptides with lengths of 22 and 42 to 44 residues were the most abundant, displaying distribution patterns similar to those of characterized AFPs (Additional file 1: Fig. S7B).

Structural characteristics and general physicochemical properties analysis of plant predicted AFPs and annotated AFPs

Finally, we conducted a comprehensive analysis of the structural characteristics and general physicochemical properties of all identified and predicted candidate AFPs (Fig. 6D–R). Secondary structure analysis revealed that AFPs possess a significantly higher proportion of alpha helices compared to non-AFPs, with predicted AFPs exhibiting even greater alpha-helix content than validated AFPs (Fig. 6D and E). This pattern aligns with the structural features observed in AtcAFP3 and AtcAFP5, which may be attributed to their classification as Plantaricin J-type AFPs.

Further analysis of amino acid composition revealed significant differences between validated AFPs and non-AFPs in the proportions of basic, acidic, polar, and charged residues (Fig. 6G, H, J, and K). In contrast, predicted AFPs exhibited similar chemical characteristics to validated AFPs, particularly in terms of the proportion of acidic residues (Fig. 6H). Regarding physicochemical properties, the isoelectric points of predicted AFPs closely resembled those of validated AFPs, but differed markedly from non-AFPs (Fig. 6L). Additionally, AFPs showed greater molecular flexibility compared to non-AFPs, with predicted AFPs displaying flexibility levels consistent with those of validated AFPs (Fig. 6R).

Collectively, these results substantiate the reliability of our predicted AFP dataset and underscore pronounced differences in protein structure and physicochemical properties between AFPs and non-AFPs, thereby laying a robust foundation for further elucidation of the mechanisms underlying such selectivity.

Discussion

Plant antifungal peptides, a class of small peptides with inherent antifungal activity, play a critical role in the natural defense mechanisms of plants [20]. Investigating these peptides enhances our understanding of plant disease resistance and facilitates the development of innovative strategies for plant disease management [48]. Therefore, the discovery of plant AFPs is crucial for improving crop disease resistance, reducing pesticide dependence, ensuring food security, and advancing sustainable agricultural practices [49].

This study aims to address the current lack of AI tools for identifying plant AFPs by developing the 'FungiGuard' project. This project integrates machine learning, particularly natural language processing (NLP) techniques, to classify plant peptides of up to 100 amino acids in length as antifungal. Compared to traditional homology-based identification methods such as BLAST, machine learning approaches offer several advantages: enhanced recognition capability enabling the detection of peptides with low sequence similarity but similar functions; efficient processing of large-scale datasets; and the ability to integrate multiple sources of information to improve model generalizability. Therefore, the development of machine learning–based identification methods hold significant research value and promising application prospects.

However, several limitations should be noted. First, the training dataset is relatively small, and the diversity of plant-derived peptides is limited, potentially introducing species-specific bias. Second, experimental validation was restricted to Botrytis cinerea, leaving an open question of whether the identified AFPs are broadly effective against other fungal pathogens.

In the future, the plant classifier can be further improved in several key areas. Firstly, with the potential discovery of a larger number of plant antifungal peptides, advanced models such as BERT and Mamba can be used to improve classification accuracy and model generalizability with larger datasets [50]. Moreover, by applying transfer learning strategies [33, 51], knowledge from large-scale protein datasets, pre-trained models can then be fine-tuned on plant antifungal peptide data. This approach is expected to significantly enhance model performance, especially when labeled plant AFP data remains limited.

In this study, we identified candidate AFPs from NCPs encoded by non-canonical ORFs in Arabidopsis thaliana using FungiGuard. In the future, given the growing number of NCPs detected by Ribo-seq in various plant species, FungiGuard will provide a powerful tool for systematically screening NCPs with potential antifungal activity across diverse plants.

Furthermore, we plan to investigate the mechanisms and biological functions of plant antifungal peptides in greater depth. In Arabidopsis, we identified the antifungal peptide PDF1.2a and its related family members [34], as well as cysteine-rich secretory proteins, all of which are potential AFPs. The antifungal peptides identified in our study predominantly exhibit simple alpha-helical structures. Notably, Plantaricin J is mainly composed of alpha-helices, consistent with the typical structural features of Class II antimicrobial peptides [46]. Moreover, previous studies have indicated that plant antifungal peptides may exhibit characteristic physicochemical properties, such as cationicity, amphipathicity, and relatively high molecular flexibility [52]. Our analysis of structural and physicochemical properties shows that the predicted AFPs are consistent with validated AFPs in terms of amino acid composition, isoelectric point, and molecular flexibility, thereby extending and corroborating existing observations. Proteins exhibit substantial diversity in both type and function, and given the limitations of the dataset, the conclusions presented here offer only a preliminary and partial perspective.

