To investigate the effects of genetic modifications on the RNAi pathway in GM plants, it is essential to consider the natural mechanisms by which RNAi can affect gene expression in plants. To this end, the cellular pathways by which siRNA molecules can be produced in plants are first described here ( Figure 1 ; Table 1 ). Precursors of siRNA are almost without exception double-stranded RNA molecules, which are either synthesized by endogenous RNA polymerases (RNA Pol) or introduced exogenously (Vazquez et al., 2004; Allen et al., 2005; Axtell et al., 2006). Endogenous precursors include natural antisense transcripts (NAT) synthesized by RNA Pol II, which base-pair with the sense mRNA of the coding gene and thus form the double-stranded substrate for corresponding RNases ( Figure 1A ) (Borsani et al., 2005; Jen et al., 2005; Zhang et al., 2012). RNA Pol II also synthesizes shRNAs encoded in the genome, which can then be processed by Dicer into siRNA ( Figure 1B ) (Wesley et al., 2001; Helliwell and Waterhouse, 2003; Senthil-Kumar and Mysore, 2011) or long non-coding (lnc) RNAs (Kim and Sung, 2012; Liu et al., 2012; Wu et al., 2012), whose secondary structures can have hairpins and can thus also be converted into siRNA by corresponding RNases ( Figure 1C ). Endogenously encoded micro (mi)RNAs are synthesized by Dicer or DCL1 in plants (Kurihara and Watanabe, 2004) from shRNAs, the miRNA precursors ( Figure 1D ), and either directly regulate the expression of target genes (by miRNA) or base-pair with the precursors of so-called trans-acting (ta)siRNAs, which are then generated by DCL from a double-stranded template ( Figure 1E ). Exogenously introduced precursors of siRNA are molecules introduced into a cell from an external source. A natural example are viral RNAs, which are either immediately present after infection and replication (RNA viruses, in plant viruses often single-stranded (ss) RNA genome) or are generated by transcription of the viral genome (DNA viruses) and are then templates for DCLs, which produce siRNA from them ( Figure 1F ) (Ruiz et al., 1998; Lu et al., 2003; Burch-Smith et al., 2004). RNA Pol IV or V can also be involved in siRNA synthesis, for example in the case of the synthesis of precursors of heterochromatic (hc)siRNAs, which are then converted to siRNA by DCL3 ( Figure 1G ) (Law and Jacobsen, 2010; Zhang and Zhu, 2011; Matzke and Mosher, 2014). In plants, the proteins DCL2, DCL3 and DCL4 generate siRNAs of different lengths mostly with 22 nt, 24 nt and 21 nt, respectively, which in turn trigger different mechanisms of silencing (Henderson et al., 2006; Mukherjee et al., 2013). Of note, DCL2-derived 22 nt siRNAs in plants are involved in a transitive and systemic spread of siRNA especially for antiviral defense, called secondary RNAi (Bouché et al., 2006; Chen et al., 2010; Garcia-Ruiz et al., 2010; Qin et al., 2017). This spread of RNAi involves the amplification and expansion of silencing signals that are mediated by RNA-dependent RNA polymerases (RdRp) (Sanan-Mishra et al., 2021). In this process, siRNAs act on longer RNAs (such as mRNA) as primers for RdRp, whereby a new, long dsRNA is synthesized, which is then eventually processed again by the RNAi machinery into siRNA triggering secondary RNAi.

Silencing mechanisms can occur in the GM plant harboring the RNAi construct, at the transcriptional level in the cell nucleus or the translational/post-transcriptional level in the cytoplasm ( Figure 2 ). In the nucleus, siRNA can pair with the nascent mRNA of the target gene, recruiting factors to the transcription machinery that inhibit the transcription elongation by RNA polymerase ( Figure 2A , left) (Guang et al., 2010). Similarly, siRNA can recruit enzymes that induce epigenetic silencing of the target gene through DNA methylation (Law and Jacobsen, 2010; Wu et al., 2010; Zhang and Zhu, 2011; Wu et al., 2012; Matzke and Mosher, 2014; Movahedi et al., 2015) or histone modification (Baulcombe, 2004; He et al., 2011; Liu et al., 2012) ( Figure 2A , right) (Verdel et al., 2009). In plants, epigenetic silencing via DNA methylation is triggered by DCL3-generated ~24 nt siRNA involving a RISC complex containing the protein Argonaute (Ago)4 (Zilberman et al., 2003; Henderson et al., 2006; Qi et al., 2006; Zheng et al., 2007; Wierzbicki et al., 2009; Havecker et al., 2010; Olmedo-Monfil et al., 2010; Sarkies and Miska, 2014; Lewsey et al., 2016). The most well-studied RNAi silencing mechanism involves the degradation of the target gene’s mRNA ( Figure 2B , left). In this process, the protein Ago recruits siRNA to the complementary mRNA sequence to form the RNA-induced silencing complex (RISC) (Wu et al., 2012). In plants, this is triggered by DCL4-generated ~21 nt siRNAs involving a RISC complex containing Ago1 (Xie et al., 2005; Qu et al., 2008; Chen et al., 2010; Wang et al., 2011). If there is perfect complementarity between siRNA and the target gene, the mRNA is degraded, leading to down-regulation of the target protein’s production (Valencia-Sanchez et al., 2006). With incomplete base pairing between siRNA and target mRNA, the RNA is not degraded; instead, ribosome-mediated translation is inhibited, resulting in reduced expression of the target gene ( Figure 2B , right) (Brodersen et al., 2008).

