Human miRNA Expression and Plant miRNA: A Statistical Approach for Cancer Therapy

1. Introduction

MicroRNAs (miRNAs) are small non-coding RNAs that modulate gene expression post-transcriptionally, primarily through interaction with messenger RNAs (mRNAs), thereby indirectly influencing DNA replication and protein synthesis[1]. They are evolutionarily conserved across plants and animals, exhibit high stability, and can be transported to specific sites to exert regulatory effects. Recent studies have suggested that exogenous plant miRNAs may compensate for dysregulated human miRNAs, provided sufficient sequence complementarity — typically a minimum of 50% — to ensure mRNA targeting capacity[2]. To substitute one plant miRNA for another, however, a higher similarity threshold of 80–90% is usually required[3,4,5]. This property, termed biological complementarity, underpins the potential interchangeability of miRNAs across species.

Alterations in miRNA concentrations are a hallmark of numerous cancers, yet most investigations have focused on individual miRNAs or small, disease-specific subsets[6,7,8]. Expression levels are typically measured in tissue and circulating blood, comparing diseased and healthy samples. Cancers frequently induce widespread perturbations in miRNA expression, leading to the hypothesis that restoring miRNA homeostasis may influence disease progression[9,10].

To explore global trends in miRNA dysregulation associated with cancer, we analyzed a comprehensive dataset containing miRNA expression profiles across multiple cancer types[11,12]. Employing machine learning, we attempted to predict miRNA expression status (under- or overexpressed) using neural networks. Despite leveraging large-scale data, classification accuracy plateaued at 70–75%, suggesting either insufficient mechanistic understanding or the influence of multiple latent variables. These results highlight the limitations of current predictive models for miRNA expression.

To bypass these constraints, we characterized each miRNA statistically, representing it by two parameters: the number of expression measurements (experiments) and a derived membership ratio (div). This enabled visualization of miRNA distributions across cancers. miRNAs with high absolute div values and extensive experimental coverage were considered strong candidates for functional relevance. These findings offer a basis for prioritizing miRNAs in targeted analyses.

We then assessed the feasibility of correcting aberrant human miRNA expression using plant miRNAs, based on average sequence complementarity. miRNAs with high div values (indicative of consistent dysregulation) were stratified by whether they were over- or underexpressed. For each plant miRNA, we computed average complementarity with both groups and identified candidates with maximal differential affinity — favoring underexpressed targets while avoiding overexpressed ones. These metrics form the foundation for evaluating therapeutic potential.

Given the vast number of dysregulated miRNAs, we further prioritized targets by identifying key genes under miRNA control. We limited our analysis to experimentally validated miRNA–gene interactions. Since one gene may be regulated by multiple miRNAs, we computed, for each gene, the number of regulating miRNAs and the aggregated div metric of these miRNAs. This allowed us to identify genes under robust miRNA-mediated regulation. Parallel expression datasets from cancer samples[17] provided direct gene-level expression data. By comparing miRNA-derived and empirical gene expression patterns, we examined consistency in regulatory trends. While many genes showed discordant profiles, others exhibited concordant regulation, suggesting differential miRNA involvement in gene expression control.

Focusing on the top 200 genes most strongly affected by both direct expression changes and miRNA-derived div metrics, we mapped back to the regulating miRNAs. From this refined miRNA set, we again computed average complementarity with plant miRNAs. Those plant miRNAs showing the largest positive differential (strong affinity for underexpressed miRNAs and weak for overexpressed) emerged as optimal candidates for therapeutic intervention. Conversely, plant miRNAs with the largest negative differential may inform mechanisms of cancer-promoting miRNA overabundance.

We also evaluated average complementarity at the plant-species level but observed minimal variation (<0.05%), suggesting that species-level generalizations are currently infeasible. This may reflect the sensitivity of plant miRNA profiles to cultivar, geography, and growth conditions.

Nevertheless, our framework provides a rational basis for the design of synthetic miRNAs with enhanced specificity — engineered to maximize complementarity with target human miRNAs while minimizing off-target interactions — opening avenues for future miRNA-based therapeutics.

2. Materials and Methods

Statistical analyses of miRNA expression data were performed utilizing a comprehensive database comprising multiple parameters, including cancer type, experimental modality, T-value/B-value metrics, and log fold change (logFC) [12]. The principal variable of interest was logFC, representing the logarithmic deviation of miRNA expression levels relative to baseline in tissue or blood samples. Each miRNA was profiled across numerous experiments, yielding a dataset encompassing 3,174 distinct miRNAs and 155,417 total experimental observations.

