Identical twins can be nearly indistinguishable by ordinary genetic markers—but their saliva microbiota may still carry a useful signal. This study asks whether selecting only the strongest microbial features can make that signal work better. Personal identification of monozygotic twins has been challenging in forensic genetics.

Microbial markers have potential value because of their specificity and long-term stability. However, previous studies used the complete information of detected microbial communities, and low-value species limited the performance of previous models.

Distinguishing monozygotic twins has long been a challenging task in individual identification, with significant implications in some cases. Genomic biomarkers can differ between monozygotic twins, but those differences are infrequent because twins share nearly identical nuclear and mitochondrial genome information, and genome-wide sequencing is typically necessary to differentiate them.

Epigenetic biomarkers can make differences between twins more significant, but using them requires strict material standards that limit application in daily work. Therefore, more suitable biomarkers for monozygotic twin identification are still required.

Advances in microbiology may provide another possible solution to the monozygotic twin identification problem. The microbiota reside on or in the human body and are known as the second genome of the human body. Recent advances in sequencing technology have helped the growth of forensic microbiology, an interdisciplinary field of forensic medicine and microbiology.

Microbial information has received particular attention for postmortem interval estimation, cause of death inference, body fluid type identification, and individual identification. The study collected 80 saliva samples from 10 monozygotic twin pairs and obtained microbiota information through 16s rRNA sequencing.

Twelve models were developed using sample collection intervals, sequence data basis, and distance selection. High-value species were selected using genetic algorithm processes, and the two models with the best distinguishing power in the training test were selected under two different sampling intervals.

The model developed for short-interval samples could completely separate monozygotic twin pairs from inner-individual sample pairs in training and test sets. The study included 10 pairs of monozygotic twins: four male pairs and six female pairs.

For each participant, four freely flowing saliva samples, one milliliter each, were collected at four different time points; the second, third, and fourth were collected 12, 13, and 14 months after the first. Each participant stated that they had not used antibiotics within six months before sampling and had not consumed food or water within two hours of sampling.

A total of 80 samples were collected and labeled based on the participant and the time point. Eight samples for each monozygotic twin pair resulted in 28 sample pairs per pair, producing 280 sample pairs across 10 pairs. Samples from the same individual were labeled Self, while samples from different individuals within a monozygotic twin pair were labeled MZT.

Sample collection intervals of less than or equal to two months were labeled short, and intervals of greater than or equal to 12 months were labeled long. The four groups were randomly divided into training and test sets with an eight-to-two ratio, and the test set was used only to validate the final models.

Table one shows how the authors divided sample pairs into training and test sets across two relationships—Self and MZT—and two collection intervals, Short and Long. Self pairs contain forty-eight training and twelve test samples in both intervals, while MZT pairs contain eighty training and twenty test samples for Short intervals, and forty-eight training and twelve test samples for Long intervals.

This organization matters because it defines the evaluation groups used to study microbial similarity across individuals and sampling times. Two types of features were used as the data basis for model construction: microbial sequencing information represented by operational taxonomic units, or OTUs, and amplicon sequence variants, or ASVs.

Three types of beta diversity distances were calculated based on these two types of features. Six possible models could be constructed from the data basis and distance choices under each sampling interval, producing a total of 12 models.

The distinguishing ability of the 12 choices was measured using all detected features, with area under the receiver operating characteristic curve, or AUC, used as the metric. AUC is a commonly used metric for evaluating binary classifiers. For each model, calculation results for all sample pairs were sorted from smallest to largest and assigned rank numbers, with average ranks assigned to pairs with identical results.

Genetic algorithms were performed 12 times on MATLAB to screen high-value OTUs or ASVs in the identification models. Genetic algorithms have been widely used in feature selection since their introduction in the 1970s. The algorithm transforms solving complex problems into processes similar to chromosome gene crossover and mutation in biological evolution.

It then searches for the optimal solution by simulating natural evolution processes. The 80 samples formed 280 Self or MZT sample pairs, divided into four groups based on the relationship of the source individuals and the time interval.

Each group was randomly divided into training and test sets in an eight-to-two ratio. A total of 12 genetic algorithm processes were performed using different feature-distance combinations in the training set. Except for two genetic algorithms that stopped after satisfying AUC equal to one, the generational-best AUC of the remaining 10 fluctuated within a relatively flat range after 1,000 to 1,500 iterations.

Table 3 indicates that the best AUC for each genetic algorithm was better than the corresponding AUC when all detected species were considered. When the time interval was short and Jaccard distance was used, the two groups could be completely separated regardless of whether OTUs or ASVs were used.

The kernel density estimation results showed that the final chosen combination of ASVs produced a larger gap between the two curves, implying better distinguishing ability. When the time interval was long, the ASV-Jaccard distance combination also provided the best AUC, so ASV-Jaccard distance combinations were selected for further model construction.

Table three reports the best area under the curve, or AUC, achieved by twelve genetic-algorithm processes across short and long time intervals, OTUs and ASVs, and Jaccard, Bray–Curtis, and Hellinger distances. It also lists AUC when all species are included and the number of selected features, showing how the authors numerically assess classification performance for each distance and species representation.

Table four reports model evaluation by data set, time interval, and threshold, using sensitivity, specificity, Youden’s index, and misjudged pairs. In the test set, the short-interval threshold with Tmax equal to 0.4 gives sensitivity, specificity, and Youden’s index of 1, with no misjudged pairs; the long-interval rows report 0.75, 0.75, and 0.5, with 3 Self and 3 MZT misjudgments.

This matters because it shows how performance is documented across intervals and validation data. Figure five shows log-ten likelihood-ratio distributions for Self pairs in blue and MZT pairs in orange, separated by short or long time intervals and training or test sets.

Across panels A through D, the dotted lines mark the reported threshold, Tmax, and likelihood ratio equal to one, while the annotations give the corresponding values for each condition. This matters because it visualizes how the final ASV combination assigns evidence across sample pairs under different time intervals and evaluation sets.

Jaccard distance was calculated for each sample pair in the test set, and likelihood ratio was calculated from the probability density curves. The two types of threshold sets were validated in the test set, as shown in Table 4 and the two lower subfigures of Figure 5.

Compared with the training set, the short-interval model provided a similar Youden index in the corresponding test set. A model using 510 ASVs made correct decisions for 95 of 96 training-set pairs. However, accuracy dropped sharply in the test set, where six of 24 pairs were misjudged, indicating the probability of overfitting.

The study has limitations, including that unrelated sample pairs from different monozygotic twin pairs were not considered during model construction. Fresh saliva samples were used, from which the host’s DNA The study used 80 saliva samples from 10 pairs of monozygotic twins to develop models for identification based on microbiota information.

Feature selection was used to build models that effectively distinguish MZT from Self pairs. When the sampling interval was less than two months, the final model could completely separate the two types of sample pairs. This highlights the potential value of microbiota information as a biomarker in monozygotic twin identification.

The feature selection process refined models and filtered out irrelevant data, resulting in more accurate and reliable models. In this dataset, selecting microbial features with a genetic algorithm improved identification models, and the short-interval ASV–Jaccard model completely separated twin-pair samples from same-person pairs in training and testing.