- Research
- Open access
- Published:
Microbial community profiling for forensic drowning diagnosis across locations and submersion times
BMC Microbiology volume 25, Article number: 244 (2025)
Abstract
Background
Drowning diagnosis has long been a critical issue in forensic research, influenced by various factors such as the environment and decomposition time. While traditional methods such as diatom analysis have limitations in decomposed remains, microbial community profiling offers a promising alternative. With the advancement of high-throughput sequencing technology, forensic microbiology has become a prominent focus in the field, providing new research avenues for drowning diagnosis. During drowning, microbial communities enter the lung tissue along with the water.
Methods
In this study, using a murine model, we collected samples from three rivers at random sites at postmortem intervals (PMI) of 1, 4, and 7 days to comprehensively evaluate the differences in microbial communities between mice subjected to drowning versus postmortem immersion.
Results
The α-diversity analysis revealed that the observed Operational Taxonomic Units (OTUs) for the drowning group on day 1 was 234.77 ± 16.60, significantly higher than the postmortem immersion group (171.32 ± 9.22), indicating greater initial microbial richness in the drowning group. Additionally, Shannon index analysis showed a significant decline in evenness in the postmortem immersion group on day 7 (1.46 ± 0.09), whereas the drowning group remained relatively stable (2.38 ± 0.15), further indicating a rapid decrease in microbial diversity in the postmortem immersion group over time. PCoA analysis demonstrated that differences in microbial community composition between drowning and postmortem immersion groups were notably stable. Key microbial taxa differentiating the groups were identified through LEfSe analysis, with Enterococcaceae (family), Escherichia-Shigella (genus), and Proteus (genus), emerging as significant markers in drowning cases. A random forest model, trained using microbial community data, exhibited high predictive accuracy (AUC = 0.96) across locations and immersion times and identified microbial markers, including Enterococcaceae (family), Lactobacillales (order), Morganellaceae (family), as critical features influencing model performance.
Conclusion
These findings underscore the potential of combining 16 S rRNA sequencing with machine learning as a powerful tool for drowning diagnosis, offering novel insights into forensic microbiology.
Introduction
Drowning ranks as the third most common cause of unintentional fatalities globally [1]. Drowning diagnosis typically involves a thorough evaluation that includes external examinations, autopsies, and laboratory tests. In cases where a body is significantly decomposed or reduced to skeleton, external signs of drowning may be absent, making laboratory testing a crucial support in the diagnosis of drowning.
The acid digestion-centrifugal enrichment-light microscopy examination method (a diatom extraction technique) remains the most commonly used method for drowning diagnosis in domestic and international forensic laboratories due to its cost-effectiveness and straightforward operational procedures, however this method has drawbacks such as violent reactions, low safety, environmental contamination, and low diatom recovery rates [2]. Furthermore, diatoms can infiltrate the gastrointestinal tract through various avenues such as living and working environments, as well as dietary intake, and the residue of diatom shells can lead to false positives in morphological methods [3,4,5,6]. Compared to diatoms, which have a particle size ranging from 2 to over 500 μm, aquatic bacteria with smaller particle sizes (0.2 ~ 2 μm) possess a greater propensity to traverse the alveolar-capillary barrier and enter the bloodstream, thereby disseminating to various organs throughout the body [7, 8]. AOYAGI et al. utilized PCR to detect the DNA fragments of Aeromonas hydrophila in the blood samples of drowned bodies for drowning diagnosis [9]; Uchiyama et al. developed a real-time fluorescence PCR system based on the TaqMan probe method and successfully detected aquatic bacteria in 266 drowned cadavers [10]; Rutty et al. employed TaqMan PCR to detect Aeromonas sp., Vibrio sp., and Photobacterium sp. in drowning organs for drowning diagnosis [11]. Nonetheless, the application of PCR detection of planktonic bacteria in drowning diagnosis is limited by the target species of the primers, which may not cover all bacterial species relevant to drowning, leading to the potential non-detection of certain aquatic bacteria.
The rapid development of high-throughput technology has brought new possibilities for drowning diagnosis. 16 S rRNA gene sequencing, which is widely used in microbial community research, can effectively reveal the overall profile of microbial populations and accurately identify the microbial species present in samples [12]. In recent years, researchers have begun to investigate the potential of 16 S rRNA sequencing in the context of drowning diagnosis. For instance, Lee et al. performed 16 S rRNA sequencing on samples from water, lung, closed organs (specifically the kidney and liver), and heart blood in drowned rats, revealing a significant presence of aquatic microorganisms in the closed organs and heart blood of the drowning group, which were absent in the postmortem immersion group [13]. Similarly, Wang et al. utilized UniFrac-based Principal Coordinates Analysis (PCoA) to analyze 16 S rRNA sequencing data from drowning and postmortem submersion rat, successfully differentiating between the drowning and postmortem submersion groups based on samples from the skin, lung, blood, and liver [14].
This study aims to reveal the characteristics of microbial communities between mice in the drowning and postmortem submersion groups through 16 S rDNA sequencing, identify characteristic microbial markers related to drowning, and develop a drowning inference model using machine learning methods, contributing novel insights to the field of forensic medicine, particularly in the context of drowning diagnosis.
Methods
Animals
A total of 324 eight-week-old male C57BL/6 N mice were purchased from Zhejiang Vital River (Zhejiang, China). All animals were group-housed with their littermates in a dedicated housing room under a 12-h light/12-h dark cycle, and food and water were available ad libitum. All animal experiments were conducted in strict accordance with the guidelines and regulations of the Guangdong Zhiyuan Biomedical Technology Co., LTD Animal Ethics Committee, ensuring animal welfare and humane care throughout the study. The experimental protocol was reviewed and approved by the Guangdong Zhiyuan Biomedical Technology Co., LTD Animal Ethics Committee (Approval No. IAEC-20211105), ensuring compliance with international animal research ethical standards.
