The skin is the largest organ of the human body, serving as a primary barrier against external stimuli and injuries^[1]. Due to its exposed location, cutaneous wounds are among the most common types of injuries encountered in forensic practice^[2]. Moreover, since skin trauma often coincides with damage to internal organs, accurately estimating the age since injury is essential not only for reconstructing events but also for inferring the age of the visceral organ trauma^[3]. Therefore, skin wound age estimation has long been a major focus of forensic research worldwide^[4-7].

Current methods for wound age estimation primarily fall into two categories: histological evaluation based on temporal changes during wound healing, typically divided into inflammatory, proliferative, and remodeling phases^[8-10], and molecular approaches relying on time-dependent biomarkers such as cytokines, chemokines, and extracellular components^[11-14]. While these markers show promise, their practical application remains limited. Postmortem degradation caused by autolysis and putrefaction significantly affects biomolecule integrity, particularly for proteins and mRNAs^[15,16]. Furthermore, most studies have focused on separate, not mechanically linked biomarkers, leading to low capacity and limited accuracy for wound age determination^[6,17,18].

Recent studies have highlighted the potential of miRNAs in forensic applications due to their remarkable stability, resistance to degradation, and critical regulatory roles in wound healing^[19-23]. These small non-coding RNAs modulate the expression of numerous target genes, influencing various phases of skin repair^[24]. However, comprehensive profiles of miRNA-mRNA regulatory networks during wound healing remain poorly characterized. In our earlier work, we identified a set of highly stable miRNA reference genes (miR-26a/30d/152) suitable for normalizing miRNA expression in the fresh wounded skin samples or those with early postmortem changes, providing a reliable foundation for subsequent quantitative analyses^[25].

Building on our previous work, we further identified differentially expressed miRNAs and mRNAs in a murine model of incised skin wounds using next-generation sequencing^[25]. Through integrated bioinformatic prediction and experimental validation, we established a three-dimensional biomarker panel encompassing miRNAs, their cognate target mRNAs, and the corresponding encoded proteins. Such functionally linked multi-layer integration is expected to provide more robust and accurate estimates than single-molecule approaches. To robustly model the complex multifactorial dynamics of wound healing, we deployed machine learning approaches including Random Forest (RF) and XGBoost, which excel at deconvolving latent patterns in high-dimensional datasets, enabling identification of optimal modeling strategies tailored to the intrinsic dynamics of each biomarker triad. Moreover, we systematically evaluated the impact of postmortem changes on the predictive performance of this system, addressing a critical practical concern in forensic casework.

This study establishes a novel, integrated framework for wound age estimation that combines multi-level molecular biomarkers with computational modeling, improving accuracy and robustness against postmortem degradation.

To explore the temporal expression patterns of miRNAs and mRNAs during skin wound healing, we performed small RNA sequencing and RNA sequencing on skin samples collected from the wound edge at different post-injury intervals (12 h, 3 d, 7 d, and 14 d), as well as from normal control skin. PCA revealed distinct clustering patterns for both miRNAs and mRNAs. For miRNAs, the normal control and 12 h groups showed partial overlap, while samples from 3 d, 7 d, and 14 d were clearly separated from the control group and from each other, indicating dynamic changes in miRNA expression during wound healing [Figure 1A]. Similarly, PCA of mRNA expression profiles demonstrated a clear separation between normal control and all wound time points, with a progressive temporal shift in the clustering pattern [Figure 1B]. Differential expression analysis was further performed, and a total of 313 unique differentially expressed miRNAs were identified in wounded skin compared to normal controls. The number of differentially expressed miRNAs increased gradually from 12 h to 7 d and decreased slightly at 14 d [Figure 1C and D]. For mRNAs, 4,737 unique differentially expressed transcripts were detected in wounded skin relative to normal controls. The number of differentially expressed mRNAs increased progressively from 12 h to 14 d, with the highest number observed at 14 d post-injury [Figure 1E and F]. The number of differentially expressed mRNAs increased progressively from 12 h to 14 d, with the highest number observed at 14 d post-injury [Figure 1D]. These results indicate that both miRNAs and mRNAs undergo substantial and time-dependent changes during skin wound healing, providing a comprehensive expression landscape for further analysis.

