Identification of Leaf Rust-Related Gene Signature in Wheat (Triticum Aestivum L.) Using High-Throughput Sequencing, Network Analysis, and Machine Learning Algorithms

Abstract

Wheat provides staple food and industrial raw material for humans and animals, but its production decreased due to leaf rust (Lr) disease caused by Puccinia triticina by up to 15%. It is challenging to identify Lr-associated genes due to the limited sample size and large genome, which hinders the breeding efforts for Lr disease. This study integrated RNA-seq data to mine the candidate genes using meta-analysis, WGCNA, and machine-learning approaches. As a result, 2153 upregulated and 1579 downregulated meta-differentially expressed genes (meta-DEGs) were identified, with four known genes (Lr13 and Lr67/Yr46/Sr55). The meta-DEGs were significantly enriched in antifungal innate immune response, glutathione metabolism, detoxification, phenylalanine, and flavonoid biosynthesis. Among these, 124 resistance (R) genes (~ 85.48% upregulated) were expressed differentially, and ~ 80% belonged to plant pattern recognition receptors (PPRs) that triggered immunity. Likewise, 162 transcription factors (TFs), including WRKY (43), ERF (30), and MYB (33), were associated with Lr disease, and 81 candidate hub genes were co-expressed for Lr. Finally, nine potential candidate genes, including TraesCS7A03G0388400 (BSP), TraesCS1A03G0869900 (PR4), TraesCS6B03G1228800 (AP2/ERF), TraesCS3B03G0088700 (MYB62), TraesCS5A03G1198800 (CYP96A10), and TraesCSU03G0129300 (LTP4), were mined via attribute weighting and machine learning model (XGBoost AUC = 0.97 and accuracy = 0.90) and validated via single-gene model, linear regression, and t-test at p ≤ 0.05. The relative expressions calculated via RT-qPCR assay of nine genes were significantly higher at different time points under Lr infection. Thus, this study reported genes under Lr infection using advanced bioinformatics and supervised machine-learning models, which provide fundamental insight and a solid foundation for understanding the molecular mechanisms of Lr resistance and offer an advanced pipeline for future breeding programs to develop superior cultivars with durable resistance.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12284-025-00839-8.

Introduction

Wheat (Triticum aestivum L.; 2n = 6x = 42; genome size: ~17Gb) is an industrial and major cereal crop, which serves as both a staple food and an industrial raw material, and also plays a crucial role in providing major dietary components for human and animal nutrition. It was domesticated approximately 10,000 years ago (Tanno and Willcox 2006), feeding more than 35% of the global population. Plant pathogens, including bacteria, viruses, and fungi, caused severe yield losses in wheat, averaging 21.5% worldwide (Savary et al. 2019; Zhao and Kang 2023). However, leaf rust (Lr), known as brown rust, is a fungal disease of wheat caused by Puccinia triticina (Pt) species in the family Pucciniaceae, phylum Basidiomycota (Johnson 1987; Liu and Hambleton 2010; Chen et al. 2014), which caused severe epidemic incidents and significant yield losses of 5–15% (Zhou et al. 2022). In China, wheat rust was recorded for the first time 4000 years ago in Gansu during the Shang Dynasty (Wei 2021), whereas its evidence is also found in the Bible and ancient Roman texts (Levine 1919; Chester 1943). Human activities and wind-dispersed spores facilitate its rapid worldwide spread. Molecular studies have reported the presence of identical or closely related strains of wheat rust in different countries, indicating international migration (Kolmer et al. 2019; Bai et al. 2021). Fungicides, early sowing, and crop rotation techniques control the Lr disease, but these practices provide limited control (Singh et al. 2023). Thus, developing rust-resistant wheat cultivars is more cost-effective and environmentally friendly in reducing yield losses (Ash 1996; Wiesner-Hanks and Nelson 2016).

Traditional breeding programs have enhanced wheat resistance to Lr over the last 100 years (Pinto da Silva et al. 2018; Pal et al. 2022; Kumar et al. 2023). However, these methods are laborious, slow, time-consuming, and require lengthy breeding cycles and multi-environment field trials due to the complex nature of Lr resistance, which is governed by multiple genes or quantitative trait loci (QTLs). Thus, marker-assisted selection (MAS) is a high-resolution strategy that links the trait with known markers. In recent years, sequencing costs have decreased, providing the basis for conducting QTL mapping, association mapping, and RNA-seq for specific traits. To date, 296 QTLs from 70 biparental mapping populations have been identified for Lr resistance, and these QTLs were distributed to all 21 chromosomes (Tong et al. 2024). For instance, Kolmer (2015), Gebrewahid et al. (2020), and Genievskaya et al. (2020) identified four, five, and 11 QTLs for Lr, respectively, using biparental populations. The QTLs identified by biparental mapping have large genomic regions, limiting candidate gene identification. Recently, association mapping has gained popularity in genomics due to its use of diverse germplasm and high-density SNP markers to uncover historical meiotic events in ancestral populations (Zhao et al. 2018; Jia et al. 2020). To date, 317 quantitative trait nucleotides (QTNs) have been identified across all 21 chromosomes for rust resistance (Tong et al. 2024). Biparental and association mapping have been widely utilized in many crops for disease traits, such as Zea mays (Kump et al. 2011; Chen et al. 2023), Gossypium hirsutum (Tian et al. 2023; Abdelraheem et al. 2024), Oryza sativa (Li et al. 2022a; Zhu et al. 2023), Glycine max (Lin et al. 2020a, 2020b; Zhao et al. 2015), and Nicotiana tabacum (Lai et al. 2021; Ikram et al. 2022). Therefore, mapping the target genes for Lr resistance in wheat is crucial for enhancing resistance.

The durable Lr resistance in wheat is challenging in plant breeding due to the complex resistance mechanism of race-specific and race-nonspecific (Pinto da Silva et al. 2018). Significant progress has been made in isolating rust resistance genes in wheat for leaf, stripe, and stem rusts. Since the first successful isolation of Lr10 and Lr21 in 2003 (Huang et al. 2003; Feuillet et al. 2003), nine race-specific Lr resistance genes, including Lr10, Lr13, and Lr21, have been cloned, many of which are annotated as nucleotide-binding domain leucine-rich repeat (NLR) immune receptors (Loutre et al. 2009). Race-specific QTLs/genes have been widely used in wheat breeding programs, providing high Lr resistance to a specific pathogen strain through programmed cell death and hypersensitive response (Flor 1956; Loutre et al. 2009; Pinto da Silva et al. 2018). However, rust pathogens are notorious for adaptation and genetic variation in the pathogen population, making these genes ineffective within a few years (Kolmer 2015; Zhao and Kang 2023; Tong et al. 2024), urging breeders to find new genetic resistance sources. In contrast, race-nonspecific genes provide an alternative strategy for durable resistance, and these genes have partial or incomplete resistance, but not all race-nonspecific genes are durable (Tong et al. 2024). For example, Lr34 has provided evidence to be durable and positioned in CIMMYT spring wheat cultivars (Singh et al. 2000, 2011), while three genes, Lr34, Lr67, and Lr34, have partial resistance to stem, leaf, and stripe rust as well as powdery mildew (Singh et al. 1998; Herrera-Foessel et al. 2012; Milne et al. 2019). Two well-characterized pleiotropic genes, Lr34/Yr18/Sr57 and Lr67/Yr46/Sr55, conferred race-nonspecific resistance to multiple pathogens under high disease stress. Lr34/Yr18/Sr57 encodes an ATP-binding cassette (ABC) transporter involved in plasma membrane remodeling and substrate transport, including phytohormone abscisic acid (ABA), which alters ABA-related processes and promotes the accumulation of antifungal phenylpropanoid metabolites (Krattinger et al. 2009, 2019; Deppe et al. 2018; Rajagopalan et al. 2020). Similarly, Lr67/Yr46/Sr55 annotates the sugar transporter STP13, which disrupts glucose transport in infected leaves, limiting nutrient availability for rust pathogens (Moore et al. 2015). These two genes provide durable resistance in wheat and barley (Risk et al. 2013; Sucher et al. 2017; Milne et al. 2019). Thus, a combination of 4–5 loci/genes could provide durable Lr resistance, and it is essential to identify more genes for the molecular breeding program against Lr infection.

