Identical twins can share the same DNA, yet the microscopic life in their mouths may still carry clues that separate one person from the other. The surprising part is that using less information can make those clues stronger.

Identifying identical twins has been challenging in forensic genetics, and previous research has shown microbial markers may offer specific and stable evidence. Microbial markers may help because they have specificity and long-term stability.

But using every detected microbial signal can let low-value species limit the performance of identification models. The samples were compared in pairs: some came from the same individual, while others came from different individuals within an identical-twin pair.

The pairs were also separated by how far apart collection times were, and then divided into training and test sets, with the test set reserved only for checking the final models. So the researchers tried a process that searches for useful signals by imitating biological evolution.

It combines and changes possible groups of signals, then searches for a strong solution. The goal was to screen the microbial signals with the highest value for the identification models, instead of automatically keeping everything detected.

The best scores for each search were better than the matching scores when all detected species were included. That means removing low-value information improved the models. For short time intervals, the two groups could be completely separated.

The chosen combination of microbial signals produced an even wider gap, meaning better distinguishing ability. The same combination also gave the best score when the time interval was long, so it was selected for the final models. The final approach was checked on the test set, not only on the data used to build the model.

For short intervals, the model gave a similar result in the test set as in the training set, suggesting that its performance carried over to new comparisons. The study also found a warning about keeping too many signals: a model could make correct decisions for almost all training pairs, yet misjudge several test pairs.

That sharp drop in the test set indicated a risk of over-fitting—learning the examples too closely instead of finding a pattern that works more broadly. The microbial approach to identifying identical twins is still in its early stages, so the study does not recommend one finished model for every case.

What it does support is the usefulness of choosing the most informative microbial signals when building models to identify identical-twin pairs. The study found that carefully choosing the most useful microbial signals improved twin identification, especially when samples were collected close together.

That could eventually add another source of evidence when ordinary genetic clues cannot distinguish identical twins.