We next visualized the global expression profiles by generating heatmaps for the top five differentially expressed miRNAs and mRNAs across normal skin and each post-injury time point [Figure 2A and B]. At 12 h post-injury, the most highly expressed mRNAs were enriched for chemokines and inflammation-related genes such as Il-1β and Cxcl10, indicating an early inflammatory response. In contrast, by 14 d, transcripts involved in tissue remodeling, including Krt28 and Tchh, became predominant, reflecting the progression of wound healing [Figure 2B]. Among the differentially expressed miRNAs, mmu-miR-15b-5p, mmu-miR-31-5p, and mmu-miR-1983 exhibited progressively increasing expression from 12 h to 14 d, displaying monotonic rather than biphasic expression patterns that are particularly suitable for wound age estimation [Figure 2A and Supplementary Figure 1A]. For mRNAs, the corresponding target genes Rsad2, Cst6, and Krt77 showed opposite trends, with expression levels decreasing over time [Supplementary Figure 1B]. Temporal expression patterns were further analyzed using STEM. The three miRNAs were clustered together in cluster 2, characterized by progressively increasing expression [Figure 2C], while their target mRNAs were enriched in cluster 9, showing a progressively decreasing trajectory (Cst6, Krt77), and Rsad2 peaking at 12 h and declining thereafter were enriched in cluster 9 [Figure 2D and E]. These coherent temporal trajectories suggested potential functional coupling between these miRNA-mRNA pairs. To systematically characterize the regulatory interactions, we constructed a miRNA-mRNA regulatory network by integrating miRNA target predictions (TargetScan, RNAhybrid, and miRanda) with differentially expressed mRNAs from RNA-seq data. Regulatory pairs with a PCC < -0.7 and P < 0.05 were considered significant. As shown in Figure 2E, the network comprised 150 miRNAs and 763 mRNAs, forming 1,573 negative regulatory edges [Table 1]. mmu-miR-15b-5p occupied a central position with the highest degree (degree = 93), and mmu-miR-31-5p also displayed extensive nodal interactions, targeting multiple downstream genes. Given the favorable monotonic expression patterns of these three miRNAs and their target mRNAs, along with their central positions in the regulatory network and their consistent inverse expression relationships, we selected mmu-miR-15b-5p/Rsad2, mmu-miR-31-5p/Cst6, and mmu-miR-1983/Krt77 for subsequent experimental validation.

Following the identification of mmu-miR-15b-5p/Rsad2, mmu-miR-31-5p/Cst6, and mmu-miR-1983/Krt77 as candidate regulatory axes with favorable temporal patterns and network centrality, we proceeded to validate their expression dynamics and regulatory relationships. RT-qPCR analysis revealed that all three miRNAs exhibited progressively increasing expression after injury, albeit with distinct peak times: mmu-miR-15b-5p peaked at day 9, mmu-miR-31-5p at day 17, and mmu-miR-1983 at day 13, followed by gradual declines [Figure 3A-]. In contrast, the corresponding target mRNAs-Rsad2, Cst6, and Krt77-displayed overall decreasing trends, with Rsad2 showing a transient elevation from 12 h to 1 d, while Cst6 and Krt77 declined progressively from 1 d onward, reaching their lowest levels between 7 d and 9 d before partial recovery [Figure 3D-F]. Dual-luciferase reporter assays confirmed direct targeting relationships: co-transfection with miRNA mimics significantly reduced luciferase activity for wild-type 3' untranslated region (3’UTR) constructs, but not for mutant constructs, validating Rsad2, Cst6, and Krt77 as direct targets of mmu-miR-15b-5p, mmu-miR-31-5p, and mmu-miR-1983, respectively [Figure 3G-I]. Finally, ELISA quantification of the corresponding proteins across the wound healing time course revealed expression patterns consistent with their respective mRNAs [Figure 3J-L]. Integrating these findings, we constructed three ‘miRNA-mRNA-protein’ triads-mmu-miR-15b-5p/Rsad2/RSAD2, mmu- miR-31-5p/Cst6/CST6, and mmu-miR-1983/Krt77/KRT77-each demonstrating coordinated inverse expression relationships, supporting their potential utility as integrated biomarkers for wound age estimation.