RNA-seq technology provides comprehensive information on all genes under specific conditions to detect gene expression and their functional annotation (Zhang et al. 2018). For example, Sharma et al. (2018) conducted RNA-seq analysis to find the Lr28 role in seedling resistance at early and late infection and reported that genes had expression changes against Lr stress. Similarly, Chauhan et al. (2024) reported 924 and 159 differentially expressed genes (DEGs) at different time points, and four potential candidate genes (TraesCS1A02G092000, TraesCS2A02G107500, TraesCS7D02G421400, and TraesCS6A02G409000) were identified for stripe rust resistance. Transcription factors (TFs), including WRKY, bZIP, NAC, MYB, and AP2/ERF, have also been reported, and their expression changes are involved in disease resistance by activating the defense-related genes in many crops (Sharma et al. 2018; Yu et al. 2023; Ma et al. 2024). Numerous transcriptomic studies have investigated the genes involved in rust resistance in wheat (Xu et al. 2011; Wang et al. 2021, 2023; Cai et al. 2023). However, these studies used limited samples and accessions in different biotic stress conditions through traditional methods or a single model, which are insufficient to identify the disease resistance genes. Thus, integrating RNA-seq data from multiple studies through meta-transcriptomics and machine-learning approaches has the potential to uncover defense-related genes and the reproducibility of results. Machine learning has the potential for reliable pattern discovery of genes in diverse and large datasets (Zhu et al. 2024), whereas meta-analysis has high statistical power to integrate the results from different studies for comprehensive and reliable results. However, these approaches have been widely applied in medical research for disease-associated signatures (Kong et al. 2023; Zhao et al. 2024; Zhu et al. 2024). However, to our knowledge, few studies have used advanced machine learning and bioinformatic methods to identify Lr resistance genes in wheat. Thus, combining meta-analysis, machine learning, advanced bioinformatics, and literature mining could expand our knowledge and provide a robust genetic foundation for wheat breeding.

To address this gap, we conducted a uniform analysis of RNA-seq data of Lr samples obtained from the SRA database. The main objectives of this study were as follows: to identify the meta-DEGs between mock and Lr treatment and their functional enrichment for pathways related to disease resistance; to identify the TFs and Plant Resistance Genes database (PRGdb) candidate genes for resistance; to identify the modules and hub genes via weighted gene co-expression network analysis; and to develop supervised machine learning models (logistic regression, XGBoost, random forest, gradient boosting, LightGBM, etc.) for the screening of Lr-associated genes. Following this, we identified four known and nine potential candidate genes for Lr resistance. These candidate genes were validated via an external dataset. Our study provides valuable insight into the effective resistance mechanism and resistance genes of Lr disease for future breeding programs.

Materials and Methods

RNA-seq Data Collection from Different Databases

The raw RNA-seq reads of wheat Lr (P. triticina) were retrieved from the SRA-NCBI database (https://www.ncbi.nlm.nih.gov/sra) and ENA database (https://www.ebi.ac.uk/ena/browser/home) using SRA toolkit v3.1.1. We queried the keywords, such as leaf rust, wheat leaf disease, Puccinia triticina, and biotic stress, in the above databases and obtained metadata based on BioProject IDs. After cleaning the metadata and removing duplicate studies, the wheat samples associated with Lr were manually reviewed to ensure their relevance to the study objectives. We ensured that all selected studies provided essential information, such as plant material details, treatment duration, control conditions, and sequencing information. Following this process, seven studies—PRJEB41456, PRJNA328385, PRJNA588134, PRJNA629995, PRJNA674985, PRJNA838495, and PRJNA718488—were selected based on the above criteria, comprising a total of 112 RNA-seq samples. The external dataset PRJNA718488 was used to validate the results and assess their consistency. The key features of these datasets are summarized in Table S1.

Quality Control and Alignment to the Reference Genome

The quality and reliability of the raw sequencing were assessed using FastQC v0.11.9 (http://www.bioinformatics.babraham.ac.uk/projects/fastqc/). Low-quality reads, defined as those with > 50% of nucleotide bases having a quality score (Q) < 20, as well as reads containing adapter sequences or > 10% unknown bases, were discarded using Trimmomatic v0.39 (Bolger et al. 2014). The raw reads were transformed into clean reads, which were again used to recalculate quality metrics, including Q30, Q20, and GC content, for downstream analysis. The RefSeq v2.1 reference genome and gene annotation files of bread wheat (T. aestivum) were obtained from the International Wheat Genome Sequencing Consortium (IWGSC) database (https://wheatgenome.org/projects/reference-genome-project/refseq-v2.1). The genome index was constructed, and the clean reads were mapped to the RefSeq v2.1 reference genome using STAR v2.7.11b (Dobin et al. 2013). Following alignment, the resulting BAM files were sorted and indexed using SAMtools v1.10 (Danecek et al. 2021). Read count values were generated from the indexed BAM files using featureCounts v2.0.0 (Liao et al. 2014). Transcripts were then reconstructed, and transcript abundance was calculated as Transcripts Per Kilobase Million (TPM) for each sample using StringTie v2.2.1 (Pertea et al. 2015).

Normalization and Batch Effect Correction

Normalized expression levels were measured for 106,914 genes, and genes with low mean expression values were filtered out, resulting in 69,441 genes retained for subsequent analysis. Principal component analysis (PCA) and k-means clustering were performed on the expression profiles of each study using R4.4.1 (http://www.r-project.org/) to assess data quality, demonstrating sufficient variation and clear distribution patterns between mock and treatment samples. To account for batch effects arising from differences in library preparation methodologies and sequencing platforms, surrogate variable analysis (SVA) was conducted using the ComBat function from the SVA package in R (Leek et al. 2012). The parametric empirical Bayes method adjusted for non-biological effects while retaining meaningful biological variation across the combined datasets.

Meta-Analysis for Differentially Expressed Genes

We employed two complementary approaches to identify positive or negative differentially expressed genes and enhance the reliability and consistency of the analysis for signature genes associated with Lr disease resistance. Firstly, we utilized batch effect correction via the limma R package (Ritchie et al. 2015) for differential expression analysis between mock and treatment samples. The significant DEGs were identified based on the following criteria: |log2FC| ≥ 0.70 and a false discovery rate (FDR) or adjusted p ≤ 0.05. Secondly, pairwise comparisons were performed between mock and treatment within each study using the DEGseq R package based on |log2FC| >1 and p ≤ 0.05 (Wang et al. 2010). Further, meta-analysis was performed using Fisher’s combined probability test (Fisher 1992) and implemented in the metaRNASeq v1.0.7 to integrate results across studies (Figs. S1-S7). Raw p-values were adjusted using the FDR method. Genes with |log2FC| >1, an adjusted p ≤ 0.05, and consistent expression direction across most studies were considered statistically significant. The results were visualized using R packages, including ggplot2 for volcano plots, pheatmap for heatmaps, and dplyr and ggrepel for data processing and annotation.

