Gaining insights into epigenetic memories through artificial intelligence and omics science in plants

Abstract

Plants exhibit remarkable abilities to learn, communicate, memorize, and develop stimulus-dependent decision-making circuits. Unlike animals, plant memory is uniquely rooted in cellular, molecular, and biochemical networks, lacking specialized organs for these functions. Consequently, plants can effectively learn and respond to diverse challenges, becoming used to recurring signals. Artificial intelligence (AI) and machine learning (ML) represent the new frontiers of biological sciences, offering the potential to predict crop behavior under environmental stresses associated with climate change. Epigenetic mechanisms, serving as the foundational blueprints of plant memory, are crucial in regulating plant adaptation to environmental stimuli. They achieve this adaptation by modulating chromatin structure and accessibility, which contribute to gene expression regulation and allow plants to adapt dynamically to changing environmental conditions. In this review, we describe novel methods and approaches in AI and ML to elucidate how plant memory occurs in response to environmental stimuli and priming mechanisms. Furthermore, we explore innovative strategies exploiting transgenerational memory for plant breeding to develop crops resilient to multiple stresses. In this context, AI and ML can aid in integrating and analyzing epigenetic data of plant stress responses to optimize the training of the parental plants.

Keywords: DNA methylation; deep learning; gene expression; machine learning; stress memory; transgenerational inheritance.

Figure 1

Schematic overview of epigenetic mechanisms in plant responses to biotic and abiotic stresses The diagram highlights the involvement of DNA methylation, histone modifications, chromatin remodeling, and non‐coding RNAs (ncRNAs; including lncRNAs, siRNAs, and miRNAs) in the regulation of defense genes and hormonal signaling pathways. These coordinated processes are crucial for establishing long‐term stress memory, facilitating adaptive responses, and enhancing plant resilience against environmental challenges. hc‐siRNAs, heterochromatic small interfering RNAs; miRNAs, microRNAs; ncRNAs, non‐coding RNAs; lncRNAs, long non‐coding RNAs; siRNAs, small interfering RNAs; TEs, transposable elements; TGS, transcriptional gene silencing; PTGS, post‐transcriptional gene silencing.

Figure 2

Epigenetic background of priming for enhanced resilience to biotic stresses This figure illustrates the main types of epigenetic modifications identified during priming and their relationship to each other in preparing plants for more effective defense responses against pathogens. Examples of affected defense‐related genes and the corresponding epigenetic marks are provided. Key examples include the following. (i) Priming of Arabidopsis thaliana against Pseudomonas syringiae pv. maculicola involved H3 modifications on the promoters of defense genes (WRKY6, WRKY9, and WRKY53), like H3K4 trimethylation at Lys 4 (H3K4me3) in npr1, sni1, and cpr1 mutants and thereby their enhanced readiness. In addition, acetylation of H3 and H4 (H3K9ac and H4K12ac) histones was also increased at the promoters of some WRKYs due to priming (Jaskiewicz et al., 2011). (ii) Sulforaphane priming in A. thaliana reduced the plant susceptibility to downy mildew (Hyaloperonospora arabidopsidis). In response to H3 modifications (H3K4me3 and H3K9ac), chromatin unpacking was detected associated with WRKY6 and PDF1.2 transcription sites (Schillheim et al., 2018). (iii) β‐aminobutyric acid (BABA) treatment primed pattern‐triggered immuniity (PTI)‐responsive genes (FRK1, NDR1, HIN1, NHL10, and CYP81F2) in A. thaliana, leading to enhanced expression upon infection with Pectobacterium carotovorum ssp. carotovorum or treatment with its epitopes (flg22, elf26 (EF‐Tu)). This primed state of PTI‐related genes was associated with enrichment of H3 acetylation (H3K9K14ac) and methylation (H3K4me2 resulting in open chromatin at their promoter regions. (iv) Histone H1 can influence H3 modifications (H3K56ac, H3K4 me3, and H3K27me3) as well as DNA methylation of defense‐related genes (Sheikh et al., 2023). (v) DNA hypomethylation can increase the expression of transposable elements (TEs) located near pathogen receptor genes (e.g., pattern recognition receptors (PRRs) and nucleotide‐binding repeats (NLRs)), while sRNAs participate in silencing these TEs via RNA‐directed DNA methylation (RdDM) (Cambiagno et al., 2018). SA, salicylic acid; sRNA, small RNA.