To evaluate the potential of these biomarkers for quantitative wound age estimation, we first assessed the performance of individual components within each triad using univariate linear regression. For each miRNA, target mRNA, and protein, we fitted linear models within their optimal post-injury time windows [Table 2]. As shown in Figure 4A-C and Table 2, mmu-miR-15b-5p exhibited a strong linear correlation with wound age within the first 9 days (R² = 0.9657), whereas Rsad2 mRNA and RSAD2 protein showed weaker negative correlations (R² = 0.5401 and 0.7100, respectively). Similar patterns were observed for the other two triads: miRNAs consistently outperformed their corresponding mRNAs and proteins in linear fit, with mmu-miR-31-5p achieving an R² of 0.8490 over 17 days [Figure 4D-F], and mmu-miR-1983 reaching 0.9498 over 13 days [Figure 4G-I]. The moderate predictive power of single mRNA or protein markers highlighted the need for multi-parameter integration.

Given the limitations of single-biomarker approaches, we next constructed multivariate models incorporating all three components within each triad. Multivariate linear regression substantially improved model fit compared to univariate approaches, with adjusted R² values reaching 0.9813 (within 9 d) for the mmu-miR-15b-5p/Rsad2/RSAD2 triad, 0.9121 (within 13 d) for mmu-miR-31-5p/Cst6/CST6, and 0.9643 (within 13 d) for mmu-miR-1983 /Krt77/KRT77. External validation confirmed acceptable predictive errors (MAE: 1.240, 1.299, and 0.872, respectively) [Table 3 and Figure 5A-C]. Given the non-linear dynamics of wound healing, we further explored machine learning approaches using RF and XGBoost [Tables 4-6]. For the mmu-miR-15b-5p/Rsad2/RSAD2 triad, both RF and XGBoost achieved optimal performance within 9 days post-injury [Figure 5D and E]. Within this window, RF achieved an R² of 0.861, RMSE of 1.20, and MAE of 0.87, while XGBoost showed comparable fit (R² = 0.861, RMSE = 1.20) with a slightly lower MAE (0.72). For the mmu-miR-31-5p/Cst6/CST6 triad, optimal performance was observed within 13 days [Figure 5F and G]. RF yielded an R² of 0.926, RMSE of 1.22, and MAE of 0.94, whereas XGBoost performed slightly better (R² = 0.942, RMSE = 1.09, MAE = 0.74). For the mmu-miR-1983/Krt77/KRT77 triad, the optimal window was 13 days [Figure 5H and I]. XGBoost substantially outperformed RF, with R² = 0.941 (vs. 0.844), RMSE = 1.10 (vs. 1.78), and MAE = 0.86 (vs. 1.28). Collectively, these results demonstrate that the triad system, when combined with appropriate multivariate modeling strategies, significantly improves the accuracy of wound age estimation. Among the two machine learning algorithms, XGBoost generally provided better or comparable performance, particularly for the mmu-miR-1983/Krt77/KRT77 triad.