Weighted Gene Co-expression Networks Analysis

In biology, WGCNA is a robust approach with high-resolution power to detect co-expressed genes and identify associations between gene networks and traits. This study performed WGCNA to identify co-expressed gene modules associated with Lr resistance using the WGCNA in R4.4.1 (Langfelder and Horvath 2008), an unsupervised method for highly correlated gene networks. Following the criteria, the top 25% of genes with the highest expression variance were chosen for further analysis. Genes and samples that did not meet the requirements of the goodSamplesGenes function were removed. A similarity matrix was constructed using Pearson’s correlation coefficients for all gene pairs, and an adjacency matrix was generated. A soft threshold power (β = 30) was chosen to achieve a scale-free topology fit (R²). Hierarchical clustering was calculated using the topological overlap matrix (TOM) to identify the group genes with similar expression profiles. The minimum module size was set to 30 genes, and a cut height threshold of 0.25 was applied to identify highly correlated gene modules. The dissimilarity between module eigengenes was calculated to refine the modules, and modules with a similarity > 0.70 were merged. This step involved selecting a cut-off point on the module dendrogram to combine closely related modules for further analysis. Furthermore, module-trait correlation analysis was performed by calculating the correlation between module eigengenes (MEs) and traits, and modules were selected based on a threshold of correlation r > 0.05 and p ≤ 0.05. The extracted significant modules were exported to Cytoscape (Otasek et al. 2019), and the hub genes were identified using the CytoHubba plugin (Chin et al. 2014) based on the Degree method. The gene interaction networks were visualized within Cytoscape.

Functional Enrichment Analysis

The eggnog-mapper database (Huerta-Cepas et al. 2019) was utilized for genome-wide functional annotation as follows: first, the protein sequences of all wheat genes were uploaded for functional annotation using default parameters; secondly, an in-house R script was implemented to construct an orgDB package. Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses were performed to identify functional annotations and pathways associated with DEGs, candidate genes, and hub genes. The GO database (https://www.geneontology.org/) and KEGG database (https://www.genome.jp/kegg/) were accessed to determine significant GO terms and metabolic pathways at an adjusted p ≤ 0.05 (Ogata et al. 1999). All enrichment analyses and visualizations of top GO terms and KEGG pathways were performed using the clusterProfiler package in R4.4.1 (Yu et al. 2012).

Attribute Weighting Models

In meta-analysis, attribute weighting techniques are powerful in assigning significance scores to genes based on their biological relevance and contribution to the target phenotype. This method enhances feature selection and improves model accuracy (Zhao et al. 2024). An integrated dataset from six studies, including gene expression data and experimental details such as cultivar type, time point, and condition, was used as numerical features for this analysis. All features (genes + study features) were evaluated using attribute weighting models, including weight by information gain, weight by rule, weight by relief, and weight by information gain ratio, implemented in Python v3.12. The weights from each model were normalized, with 1 indicating high importance and 0 indicating low importance of a gene concerning disease response. Genes with a cumulative weight ≥ 3.00 were selected as significant signature genes.

Machine Learning and Deep Learning Models for Candidate Gene

Machine learning (ML) and deep learning (DL) models were employed to validate selected feature genes and identify potential candidate genes associated with Lr resistance. The input data for the models included gene IDs, sample IDs, binary class labels indicating control and disease samples, and corresponding gene expression levels. The genes identified through attribute weighting techniques were used as input features for machine learning models to predict potential candidate genes. Six ML models—Random Forest (RF), XGBoost, LightGBM, Gradient Boosting, Logistic Regression (LR), and Support Vector Machine (SVM)—along with one DL model (Neural Network), were implemented for feature selection (Cox 1958; Rumelhart et al. 1986; Cortes and Vapnik 1995; Breiman 2001; Friedman 2001; Chen and Guestrin 2016; Lundberg and Lee 2017; Singh 2021). The dataset was split into a 70:30 ratio for training and testing sets, with a random seed 42 for reproducibility. The models were trained on the training dataset with hyperparameter tuning via gridSearch to identify optimal parameters using five-fold cross-validation. The hyperparameters were set as subsample = 0.8, min_child_weight = 1, max_depth = 8, min_samples_split (2), learning_rate = 0.02, n_estimators = 100, colsample_bytree = 1.0, and eval_metric=’logloss’. Model performance was evaluated using key metrics, including the Area under Curve (AUC) or Receiver Operating Characteristic (ROC) curve, Accuracy, Specificity (True Negative Rate, TNR), Recall (True Positive Rate, TPR), Precision (Positive Predictive Value, PPV), F1-score, and Negative Predictive Value (NPV). All models were implemented in Python v3.12 and executed on a Linux-based server with an Intel Xeon Gold 6138 processor (40 cores and 80 threads) and 256 GB of RAM. SHapley Additive exPlanations (SHAP) analysis was conducted to interpret the black-box nature of these models (Lundberg and Lee 2017). SHAP bar plot was used to rank the top 20 genes by mean absolute SHAP values, highlighting their contribution to Lr resistance prediction. Waterfall plots illustrated how individual genes influenced prediction in each sample.

Potential Candidate Genes Verification

For verification, a single-gene-based machine learning model (XGBoost) was trained on 80% of the dataset, while the remaining 20% was used to evaluate the performance using the AUC and accuracy metrics. Five-fold cross-validation was applied to each potential candidate gene to ensure robustness. The analysis was implemented in Python v3.12. An external dataset (PRJNA718488) was utilized for further validation through the linear model and correlation analysis (p ≤ 0.05) to assess the consistency of potential candidate genes. Furthermore, normalized expression values were analyzed to identify significant differences between mock and treatment conditions for each gene using a student’s t-test at a significance threshold of p ≤ 0.05.

Plant Materials and RT-qPCR Analysis of Candidate Genes

Wheat cultivar Ujala-16, resistant to Lr, was used for the experiment. Seeds were surface-sterilized by soaking in 70% ethanol for 1 min, followed by 2% sodium hypochlorite for 5 min, and rinsed seven times with distilled water. Approximately 20 sterilized seeds were planted in a pot (10 cm diameter) in a growth chamber at 20 ± 2 °C with a 16 h light/8 h dark photoperiod, 60% relative humidity. Plants were grown for 14 days (two-leaf stage) before Lr inoculation. The THTT strain of fungal pathogen-caused Lr was maintained on susceptible wheat plants in a separate growth chamber. The water-based spore suspension was sprayed onto the leaves of 14-day-old plants. Control plants (mock treatment) were sprayed with sterile water. Inoculated plants were incubated at 15 °C and 100% relative humidity for 24 h in the dark to promote infection. Leaf samples were collected at 0 h, 24 h, 48 h, 72 h, and 96 h from three biological replicates per time point. The leaf samples were frozen in liquid nitrogen and stored at − 80 °C until RNA extraction.

For the RT-qPCR assay, total RNA was extracted using a Qiagen RNA extraction kit according to the manufacturer’s protocol. A Takara reverse transcription kit was used for the synthesis of first-strand cDNA. The qPCR was conducted in a 10 µL reaction volume containing 5 µL UltraSYBR Mixture (CWBIO), 0.4 µM of each gene-specific primer, and 2 uL of 1:20 diluted cDNA. Amplification was performed on a CFX96 Manager instrument (Bio-Rad) under the following conditions: 95 °C for 3 min; 39 cycles of 95 °C for 10 s, 60 °C for 10 s, and 72 °C for 10 s; followed by a melt curve analysis at 95 °C for 20 s, 60 °C for 1 min, and 95 °C for 15 s. The TaActin (GenBank accession: AB181991) gene was used for normalization as the internal reference gene (Paolacci et al. 2009). Gene-specific primers were designed using the qPrimerDB database (https://biodb.swu.edu.cn/qprimerdb/), and primers are listed in Table S2. The relative expression of potential candidate genes was calculated using the 2^−ΔΔCt method (Livak and Schmittgen 2001). Three biological replicates were included per sample. Expression data were analyzed using one-way analysis of variance (ANOVA), followed by post-hoc Tukey’s test to determine the significant differences between time points (p ≤ 0.05) via an in-house R (v4.3.0) script and the ggplot2 package.