The mice were randomly divided into two groups: the drowning group (N = 162) and the postmortem submersion group (N = 162). Each mouse was individually numbered and marked to ensure accurate tracking during subsequent experiments. Both drowning and postmortem submersion groups were further divided into three subgroups (N = 54), corresponding to the three water samples (L1, L2, L3) from the Northwestern River (Figure S1). Lung samples were collected from both groups at three postmortem time points: 1, 4, and 7 days. Each subgroup consisted of 18 mice per time point, ensuring consistent sample sizes for comparative analysis.
Drowning and tissue Preparation
The drowning model was established as previously reported [13]. The experimental drowning model was conducted as follows: three water samples (L1, L2, L3) from the Northwestern River were used for the drowning procedure. For both the drowned and postmortem submersion groups, sterile plastic cages were filled with 100 L of water. Drowned mice (n = 164) were placed into sterile cages at the corresponding locations and submerged in water until death. Mice in the postmortem submersion group (n = 164) were anesthetized with 2% chloral hydrate (0.18 mL/20 g) and euthanized by cervical dislocation and then submerged in sterile cages at the corresponding sites.
Mice were retrieved from the water at three distinct time points: 1 day, 4 days, and 7 days post-submersion. Lung tissue samples were collected and labeled as L1-D, L2-D, and L3-D for the drowning group and L1-PS, L2-PS, and L3-PS for the postmortem submersion group. These labels were further categorized by immersion time: 1 day, 4 days, and 7 days. Each subgroup consisted of 18 mice, ensuring consistent sample sizes across experimental conditions.
All lung tissue samples were immediately flash-frozen in liquid nitrogen upon removal to minimize tissue degradation. The frozen samples were stored at − 80 °C for subsequent histological and molecular biological analysis [13].
DNA extraction
For DNA extraction from mouse lung tissue, the tissue samples were first ground in a mortar filled with liquid nitrogen to ensure thorough cell disruption at low temperatures, thereby minimizing the risk of nucleic acid degradation. The ground homogenized tissue samples were then processed using the EZNA Water DNA Kit (Omega, USA) according to the manufacturer’s instructions. This kit is designed for efficient DNA isolation from tissues and biological samples, ensuring the recovery of high-quality genomic DNA suitable for subsequent molecular biology experiments. All DNA extractions were performed under sterile conditions with stringent precautions to prevent contamination.
PCR amplification and sequencing
The V3-V4 region of the 16 S rDNA from mouse samples was amplified using primers 340 F (5’-CCTACGGGNBGCASCAG-3’) and 805R (5’-GACTACNVGGGTATCTAATCC-3’) to analyze microbial communities. PCR amplification followed the protocol of previously reported [15].The first PCR reaction was performed in a 25 µL system containing 12.5 µL of 2× KAPA HiFi HotStart Ready Mix Buffer (Kapa Biosystems, Wilmington, MA, USA), 2.5 µL of extracted DNA (5 ng/µL), and 5 µL of each primer (1 µM). The amplification conditions included an initial denaturation and enzyme activation step at 95 °C for 3 min, followed by 25 cycles of 95 °C for 30 s, 55 °C for 30 s, and 72 °C for 30 s, with a final extension at 72 °C for 5 min. PCR products were confirmed by 2% agarose gel electrophoresis, and bands matching the expected size were purified using magnetic beads to remove impurities. The purified products served as templates for a second round of PCR under the same conditions. Following purification, the amplicons were further cleaned using the MoBio UltraClean PCR purification kit. A total of 300 ng of amplicons from each sample were pooled to ensure consistent amplification. Sequencing was performed on the Illumina NovaSeq 6000 platform using PE250 paired-end sequencing at Lianchuan Biotechnology Co., Ltd (Hangzhou, China).
Bioinformatics analysis and statistical methods
The raw sequencing data were first subjected to quality control using fastp software (version 0.20.0) to ensure high-quality sequences [16]. After quality control, sequences were merged using FLASH software (version 1.2.7) to generate high-quality paired-end sequences [17]. The merged sequences were then denoised using the Deblur tool [18] and imported into the QIIME2 platform (version 2023.2) for further analysis [19]. Taxonomic classification was performed using the SILVA database (version 138) downloaded from the QIIME2 website, with classification done using the classify-sklearn method. The 16 S rDNA copy number was adjusted accordingly. Following taxonomic annotation, data were collapsed to various taxonomic levels, and pairwise comparisons of Observed OTU and Shannon indices were performed using the Kruskal-Wallis test. Beta diversity was assessed using the Bray-Curtis distance, and statistical testing was performed through Permutational Multivariate Analysis of Variance (PERMANOVA). A p-value of less than 0.05 was considered statistically significant (*p < 0.05, **p < 0.01, and ***p < 0.001).
Data visualization and model analysis
Data visualization of microbial community structures was conducted using Principal Coordinates Analysis (PCoA) in R software v4.4.1, combining qiime2R and ggplot2 packages. PCoA was used to capture differences in community structures between groups, and the results were displayed in a two-dimensional plot, highlighting group similarities and differences. Further analysis of significant taxa across groups was conducted using the microeco package, with Linear Discriminant Analysis Effect Size (LEfSe) used to identify significantly different taxa. LEfSe was run with default parameters to identify biomarkers and rank them by effect size. For classification model construction, the Random Forest algorithm in the H2O package in R software v4.4.1 was used. The dataset was split into training sets (samples from L1 and L2) and a test set (samples from L3). The model underwent 100 iterations of optimization, and key microbial features were identified based on their importance in the classification. A 10-fold cross-validation was applied to enhance the model’s stability and accuracy, minimizing error and identifying biomarkers. The final model’s performance was evaluated through classification accuracy and the Area Under the Receiver Operating Characteristic (ROC) curve, with the Area Under the Curve (AUC). The ROC curves were plotted using the pROC package in R software v4.4.1 and provided a visual representation of the model’s classification capability across different decision thresholds. A larger AUC indicated a stronger model performance in distinguishing between groups.