Initial analytical steps involved deploying a neural network model to predict both the quantitative logFC values and the qualitative directionality of expression changes (upregulation or downregulation). Predictions were generated for diverse sample representations, including individual miRNA measurements and mean expression profiles. Model performance was evaluated through training on input features comprising miRNA sequence data (mature and precursor forms), disease classification, experimental context, and ancillary variables.

Data processing, sorting, and visualization were conducted using custom software implemented in Python 3 within the PyCharm IDE (version 2022.1). The membership ratio was computed following the formula described in [106].

$$div=(\mathrm{n}\mathrm{u}\mathrm{m}\_\mathrm{o}\mathrm{f}\_\mathrm{d}\mathrm{o}\mathrm{w}\mathrm{n}-\mathrm{n}\mathrm{u}\mathrm{m}\_\mathrm{o}\mathrm{f}\_\mathrm{u}\mathrm{p})/(\mathrm{n}\mathrm{u}\mathrm{m}\_\mathrm{o}\mathrm{f}\_\mathrm{d}\mathrm{o}\mathrm{w}\mathrm{n}+\mathrm{n}\mathrm{u}\mathrm{m}\_\mathrm{o}\mathrm{f}\_\mathrm{u}\mathrm{p})$$

(1)

where num_of_down ​ and num_of_up denote the number of experiments in which the miRNA was significantly down- or upregulated, respectively.

It is established that each miRNA regulates a large number of target genes, and elucidating the specific mechanisms of these interactions is a crucial task, particularly regarding the interdependencies of factors, i.e., correlations of quantitative features. Databases of miRNAs and their target genes, such as miRDB [103] and miRTarBase [104], provide useful resources for this purpose. The former database includes both experimentally validated target genes and predicted miRNA-gene relationships, thereby expanding the dataset. The latter contains only experimentally validated miRNA-gene interactions. We utilized miRTarBase to identify genes associated with varying numbers of critically up- and downregulated miRNAs, as determined by the third (III) nonlinear critical threshold (Figure 2). Gene expression data for multiple cancer types were obtained from the database in [17]. Additionally, the plant miRNA database [105], comprising 10,898 plant miRNAs, was employed.

3. Results

3.1. Identification of Critical miRNAs

Neural network training results indicate that prediction accuracy for the direction of expression achieves a minimum of approximately 50% and does not exceed ~65% when considering all miRNAs. Numerical experiments applying various filters—such as restricting analysis to specific tumor types or experimental modalities—yielded similar accuracy. When focusing solely on individual miRNAs characterized by a predominance of either underexpression or overexpression (based on percentage ratio) and supported by a sufficient number of experiments to establish a trend, accuracy improves to 70–80% [16].

From these findings, several assumptions can be made:

• Input variables—such as miRNA sequences, disease type, and experiment type—play a role but do not fully capture the overall miRNA expression profile, indicating that not all relevant information influencing expression prediction is accounted for.
• The number of distinct miRNAs studied may be insufficient.
• The expression data in the database may contain measurement errors.

Thus, a statistical approach is warranted to identify existing dependencies within the expression data and uncover opportunities to extract useful information about miRNAs, particularly regarding specific miRNA sequences. Given the multitude of parameters influencing miRNA expression, it is necessary to analyze expression data collectively rather than on an experiment-by-experiment basis.

Accordingly, each miRNA was evaluated in the context of all experiments in which it was observed. An “average” expression value for each miRNA was estimated by summing all logFC values across all experiments. The resulting data were organized into a new table containing the following columns: miRNA identifier, cumulative sum of logFC, number of experiments with negative logFC, and number of experiments with positive logFC. This procedure was applied under various filtering conditions, such as specific cancer types or experimental designs. Although the values vary slightly depending on the applied filters, the overall trend remains consistent.

The number of miRNAs predominantly exhibiting overexpression and underexpression was found to be nearly equal (1606 versus 1568). It is important to note that many miRNAs are represented by a small number of experiments. Some miRNAs demonstrate membership values around or above 95%, although the number of associated experiments does not exceed several dozen.

To summarize the classification of miRNAs based on cumulative logFC expressions across all cancer types, the data are presented in the form of a truncated table (Table 1).