Having established the predictive performance under optimal conditions, we next assessed whether this triad system could withstand postmortem degradation - a critical consideration for forensic applications. To evaluate the robustness of the triad system against postmortem degradation, we assessed model performance using samples subjected to different postmortem intervals (PMI). Skin tissues were collected from normal controls and at 1 d, 5 d, and 13 d post-injury. Samples were maintained at 25 ± 2 °C and 55% ± 3% humidity, and collected at PMI of 0, 1.5, 3, 4.5 d and 6d (PMI = 6 d samples were collected but not included in data analysis due to serious degradation. All results below are based on PMI ≤ 4.5d). Expression levels were input into the established multivariate models (multivariate linear regression, RF, and XGBoost) to evaluate the impact of postmortem changes on estimation accuracy [Figure 6 and Tables 7-9]. The mmu-miR-31-5p/Cst6/CST6 triad showed the greatest resistance to postmortem interference. At PMI = 0, all three models performed well (R² = 0.858-0.876). Performance remained stable with increasing PMI: at PMI = 1.5 d, RF and XGBoost achieved R² = 0.956 and 0.960 with MAE as low as 0.77 and 0.60; at PMI = 3 d, XGBoost maintained R² = 0.932 (MAE = 1.04); and even at PMI = 4.5 d, XGBoost retained R² = 0.923 (MAE = 1.16). When all PMI data were combined, XGBoost yielded the best overall performance (R² = 0.908, RMSE = 1.56, MAE = 1.05), indicating that this triad retains considerable predictive power under postmortem conditions. In contrast, the other two triads exhibited substantial performance degradation with increasing PMI, with predictive performance declining notably even at early PMI time points [Tables 7-9]. Collectively, these results indicate that the robustness of the triad system against postmortem changes varies considerably across triads. The mmu-miR-31-5p/Cst6/CST6 triad demonstrated superior resistance to postmortem degradation, particularly when combined with XGBoost, whereas the other two triads were more susceptible to postmortem interference. These findings underscore the importance of selecting appropriate biomarker sets and modeling strategies for wound age estimation in postmortem settings.

Accurate estimation of skin wound age remains a critical challenge in forensic pathology, as traditional single-marker approaches are often compromised by various pre- and post-mortem factors^[26,27]. This study aimed to overcome these limitations by integrating multi-omics technologies with machine learning algorithms to develop a novel, reliable method for wound age estimation. We successfully constructed and validated a “miRNA-mRNA-protein triad system” based on mmu-miR-15b-5p/Rsad2/RSAD2, mmu-miR-31-5p/Cst6/CST6, and mmu-miR-1983/Krt77/KRT77. For each biomarker set, we identified optimal mathematical models, and systematically evaluated the impact of postmortem changes on estimation accuracy. Our results demonstrate that this integrated strategy enhances both accuracy and robustness, offering a new methodological framework for forensic practice.

Skin wound healing is a complex, multi-phase process involving coordinated regulation of numerous genes and signaling pathways. Using next-generation sequencing, we characterized the temporal expression landscapes of miRNAs and mRNAs during wound healing, identifying 313 differentially expressed miRNAs and 4,737 differentially expressed mRNAs. Consistent with previous studies^[28,29], both miRNA and mRNA expression exhibited time-dependent patterns, with monotonic expression trajectories - such as those exhibited by mmu-miR-15b-5p, mmu-miR-31-5p, and mmu-miR-1983 - being particularly advantageous for wound age estimation. By integrating three complementary miRNA target prediction algorithms with our RNA-seq data, we constructed a robust miRNA-mRNA regulatory network. The centrality of mmu-miR-15b-5p (degree = 93) and mmu-miR-31-5p within this network, together with their monotonic expression patterns, supported their selection as candidate biomarkers. This network-based approach not only improved prediction accuracy but also provided insight into the molecular architecture underlying wound healing.

The biological relevance of the selected miRNA-mRNA axes is supported by existing literature. mmu-miR-15b-5p, a member of the miR-15 family, has been implicated in regulating apoptosis, proliferation, angiogenesis, and inflammation^[30-32]. Its target, Rsad2, is an interferon-stimulated gene involved in antiviral responses and modulation of type I interferon signaling^[33-36]. The inverse expression pattern observed in our study suggests a potential regulatory axis linking miRNA-mediated suppression of inflammatory responses to wound repair dynamics. Similarly, mmu-miR-31-5p has been shown to promote re-epithelialization via RAS/MAPK signaling and is associated with hypertrophic scar formation through HIF-1α pathway activation^[37]. Its target, Cst6 (cystatin M/E), is predominantly expressed in skin and localizes to the suprabasal layer of healing epidermis^[38], consistent with our observed expression pattern. Although mmu-miR-1983 and its target Krt77 have been less characterized in wound healing contexts, their monotonic expression trajectories and network centrality suggest functional relevance warranting further investigation.