Results
Differences in alpha diversity between drowning and postmortem submersion groups
We analyzed the alpha diversity of lung tissues between the drowning and postmortem submersion groups and found significant differences in the number of observed OTUs and the Shannon diversity index (Fig. 1). On Day 1, the drowning group had an OTU count of 234.77 ± 16.60, whereas the postmortem submersion group had a count of 171.32 ± 9.22. By Day 4, the OTU count in the drowning group had decreased to 79.73 ± 7.84, compared to 110.08 ± 8.27 in the postmortem submersion group. On Day 7, the OTU count in the drowning group further declined to 50.38 ± 1.90, while the postmortem submersion group had an OTU count of 65.89 ± 4.46 (Fig. 1A). Initially, on Day 1, the drowning group had a significantly higher OTU count than the postmortem submersion group; however, this pattern shifted by Day 4. The drowning group showed a sharp decrease in OTU count with increasing postmortem interval (PMI), while the postmortem submersion group exhibited a more gradual decline. Regarding the Shannon diversity index, on Day 1, the drowning group had a value of 3.60 ± 0.15, compared to 4.01 ± 0.10 in the postmortem submersion group. By Day 4, the index for the drowning group decreased to 2.34 ± 0.14, whereas the postmortem submersion group’s index dropped to 1.85 ± 0.16. On Day 7, the Shannon index for the drowning group was 2.38 ± 0.15, while the postmortem submersion group had an index of 1.46 ± 0.09 (Fig. 1B). On Day 1, the Shannon index was significantly lower in the drowning group compared to the postmortem submersion group, but this trend shifted by Day 4. Both groups experienced a decline in the Shannon index from Day 1 to Day 4, with a more pronounced decrease observed in the postmortem submersion group.
Principal coordinate analysis (PCoA) of microbial communities
To investigate differences in microbial communities in mouse lung tissues between the drowning and postmortem submersion groups, we performed PCoA using Bray-Curtis distances, as shown in Fig. 2. Different colors in the scatter plots represent lung samples collected between the drowning and postmortem submersion groups, different sizes indicate different PMI, and various shapes denote different collection points. The first three principal coordinates accounted for 14.56%, 12.35%, and 8.44% of the variation between these microbial communities. Under the separation of PCo1 and PCo2, drowning samples predominantly cluster in the lower right quadrant (Fig. 2A). When distinguishing between PCo1 and PCo3, drowning samples are primarily positioned on the right side (Fig. 2B), whereas separation based on PCo2 and PCo3 shows a mix between drowning and postmortem submersion groups (Fig. 2C). These results suggest that microbial communities can differentiate between the drowning and postmortem submersion groups. Additionally, PERMANOVA analysis based on Bray-Curtis distances revealed significant variation between the drowning and postmortem submersion groups (pseudo-F = 32.79, **p = 0.001), providing a clearer visualization of these differences.
Principal Coordinate Analysis (PCoA) of lung samples (different colors represent different groups, different sizes represent different PMIs, and different shapes represent different water collection points). (A) PCoA results based on PCo1 and PCo2; (B) PCoA results based on PCo1 and PCo3; (C) PCoA results based on PCo2 and PCo3. D, drowning group; PS, postmortem submersion group
LEfSe analysis reveals differential species between groups
We utilized LEfSe to identify differential species within microbial communities between the drowning and postmortem submersion groups, as depicted in Fig. 3. In this figure, different colors represent different stages of decomposition. The cladogram illustrates the taxonomic hierarchy from phylum to genus (from the inner to the outer circles), with node sizes indicating the average relative abundance of the taxonomic units. Notably, hollow nodes indicate taxonomic units with no significant differences between groups, while colored nodes highlight units with significant inter-group variability and higher relative abundance in the samples represented by the corresponding color.
Several microbes with high Linear Discriminant Analysis (LDA) scores were identified (Fig. 4). At the phylum level, Firmicutes and Proteobacteria were notable; at the class level, Bacilli and Alphaproteobacteria; at the order level, Aeromonadales, Lactobacillales, Clostridiales, Burkholderiales, and Rhizobiales; at the family level, Enterobacteriaceae, Enterococcaceae, Clostridiaceae, Fusobacteriaceae, Morganellaceae, Burkholderiaceae, Yersiniaceae, and Rhizobiaceae; and at the genus level, Escherichia-Shigella sp., Aeromonas sp., Enterococcus sp., Clostridium sensu stricto 1 sp., Enterobacter sp., Proteus sp., Serratia sp., Ralstonia sp., and Burkholderia-Caballeronia-Paraburkholderia sp.. At the species level, Clostridium botulinum and Clostridium butyricum were identified. Interestingly, Firmicutes (phylum), Bacilli (class), Lactobacillales (order), Enterococcaceae (family), and Enterococcus sp. (genus) not only exhibited the highest LDA scores in drowning samples but also belong to the same taxonomic branch from a classification perspective. These markers have the potential to serve as distinct microbial signatures for differentiating between the drowning and postmortem submersion groups.