It was observed that across all expression data, a pronounced tendency exists for the majority of miRNAs (over 70% of the total distinct miRNAs) to be classified as either predominantly overexpressed or underexpressed, independent of disease type or experimental conditions (e.g., blood-derived or cell-derived samples). This trend remains consistent even when the dataset is restricted by applying various filters.

We further compared the membership ratios derived from high-throughput expression data with those obtained from low-throughput methods [12], revealing an approximate concordance of 90% (±5%), contingent on the applied filters such as total expression counts and membership thresholds.

In the following table, we present a comparative analysis of miRNAs identified in Table 1 alongside those reported in the literature, listing miRNAs classified as overexpressed or underexpressed and comparing their corresponding statistical metrics.

The following Figure 1 illustrates the distribution of all 3,173 miRNAs according to the number of experiments in which they were observed and their respective membership ratios. All miRNAs listed in Table 2 are highlighted within this overall distribution.

As observed, many studied indicators exhibit a clear trend toward either upregulation or downregulation. However, numerous miRNAs remain poorly characterized with respect to the molecular mechanisms in which they participate. Additionally, as the number of experiments per miRNA increases, its position on the distribution converges toward an average value.

A key challenge lies in defining the boundary between "critical" and "non-critical" miRNAs. Multiple factors influence miRNA expression, a conclusion supported by neural network analyses. This underscores the necessity for additional contextual information regarding the experimental conditions under which miRNA expression is measured, including potentially relevant molecular mechanisms and regulatory factors.

It is evident that increasing the number of experiments enhances the robustness of correlations between miRNA expression patterns in diseased versus healthy tissues. Furthermore, a higher ratio between the number of experiments showing negative versus positive miRNA expression strengthens the statistical significance of associations between miRNA dysregulation and cancer. Balancing these parameters is complicated by substantial variability in experiment counts across individual miRNAs.

Figure 2 presents the relationship between the cumulative difference metric (div) and the number of experiments, delineating three boundary regions corresponding to nonlinear thresholds. These critical div values serve to classify general miRNA expression trends across cancer types. A color-coded scheme visually represents miRNA concentration (dots) within distinct graph regions.

Panel (b) of Figure 2 illustrates the upper threshold (I) alongside miRNAs from the table whose expression behaviors have been characterized in greater detail.

It should be noted that only a small subset of miRNAs has been studied in detail (see references in Table 2), indicating that a considerable number of human miRNAs remain insufficiently characterized with respect to their biological functions. Notably, within the nonlinear threshold I, Table 2 lists 17 out of 165 overexpressed miRNAs and only 6 out of 414 underexpressed miRNAs. This suggests that miRNAs with higher abundance have been more extensively investigated compared to those present at lower levels.

Figure 3 presents expression distributions as a function of the number of experiments for the 20 most frequently studied miRNAs within the nonlinear zone I, comprising both underexpressed and overexpressed groups.

It is evident that as the number of experiments assessing miRNA expression increases, the average expression value tends to converge toward a membership ratio of 0. However, there are miRNAs with between 20 and 50 experiments that maintain a membership ratio exceeding 0.7, as illustrated in Figure 4.

Table 3 presents selected statistical parameters for overexpressed and underexpressed miRNAs. For all miRNAs exceeding the nonlinear thresholds, the table reports the number of unique miRNAs (data points), the total number of experiments aggregated across these miRNAs, and the cumulative expression values. Across all metrics, underexpressed miRNAs predominate over overexpressed ones. This disparity is especially pronounced at the primary (I) nonlinear threshold, where miRNAs above the threshold are considered more biologically relevant. Here, the dominance of underexpressed miRNAs is markedly clear.

Overall, these findings suggest that in cancerous conditions, the repertoire of underexpressed miRNAs substantially exceeds that of overexpressed miRNAs. This implies that many miRNAs exhibit markedly reduced concentrations in the presence of cancer compared to those whose levels are elevated. The underlying causes may vary: cancer could impair miRNA biogenesis pathways, reflect a host protective response to tumorigenesis, or result from a complex interplay of these and other factors.

3.2. Statistical Relationship Between miRNA and Target Genes

Statistical associations between miRNAs and their target genes were analyzed by calculating, for each gene, the number of underexpressed and overexpressed miRNAs (according to threshold III), as well as the difference between these counts (see Table 4).