The current literature suggests that siRNA molecules produced via different biogenesis pathways can differ in terms of their length, triggering different types of mechanisms of gene expression regulation described. While DCL4-generated 21 nt siRNA predominantly triggers PTGS via mRNA degradation, DCL3-generated 24 nt siRNA triggers TGS via epigenetic silencing and DCL2-generated 22 nt siRNA induces secondary siRNA. However, all DCL may act on long dsRNA molecules introduced into the plant. Therefore, both TGS and PTGS need to be considered when analyzing RNAi-induced effects in GM plants ( Table 1 ).

With regard to the knowledge about mechanisms by which RNAi can potentially induce unintended effects in the GM plant, the existing literature shows clear bias towards PTGS. For example, a PubMed search with the search term “RNAi AND PTGS NOT TGS” in title and abstract returned 118 hits, whereas the search term “RNAi AND TGS NOT PTGS” only returned 25 hits (as of 18.10.2024). Hence, most published studies are concerned with the investigation of effects resulting from the inhibition of translation or degradation of mRNA (possibly resulting from incomplete complementarity). Relatively fewer studies deal with TGS, possibly because here effects, for example via epigenetic silencing, could also arise upstream or downstream of the gene with sequence complementarity and these cannot be clearly determined on the basis of the pure small RNA sequence by analyzing complementary sequences in the genome.

Unintended effects of RNAi applications in GM plants themselves are a critical focus in the safety assessment of food and feed. Consequently, the Food and Agriculture Organization of the United Nations, for example, has issued guidelines for conducting food safety assessments of food derived from recombinant DNA plants (Food and Agriculture Organization of the United Nations, 2003). Also the OECD publishes science-based consensus documents offering information for the regulatory assessments of specific food and feed products, including those derived from transgenic organisms (Organisation for Economic Co-operation and Development, 2021). These documents gather data on the product’s nutrients, anti-nutrients and toxicants, its use as food or feed, and other factors relevant to food and feed safety. Here and in various review articles on the topic of risk assessment of RNAi-based GM crops, primarily untargeted methods for analyzing gene products and their metabolites, such as proteomics and metabolomics, are proposed to investigate RNAi-induced effects in the GM crop itself (Senthil-Kumar and Mysore, 2011; Kleter, 2020; Papadopoulou et al., 2020; Chaudhary et al., 2024).

The mechanisms by which the RNAi pathway can trigger specific gene expression changes in plants include both transcriptional and post-transcriptional regulation. These processes rely on specific base pairing, either with the nascent transcript (TGS) or with the mature target mRNA or a sequence-like mRNA (PTGS). While 21 nt siRNAs are predominantly involved in PTGS, 24 nt siRNAs often trigger TGS via epigenetic changes. In PTGS the target gene is directly known based on the sequence, whereas TGS can also affect genes located in close or distant proximity to the gene with sequence homology, making sequence-based prediction of TGS induced effects more difficult. PTGS is by far the most investigated mechanism in scientific studies to date, while the literature on RNAi-induced TGS is relatively limited. Therefore, the sheer number of scientific studies and the focus on PTGS to date does not necessarily reflect the actual relevance of the respective mechanisms in the plant, making it difficult to rank them according to their potential for causing unintended effects in plants.

Scientific literature on case studies investigating unintended effects in RNAi-based GM crops is currently scarce. However, bioinformatic tools are being dynamically developed to predict intended target genes and potential unintended effects on off-target genes in the GM crop or NTOs in case of HIGS, leveraging sequence homology to enhance the accuracy and scope of these predictions (Chen et al., 2019). While these tools often reach their limits in NTOs due to the lack or deficient annotated-genomes, high quality annotations are available for model plants or major crops, enabling such tools to predict PTGS effects on plant off-target genes with a higher probability. However, there are also mechanisms (such as TGS) that are not based on direct sequence homology to the target and whose unintended effects cannot be easily predicted bioinformatically. In most cases, it can be assumed that off-target effects manifest themselves at the transcriptome level and can be measured using sufficiently sensitive methods.