Results

Integrated RNA-seq Dataset of Different Studies

Seven studies, six for meta-analysis and one for validation, were integrated for meta-analysis of RNA-seq, and these studies had 112 samples, including 41 mock and 71 Lr infections (Table S2). The average raw reads were 34.86 million, with 33.23 million cleaned reads. The mapped reads with reference genome were 31.02 million, and the alignment rate ranged from 73.70 to 98.73%, averaging 93.70% (Table S2). After preprocessing, 69,441 genes associated with the response to Lr infection were identified and included in downstream analysis (Fig. 1).

Fig. 1.

The flowchart illustrates the workflow of all processes, including data preprocessing, differentially expressed genes, co-expressed genes network, supervised models, machine learning approaches, and validation of potential candidate genes in this study

Identification of Consistently Differentially Expressed Genes

In this study, we quantified the expression profiles of 69,441 coding genes using the wheat RefSeq2.1 reference genome. After preprocessing and adjusting for non-biological batch effects, the normalized gene expression data from six studies were used to identify differentially expressed genes (Fig. 1). A meta-analysis revealed 335 DEGs, comprising 210 upregulated and 124 downregulated genes (Fig. 2a). Further, we employed an alternative meta-analysis approach using Fisher’s combined probability test to integrate results from multiple studies and identify consistent gene expression patterns across studies (Figs. S1–S7). Ultimately, 3732 meta-DEGs were identified, including 2153 upregulated and 1579 downregulated genes, based on the criteria of FDR ≤ 0.05 and |log2FC| ≥ 1 (Fig. 2b–d and Table S3). The gene expression patterns were almost consistent in both approaches, and Fisher’s combined probability test had a higher resolution power (detected high number of DEGs) than the meta-analysis via limma (Fig. 2). However, the heatmap revealed a consistent pattern of upregulated and downregulated DEGs across samples and studies (Fig. 2e).

Fig. 2.

Gene expression profile and differentially expressed genes associated with leaf rust disease. a, b volcano plots of DEGs identified using limma and Fisher’s combined probability test, respectively; c Venn plot showed common genes between two approaches; d Number of upregulated and downregulated genes using meta-analysis; e Heatmap represents expression patterns of top 50 upregulated and top 50 downregulated genes in resistant and susceptible cultivars; f Top pathways in KEGG enrichment analysis at adjusted p ≤ 0.05; and g GO enrichment analysis of meta-DEGs at adjusted p ≤ 0.05

Moreover, among meta-DEGs, four genes—TraesCS2B03G0442800 (Lr13/Yr27), TraesCS4A03G0132000 (Lr67/Yr46/Sr55), TraesCS4B03G0663100 (Lr67/Yr46/Sr55), and TraesCS4D03G0585200 (Lr67/Yr46/Sr55)—were previously known genes for Lr disease resistance (Fig. 3 and Tables S3). A total of 162 TFs associated with disease resistance were identified from six different gene families, such as bHLH (19), four bZIP (4), MYB (33), AP2/ERF (30), WRKY (43), and NAC (33) (Fig. 3a–c and Table S4). Among 43 WRKY, 41 were upregulated, two were downregulated, and 20 NAC TTs were upregulated (Fig. 3a and Table S4). In addition, 124 R-genes were identified using the Pathogen Receptor Genes (PRGs) database (Fig. 3d, e and Table S5), which belonged to major classes of kinases (KIN) domain, pattern recognition receptors (PRRs)/ receptor-like protein (RLP), and nucleotide-binding leucine-rich repeat (NLR) receptors/RLK. Among 124 genes, 106 were upregulated (~ 85.48%), and 18 were downregulated (~ 14.56%). Of these, 71 genes belonged to the KIN class, followed by the RLP class (14), TRAN class (6), RLK (5), while 28 belonged to the N, NL, C, CK, CNL, and CL minor classes (Fig. 3e and Table S5).

Fig. 3.

Heatmaps of resistant genes and transcription factor families involved in leaf rust disease resistance in wheat. The upregulated and downregulated genes TF families in resistant and susceptible cultivars: a WRKY; b AP2/ERF, bHLH, and bZIP; c MYB and NAC. d, e The R-genes showed differential expression of different pattern recognition receptor (PRRs) classes. f Four previously known genes have been reported for Lr disease

Functional Enrichment Analysis of Meta-DEGs

The biological interpretation of upregulated and downregulated genes derived from meta-analysis was used for GO and KEGG enrichment analyses. GO enrichment analysis revealed that 354 upregulated and 348 downregulated meta-DEGs were significantly enriched in GO terms (adjusted p ≤ 0.05), providing functional annotations of the genes to understand better their biological roles (Fig. 2f and Table S6). In brief, upregulated meta-DEGs were primarily overrepresented in processes such as antifungal innate immune response, photosynthesis, regulation of antifungal innate immune response, oxidoreductase activity, circadian rhythm, defense response to insect, and sucrose biosynthetic process. In comparison, downregulated genes were enriched in terms related to the lignin biosynthetic process, lipid transport, carbohydrate transport, glutathione metabolic process, regulation of phenylpropanoid metabolic process, flavone biosynthetic process, and detoxification (Fig. 2f and Table S6). Additionally, terms such as response to hydrogen peroxide, positive regulation of response to salt stress, negative regulation of circadian rhythm, cellular response to sulfate starvation, and regulation of leaf senescence were simultaneously identified among the top 30 terms for upregulated and downregulated genes (Fig. 2f and Table S6). KEGG enrichment analysis revealed that upregulated genes were mainly enriched in pathways (adjusted p ≤ 0.05), such as linoleic acid metabolism, phenylalanine metabolism, zeatin biosynthesis, steroid biosynthesis, and galactose metabolism pathways. In contrast, downregulated genes were enriched in 10 pathways (adjusted p ≤ 0.05), including photosynthesis, circadian rhythm–plant, carotenoid biosynthesis, stilbenoid, diarylheptanoid and gingerol biosynthesis, nitrogen metabolism, and flavonoid biosynthesis (Fig. 2g and Table S7).

Construction of Co-expressed Gene Modules

We conducted a weighted co-expression network analysis to obtain modules significantly associated with disease responsive using resistant and susceptible datasets (Figs. 4 and S8). As a result, the soft threshold β was 12 for both resistant and susceptible groups, with correlation coefficients of 0.87 and 0.86, respectively, confirming a scale-free network relevant to gene expression data (Fig. 4a-h). We identified 16 and 17 co-expression modules in the resistant and susceptible groups, respectively, using dynamic tree-cutting and merging modules at a threshold of similarity > 0.70 (Fig. 4a-h). Further, TOM was utilized to identify the gene modules’ association with disease resistance features, with the grey module representing genes not assigned to any module. Among the 16 modules in resistance group, light-yellow (r = 0.57; p = 2.00 × 10^− 4), magenta (r = 0.63; p = 8.00 × 10^− 4), yellow (r = 0.55; p = 0.002), and black (r = 0.59; p = 0.02) modules showed the highest correlation with disease resistance, containing a total of 4,216 genes (light-yellow: 67, magenta: 561, yellow: 1,234, black: 2,351) (Fig. 4i-j). Similarly, in the susceptible group, light-yellow (r = 0.54; p = 0.03), brown (r = 0.61; p = 0.01), and black (r = 0.57; p = 0.002) modules demonstrated highly associated with disease, containing a total of 3,538 genes (light-yellow: 71, brown: 1,043, black: 2,424) (Fig. 4i-j and Table S8). We selected these seven modules for further downstream analysis to identify hub genes and explore their functional roles and regulatory mechanisms.

Fig. 4.