Similarly, a graph depicting the relationship between the membership ratio and the number of targeting miRNAs (rather than the number of experiments as in the previous analysis) was constructed. Only genes targeted by more than 20 miRNAs were included, reducing the dataset from 7,371 to 3,495 genes. A nonlinear threshold, approximating the critical threshold III, was applied to distinguish genes classified as “downregulated” or “upregulated.” The resulting gene distribution is presented in Figure 5. Additionally, the figure highlights a subset of genes (10 per category) from database [8], for which experimentally validated interactions relevant to cancer development have been documented.

Table 5, analogous to the miRNA table, presents numerical parameters derived from the graph, including the number of genes above the nonlinear threshold curve for overexpressed and underexpressed categories, the total number of miRNAs targeting these genes, and the cumulative membership ratio for all genes.

The data presented in Figure 5 can be utilized for various analyses: investigating individual genes, gene sets, gene categories, or providing a general characterization of gene expression patterns in cancer. Genes positioned above the nonlinear critical threshold, exhibiting a high number of targeting miRNAs (either up- or downregulated) and a membership ratio approaching unity, are of particular interest regarding their involvement in cancer-related mechanisms.

To elucidate the influence of miRNA expression on gene regulation in cancer, gene expression data from database [17], encompassing various cancer types, were analyzed. For each gene, the average expression across all cancer types was calculated, yielding a single representative value (mean log_expression). Concurrently, “critical” genes identified through miRNA expression patterns in Figure 5 were assigned a cumulative membership ratio (div), enabling a two-dimensional representation of gene status. This distribution is illustrated in Figure 6.

As illustrated in Figure 6, genes are distributed broadly across the two-dimensional parameter space defined by mean logarithmic expression and membership ratio. Each quadrant corresponds to distinct biological interpretations and potential regulatory mechanisms. Table 6 summarizes the gene counts stratified by cancer type. The first column lists the cancer type, the second column indicates the number of genes with mean log_expression values between –0.1 and 0.1, representing near-zero expression. The third column shows the total number of genes measured for each cancer type, and the fourth column provides the ratio of near-zero expression genes to total measured genes.

These genes exhibit negligible expression changes during cancer but are targeted by miRNAs with non-zero membership ratios, suggesting potential miRNA-mediated regulation that is balanced or compensated by other factors. Alternatively, it may indicate that the miRNAs do not directly regulate these genes under the studied conditions.

The above table provides useful information about genes whose expression remains within the normal range of miRNA concentration. Our main objective is to influence specific genes by modulating the concentration of miRNAs in order to correct the expression levels of a limited set of genes of interest.

At the same time, many genes show clear correlation patterns, as exemplified by the 20 genes located farthest from the origin in the coordinate system.

A subset of 200 individual genes was selected based on having the largest sum of the absolute values of mean log_expression and membership ratio (div). These represent critical genes whose expression may potentially be regulated by changes in the expression of their corresponding miRNAs. Such genes exhibit a clear relationship between their expression in disease and the expression levels of their targeting miRNAs. Of course, it should be noted that gene expression is influenced not only by miRNAs but also by other regulatory factors, so some coincidences may occur.

We then selected miRNAs corresponding to the III non-linear boundary (Figure 2a) that target these 200 critical genes. Among them, there are 634 underexpressed and 462 overexpressed miRNAs, with most miRNAs targeting only one gene from the critical list. For each miRNA from this set, we calculated the number of critical genes it targets. Restricting the analysis to miRNAs targeting at least 10 critical genes, we identified 115 underexpressed and 93 overexpressed miRNAs, which are presented in Table 7 and Figure 7 as examples.

It is assumed that these miRNAs (115 underexpressed and 93 overexpressed) may play a significant role in regulating key mechanisms of cancer development, as they are clearly underexpressed or overexpressed and are likely to influence critical genes whose expression changes have been observed in the presence of cancer and which are targeted by these miRNAs.

We are primarily interested in genes with a positive membership ratio (div), as these correspond to genes targeted by miRNAs that are significantly underexpressed. These genes have greater practical significance because it is generally more feasible to compensate for insufficient miRNA concentrations (underexpression) than to reduce the levels of overexpressed miRNAs. Among the 115 underexpressed miRNAs identified, only 12 appear in Table 2, whereas 35 of the 93 overexpressed miRNAs are present there. This indicates that changes in the expression of miRNAs and their target genes have been studied more extensively for overexpressed miRNAs compared to underexpressed ones (see Table 8).

