Machine learning-enhanced noninvasive prenatal testing of monogenic disorders

Area of Science:

• Genetics
• Genomics
• Molecular Biology

Background:

• Single-nucleotide variants (SNVs) are a primary cause of inherited single-gene disorders (SGDs).
• Non-invasive prenatal screening (NIPS) for maternally inherited SNVs presents significant challenges.
• Accurate identification of fetal SNVs is crucial for prenatal diagnosis.

Purpose of the Study:

• To develop and validate an improved genome-wide method for predicting inherited SNVs from both maternal and paternal origins.
• To enhance the accuracy of NIPS, particularly for challenging maternally inherited SNVs.
• To assess the clinical utility of the refined SNV-NIPS method across diverse fetal fraction levels.

Main Methods:

• Employed a combination of cell-free DNA (cfDNA) fragment features, Bayesian inference, and random forest (RF) machine learning classifiers.
• Trained RF models on millions of non-pathogenic variants for refined prediction.
• Evaluated the method in 16 families with singleton pregnancies and varying fetal fraction (FF) levels, validating against millions of inherited variants per fetus.

Main Results:

• Achieved high average area under the ROC curve (AUC) values: 0.996 for paternal, 0.81 for maternal, 0.86 for homozygous biallelic, and 0.95 for compound heterozygous variants.
• Demonstrated effective discrimination even with low fetal fraction (FF).
• Showcased high accuracy in predicting pathogenic SNVs, correctly identifying unaffected/affected status in 100% of relevant families.

Conclusions:

• Successfully developed a genome-wide NIPS method capable of detecting maternal and homozygous biallelic SNVs.
• Validated the method's performance in a real-world clinical setting.
• The improved SNV-NIPS approach shows significant promise for comprehensive prenatal diagnosis of inherited disorders.