Random forest model performance and evaluation
The H2O random forest model, trained on samples from L1 and L2 using 10-fold cross-validation, exhibited robust performance across multiple evaluation metrics. The model demonstrated a rapid learning curve, with error classification rates significantly decreasing as the number of trees increased (Fig. 5A). It achieved a mean squared error (MSE) of 0.0497 and a root mean squared error (RMSE) of 0.2230, indicating a low prediction error on the training data. The log loss was 0.2132, reflecting strong probabilistic prediction performance. The mean per-class error was 0.0185, showing a low classification error rate. The model also demonstrated high accuracy, with an area under the curve (AUC) of 0.9919 and an area under the precision-recall curve (AUCPR) of 0.9923. Additionally, the Gini coefficient was 0.9839, and the R² value was 0.8011, indicating that the model explains 80.11% of the variance in the data. These results, derived from the cross-validated model, highlight its excellent predictive power and generalization capability.
Random forest model establishment and prediction. (A) Learning curve of the RF model. (B) Variable importance ranking of the RF model (top 10). (C) ROC curve for predicting L3 samples with the gray area representing the 95% confidence interval. (D) Graphical representation of SHAP feature importance. The labels (c, o, f, g) before each taxon represented class, order, family, and genus respectively
Feature importance analysis of the H2O random forest model identified several key variables that significantly impacted the model’s performance. The most influential feature was Enterobacteriaceae (family), followed by Enterococcaceae (family), Lactobacillales (order), Morganellaceae (family), Burkholderiales (order), Enterococcus (family), Proteus sp. (genus), Rhizobiales (order), Alphaproteobacteria (class), and Yersiniaceae (family) (Fig. 5B).
To further validate the effectiveness of the H2O random forest model, we evaluated its performance by predicting results from 108 samples from L3 (Table S1). The model achieved an accuracy of 90.74% and an AUC of 0.96, demonstrating high predictive accuracy and confirming its effectiveness on new data. This result underscores the model’s robust generalization capability and applicability across different datasets (Fig. 5C). The model’s performance varied over time: on Day 1, it achieved an impressive accuracy of 98.22% and an AUC of 1.00, reflecting near-perfect identification capability in the early stages. By Day 4, the model maintained a strong performance with an accuracy of 88.89% and an AUC of 0.96. On Day 7, accuracy slightly declined to 86.11%, but the AUC remained stable at 0.96. These results indicate that the model’s predictive ability is strongest in the early stages, with high accuracy and AUC scores, though accuracy diminishes over time (Figure S2).
To interpret the model’s predictions, SHAP (SHapley Additive exPlanations) values were computed, providing a detailed understanding of feature contributions. SHAP analysis identified Morganellaceae (family), Enterobacteriaceae (family), Lactobacillales (order), Alphaproteobacteria (class), and Burkholderiales (order) as the most influential features affecting model predictions (Fig. 5D). These findings align with the model’s feature importance analysis, highlighting the critical role of these taxes in driving the model’s output. The results emphasize the significance of these microbial taxa in shaping model predictions and provide valuable insights into feature importance within the model.
Discussion
The application of Next-Generation Sequencing (NGS) in forensic microbiology has emerged as a powerful tool for differentiating between drowning and postmortem submersion groups [13]. Recent studies have demonstrated that NGS, through the analysis of bacterial community profiles in various tissues such as lungs, blood, and liver, provides forensic scientists with a more precise method for diagnosing drowning. For instance, research conducted in 2020 utilized unweighted UniFrac-based PCoA to clearly distinguish drowned rats from those submerged postmortem based on their microbial communities [14]. Our findings further reveal significant differences in the microbial communities of lung tissues from mice across different locations and PMI between the drowning and postmortem submersion groups. This approach enhances the accuracy of forensic diagnoses, complementing traditional methods like diatom analysis.
Our alpha diversity analysis revealed distinct microbial community dynamics between the two groups. The drowning group exhibited a significantly higher initial OTU count on Day 1 (234.77 ± 16.60 vs. 171.32 ± 9.22 in the postmortem submersion group), suggesting an influx of waterborne microbes during active drowning. This aligns with previous studies where aspiration of water introduces exogenous microorganisms into the lungs, transiently increasing microbial richness [13]. However, the sharp decline in OTUs in the drowning group over time (79.73 ± 7.84 on Day 4 and 50.38 ± 1.90 on Day 7) contrasts with the slower decline in the postmortem submersion group (110.08 ± 8.27 on Day 4 and 65.89 ± 4.46 on Day 7). This pattern may reflect differential decomposition dynamics: in drowning, initial microbial enrichment from water aspiration is followed by rapid host-mediated clearance and tissue degradation, whereas postmortem submersion involves gradual colonization by environmental bacteria without prior host immune activity. The Shannon index, which accounts for both richness and evenness, showed a reversal in trends: lower in the drowning group on Day 1 (3.60 ± 0.15 vs. 4.01 ± 0.10) but higher by Day 7 (2.38 ± 0.15 vs. 1.46 ± 0.09). This suggests that drowning creates a transiently diverse but unstable microbial community, while postmortem submersion allows for slower but more uniform colonization by environmental taxa. Similar shifts in diversity indices have been observed in decomposition studies, where early-stage microbial heterogeneity gives way to specialized decomposer communities [20, 21].
Principal Coordinate Analysis (PCoA) further corroborated these findings, with drowning samples clustering separately from postmortem submersion groups along PCo1 (14.56% variance) and PCo3 (8.44% variance). The distinct clustering (pseudo-F = 32.79, **p = 0.001) aligns with prior work using UniFrac distances to distinguish drowning-associated microbiomes [14]. Notably, the drowning group’s positioning in the lower right quadrant of the PCo1-PCo2 plot (Fig. 2A) suggests unique taxonomic signatures driven by water aspiration. This spatial separation mirrors findings in marine vs. freshwater drowning studies, where environmental microbial inputs dominate early community structure [15, 22]. The temporal dispersion of samples (indicated by size gradients) highlights PMI-dependent succession patterns, consistent with established models of postmortem microbial ecology [20, 23].