To detect unintended RNAi-induced effects in GM plants for risk assessment, knowledge about the siRNAs processed in the GM plant, such as size and sequence, compared to the wild type is necessary. Since both intended and possible secondary siRNAs (such as tasiRNA) can play a role, untargeted analyses, such as small RNA sequencing, should be used to identify the sequences of all siRNAs. With this knowledge, bioinformatic tools can be used to predict both intended and unintended effects mediated by sequence homology, primarily through PTGS, and these predictions can be validated using targeted methods such as RT-qPCR. However, a comprehensive bioinformatic search for homologies requires access to the plant´s complete genome, whereas RT-qPCR analyses can also be managed with knowledge of shorter sequence segments. Unintended effects mediated by TGS, on the other hand, are not directly linked to the actual sequence of the siRNA and therefore cannot be adequately detected with targeted methods, but only with untargeted methods. RNA sequencing, for example, can be used for the direct, untargeted investigation of gene expression changes, changes in histone modifications can be detected using ChIP-Seq or altered DNA methylation patterns can be detected using bisulphite sequencing. However, all these methods require the availability of the plant´s genome for accurate analysis. Additionally, there are currently no studies that specifically address the importance of selecting appropriate plant material such as tissue type, developmental stage, and sampling time points or the sensitivities required for untargeted analyses to effectively capture RNAi-induced changes (e.g. alterations in gene expression). Most published studies have focused on using plant tissues, like leaves, without a detailed exploration on how these factors might influence the detection and interpretation of RNAi-induced effects. Likewise, unintended off-target genes may be expressed, for example, in certain tissue types and not in others. These gaps highlight the need for more comprehensive research to optimize experimental designs in RNAi studies aiming to identify unintended effects.

Among the few studies assessing unintended effects of RNAi in GM plants, some have employed untargeted omics methods to analyze changes in gene expression and metabolite profiles. For example, Huang et al. (2022) compared the leaves of three transgenic maize RNAi lines resistant to Apolygus lucorum with those of three conventionally bred maize lines. Using untargeted omics methods at the levels of small RNAs, the transcriptome and the metabolome, the authors observed that the number of differentially expressed genes (DEGs) and differentially accumulated metabolites (DAMs) were greater in RNAi lines than in conventional lines. Additionally, Zörb et al. (2013) using GC-MS-based metabolite profiling showed that RNAi-mediated silencing of the sulfur-rich alpha-gliadin storage protein family in wheat grains did not induce changes in any of the 109 metabolites analyzed. Similarly, Zhang et al. (2020) investigated transcriptomic and metabolomic changes in RNAi-based GM maize resistant to Monolepta hieroglyphica compared to its unmodified variant. This study only identified a single DEG at the transcriptome level and 8 out of 5787 metabolites as DAMs, leading the authors to conclude that the RNAi variant exhibited negligible changes compared to the wild type.

Building on the insights gained from studies exploring off-target effects in RNAi-based GM plants, these findings have helped to inform regulatory approaches, including the one of the first authorization-relevant risk assessments for an RNAi-based genetically modified crop was carried out by the US Environmental Protection Agency (US EPA) for SmartStax Pro (MON 87411/Unique ID: MON-87411-9) (EPA Reg. Number: 62719-707). As part of the product characterization and human risk assessment, in 2016 the US EPA recommended a number of methods to rule out unintended side effects. These include transcriptome analyses using microarray or RNA sequencing, proteome analyses, GC-MS-based metabolomics, and the global detection of changes in DNA methylation patterns. It should be noted that certain recommended methods, such as microarray analyses for transcriptome studies or 2D gel electrophoresis coupled with MS for transcriptome analysis, are no longer state-of-the-art and should be replaced by more up-to-date methods such as RNA sequencing and LC-coupled MS, respectively. The US EPA advised that these analyses should be carried out comparatively between the GM plant containing all modification events (SmartStax Pro), the GM plant lacking the dsRNA cassette (SmartStax) as well as non-genetically modified lines across several generations. Furthermore, they recommended using a combination of different omics methods and to combine them with more sensitive methods such as RT-qPCR, to thoroughly exclude unintended effects.

In summary, the challenges in detecting unintended RNAi effects in GM plants lie in the diversity of siRNAs that can be formed from corresponding precursor molecules and in the fact that TGS (especially via epigenetic mechanisms) can also affect the expression of nearby genes without sequence homology, indicating that targeted/biased bioinformatic methods alone are not sufficient for excluding unintended effects. The few available studies indicate that the RNAi method appears to be relatively specific with minimal unintended effects expected (Zörb et al., 2013; Zhang et al., 2020; Huang et al., 2024).

Untargeted approaches, such as RNA sequencing for transcriptome analysis, LC-MS-based proteomics or GC-MS-based metabolome profiling, offer a promising and increasingly sensitive means of investigating these effects. The current state of well-annotated plant genomes varies significantly across species, with high-quality annotations available for some model plants and major crops, while others remain underrepresented. This variability poses challenges for accurately mapping RNAi-induced changes, as comprehensive and well-annotated reference genomes are crucial for identifying both target and off-target effects, as well as for understanding the broader biological impact of RNAi in diverse plant species. One way around this problem is to perform a de novo assembly of the transcriptome of unannotated plants (Surget-Groba and Montoya-Burgos, 2010; Narzisi and Mishra, 2011). However, this depends on the quality and depth of the sequencing. In combination, bioinformatic approaches with untargeted methods, such as various omics, offer the possibility to detect specific off-target effects in GM plants.

Future research on detecting RNAi-induced effects in GM plants should focus on improving sensitivity and specificity with advanced sequencing technologies, better off-target detection through CRISPR, and more accurate quantification using methods like RT-qPCR and proteomics. Environmental impact studies, long-term monitoring, and standardizing protocols will be key for regulatory safety assessments.