Weighted gene co-expression network analysis of resistant and susceptible cultivars. a–d Soft threshold power and cluster dendrograms indicate the colors of the merged module based on Topological Overlap. Each row with a different color shows a group of highly connected genes. e, f The trait-modules relationship between co-expressed modules and disease resistance. The Figures exhibit black, light-yellow, magenta, and yellow modules significantly associated with disease resistance in resistance cultivars, while brown, light-yellow, and black modules are highly associated with Lr in susceptible cultivars. g, h Number of genes distributed in each significant module. Red indicates a positive correlation, and blue indicates a negative correlation with significance level at p ≤ 0.05 = *; p ≤ 0.01 = **; and p ≤ 0.001 = ***

Hub Genes Associated with Leaf Rust and Functional Enrichment Analysis

We subjected seven significant resistant and susceptible group modules to identify the top 30 hub genes (Fig. 5 and Table S9). In the resistant group, the black module contained 21 genes from the ribosomal protein family and five ubiquitin-related genes, while the light-yellow module included eight pathogenesis-related (PR) hub genes (Fig. 5a–d and Table S9). The magenta module comprised nine significant hub genes, including CYP721A1; TraesCS7D03G0485000, annotated as a leucine-rich repeat protein kinase family protein; TraesCS3A03G1130300 and TraesCS3B03G1306400, both annotated as pathogenesis-related 2 (PR2); three genes (TraesCS5A03G0649100, TraesCS5D03G0615200, and TraesCS3D03G0953100) associated with wall-associated kinases 3 (WAK3); and TraesCS5D03G0884400, annotated as a disease resistance-responsive gene. The yellow module identified hub genes, including TraesCS3B03G0429800 (CYP71A23) and TraesCS5A03G1198800 (CYP96A10), which belong to the cytochrome P450 family (Fig. 5a–d and Table S9). Additionally, it contained three cytokinin oxidase-related genes (TraesCS3A03G0238600, TraesCS3B03G0305700, and TraesCS3D03G0230100), two MYB TFs (TraesCS3B03G0088700 and TraesCS3D03G0065300), and TraesCS3A03G0651000, annotated as a nitrate transporter 2.5 (NRT2.5). Similarly, in the susceptible group, the brown modules contained hub genes as TraesCS6B03G1228800, TraesCS5A03G0123600, TraesCS5B03G0456000, TraesCS4A03G0646200, and TraesCS1B03G0574600 (GSTU18), and these genes were annotated as follows: AP2/B3-like transcriptional factor family protein, flavonol synthase/flavanone 3-hydroxylase, heavy metal transport/detoxification superfamily protein, auxin-associated family protein, and glutathione S-transferase TAU 18, respectively (Fig. 5e–g and Table S9). In the black module, genes—TraesCS3B03G1129700 (GRP3), TraesCS3A03G0980900 (GRP4), and TraesCS3D03G0916100 (GRP4)—were identified as members of the glycine-rich RNA-binding protein family, which is known to play a significant role in stress responses in different crops (Fig. 5e–g and Table S9). The light-yellow module contained 22 genes annotated as lipid transfer proteins (LTPs), and these proteins are involved in plant defense mechanisms, acting as pathogenesis-related proteins against bacteria, fungi, and viruses (Fig. 5e–g and Table S9). Among these hub genes, 89 were interconnected with meta-DEGs, including five downregulated and 84 upregulated genes, and mapped within module networks along with their expression values (Fig. 5 and Table S9).

Fig. 5.

Visualization of WGCNA network connections identifies hub genes associated with leaf rust resistance in wheat. Hub genes in modules, such as black a, light-yellow b, magenta c, and yellow d, show the top 30 hub genes in resistant cultivars, while black e, brown f, and light-yellow g in susceptible cultivars. h, i GO and KEGG enrichment analysis of hub genes identified for disease resistance in wheat. Candidate genes, marked in each module, were identified based on comparative genomics and literature mining. The color of each gene circle is based on log₂FC values, where red indicates upregulated and blue indicates downregulated genes

Furthermore, selected hub genes from seven modules were used for GO and KEGG enrichment analysis to elucidate their potential functional characteristics. GO enrichment analysis revealed significant enrichment in BP terms such as lipid transport and localization, CC terms including cytosolic ribosome, cytosolic large ribosomal subunit, and large ribosomal subunit, and MF terms related to purine nucleobase transmembrane transporter activity, ATP binding, adenyl ribonucleotide binding, and RNA helicase activity (Fig. 5i). Similarly, KEGG pathway analysis highlighted enrichment in ribosome biogenesis, zeatin biosynthesis, galactose metabolism, and carotenoid biosynthesis (Fig. 5h). Thus, these genes showed higher expression profiles in response to Lr disease, suggesting that they may play a significant role in biotic stress responses in wheat.

Top-Ranked Genes Based on Attribute Weighting

Different attributing weight models were employed on all features (genes + study characteristics) using normalized data, and the features were ranked based on the sum of weight scores. Thus, four attribute weighting models (Info Gain, Info Gain Ratio, Rule, and Relief) detected top-ranked genes that discriminate disease from control samples. Finally, 4,741 gene signatures were identified for disease resistance at a threshold sum of weights ≥ 3.00. Further, we interconnected these genes with meta-DEGs and co-expressed modules hub genes and identified 79 key genes associated with disease response (Table S10). These findings provide evidence for the consistency of transcriptomic data, demonstrating similar gene expression profiles across multiple studies.

Mining of Disease Resistance Hub Genes Through Machine Learning

One hundred twenty-nine genes, including 79 common between meta-DEGs, co-expressed modules hub genes, and attributes weight, and 89 candidates between meta-DEGs and hub genes, were incorporated into seven supervised machine learning and deep learning models. Among the seven models, XGBoost demonstrated superior performance in feature selection, with an AUC of 0.97, accuracy of 0.90, recall (TPR) of 1, specificity (TNR) of 0.75, precision (PPV) of 0.86, and F1-score of 0.92, followed by gradient boosting, LightGBM, RF, LR, SVM, and neural network (Fig. 6a; Table 1). Furthermore, the SHAP analysis was performed for feature selection due to the Blackbox nature of these models, and the top 20 candidate genes were selected with the highest impact on model prediction (Fig. 6b and Table S11). These genes were also identified as interconnected in a prior analysis of this study and regarded as candidate hub genes. Among these, MYB62, AP2/ERF, cytochrome P450, NRT2.5, PR4, and LTP4 exhibited the highest mean absolute SHAP values (Fig. 6b and Table S11). The machine learning model predicted the true positive class due to the higher expression patterns in treatment, and the true negative class predicted a lower expression pattern in the control condition. For example, the genes in Fig. 6c-d, including TraesCS4D03G0450000, had lower expression in control and higher in treatment conditions, pushing the model to predict their actual class (Figs. 6c-d and S9-10). Thus, attribute weighting-derived transcriptomic genes significantly increased the prediction performance of models to develop the strategy for identifying Lr-responsive candidate genes.

Fig. 6.

Analysis of supervised machine learning models to identify the patterns and candidate genes related to leaf rust. a The AUC graph demonstrates the performance of seven machine learning models, and XGBoost represents higher performance. b The top 20 feature genes were ranked based on SHAP values, which pushed the model for the prediction of Lr disease resistance. c, d Waterfall treatment plots and mock individual samples represent how machine learning predicted the outcomes. e List of potential candidate genes based on feature relative importance, indicating the highest predictive values, which were biomarkers for Lr

Table 1.