One of the significant challenges in modeling microbial communities for forensic analysis is the variability introduced by different PMI and drowning locations. Microbial communities are highly sensitive to environmental factors such as temperature, water composition, and geographic location, which can lead to considerable variation in community structure over time and across different sites [15, 24]. This variability complicates the construction of robust machine learning models, as the input data may exhibit substantial heterogeneity, potentially reducing the accuracy and generalizability of predictive models. Environmental factors significantly influence postmortem microbial succession, making it difficult to standardize forensic models [25, 26]. Additionally, microbial communities undergo dynamic shifts depending on the decomposition stage and external conditions, adding complexity to forensic microbiome analyses [8, 27]. Similarly, Christian et al. emphasized that the temporal and spatial variability in microbial community composition requires tailored approaches to model building in forensic contexts [28]. Sheree et al. also observed that the microbial communities associated with cadavers exhibit considerable diversity depending on the surrounding environment, which poses challenges for creating universally applicable forensic tools [29]. Finally, our recent study highlighted the importance of considering site-specific data when developing predictive models, as microbial responses to environmental conditions can significantly alter the data used for model training [15]. Our study’s random forest model further expands the application of microbial analysis in forensic science by demonstrating its ability to predict outcomes across different locations and time points. The model’s performance, characterized by high accuracy and stability, underscores its robustness in varying environmental conditions. Temporal variation in model performance, particularly the highest accuracy of 98.22% on Day 1 and a slight decline to 86.11% by Day 7, underscores the significance of early detection. Microbial communities often exhibit rapid shifts in composition shortly after disturbances, such as infections or environmental changes [30]. The integration of SHAP values provided a transparent and interpretable framework for understanding the RF model’s predictions. SHAP analysis corroborated the feature importance findings, identifying Morganellaceae, Enterobacteriaceae, and other microbial taxa as major drivers of the model’s predictions. This reinforces the model’s biological plausibility and strengthens confidence in its predictions. The use of SHAP in microbiome studies is emerging as a robust tool for understanding feature contributions, offering a level of interpretability that is often missing in black-box models [31]. Future studies should aim to integrate SHAP and other interpretability tools to further enhance model transparency and applicability.
Traditional bacterial and diatom indicators like Aeromonas sp. and other planktonic bacteria have indeed been used in forensic science to help diagnose drowning [10, 32, 33]. For example, Aeromonas sp. are commonly found in freshwater environments and can be used to indicate that a person drowned in such a setting [9, 10, 34]. This is consistent with our analysis results in LEfSe, where Aeromonas sp. was highly enriched in the drowning group. Similarly, the presence of other planktonic bacteria can help forensic scientists determine the type of water (freshwater, saltwater, etc.) and even the specific location where the drowning took place [15, 35]. Our model’s integration of comprehensive microbial community analysis offers a more refined approach. It aligns with the growing body of research emphasizing the importance of assessing the entire microbiome rather than relying solely on individual bacterial markers. Furthermore, previous studies have shown that marine and freshwater bacterioplankton do not easily invade bodies postmortem, helping to distinguish drowning from other causes of death [35]. This supports our findings, where specific microbial taxa were identified as key predictors in our model.
The random forest model’s analysis identified key microbial markers that were strongly associated with drowning cases, shedding light on their diagnostic potential. The presence of specific microbial taxa, such as Enterobacteriaceae, Morganellaceae, and Lactobacillales, plays a crucial role in the model’s predictive performance, reinforcing their potential use in postmortem microbiome analysis for drowning diagnosis. Enterobacteriaceae, for example, was the most influential feature in the model’s predictions. This family is commonly found in aquatic environments and can be introduced into the body through water inhalation during drowning [36]. Morganellaceae, another important microbial family identified in the model, is also associated with aquatic environments and has been linked to waterborne infections [37]. Its detection in the model’s feature importance and SHAP analysis indicates that Morganellaceae may serve as a reliable marker for postmortem submersion, given its frequent occurrence in water-related fatalities [23]. The identification of Lactobacillales, a group of lactic acid bacteria, as a key feature in the model is particularly intriguing [38]. While Lactobacillales are typically associated with the gut and mucosal surfaces, their presence in forensic samples could be indicative of postmortem translocation or contamination from the environment [39,40,41,42]. Studies have shown that Lactobacillales can survive in diverse environments, including aquatic settings, and their identification in postmortem samples might signal immersion in water [43, 44]. The significant expression of Alphaproteobacteria and Burkholderiales in the postmortem submersion group points to their potential role as microbial markers for drowning and other submersion-related deaths. Alphaproteobacteria are commonly found in aquatic ecosystems and play roles in nitrogen fixation and symbiosis [45]. Burkholderiales, an order of bacteria known for thriving in water and moist environments, includes species with both pathogenic and environmental roles, such as Burkholderia cepacia, which is commonly associated with aquatic habitats [46, 47]. The distinct microbial signatures of these bacterial groups in postmortem submersion cases likely reflect their adaptability to aquatic conditions, further reinforcing their potential utility as markers for deaths involving prolonged water exposure.
Conclusions
In conclusion, our random forest model represents a significant advancement in forensic microbiology by demonstrating robust predictive power across different locations and time points. This capability not only enhances the accuracy of drowning diagnoses but also broadens the applicability of microbial community analysis in forensic investigations. Future research should continue to refine these models and explore the integration of NGS with traditional forensic methods to develop comprehensive diagnostic tools. Additionally, efforts to create detailed microbial databases will be essential in distinguishing between decomposition-related microbial invasion and drowning-specific signatures, further improving the reliability of forensic microbiology.