Given the limitations of single-marker approaches, we next evaluated the performance of individual biomarkers and then constructed multivariate models incorporating all three components within each triad. Univariate linear regression revealed that while miRNAs exhibited strong linear correlations with wound age (R² = 0.849-0.966), their corresponding mRNAs and proteins showed weaker predictive power (R² = 0.540-0.808). Multivariate linear regression substantially improved model fit, with adjusted R² values reaching 0.912-0.981, supporting the notion that functionally linked biomarkers provide more reliable estimates when analyzed collectively. Considering the non-linear dynamics of wound healing, we further explored machine learning approaches (RF and XGBoost). Notably, the optimal model varied across triads, reflecting distinct molecular dynamics. For the mmu-miR-15b-5p/Rsad2/RSAD2 triad, both models performed comparably within the 9 d window. For the mmu-miR-31-5p/Cst6/CST6 triad, XGBoost yielded better fit (R² = 0.942 vs. 0.926), while for the mmu-miR-1983/Krt77/KRT77 triad, XGBoost substantially outperformed RF (R² = 0.941 vs. 0.844). These differences can be attributed to the distinct characteristics of the two algorithms: RF may offer greater robustness against noise and overfitting through bootstrap aggregation and random feature selection, making it well-suited for datasets with high variability. In contrast, XGBoost iteratively optimizes residuals and incorporates regularization, which may enable higher predictive accuracy when underlying patterns are relatively consistent, particularly within shorter, well-defined time windows (9-13 d). These findings suggest that algorithm selection should be tailored to the intrinsic dynamics of the molecular markers rather than seeking a universal model.

A critical consideration for forensic applications is biomarker stability under postmortem conditions. To our knowledge, this study represents the first systematic evaluation of how postmortem changes affect the predictive accuracy of a multi-marker wound age estimation system. We found that the mmu-miR-31-5p/Cst6/CST6 triad exhibited reliable resistance to postmortem degradation, with XGBoost maintaining R² > 0.92 even at PMI = 4.5 d. In contrast, the other two triads showed substantial performance deterioration with increasing PMI. The greater stability of the mmu-miR-31-5p/Cst6/CST6 triad may reflect protective mechanisms such as the sequestration of miR-31-5p by RNA-binding proteins or extracellular vesicles, and the unique epidermal localization of CST6, which may limit its exposure to degradative enzymes^[39-41]. Further mechanistic studies are needed to confirm these possibilities. Nevertheless, our findings provide a proof-of-concept that integrating multiple molecular layers with appropriate modeling strategies can partially mitigate postmortem effects.

Several limitations should be acknowledged. First, our study was conducted in a mouse model, and the molecular timeline of skin wound healing differs from that in humans. Although the selected triads are evolutionarily conserved, the specific time windows derived from this model may not directly translate to human cases; future validation using human skin wound samples is therefore necessary. Second, although the mmu-miR-31-5p/Cst6/CST6 triad demonstrated considerable stability under the tested postmortem conditions (25 ± 2 °C, 55% ± 3% humidity), these conditions represent only one controlled environmental scenario. In forensic practice, variable temperature, humidity, and other factors (e.g., insect activity, clothing) may differentially affect biomarker stability, and further validation under a broader range of postmortem conditions is needed. Third, although our sample sizes (n = 3-5 per time point) are consistent with exploratory studies, they are modest for training XGBoost. To mitigate overfitting, we used five-fold cross-validation. However, confidence intervals for performance metrics are expected to be wide due to the small dataset. Therefore, the current models should be considered proof-of-concept, and validation in larger cohorts is needed. Fourth, beyond 4.5 d postmortem, the “predictive performance” of all triads was compromised remarkably due to postmortem degradation, underscoring that tissue decomposition remains a major interfering factor for wound age estimation.

In summary, this study demonstrates that forensic wound age estimation is transitioning from traditional morphological and single-marker analyses toward a multi-marker, data-driven intelligent computational paradigm. By integrating high-throughput sequencing, systematic experimental validation, and advanced machine learning algorithms, we established a triad system that significantly improves estimation accuracy. Notably, we identified that different biomarker triads are best suited to different modeling approaches, and that postmortem robustness varies substantially across triads. These findings provide both an experimental basis for enhancing wound age estimation accuracy and a methodological framework for selecting appropriate biomarker-model combinations under challenging conditions.