Figure 3

Priming‐mediated modification of the plant epigenome for the production of epigenetically trained (Epi‐Trained) or bred (Epi‐Bred) plants Transgenerational inheritance of stress memory, involving the transmission of epigenetic marks across generations, presents a novel opportunity in plant production and breeding. The mechanism of this transmission is influenced by the plant's mode of reproduction. Epi‐Trained Plants (Asexual Propagation): in asexually propagated plants, stress‐induced epigenetic states are passed on to the next generation via mitosis. This offers an innovative strategy to “train” parental plants, enhancing the tolerance of their clonal offspring to single or multiple stresses. Epi‐Bred Plants (Sexual Reproduction): epigenetic inheritance through sexual reproduction requires the stress‐induced epigenetic state to withstand epigenetic reprogramming during gametogenesis and seed development, allowing acquired resistance to be inherited by the progeny. This process relies on epialleles, which are genomic regions exhibiting heritable differences in their epigenetic state. These offspring are termed Epi‐bred plants. VOC, volatile organic compound; sRNA, small RNA; BABA, β‐aminobutyric acid; Me, methylation; AC, acetylation; Uq, ubiquitination; Ph, phosphorylation.

Figure 4

Machine learning pipeline for integrating and analyzing multi‐omics epigenetic data in plant stress responses This workflow outlines the steps involved in using machine learning to understand how epigenetic modifications relate to plant responses to stress. I. Input and Preprocessing: The pipeline begins with three types of omics data: DNA methylation profiles (methylome data), non‐coding RNA (ncRNA) expression levels, and chromatin modification patterns. Raw data undergo quality control using FastQC and normalization with Trim Galore. II. Feature Engineering and Selection: to prepare the data for machine learning, feature engineering involves dimensionality reduction using principal component analysis (PCA), retaining 95% of the variance, and imputation to handle missing values. Feature selection employs a combination of three approaches: (i) Filter Methods: analysis of variance (P < 0.05) to identify statistically significant features; (ii) Wrapper Methods: Recursive Feature Elimination (RFE) to select the top 1,000 most informative features; (iii) Embedded Methods: Least Absolute Shrinkage and Selection Operator (LASSO) with α = 0.01 for sparse feature selection. III. Model Training and Evaluation: the selected features are then used to train three machine learning models: Random Forest, Support Vector Machines (SVM), and Neural Networks. Model robustness is assessed using 10‐fold cross‐validation (CV). Random Forest is chosen for its ability to handle numerous features and evaluate feature importance. SVM is included for its effectiveness with high‐dimensional and sparse data sets. Neural Networks are used to model complex, non‐linear relationships inherent in omics interactions. IV. Output and Analysis: the pipeline culminates in two parallel, interconnected analyses: (i) Characterization of Plant Responses to Stresses: identifying key epigenetic features and their patterns associated with different stress responses and (ii) Modeling of Epigenetic Memory: predicting and understanding the epigenetic signatures that contribute to long‐term stress memory. The outputs of these analyses can iteratively inform each other, leading to a deeper understanding of the complex interplay between epigenetics and plant stress responses.

Figure 5

Integration of RNA modifications in plant gene expression, stress memory, and artificial intelligence (AI)‐driven multi‐omics frameworks This figure illustrates how RNA modifications, specifically N6‐methyladenosine (m⁶A), 5‐methylcytosine (m⁵C), and pseudouridine (Ψ), influence gene expression and contribute to stress memory in plants. RNA Modifications and Gene Expression: these modifications generally enhance transcript stability and translation, particularly when located in the 3′ UTR (untranslated region), leading to the activation of stress‐responsive genes. The deposition of these modifications is mediated by “writer” proteins (e.g., MTA, MTB, FIP37), their recognition by “reader” proteins (e.g., YTH‐domain proteins ECT2–ECT4), and their removal by “eraser” proteins (e.g., ALKBH10B). Role in Stress Memory: studies in Triticum aestivum (wheat), Arabidopsis thaliana, and Oryza sativa (rice) have shown that persistent or recurrent m⁶A marks, particularly at 3′ UTR sites, are involved in plant stress memory. AI‐Driven Multi‐Omics Framework: an AI‐driven computational framework, utilizing tools such as 6mAboost and iRNA‐Methyl, facilitates the prediction of RNA modification sites and their functional annotation. This framework enables stress response modeling and the integration of RNA modification data with other multi‐omics data sets. Diagram key: solid arrows represent molecular pathways, while dashed arrows indicate feedback loops and computational inferences.