Patients were prospectively enrolled based on the following criteria: (1) abnormalities in routine neonatal screening testing (NST); (2) unexplained neonatal hypotonia or neonate-onset seizures without exponential causes such as prematurity, malformations, and perinatal insults; (3) unexplained and abnormal laboratory findings, such as liver enzymes, metabolic acidosis, and hyperammonemia, suggesting hidden metabolic diseases; (4) skeletal dysplasia or joint problems suggestive of storage diseases; and (5) neurodevelopmental delay with abnormal results in metabolic screening testing. This study was approved by the Institutional Review Board of Seoul National University Hospital, Seoul, Korea (No. H-2002-069-1101). Informed consent for genetic testing was obtained from all patients and their guardians in the case of minor patients.

The NEOseq_ACTION panel was designed to diagnose genetic diseases in neonates or during early infancy that require a rapid and precise diagnosis to ensure prompt and efficient treatment. The NEOseq_ACTION is a comprehensive NGS panel that includes 254 genes that were selected based on consultations with clinical and metabolic geneticists. The selected genes include potentially actionable genes associated with diseases exposed by neonatal metabolic screening performed in Korea and diseases listed on the US Recommended Uniform Screening Panel Core and Secondary Conditions of Advisory Committee on Heritable Disease in Newborns and Children (https://www.hrsa.gov/advisory-committees/heritable-disorders/rusp/index.html).

The diseases targeted in this panel can be classified as: glycogen storage diseases; adrenal insufficiency; galactosemia; inborn errors of amino acid, organic acid, and fatty acid metabolism; treatable or curable lysosomal storage diseases; immunodeficiency; hemoglobinopathy; genetic hyperbilirubinemia; congenital and genetic sensorineural hearing loss; and other genetic diseases, such as cystic fibrosis, mineral metabolism disorders, and congenital myasthenic syndromes (curable diseases: Table 1; full list: Supplemental Data Table S1).

The validation set included data from 24 patients with known pathogenic single nucleotide variants (SNVs) and small insertion or deletions (indels) across 22 genes (Supplemental Data Table S2). All validation samples were confirmed through biochemical testing. In total, 99.8% and 98.96% of bases were covered by at least 1- and 20-fold coverage, respectively. The median percentage of on-target reads across samples was 78.2%. The mean depth of coverage was 142×. All 40 variants in the 24 cases were detected using the NEOseq_ACTION panel.

Genomic DNA was extracted from two out of three DBSs (Fig. 1) using a Chemagic 360 instrument (PerkinElmer, Baesweiler, Germany) and fragmented using a Nextera DNA library preparation kit (Illumina, San Diego, CA, USA). Libraries were prepared from 100 ng of total genomic DNA using the Nextera DNA library preparation kit, according to the manufacturer’s instructions. Paired-end 150-bp sequencing was performed on the MiSeqDx platform (Illumina). Raw sequencing data were obtained in the FASTQ format.

The sequencing reads were aligned to the human genome reference sequence (GRCh37) using the Burrow–Wheeler Aligner (BWA-MEM, version 0.7.5) [12]. The aligned sequence reads (SAM file) were sorted, indexed, and converted to BAM files using SAMtools (version 1.9) [13]. Duplicate sequences in the BAM files were removed using Picard tools (version 2.18.2) (https://broadinstitute.github.io/picard/). Next, local realignment and base recalibration of the BAM files were performed using the Base Recalibrator in GATK (version 4.0.3.0) [14] based on the dbSNP (version 138), Mills, and 1000G gold standard indel databases. SNVs and indels were identified using the GATK Haplotype Caller tool. Called variants were annotated using SnpEff [15].

Called variants were interpreted using the MedyCVi software (MedySapiens, Seoul, Korea) based on InterVar [16]—a bioinformatics tool for the interpretation of genetic variants according to the American College of Medical Genetics and Genomics and Association for Molecular Pathology guidelines [17] based on databases of disease-associated genetic variants such as ClinVar [18] and Human Genome Mutation Database (HGMD) [19]. A machine learning-based prediction model was developed to support clinical decisions in variant interpretation based on estimation of the pathogenicity of each VUS.

The MedyPatho (MedySapiens) prediction model predicts the pathogenicity of non-synonymous SNVs based on nine functional scores (PolyPhen [20], MutPred [21], Sorting Intolerant From Tolerant [22], Protein Variant Effect Analyzer [23], Variant Effect Scoring Tool [24], Likelihood Ratio Test [25], MutationAssessor [26], MutationTaster [27], Fathmm [28]), two conservation scores (GERP [29] and PhastCons [30]), and the population allele frequency from four databases (Genome Aggregation Database [31], Exome Aggregation Consortium [32], Korean Reference Genome [33], and Korean Variant Archive [34].

The prediction model was trained and evaluated using a dataset compiled from the variant databases of ClinVar and HGMD. Specifically, only missense variants were selected, and for the ClinVar database, variants with a review status labeled as “criteria provided, multiple submitters, no conflicts; criteria provided, single submitter; or reviewed by expert panel” were considered. In addition, pathogenic variants marked as disease-causing mutation or disease-associated polymorphism in the HGMD were included. The final dataset comprised 143,555 pathogenic and 86,326 benign variants.

The variants added to ClinVar since 2020 were considered as test data. Variants for which protein sequences were similar to those in the training set were excluded using the CD-HIT program [35] based on a sequence identity threshold of 80%. The prediction model was trained with 43,590 pathogenic and 26,499 benign variants and was subsequently evaluated with 8,295 pathogenic and 8,630 and benign variants.