Performance comparison of seven machine-learning and deep-learning models for identifying leaf rust-responsive candidate genes

Model	Accuracy	Recall (TPR)	Specificity (TNR)	Precision (PPV)	Negative predictive value (NPV)	F1-score
Logistic Regression	0.70	0.75	0.63	0.75	0.63	0.75
Random Forest	0.80	0.92	0.63	0.79	0.83	0.85
Gradient Boosting	0.85	0.92	0.75	0.85	0.86	0.88
SVM	0.60	0.58	0.63	0.70	0.50	0.64
LightGBM	0.80	0.92	0.63	0.79	0.83	0.85
XGBoost	0.90	1.00	0.75	0.86	1.00	0.92
Neural Network	0.70	0.83	0.50	0.71	0.67	0.77

Literature Mining and Potential Candidate Genes Associated with Leaf Rust

The 20 candidate genes identified via machine learning were further investigated for functional annotation and subjected to literature mining to identify potential candidates. Based on gene annotation, comparative genomics, PRGdb, enrichment analysis, and literature review, nine potential candidate genes were identified as directly or indirectly involved in wheat Lr resistance (Fig. 6e; Table 2). For example, TraesCS1A03G0869900 (PR4) belongs to the pathogenesis-related protein family, and these genes are activated in response to different biotic threats and are involved in antifungal activity to protect the crop plant from pathogens (Zribi et al. 2021). Similarly, two TF genes, TraesCS6B03G1228800 (AP2/ERF) and TraesCS3B03G0088700 (MYB62), were identified as potential candidate genes due to their association with plant defense mechanisms against pathogens (Yu et al. 2023; Ma et al. 2024). TraesCS5D03G0884400 belongs to the disease resistance-responsive (dirigent-like protein) family, a well-known family associated with disease resistance (Yadav et al. 2021). The remaining genes, TraesCS5A03G1198800 (CYP96A10), TraesCS7A03G0388400, TraesCS7D03G0485000, TraesCS2B03G1167400 (LTP), and TraesCSU03G0129300 (LTP4), were detected as putative candidate genes, and their functional annotations as cytochrome P450 (family 96, subfamily A, polypeptide 10), plant basic secretory protein (BSP) family protein, leucine-rich repeat protein kinase family protein, and lipid transfer protein, respectively (Table 2). The relative importance of these nine genes ranged from 1.66 to 5.05%, suggesting their potential role in disease resistance under stress conditions compared to normal conditions (Fig. 6e; Table 2). However, further molecular or functional validation of these nine potential candidate genes is required to elucidate their fundamental roles in regulating disease resistance.

Table 2.

List of potential candidate genes involved in leaf rust disease resistance in wheat

Gene id	Sum weight	Meta log2FC	Meta FDR	Regulation	Gene name	Annotation
TraesCS2B03G1167400	3.38	2.34	3.3E-95	Up	LTP	Lipid-transfer protein
TraesCS3B03G0088700	3.64				MYB62	myb domain protein 62
TraesCS1A03G0869900	3.03	3.83	0.0E + 00	Up	PR4	Pathogenesis-related 4
TraesCSU03G0129300	3.41	1.04	0.0E + 00	Up	LTP4	Lipid transfer protein 4
TraesCS5A03G1198800	3.64	1.05	3.3E-46	Up	CYP96A10	Cytochrome P450, family 96, subfamily A, polypeptide 10
TraesCS6B03G1228800	3.53					AP2/B3-like transcriptional factor family protein
TraesCS7A03G0388400	3.34	2.31	1.8E-225	Up		Plant basic secretory protein (BSP) family protein
TraesCS7D03G0485000	3.24	1.15	1.1E-83	Up		Leucine-rich repeat protein kinase family protein
TraesCS5D03G0884400	3.22	2.38	1.8E-203	Up		Disease resistance-responsive (dirigent-like protein) family protein

Validation of Key Disease-Responsive Genes

To verify the nine potential candidate genes, we employed the XGBoost model to identify their performance under mock and Lr inoculum conditions. We trained the model and evaluated its performance on the test dataset via a single genes model (Fig. 7). The model showed expression differences in the nine genes, with AUC values ranging from 0.67 to 0.97 and accuracy values ranging from 0.60 to 0.95, and three genes, TraesCS6B03G1228800, TraesCS7A03G0388400, and TraesCS1A03G0869900 (PR4), had the highest accuracy (Fig. 7a-b). These results are similar to combined gene models (AUC = 0.69 to 0.97), indicating the reliability of these potential candidate genes. The external dataset was used to further validate these nine genes through linear regression and correlation analysis (Fig. 7c). The results revealed high correlation coefficients between meta-analysis and susceptible genotype (r = 0.62, P = 0.011), meta-analysis and resistant genotype (r = 0.68, P = 0.001), and resistant and susceptible (r = 0.80, P = 0.001), demonstrating strong consistency in gene expression (Fig. 7c). Additionally, the normalized expression of these genes showed significant differences between mock and stress conditions at a p ≤ 0.05 (Fig. 7d).

Fig. 7.

Validation of potential candidate genes using single-gene model and external dataset. a, b Single-gene model determines the AUC and accuracy values of each gene. c The expression profile values of potential candidate genes were plotted as log₂FC using linear regression analysis. d The comparison of normalized expression of each gene between mock and treatment using the student t-test at p ≤ 0.05

Moreover, we performed RT-qPCR analysis of nine potential candidate genes, revealing significant temporal gene expression changes across five-time points (0 h, 24 h, 48 h, 72 h, and 96 h) under Lr infection (Fig. 8). Among these genes, most exhibited a significantly high expression profile at 24 h compared to mock and other time points (Fig. 8a–i). For example, TaPR4 and TaCYP96A10 had higher relative expression at 24 h. Three genes, TraesCS7A03G0388400 (plant basic secretory protein family protein), TaLTP4, and TraesCS7D03G0485000 (leucine-rich repeat protein kinase family protein), showed stable expression profiles at all time points compared to mock (Fig. 8g–i). TraesCS5D03G0884400 encodes a disease resistance-responsive (dirigent-like protein) family protein with higher expression at 24 h, 48 h, and 72 h compared to 96 h. Further, correlation analysis was conducted between RNA-seq and RT-qPCR, which revealed a high correlation coefficient, suggesting the consistency of RNA-seq data and analysis (Fig. 8j). Thus, findings indicate that these candidate genes might be involved in Lr resistance, and their predictions highlight their importance for wheat disease resistance and molecular breeding programs.

Fig. 8.

Confirmation of potential candidate genes identified via machine learning models in response to leaf rust resistance through RT-qPCR. a–i Candidate genes showed different expression profiles, and letters on the bars indicate significant differences between time points based on Tukey’s post-hoc test following a one-way analysis of variance at p ≤ 0.05. Bar plots represent the mean ± SD based on three replicates (n = 3). j The RNA-seq and RT-qPCR log₂FC values were compared using correlation analysis and showed a significant positive association

Discussion

Leaf rust (Lr) is a devastating fungal disease from centuries that has significantly decreased wheat yield by up to ~ 15% (Zhou et al. 2022) and seriously threatens global food security. The world population is increasing yearly, and global demand for food also increases (Godfray et al. 2010), putting pressure on wheat as a staple food worldwide. Thus, Lr resistance cultivars that showed durable resistance for different strains urgently need to mitigate yield losses and ensure sustainable food production. In recent years, sequencing technologies and bioinformatics advancements have improved the resolution power to identify the key genes or markers linked to important traits, and these markers provide an efficient and precise approach compared to traditional methods for molecular breeding programs. Thus, in this study, we integrated data from RNA-seq experiments with corrected batch effect and employed advanced machine learning models, meta-analysis approaches, genes co-expressed analysis, and literature mining to detect the potential candidate genes responsive to Lr disease resistance. Through these comprehensive approaches, 3,732 DEGs (2153 up- and 1,579 down-regulated), 162 differentially expressed TFs (Figs. 2 and 3 and Table S3), 124 R-genes (Fig. 3 and Table S5), 81 hub genes (Figs. 4 and 5 and Table S3), and nine potential candidate genes (Predictive accuracy: ~75.55) (Fig. 6) were identified, and also enriched pathways, including flavonoid biosynthesis, nitrogen metabolism, phenylalanine metabolism, photosynthesis, steroid biosynthesis, and stilbenoid, diarylheptanoid and gingerol biosynthesis related to biotic stress (Fig. 6), and these genes are involved in resistance that could be key markers to develop cultivars for Lr resistance. The identified genes exhibited differential regulation in response to different biotic stresses such as Lr, strip rust, powdery mildew, head blight, etc., indicating that these genes might be involved in multiple stress responses. This study provides s cross-talk for different stress regulators and logical for researchers to implement the strategies for crop improvement. The potential candidate genes were validated through a single-gene model, a linear regression model, and an external expression dataset, which indicates their reliability and reproducibility.