In this study, we confirmed that AtcAFP5 suppresses B. cinerea infection. In addition to directly inhibiting pathogen growth, AtcAFP5 may enhance plant disease resistance by activating innate immune responses. This dual functionality suggests that AtcAFP5 not only serves as a natural antifungal peptide but may also act as a signaling molecule in plant–pathogen interactions. However, the underlying mechanisms remain unclear, and it is unknown whether the induced immunity depends on key signaling components such as BAK1 or SGT1 [53, 54]. By comparison, the dual-function antimicrobial peptide SAMP uses its α-helical domain to form stable hexameric pores that disrupt bacterial membranes and simultaneously activates plant immunity via NPR1- and SGT1-dependent pathways [44]. Further mechanistic studies are needed to clarify and confirm the dual antifungal and immunomodulatory functions of AtcAFP5.

Conclusions

FungiGuard provides a robust computational-experimental framework for plant AFP discovery, combining machine learning, structural analysis, and functional validation. By addressing technical limitations, situating findings within existing literature, and exploring mechanistic implications, this study contributes both practical tools and conceptual insights to plant peptide research and sustainable agriculture.

Methods

Data collection

In our study, we collected a single dataset sourced from PlantPepDB with functionally annotated plant small peptides. From PlantPepDB, we obtained 529 AFPs and 5,643 non-AFPs. These peptides ranged in length from 2 to 1,295 amino acids, and we filtered out peptides longer than 100 amino acids. Based on their classification features, 506 plant AFPs were shorter than 100 amino acids, representing 93.53% of the total AFPs. Among the non-AFPs, 2,638 sequences were shorter than 100 amino acids, accounting for 46.75% of the total non-AFPs. We obtained sequences from PPepDB_1 to PPepDB_6173 in the PlantPepDB database. Since PlantPepDB is a website whose content may change over time, we have included the retrieved sequence information in Additional file 2: Table S1.

The proteins used for prediction in this study were mainly derived from three sources. First, we utilized reference proteomes of Arabidopsis thaliana (TAIR10), rice (Oryza sativa Japonica Group IRGSP-1.0), wheat (Triticum aestivum IWGSC RefSeq v1.0), and maize (Zea mays Zm-B73-REFERENCE-NAM-5.0) obtained from Ensembl Plants, from which protein sequences shorter than 100 amino acids were extracted for subsequent analysis. Second, based on Ribo-seq data, unconventional peptides in Arabidopsis were identified, and sequences shorter than 100 amino acids were retained [4]. Finally, to support model training and evaluation, 10,000 protein sequences were randomly generated based on observed peptide length and amino acid frequency distributions of annotated plant AFPs, using a random seed of 1 to ensure reproducibility.

Bioinformatics analysis

For signal peptide prediction, the local version of SignalP 6 was used [55]. To analyze the physicochemical properties of the proteins, we employed the EMBOSS Pepstats web tool [56]. Secondary structure composition was determined using the PSRSM-Server (http://qilubio.qlu.edu.cn:82/protein_PSRSM/default.aspx). For obtaining tertiary structures, AlphaFold3 was utilized [45]. To construct the phylogenetic tree, sequences were first aligned using MUSCLE to produce a comprehensive multiple sequence alignment [57].

To investigate the expression of candidate AFPs following pathogen infection, we conducted transcriptome analysis. Gene expression under various pathogen challenges was assessed using a comprehensive online database containing approximately 20,000 publicly available Arabidopsis RNA-Seq libraries [41]. For Botrytis cinerea infection specifically, published FPKM data were referenced for targeted screening and analysis [58].

Prediction pipeline by combination of multiple models

We used the dataset that is composed of 2,638 non-AFPs and 506 AFPs as input data and made the results output as vectors. We developed six models for sequence classification, including four natural language processing (NLP)-based models and traditional machine learning models (Fig. 1B).