3.3. Connection with Plant miRNAs

The next step involved investigating the relationship between plant miRNAs and specific human miRNAs identified in the previous section. Using the critical human miRNAs—607 overexpressed and 786 underexpressed miRNAs defined by the (III) nonlinear separation threshold—we performed a complementarity analysis between plant and critical human miRNAs.

Each plant species has a characteristic set of miRNAs. For each plant, we calculated the average complementarity score, defined as the sum of all complementarity scores with critical human miRNAs divided by the total number of these comparisons [105]. Additionally, average complementarity scores were computed separately for critical overexpressed and critical underexpressed human miRNAs.

Our results show that the average complementarity across all 127 plant species analyzed is approximately ±0.5 for both overexpressed and underexpressed miRNAs. The difference in complementarity between the critical overexpressed and underexpressed groups does not exceed 0.07.

Furthermore, we compared all plant miRNAs against the subset of critical human miRNAs that target critical genes identified previously—specifically, 115 underexpressed and 93 overexpressed miRNAs. The average complementarity-similarity for all plant miRNAs, calculated separately for these two groups of human miRNAs, is summarized in Table 9.

If the same complementarity analysis is performed for underexpressed and overexpressed miRNAs defined by the III nonlinear threshold, the difference between their average complementarities with plant miRNAs is significantly smaller — see Table 10.

These plant miRNAs may play a crucial role in compensating for the deficiency of underexpressed human miRNAs observed in cancer. Conversely, many plant miRNAs also show average complementarity to overexpressed human miRNAs. From a therapeutic perspective, targeting the restoration of underexpressed human miRNAs appears more promising, as supplementing deficient miRNAs with exogenous counterparts is generally more feasible than attempting to reduce the levels of overexpressed miRNAs. It is also possible to design synthetic miRNAs that maximize complementarity to underexpressed human miRNAs while minimizing binding to overexpressed ones. Furthermore, some plant miRNAs identified here could be key candidates for replacing deficient human miRNAs, although their therapeutic potential requires experimental validation.

4. Discussion

The question arises as to why, despite using a universal approximator like a neural network, we are unable to reliably predict human miRNA expression based solely on its nucleotide sequence. More precisely, predictions can be made, but with an error margin large enough to cast doubt on their practical utility. This likely reflects the fact that miRNA expression depends not only on its sequence but also on a range of other factors—such as interacting molecules or context-specific mechanisms—that significantly influence its regulation. Consequently, the application of neural networks is limited by the heterogeneity of the training data: both the inputs (features) and outputs (expression profiles) vary across individual miRNAs and their specific regulatory contexts, making it difficult to capture universal patterns. Notably, similar levels of prediction error are observed when using purely statistical analyses of miRNA expression data. However, statistical approaches are often more effective at leveraging heterogeneous datasets to uncover robust parameter relationships with a sufficient degree of confidence.

5. Conclusions

The conducted analysis demonstrates that the use of exogenous miRNAs, particularly those derived from plants, represents a promising approach to compensate for deficiencies in human miRNA expression associated with cancer. Our findings lay the groundwork for the development of targeted microRNA panels designed for therapeutic and diagnostic applications across a broad spectrum of cancer types. These panels consist of carefully selected sets of plant miRNAs that can potentially be introduced into the body of animals or humans with cancer to restore the balance of miRNA expression.

The methodology for compiling these panels is based on identifying differences between critical and non-critical human miRNAs, considering all types of cancer. Critical miRNAs are defined as those whose expression deviates significantly from normal levels in the presence of cancer, reflecting their important regulatory roles. By targeting these critical miRNAs, the proposed panels aim to modulate gene expression more effectively, addressing the imbalance caused by underexpressed or overexpressed endogenous miRNAs.

Furthermore, our study emphasizes that restoring underexpressed miRNAs is a more feasible and promising therapeutic direction than attempting to suppress overexpressed miRNAs. The complementarity analysis between plant and human miRNAs supports the potential for plant miRNAs to serve as functional substitutes or modulators for human miRNAs with diminished expression in cancer.

Overall, this work provides a comprehensive statistical and bioinformatic foundation for advancing miRNA-based interventions and encourages further experimental validation to confirm the therapeutic efficacy of plant miRNA panels in cancer treatment.