From genetic data to kinship clarity: employing machine learning for detecting incestuous relations.

Introduction: The aim of the study was to develop a predictive model based on STR profiles of mothers and children for the detection of incestuous conception.

Methods: Based on allele frequency data from the USA and Saudi Arabia, STR profiles were generated and used to simulate offspring profiles corresponding to father-child and brother-sister incest scenarios. Model training and evaluation were performed using the STR profiles of the mother and child. In addition to the baseline model, we examined its performance under a one-step mutation model, as well as its ability to detect incestuous relationships based solely on the child's STR profile. Several machine learning algorithms and neural networks were tested for classification accuracy.

Results: The CatBoost algorithm performed best in the binary classification of Normal Paternity vs. Incest Kinship. For the USA, we achieved the following results: 96.94% for 29 markers and 95% for 21 markers. The same accuracy was obtained with a single-step mutation, while prediction based on child profiles exclusively yielded an accuracy of 90.37% in the U.S. population. When analysing profiles from Saudi Arabia and modified Saudi frequencies, an accuracy of 94% was achieved.

Discussion: It was established that population structure does not affect the model's accuracy and that it can be applied even in isolated populations.