PMI postmortem intervals.
Data availability
The sequence data generated during the current study are available in the Sequence Read Archive (SRA) at NCBI under Bioproject: PRJNA1161220.
Abbreviations
- OTUs:
-
Operational Taxonomic Units
- PCoA:
-
Principal Coordinates Analysis
- PERMANOVA:
-
Permutational Multivariate Analysis of Variance
- LEfSe:
-
Linear Discriminant Analysis Effect Size
- ROC:
-
Receiver Operating Characteristic
- AUC:
-
Area Under the Curve
- LDA:
-
Linear Discriminant Analysis
- MSE:
-
Mean squared error
- RMSE:
-
Root mean squared error
- AUCPR:
-
Area under the precision-recall curve
- NGS:
-
Next-Generation Sequencing
References
Hansen RM, Agana-Norman DFG, Hufton A, Hansen MA. Submersion injuries and the cost of injury associated with drowning events in the united States, 2006–2015. J Community Health. 2024;49(3):549–58. https://doiorg.publicaciones.saludcastillayleon.es/10.1007/s10900-023-01323-4.
Ming M, Meng X, Wang E. Evaluation of four digestive methods for extracting diatoms. Forensic Sci Int. 2007;170(1):29–34. https://doiorg.publicaciones.saludcastillayleon.es/10.1016/j.forsciint.2006.08.022.
Peabody AJ. Diatoms and drowning–a review. Med Sci Law. 1980;20(4):254–61. https://doiorg.publicaciones.saludcastillayleon.es/10.1177/002580248002000406.
Marella GL, Feola A, Marsella LT, Mauriello S, Giugliano P, Arcudi G. Diagnosis of drowning, an everlasting challenge in forensic medicine: review of the literature and proposal of a diagnostic algorithm. Acta Med. 2019;35:900–19.
Tsuneya S, Nakajima M, Yoshida M, Hoshioka Y, Chiba F, Inokuchi G, et al. Detection of diatoms in a case of mud aspiration at a coastal area. Legal Med (Tokyo Japan). 2024;66:102354. https://doiorg.publicaciones.saludcastillayleon.es/10.1016/j.legalmed.2023.102354.
Du Y, Xiang Q, Niu Y, Liu L, Liu J, Su Q, et al. Diatoms pass through the Gastrointestinal barrier and lead to false-positive: an animal experiment. Forensic Sci Med Pathol. 2024;20(2):351–9. https://doiorg.publicaciones.saludcastillayleon.es/10.1007/s12024-023-00593-7.
Kakizaki E, Kozawa S, Imamura N, Uchiyama T, Nishida S, Sakai M, et al. Detection of marine and freshwater bacterioplankton in immersed victims: Post-mortem bacterial invasion does not readily occur. Forensic Sci Int. 2011;211(1–3):9–18. https://doiorg.publicaciones.saludcastillayleon.es/10.1016/j.forsciint.2011.03.036.
Oliveira M, Amorim A. Microbial forensics: new breakthroughs and future prospects. Appl Microbiol Biotechnol. 2018;102(24):10377–91. https://doiorg.publicaciones.saludcastillayleon.es/10.1007/s00253-018-9414-6.
Aoyagi M, Iwadate K, Fukui K, Abe S, Sakai K, Maebashi K, et al. A novel method for the diagnosis of drowning by detection of Aeromonas sobria with PCR method. Leg Med (Tokyo). 2009;11(6):257–9. https://doiorg.publicaciones.saludcastillayleon.es/10.1016/j.legalmed.2009.07.003.
Uchiyama T, Kakizaki E, Kozawa S, Nishida S, Imamura N, Yukawa N. A new molecular approach to help conclude drowning as a cause of death: simultaneous detection of eight bacterioplankton species using real-time PCR assays with TaqMan probes. Forensic Sci Int. 2012;222(1–3):11–26.
Rutty GN, Bradley CJ, Biggs MJ, Hollingbury FE, Hamilton SJ, Malcomson RD, et al. Detection of bacterioplankton using PCR probes as a diagnostic indicator for drowning; the Leicester experience. Legal Med (Tokyo Japan). 2015;17(5):401–8. https://doiorg.publicaciones.saludcastillayleon.es/10.1016/j.legalmed.2015.06.001.
Kuiper I. Microbial forensics: next-generation sequencing as catalyst: the use of new sequencing technologies to analyze whole microbial communities could become a powerful tool for forensic and criminal investigations. EMBO Rep. 2016;17(8):1085–7. https://doiorg.publicaciones.saludcastillayleon.es/10.15252/embr.201642794.
Lee SY, Woo SK, Lee SM, Ha EJ, Lim KH, Choi KH, et al. Microbiota composition and pulmonary surfactant protein expression as markers of death by drowning. J Forensic Sci. 2017;62(4):1080–8. https://doiorg.publicaciones.saludcastillayleon.es/10.1111/1556-4029.13347.
Wang LL, Zhang FY, Dong WW, Wang CL, Liang XY, Suo LL, et al. A novel approach for the forensic diagnosis of drowning by Microbiological analysis with next-generation sequencing and unweighted UniFrac-based PCoA. Int J Legal Med. 2020;134(6):2149–59. https://doiorg.publicaciones.saludcastillayleon.es/10.1007/s00414-020-02358-1.
Su Q, Yang C, Chen L, She Y, Xu Q, Zhao J, et al. Inference of drowning sites using bacterial composition and random forest algorithm. Front Microbiol. 2023;14:1213271. https://doiorg.publicaciones.saludcastillayleon.es/10.3389/fmicb.2023.1213271.