Global Pattern of Transcription Under Leaf Rust Infection

This study identified 3,732 meta-DEGs, such as 2,153 upregulated and 1,579 downregulated, through combined approaches (Fig. 2 and Table S3). Among these, TraesCS2B03G0442800 encodes NBS-type disease resistance protein, previously known as Lr13, and has been reported for Lr resistance in wheat (Yan et al. 2021). Three genes—TraesCS4A03G0132000, TraesCS4B03G0663100, and TraesCS4D03G0585200 belonged to the transporter family protein and have been published in wheat (Lr67/Yr46/Sr55/Pm46/Ltn3) for Lr resistance (Moore et al. 2015). These results indicate that our trimming, genome alignment, and meta-analysis pipeline achieved high accuracy, as evidenced by the identification of known genes, which validated the reliability of our approach.

Pathogen Recognition-Related Genes for Leaf Rust

Uncovering the Lr R-genes or defense-related genes is challenging in wheat due to its large ~ 17 Gb and polyploid genome size (Wu et al. 2019). This study detected 124 differentially expressed putative R-genes using the PRG database (Fig. 3 and Table S5), and these genes play a role in recognizing proteins expressed by avirulence genes of pathogens (Sanseverino et al. 2010). These 124 (106 up- and 18 down-regulated) identified genes were categorized into PRG classes, such as KIN, RLP, CNL, TRAN, RLK, etc. Almost ~ 80% of upregulated genes were identified as plant pattern recognition receptors (PPRs) from RLKs and RLPs classes that regulated pattern-triggered immunity (PTI) via microbe/pathogen-associated molecular patterns (MAMP/PAMP) (Chisholm et al. 2006; Boller and Felix 2009; Tor et al. 2009). For example, CRK18, CRK23, CRK25, CRK26, CRK34, and CRK40 encode receptor-like kinases with serine/threonine or cysteine-rich kinase domains, and their homologous genes have been characterized in rice and tomato for Cladosporium fulvum (KRUIJT et al. 2005; Tang et al. 2010), grapevine for multi-biotic stress (Di Gaspero and Cipriani 2003), and A. thaliana for Pseudomonas syringae (Zhang et al. 2013). Similarly, the RLK10, as Lr10, has been reported in wheat (Feuillet et al. 1997). The RNA-seq study of Lactuca sativa revealed several differentially expressed RLK genes against Botrytis cinerea (DE CREMER et al. 2013).

Thirteen wall-associated kinases (WAKs) genes (Fig. 3 and Table S5), including WAK3, WAK32, WAK4, WAK5, WAK55, TraesCS5A03G0623200 (WAK80), and TraesCS5A03G0649100 (WAK84), were detected as damage-associated molecular patterns (DAMPs) such as oligogalacturonides (OGs). The WAKs genes were involved in biotic stress in A. thaliana (He et al. 1998; Brutus et al. 2010), such as WAK1 controlled the blast in rice (Li et al. 2009), nematode infection in potato (Chen et al. 2022), and downy mildew in Brassica rapa (Zhang et al. 2023). In (A) thaliana, the brassinosteroid insensitive 1 (BRI1) associated receptor Kinase 1 (BAK1) mutant was identified as a negative regulator of (B) cinerea (Kemmerling et al. 2007). Four SERK genes were differentially expressed in this study, which enhances PTI via brassinosteroid signaling and pathogen recognition (Lin et al. 2020a, 2020b). TraesCS2B03G1206300, homologous to AT5G06740, may bind carbohydrate elicitors to enhance LR resistance similar to CaLecRK-S.5 in pepper (Woo et al. 2016). These findings suggest that upregulation of these PPR genes might be involved in early resistance against Lr infection (Fig. 3 and Table S5), and functional validation via overexpression or gene silencing could elucidate their specific pathogen targets and signaling partners.

Transcription Factors Involved in Disease Resistance

TFs involve Lr disease resistance and responses to biotic and abiotic stresses (He et al. 2005; Lu et al. 2011; Chen et al. 2023). For example, MYB, bHLH, AP2/ERF, WRKY, NAC, and GRAS TFs were previously identified for A. tenuissima infection (Li et al. 2014) and biotic stress in grapevine (Zinati et al. 2024). Our study identified 30 AP2/ERF, four bZIP, 19 bHLH, 33 NAC, 33 MYB, and 43 WRKY TFs, with ~ 70% upregulated, indicating their potential role in the Lr resistance (Fig. 3 and Table S4). Identified TFs, such as TraesCS1A03G0141400 and TraesCS5A03G1116700, are homologous to AT2G28550 (RAP2.7), known to regulate stress responses in A. thaliana (Gutterson and Reuber 2004) and enhance blast fungus resistance in rice (Liu et al. 2012), suggesting the AP2/ERF family may regulate biotic stress in wheat (Zhao et al. 2019). TraesCS1A03G0574000 annotates ERF6, and its homologous gene (AT4G17490) plays a role in defense against Botrytis cinerea by regulating reactive oxygen species in A. thaliana (Meng et al. 2013) and blast resistance in cereals (Liu et al. 2012). Similarly, genes annotating ERF4, ERF114, and ERF1 showed resistance against pathogens in multiple crops for Pseudomonas syringae, necrotrophic, blast, Stemphylium lycopersici, and stem rot (Lu et al. 2011; Dong et al. 2015; Yang et al. 2021; Li et al. 2022b). WRKY TFs regulate defense via salicylic acid, jasmonic acid, and ethylene signaling pathways (Lu et al. 2011; Sharma et al. 2018; Wang et al. 2020), while MYB in A. thaliana trigger apoptosis against Pseudomonas syringae (Vailleau et al. 2002). These findings highlight the role of WRKY, MYB, ERF, etc., genes in disease resistance in many crops, and these might also be involved in Lr resistance in wheat.

Phenylalanine Metabolism Pathway-Related Genes Involved in Response to Leaf Rust Infection

We identified 16 upregulated candidate genes associated with the phenylalanine pathway under the P. triticina infection (Table S7). The phenylalanine pathway has been previously known for its contribution to plant immunity through the biosynthesis of secondary metabolites, including lignin, flavonoids, and phytoalexins, which provide antimicrobial effects (Dixon et al. 2002; NAOUMKINA et al. 2010; Yadav et al. 2020; Huang et al. 2021). The upregulation of genes suggests an enhanced production of these defense compounds and chemical defenses against the rust pathogen. These findings are consistent with previous studies (Maher et al. 1994; Yadav et al. 2020; Huang et al. 2021), which indicate that phenylpropanoid-derived metabolites are crucial in disease resistance. Similarly, 11 meta-DEGs encode the cytochrome P450 gene family and are involved in steroid biosynthesis (Table S7). Steroids, such as brassinosteroids and glycoalkaloids, have been reported to enhance resistance by regulating gene expression (Li and Chory 1997; Kim et al. 2022)—for example, race-nonspecific genes like Lr34 influence secondary metabolism (Krattinger et al. 2009). The functional validation of these genes is needed to uncover the exact role and potential synergistic effects for Lr resistance.