The NLP-based models utilize various architectures, including standard Long Short-Term Memory (LSTM) layers, attention mechanisms, bidirectional LSTM, and combinations of bidirectional LSTM with attention mechanisms [59]. These models utilize Gaussian Error Linear Units (GELU) activation functions and dropout regularization [60]. The models are optimized using the Adam optimizer, which adjusts the network weights, and use cross-entropy as the loss function. The cross-entropy loss function is defined as follows:

$$Loss=-{\sum }_{i=1}^C{y}_{i}log(p_i)$$

where $y_i$ represents the true label, $p_i$ is the predicted probability for class $i$, and $C$ is the total number of classes. In classification tasks, the softmax function converts raw output scores into class probabilities. For an input $x$, the probability of class $k$ is given by:

$$P\left(y,=,k,|,x\right)=\frac{e^{z_k}}{{\sum }_{i=1}^C{e}^{z_j}}$$

where $z_k$ is the raw score for class $k$, and $C$ is the total number of classes. The softmax function normalizes the exponentiated scores so that the probabilities sum to 1 and fall within the range of 0 to 1, thus providing a valid probability distribution for classification.

The traditional machine learning models employed in this study include Random Forest Classifier (RFC) and Support Vector Classifier (SVC). To ensure accurate probability estimation, hyperparameter tuning for both models was conducted using GridSearchCV to determine the optimal parameters. Probability estimation for these models adhered to established methodologies grounded in prior research. Specifically, the RFC model calculates probabilities by averaging the class probabilities predicted by each tree in the ensemble [61]. In contrast, the SVC model employs Platt scaling, which applies a sigmoid function to the decision function output, thereby converting it into probability estimates [62].

Class imbalance is common in real-world datasets, such as the disproportionate ratio of non-AFPs to AFPs. To address this, we used a weighted random sampler. Each data point was assigned a weight inversely proportional to its class size, allowing the sampler to ensure balanced representation during data selection.

Model training and evaluation

The models were trained using the Adam optimizer with a learning rate set to 0.001. Cross-entropy loss function was employed to update model parameters during training. To prevent overfitting, the training was halted once the model performance on the validation set stabilized. We conducted a series of validation tests to rigorously assess the models' performance. For initial evaluation, the dataset was randomly split into a training set and a test set with a 4:1 ratio. The training set was used to train the models, while the test set was reserved for evaluation.

To further assess the models' generalization capabilities, we performed cross-dataset validation. Initially, the models were trained on the primary dataset and validated on an independent dataset to determine their robustness. The final evaluation included testing the models on additional independent datasets, further ensuring their reliability across various data sources.

The classification performance was evaluated using several key metrics: accuracy, precision, recall, and F1 score. The formulas used for these calculations are as follows:

$$\begin{array}{c}Accuracy=\frac{TP+TN}{TP+TN+FP+FN}\\ \begin{array}{c}Precision=\frac{\mathit{TP}}{TP+FP}\\ \begin{array}{c}Recall=\frac{\mathit{TP}}{TP+FN}\\ F1Score=\frac{2x\left(P,r,e,c,i,s,i,o,n,\times ,R,e,c,a,l,l\right)}{\left(P,r,e,c,i,s,i,o,n,+,R,e,c,a,l,l\right)}\end{array}\end{array}\end{array}$$

Here, TP, FP, TN, and FN represent true positive, false positive, true negative, and false negative instances, respectively.

In vitroantifungal activity

Fresh conidia of Botrytis cinerea were obtained by culturing the fungus on potato dextrose agar (PDA) plates at 25 °C for 7 days. The conidial suspension was filtered through two layers of Miracloth (Merck, Germany), counted using a hemocytometer, and diluted with sterile distilled water to the desired concentration. The antifungal activity of AFPs against B. cinerea was evaluated using a spectrophotometric approach [63, 64]. Briefly, AFPs were serially twofold diluted in sterile distilled water, and 20 μL of each dilution (final concentrations: 0.25, 0.5, 1, 2, 4, 8, 16, and 32 μM) was added to the wells of a 96-well microtiter plate containing 180 μL of spore suspension (~ 10⁴ spores/mL) prepared in 20-fold diluted potato dextrose broth (PDB). The plates were incubated at 25 °C for 48 h, and fungal growth inhibition was quantified by measuring absorbance at 595 nm using a Spark 175 microplate reader (Tecan, Switzerland). The minimum inhibitory concentration (MIC) of each peptide against B. cinerea was defined as the lowest peptide concentration that completely inhibited fungal growth in all replicate experiments. Each concentration was tested in three technical replicates.