Chen S, Zhou Y, Chen Y, Gu J. Fastp: an ultra-fast all-in-one FASTQ preprocessor. Bioinformatics. 2018;34(17):i884–90. https://doiorg.publicaciones.saludcastillayleon.es/10.1093/bioinformatics/bty560.
Magoč T, Salzberg SL. FLASH: fast length adjustment of short reads to improve genome assemblies. Bioinf (Oxford England). 2011;27(21):2957–63. https://doiorg.publicaciones.saludcastillayleon.es/10.1093/bioinformatics/btr507.
Amir A, McDonald D, Navas-Molina JA, Kopylova E, Morton JT, Zech Xu Z, et al. Deblur rapidly resolves Single-Nucleotide community sequence patterns. mSystems. 2017;2(2). https://doiorg.publicaciones.saludcastillayleon.es/10.1128/mSystems.00191-16.
Bolyen E, Rideout JR, Dillon MR, Bokulich NA, Abnet CC, Al-Ghalith GA, et al. Reproducible, interactive, scalable and extensible Microbiome data science using QIIME 2. Nat Biotechnol. 2019;37(8):852–7. https://doiorg.publicaciones.saludcastillayleon.es/10.1038/s41587-019-0209-9.
Zhang F, Wang P, Zeng K, Yuan H, Wang Z, Li X, et al. Postmortem submersion interval Estimation of cadavers recovered from freshwater based on gut microbial community succession. Front Microbiol. 2022;13:988297. https://doiorg.publicaciones.saludcastillayleon.es/10.3389/fmicb.2022.988297.
Liu R, Wang Q, Zhang K, Wu H, Wang G, Cai W, et al. Analysis of postmortem intestinal microbiota successional patterns with application in postmortem interval Estimation. Microb Ecol. 2022;84(4):1087–102. https://doiorg.publicaciones.saludcastillayleon.es/10.1007/s00248-021-01923-4.
Nakanishi H, Takada A, Yoneyama K, Kodama S, Sakai K, Saito K. The utility of drowning site inference through metagenomic diatom analysis. Leg Med (Tokyo). 2024;71:102548. https://doiorg.publicaciones.saludcastillayleon.es/10.1016/j.legalmed.2024.102548.
Wang L, Zhang F, Zeng K, Dong W, Yuan H, Wang Z, et al. Microbial communities in the liver and brain are informative for postmortem submersion interval Estimation in the late phase of decomposition: A study in mouse cadavers recovered from freshwater. Front Microbiol. 2022;13:1052808. https://doiorg.publicaciones.saludcastillayleon.es/10.3389/fmicb.2022.1052808.
Javan GT, Finley SJ, Can I, Wilkinson JE, Hanson JD, Tarone AM. Human thanatomicrobiome succession and time since death. Sci Rep. 2016;6:29598. https://doiorg.publicaciones.saludcastillayleon.es/10.1038/srep29598.
Moitas B, Caldas IM, Sampaio-Maia B. Microbiology and postmortem interval: a systematic review. Forensic Sci Med Pathol. 2024;20(2):696–715. https://doiorg.publicaciones.saludcastillayleon.es/10.1007/s12024-023-00733-z.
Cui C, Song Y, Mao D, Cao Y, Qiu B, Gui P, et al. Predicting the postmortem interval based on gravesoil Microbiome data and a random forest model. Microorganisms. 2022;11(1). https://doiorg.publicaciones.saludcastillayleon.es/10.3390/microorganisms11010056.
Yuan H, Wang Z, Wang Z, Zhang F, Guan D, Zhao R. Trends in forensic microbiology: from classical methods to deep learning. Front Microbiol. 2023;14:1163741. https://doiorg.publicaciones.saludcastillayleon.es/10.3389/fmicb.2023.1163741.
Lauber CL, Ramirez KS, Aanderud Z, Lennon J, Fierer N. Temporal variability in soil microbial communities across land-use types. Isme J. 2013;7(8):1641–50. https://doiorg.publicaciones.saludcastillayleon.es/10.1038/ismej.2013.50.
Finley SJ, Pechal JL, Benbow ME, Robertson BK, Javan GT. Microbial signatures of cadaver gravesoil during decomposition. Microb Ecol. 2016;71(3):524–9. https://doiorg.publicaciones.saludcastillayleon.es/10.1007/s00248-015-0725-1.
Faust K, Raes J. Microbial interactions: from networks to models. Nat Rev Microbiol. 2012;10(8):538–50. https://doiorg.publicaciones.saludcastillayleon.es/10.1038/nrmicro2832.
Scott M, Su-In L. A unified approach to interpreting model predictions. Adv Neural Inf Process Syst. 2017;30:4765–74.
Li Q, Ma KJ, Zhang XD, Yu YA, Xu SG, Zhao H, et al. [Diatom test in lung tissue of corpses in water and causes of death]. Fa Yi Xue Za Zhi. 2011;27(5):324–6.
Yu Z, Xu Q, Xiao C, Li H, Wu W, Du W, et al. SYBR green real-time qPCR method: diagnose drowning more rapidly and accurately. Forensic Sci Int. 2021;321:110720. https://doiorg.publicaciones.saludcastillayleon.es/10.1016/j.forsciint.2021.110720.
Ender PT, Dolan MJ, Dolan D, Farmer JC, Melcher GP. Near-drowning-associated Aeromonas pneumonia. J Emerg Med. 1996;14(6):737–41. https://doiorg.publicaciones.saludcastillayleon.es/10.1016/s0736-4679(96)00183-7.