The prediction model was built using LightGBM [36]—a tree-based gradient boosting model. Performance was evaluated in comparison with two in silico prediction tools (Rare Exome Variant Ensemble Learner [REVEL] [37] and Combined Annotation-Dependent Depletion [CADD] [38]). Fig. 2 shows the performance comparison of the pathogenicity prediction tools based on test data from ClinVar release Nov 29, 2020.

All mutations filtered by MedyPatho (MedySapiens) were manually curated by genetic experts.

To evaluate the usefulness of the NEOseq_ACTION panel for patients, prospective panel testing was conducted in 111 patients. The clinical characteristics of the patients are shown in Table 2. The mean age at symptom or sign recognition was 10.4 months (range, 1–188 months). Around three quarters of the patients presented with symptoms or showed laboratory abnormalities within a month after birth.

Seventy patients were clinically asymptomatic, and 57 patients were referred to the outpatient clinic because of initial NST abnormality. Patients with NST abnormality were directly enrolled without further investigation, although most of them had simultaneous metabolic screening or NST repetition. Thirteen patients visited the clinic for incidentally detected laboratory abnormalities, such as elevated liver enzymes, creatinine phosphokinase (CPK), or unexplained acidosis. The patients had no perinatal history and showed normal development.

Forty-one patients had one or more clinical symptoms. Neonatal hypotonia and unexplained respiratory failures were the most frequent presentations. However, NST could not provide reliable data because the patients often required mechanical ventilation and total parenteral nutrition because of frequent infections. Among 10 patients with developmental delay, four showed abnormal metabolic screening results, and another four had additional features, including coarse face, skeletal abnormalities, or extensive Mongolian spots suggestive of storage diseases.

The average time from patient visit to result reporting was 8.3 days, and the average time from sample preparation to result reporting was 3.4 days. As the MiSeq Reagent Kit v2 Nano was used for less than three samples, sequencing was performed even if only one sample was submitted. We identified significant variants in 10 genes in 15 patients; detailed information is provided in Table 3. Among them, 11 patients were asymptomatic (10 with NST abnormality, and one with abnormal liver function). Four symptomatic patients (neonatal hypotonia, early-onset seizures, skeletal abnormalities, or abdominal distension) were genetically confirmed. Suspected compound heterozygous or homozygous variants were detected in 13 patients (genes related to autosomal recessive diseases), and two hemizygous variants were detected in two patients. Among the 15 patients with suspected recessive diseases, both alleles were initially classified as pathogenic or likely pathogenic in nine patients (P2, P5, P14, P41, P45, P46, P72, P74, and P83). The other patients carried one pathogenic or likely pathogenic allele and one VUS (P6, P10, P25, P36, P40, and P111). Parental testing was performed in four families, and further biochemical testing was performed for all except two patients. Therefore, the final classification was changed for four patients (P6, P25, P36, and P40). The second variant in P10 and P111 remained of unknown significance after retesting. However, the patients started treatment as soon as the biochemical testing results provided strong evidence for the final diagnosis. As P10 and P111 showed a mildly elevated serum phenylalanine level when compared with actual patients with phenylketonuria on repeated testing, their second variants were considered not significant.

All positive results in 15 cases affected the treatment decisions (Table 4). Two patients received a final diagnosis, although they had received appropriate treatment based on clinical suspicion. Eleven patients started medication or changed their daily diet. All families received counseling based on the patients’ results. The average time from testing to medical decision making was 20.8 days (2–68 days). Below, we describe some representative cases.

P2 was a 3.5-year-old girl with no perinatal history. At 12 months, her growth was retarded. At 2 years, she complained of bilateral knee pain, which was followed by gross deformation of several joints. She underwent some tests and received regular rehabilitation, but without clinical improvement. She was referred to our clinic 2 years after the onset of the symptoms and underwent simultaneous NEOseq_ACTION panel sequencing and enzyme testing on suspicion of mucopolysaccharidosis (MPS) type IVA. Compound heterozygous variants in GALNS were identified in 5 days, and glucose 6-sulfatase activity was 0.37 nmol/17 hr/mg protein (reference range [ref], 18.6–61.8). The patient was confirmed to have MPS IVA and started enzyme replacement therapy 3 weeks after the first visit.

P14 was a 3-week-old boy whose NST results indicated low arginine (92.6 μmol/L; ref, 120–600) and high citrulline (172.6 μmol/L; ref, 8–47) levels. The patient was admitted to the neonatal intensive care unit and started a diet with special milk and carglumic acid for 2 weeks. The patient was referred for prompt diagnosis, and panel sequencing demonstrated pathogenic compound heterozygous variants in SLC25A13. He was diagnosed as having citrullinemia type II and started oral arginine therapy. He is currently 13 months old and shows normal development.

P41 was a 4-month-old boy who presented poor suckling and respiratory difficulty after birth. He was lethargic and depended on the nasogastric tube. He could make eye contact with his family, but not control his head. Panel testing indicated pathogenic compound heterozygous variants in PREPL associated with congenital myasthenic syndrome. We attempted repetitive nerve stimulation tests; however, the results were unreliable because of extremely low muscle mass and immature myelination. He started on an acetylcholine esterase inhibitor and salbutamol and showed remarkable clinical improvement. A month after starting the treatment, he could roll and control his head. He started to crawl at 6 months and could stand without support at 13 months. He has speech and eats ordinary meals orally.

P111 was a 1-month-old boy who showed NST abnormality indicating an elevated phenylalanine level (240 μmol/L; ref, 38–137). He underwent NST thrice because of abnormal results at different hospitals. The panel test indicated a pathogenic variant and a VUS in PAH. As detailed metabolic testing indicated slightly increased phenylalanine (100 µmol/L; ref, 25–74) and normal tyrosine (89 μmol/L; ref, 13–91) levels, we concluded that the patient had hyperphenylalaninemia, rather than phenylketonuria. The patient continued on breast milk and showed normal development until the last follow-up.