Agrobacterium tumefaciens infiltration

The full-length AFPs were cloned from Arabidopsis cDNA, with the relevant primers detailed in Table S3. The constructs were introduced into the A. tumefaciens strain GV3101. Following selection with antibiotics, individual colonies were confirmed by PCR and cultured in LB medium at 28 °C with shaking at 220 rpm for 36 h. The bacteria were pelleted by centrifugation, resuspended in an MES buffer (10 mM MgCl₂, 10 mM MES, 200 mM acetosyringone, pH 5.7) to a final optical density at 595 nm (OD₅₉₅) of 0.6, and kept in the dark at room temperature for 3 h before infiltration. The A. tumefaciens cell suspension was infiltrated into plant leaves.

Pathogen inoculation and symptom assessment

To investigate the effects of AFPs on B. cinerea infection in planta, leaves of 5- to 7-week-old Nicotiana benthamiana were subjected to two independent treatments. In the first treatment, synthetic AFP candidate peptides (GenScript, China) were sprayed onto the leaf surface, followed by inoculation with B. cinerea. In the second treatment, AFPs were transiently overexpressed in N. benthamiana leaves via Agrobacterium-mediated transformation, and the leaves were inoculated with B. cinerea at 36 h post-infiltration.

B. cinerea was cultured on PDA plates, and PDA plugs containing actively growing mycelia were used for inoculation. The inoculated leaves were maintained in a humidified chamber at 25 °C for 3 days. Symptom development was visually assessed at 3 days post-inoculation (dpi). Lesion areas were quantified using Adobe Photoshop (v23.0) by threshold-based selection of necrotic regions, and values were normalized to those of the GFP control (set to 1). Data from three biological replicates were analyzed using Student’s t-test.

Changes in mycelial morphology and membrane integrity of B. cinerea

To investigate the effects of AFPs on B. cinerea, 100 μL of conidial suspension (~ 10⁴ spores/mL) was inoculated into 20 mL PDB and incubated at 25 °C with shaking for 48 h. The resulting mycelia were collected, washed, and resuspended in an equal volume of phosphate-buffered saline (PBS, 0.01 M, pH 7.2). AFPs were added to the suspensions (1/2, 1, and 2 × MIC) and incubated at 25 °C for an additional 3 h, with sterile water as the control. Mycelial morphology and membrane integrity were assessed using a Zeiss fluorescence upright microscope (Zeiss, Germany). For membrane integrity, samples were stained with 0.05 mg/mL propidium iodide (PI, Yeasen, China) for 15 min in the dark before observation.

Protein extraction and Western blotting

Total protein was extracted from plant tissues using a plant protein extraction kit (Solarbio, China). Equal amounts of protein were separated by SDS-PAGE, transferred to PVDF membranes, and blocked with 5% non-fat milk in TBST. Membranes were incubated with anti-GFP antibody (AbinScience, China) overnight at 4 °C, followed by HRP-conjugated secondary antibody for 1 h at room temperature. Protein bands were visualized using an ECL Super Kit (Abclonal, China).

RNA extraction and RT-qPCR

The RT-qPCR method was used to define temporal expression patterns of the target genes. Specific RT-qPCR primers (Additional file 3: Table S3), were designed to identify each gene. The FastPure Universal Plant Total RNA Isolation Kit (Vazyme, China) was used to extract RNA. RT-qPCR was performed on the qTOWER³ touch/qTOWER³ G touch Real-Time PCR Thermal Cycler (analytik jena, GER). The relative expression of target genes in N. benthamiana was determined using the 2^−ΔΔCt method [65, 66]; expression levels of all these genes were normalized to NbActin. Three biological replicates of each sample were analyzed.

Authors’ contributions

X.Y., and X.L. designed this study. X.L. and Y.F. performed the computational analysis. X.L. performed the validation experiments. X.L., X.Y., Y.F. and Y.W. wrote the manuscript. All authors reviewed and approved the final version of the manuscript.

Funding

This work was funded by grants from National Natural Science Foundation of China (Grant No. 32370587, No. 32170581 and 32500568), Shanghai Municipal Education Commission (No. 2024AIYB005), and Shanghai Municipal Science and Technology Commission (No. 25JS2850100 and 25ZR1402261).

Data availability

The FungiGuard codes used in this study are available under the MIT license at the GitHub repository: https://github.com/yulab2021/FungiGuard and at Zenodo: 10.5281/zenodo.18242095 [67, 68]. Published RNA-seq data for Botrytis cinerea infection are available in GEO under accession GSE242932 [58, 69].

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

All authors have read and approved the final manuscript and give their consent for its publication.

Competing interests

The authors declare no competing interests.