Kakizaki E, Takahama K, Seo Y, Kozawa S, Sakai M, Yukawa N. Marine bacteria comprise a possible indicator of drowning in seawater. Forensic Sci Int. 2008;176(2–3):236–47. https://doiorg.publicaciones.saludcastillayleon.es/10.1016/j.forsciint.2007.09.018.
Lucci A, Campobasso CP, Cirnelli A, Lorenzini G. A promising Microbiological test for the diagnosis of drowning. Forensic Sci Int. 2008;182(1–3):20–6. https://doiorg.publicaciones.saludcastillayleon.es/10.1016/j.forsciint.2008.09.004.
Janda JM, Abbott SL. The changing face of the family Enterobacteriaceae (Order: Enterobacterales): New Members, Taxonomic Issues, Geographic Expansion, and New Diseases and Disease Syndromes. Clin Microbiol Rev. 2021;34(2). https://doiorg.publicaciones.saludcastillayleon.es/10.1128/cmr.00174-20.
Zheng J, Wittouck S, Salvetti E, Franz C, Harris HMB, Mattarelli P, et al. A taxonomic note on the genus Lactobacillus: description of 23 novel genera, emended description of the genus Lactobacillus Beijerinck 1901, and union of lactobacillaceae and Leuconostocaceae. Int J Syst Evol Microbiol. 2020;70(4):2782–858. https://doiorg.publicaciones.saludcastillayleon.es/10.1099/ijsem.0.004107.
Van den Abbeele P, Roos S, Eeckhaut V, MacKenzie DA, Derde M, Verstraete W, et al. Incorporating a mucosal environment in a dynamic gut model results in a more representative colonization by lactobacilli. Microb Biotechnol. 2012;5(1):106–15. https://doiorg.publicaciones.saludcastillayleon.es/10.1111/j.1751-7915.2011.00308.x.
Paone P, Cani PD. Mucus barrier, mucins and gut microbiota: the expected slimy partners? Gut. 2020;69(12):2232–43. https://doiorg.publicaciones.saludcastillayleon.es/10.1136/gutjnl-2020-322260.
Cláudia-Ferreira A, Barbosa DJ, Saegeman V, Fernández-Rodríguez A, Dinis-Oliveira RJ, Freitas AR, et al. The future is now: unraveling the expanding potential of human (Necro)Microbiome in forensic investigations. Microorganisms. 2023;11(10). https://doiorg.publicaciones.saludcastillayleon.es/10.3390/microorganisms11102509.
Mesli V, Neut C, Hedouin V. Postmortem bacterial translocation. In: Forensic Microbiology. 2017. pp. 192–211.
Duar RM, Lin XB, Zheng J, Martino ME, Grenier T, Pérez-Muñoz ME, et al. Lifestyles in transition: evolution and natural history of the genus Lactobacillus. FEMS Microbiol Rev. 2017;41(Supp1):S27–48. https://doiorg.publicaciones.saludcastillayleon.es/10.1093/femsre/fux030.
Dash HR, Das S. Microbial community signatures for Estimation of postmortem time intervals. Adv Appl Microbiol. 2022;118:91–113. https://doiorg.publicaciones.saludcastillayleon.es/10.1016/bs.aambs.2022.02.002.
Lema KA, Willis BL, Bourne DG. Corals form characteristic associations with symbiotic nitrogen-fixing bacteria. Appl Environ Microbiol. 2012;78(9):3136–44. https://doiorg.publicaciones.saludcastillayleon.es/10.1128/aem.07800-11.
Zhang L, Li Y, Sun X, Adams JM, Wang L, Zhang H, et al. More robust Co-Occurrence patterns and stronger dispersal limitations of bacterial communities in wet than dry seasons of riparian wetlands. mSystems. 2023;8(2):e0118722. https://doiorg.publicaciones.saludcastillayleon.es/10.1128/msystems.01187-22.
Govan JR, Hughes JE, Vandamme P. Burkholderia Cepacia: medical, taxonomic and ecological issues. J Med Microbiol. 1996;45(6):395–407. https://doiorg.publicaciones.saludcastillayleon.es/10.1099/00222615-45-6-395.
Acknowledgements
We thank the Lianchuan Biotechnology Co., Ltd (Hangzhou, China) for technical support in sequencing.
Funding
This work was supported by the National Natural Science Foundation of China (82371901) and Grant-in Aids for Scientific Research from Ministry of Public Security of the People’s Republic of China (2022JC35).
Author information
Authors and Affiliations
Contributions
Q.S., Q.X., and C.L. conceived and designed the project. Q.S., X.Z., C.Y., and W.W. Collect water samples and animal models. Q.S., X.C., Z.Y., and J.Z. Lung tissue extraction and DNA extraction were performed. Q.S., Q.X., J.Z., and L.C. analyzed the data and generated the figures. Q.S., Q.X., and C.L. wrote the paper. All of the authors agreed to submit the final manuscript.
Corresponding authors
Ethics declarations
Ethics approval and consent to participate
The Experimental Animal Welfare and Ethics Committee of the Guangdong Zhiyuan Biomedical Technology Co., LTD (Guangzhou, China) approved the experimental protocols.
Consent for publication
Not applicable.
Competing interests
The authors declare no competing interests.
Additional information
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Electronic supplementary material
Below is the link to the electronic supplementary material.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.
About this article
Cite this article
Su, Q., Zhang, X., Chen, X. et al. Microbial community profiling for forensic drowning diagnosis across locations and submersion times. BMC Microbiol 25, 244 (2025). https://doiorg.publicaciones.saludcastillayleon.es/10.1186/s12866-025-03902-y
Received:
Accepted:
Published:
DOI: https://doiorg.publicaciones.saludcastillayleon.es/10.1186/s12866-025-03902-y