Introduction:
Newborn screening (NBS) using dried blood spots (DBS) identifies infants with a metabolic disorder before the onset of life-threatening symptoms, allowing for early intervention. At present, mass spectrometry (MS/MS) screening identifies most affected newborns and additional biochemical and/or DNA testing of all screen-positive cases is needed to confirm or exclude a final diagnosis. This two-tier strategy can lead to iterative testing rounds and diagnostic delays, placing undue burden on healthcare systems and on the patients and their families. Here we developed a novel screening approach combining genome sequencing and expanded metabolite profiling from DBS with artificial intelligence/machine learning (AI/ML)-based data mining to enhance the accuracy of NBS.

Methods:
DNA sequencing and expanded metabolite analysis from a single 3 mm DBS punch was performed in 120 screen-positive cases reported by the California NBS program, consisting of true- and false-positives from four disorders: Glutaric aciduria type 1 (GA1), Methylmalonic acidemia (MMA), Ornithine transcarbamylase deficiency (OTCD), and Very long-chain acyl-CoA dehydrogenase deficiency (VLCADD). Using the dbRUSP database and an AI/ML-driven approach, we investigated metabolic differences across diverse populations and examined the impact of clinical and demographic factors (e.g., gestational age, birth weight, parent-reported ethnicity) on blood analyte levels. This analysis aimed to enhance the separation of true and false-positive screens to improve NBS performance.

Results:
DNA sequencing identified 31 of 36 true-positives (86%) with two reportable variants in a gene confirming the first-tier MS/MS screening results. Of 84 false-positives, 63 (75%) had no variant and 21 (25%) carried either a known P/LP variant (N=16) or a rare VUS (N=5) in a gene corresponding to the NBS result. Notably, we identified a tendency for VLCADD false-positives to be carriers of a pathogenic ACADVL variant (Wilcoxon test, P=3.46e-07), and for MMA false-positives to be heterozygous for a pathogenic MMA gene variant (P= 0.0391). We did not detect enrichment for pathogenic GCDH or OTC gene variants among false positives for these respective disorders. In comparison, expanded metabolomic profiling combined with AI/ML accurately confirmed each true-positive case, while the effectiveness in reducing false positives varied by disorder (51-100%). This performance was driven by a combination of known disease markers and novel metabolites identified to be isobaric to acylcarnitine species. Furthermore, our analysis of blood metabolite levels in screen-positive cases revealed that VLCADD false-positives carrying heterozygous pathogenic ACADVL variants showed significantly higher metabolite marker levels compared to those without such variants. Similarly, we observed that MMA false positives with MMA gene variants had elevated blood levels of MMA metabolite markers, potentially explaining their increased likelihood of being flagged as false positives in newborn screening.

Conclusion:
Our results suggest there is not a one-size-fits-all solution for reducing false positives in newborn screening, and each disorder may be influenced by a combination of genetic heterogeneity, physiologic features, exogenous interferences, and technical limitations inherent to genetic and metabolite analysis. Our finding of increased false-positive screens in healthy infants carrying pathogenic disease gene variants highlights the potential value of prenatal and parental carrier testing in informing newborn screening outcomes.