Hub Genes Under Leaf Rust Infection

In our study, 81 hub genes, encoding cytochrome P450, cytokinin oxidase, leucine-rich repeat protein kinase, pathogenesis-related, ribosomal proteins, wall-associated kinase, etc., were identified through a co-expressed genes network for Lr resistance (Fig. 5 and Table S9). Of these, 10 genes were annotated as pathogenesis-related, such as PR3, PR4, PR1, and PR2; these genes are well-known for disease resistance (Zribi et al. 2021; Islam et al. 2023). For example, PR1, PR2, PR4, and PR5 genes were reported for fusarium infection in Allium sativum (Anisimova et al. 2021), LcPR4a for Ascochyta lentis in lentils (Vaghefi et al. 2013), and silencing of PR-2 increased disease susceptibility (Graham et al. 2007). Further, 21 hub genes encoding ribosomal proteins were identified for Lr resistance, and these proteins play an essential role in plant development, growth, and defense responses (Nagaraj et al. 2016; Ramu et al. 2020). Six hub genes— TraesCS4A03G0155600, TraesCS4B03G0015700, TraesCS4D03G0012000, TraesCS4B03G0015700, TraesCS4A03G0769500, and TraesCS4D03G0564000—were belonged to ubiquitin supergroup gene family, and the ubiquitination is crucial for many cellular processes, including development, homeostasis, cell division, growth, hormone, and stress responses (Smalle and Vierstra 2004; Popovic et al. 2014). Hub genes—TraesCS3B03G1129700 (GRP3), TraesCS3A03G0980900 (GRP4), and TraesCS3D03G0916100 (GRP4)—were identified as members of the glycine-rich RNA-binding protein family, which is known to play a significant role in stress responses in different crops (Ma et al. 2021; Jose et al. 2024). We also identified 22 genes annotated as lipid transfer proteins, and these proteins are involved in plant defense mechanisms via pathogenesis-related proteins against fungal and bacterial pathogens (Deppe et al. 2018; Islam et al. 2023). These results align with studies that reported multiple defense mechanisms in response to fungal diseases (Graham et al. 2007; Vaghefi et al. 2013; Zribi et al. 2021; Anisimova et al. 2021; Islam et al. 2023), indicating that these hub genes could be candidate targets for Lr disease resistance in wheat.

Potential Candidate Genes for Leaf Rust Resistance

We employed supervised machine learning and deep learning models to identify the target genes in response to Lr stress. Twenty genes were identified based on the relative importance of features using attribute weighting and machine learning models. Finally, based on functional annotation, enrichment analysis, comparative genomics, and literature mining, 9 out of 20 potential candidate genes were related to Lr disease resistance in wheat (Figs. 6 and 7; Table 2 and S11). For example, TraesCS5D03G0884400 belongs to the disease resistance-responsive (dirigent-like protein) family, and the homologous gene (AT1G58170) is associated with disease resistance (Khan et al. 2018; Yadav et al. 2021). Likewise, TraesCS1A03G0869900 (PR4) is involved in antifungal activity to protect the crop plant from pathogens (Vaghefi et al. 2013; Zribi et al. 2021). Two TF genes, TraesCS6B03G1228800 (AP2/ERF) and TraesCS3B03G0088700 (MYB62), were associated with plant defense mechanisms against fungal pathogens in different crop plants (Liu et al. 2012; Chen et al. 2023; Li et al. 2023; Yu et al. 2023; Ma et al. 2024). Cytochrome P450 family gene, TraesCS5A03G1198800 (CYP96A10), was identified via meta-analysis, hub genes, and machine learning, and cyp96a4 mutant is associated with wax biosynthesis and participates in the JA signaling pathway that provides diverse abiotic and biotic stresses in A. thaliana (Huang et al. 2024). TraesCS7D03G0485000 annotates to leucine-rich repeat protein kinase family protein, which is involved in pathogen recognition and signaling (Tor et al. 2009; Loutre et al. 2009). Two potential candidate genes, TraesCS2B03G1167400 (LTP) and TraesCSU03G0129300 (LTP4), belonged to lipid transfer protein, and these proteins play an essential role in biotic stresses (Iqbal et al. 2023; Islam et al. 2023). TraesCS7A03G0388400 was detected as a putative candidate gene, with functional annotations as a plant basic secretory family protein, which provides resistance to powdery mildew in G. max (Xian et al. 2022). This study identified potential candidate genes in response to Lr infection via multiple approaches and developed model for the identification. However, future studies are required to elucidate their exact role in Lr disease resistance and breeding programs to develop superior cultivars with durable resistance.

Statistical Power of Machine Learning Approaches

The present study identified disease resistance genes through traditional statistical models based on predefined thresholds and machine learning models based on ranking top features. Of the 20 top feature genes identified by machine learning, 15 (~ 75%) overlapped with those detected by traditional models, while five were novel and identified by machine learning (Table S10-11). Among these, nine were potentially linked with Lr disease resistance: seven (~ 77.80) were consistent with meta-DEGs and hub genes analysis, and two were identified by machine learning models, validating the reliability of machine learning. The performance matrix was varied for single-gene or gene-specific models, with AUC ranging from 0.67 to 0.97 and the accuracy from 0.60 to 0.95 (Figs. 6 and 7), indicating the heterogeneity in gene-specific contribution to the disease phenotype. These findings demonstrate that machine learning models provide a higher statistical power to detect the novel candidate genes associated with Lr that were missed by traditional approaches. Thus, by integrating conventional models with machine learning approaches, this study provides deeper insights and genetic foundations for mining genes associated with Lr disease, and it also provides a high-resolution approach for advanced research in pathology and disease resistance.

Conclusion

This study integrated high-throughput sequencing data from multiple studies to identify the candidate genes and pathways associated with Lr resistance. After preprocessing and batch effect correction, 2,153 upregulated and 1,579 downregulated genes were identified via meta-analysis. Further, 162 TFs, such as bHLH, NAC, MYB, WRKY, and AP2/ERF, were expressed differently and are involved in activating defense-related genes. Among 124 R-genes, ~ 80% of upregulated genes were plant pattern recognition receptors PPRs of RLKs and RLPs classes that triggered immunity through MAMP. These genes were significantly enriched in phenylalanine metabolism, steroid biosynthesis, photosynthesis, stilbenoid, diarylheptanoid, gingerol biosynthesis, and flavonoid biosynthesis pathways. Through a co-expressed network, 81 hub genes were associated with Lr infection, and their homologous genes have been reported for plant defense mechanisms and biotic stresses. Finally, nine potential candidate genes were identified using machine learning, and these genes were associated with Lr disease based on comparative genomics, literature mining, and enrichment analysis, which further required functional studies to determine their exact role in disease resistance. These findings provide valuable genetic resources to understand the molecular disease resistance mechanism and genetic basis for controlling and breeding wheat plants resistant to Lr infection.

Supplementary Information

Below is the link to the electronic supplementary material.

Author Contributions

YW and SY conceived and designed the experiments; MF, YW, SY, and MI performed the experiments and analyzed data; MF, MI, JS, SY, and YW contributed reagents/materials/analysis tools; MF, YW, SY, and MI wrote the paper. All authors have read and agreed to the published version of the manuscript.

Funding

This research is supported by the following projects: Construction of high quality and efficient mechanized scientific and technological innovation talent team of characteristic coarse cereals in Guizhou Province (qiankehepingtairencai-BQW[2024]009) and Research and integrated application of key technologies of green and high yield in characteristic mountain agriculture (guidalingjunhezi[2023]07).

Data Availability

No datasets were generated or analysed during the current study.

Declarations

Ethics Approval and Consent to Participate

No specific permissions were required to conduct this study. The data used in this study were obtained from the SRA database and comply with relevant institutional, national, and international guidelines and legislation.

Consent for Publication

Not applicable.

Competing Interests

The authors declare no competing interests.