We re-used the raw sequencing data previously obtained from our AS PromethION sequencing runs (Deserranno et al., 2023). No additional sequencing was performed. In brief, we previously performed targeted ONT sequencing by enriching for 1,036 PGx genes in five Genome in a Bottle (GIAB) reference samples. Rather than physically designing a hybridization capture panel, we used ONT’s AS to selectively enrich library fragments corresponding to the genes of interest during sequencing. On a first PromethION R10.4.1 flow cell, the NA24385 (HG002) and NA24631 (HG005) samples were multiplexed. On a second PromethION flow cell, the NA12878 (HG001), HG01190, and NA19785 were sequenced. The raw .fast5 data were first converted to .pod5 using pod5 convert fast5 (Githup, 2024). Rebasecalling and demultiplexing were performed using Dorado (v0.9.0) (Oxford Nanopore Technologies, 2025). For the first flow cell with HG002 and HG005, the dna_r10.4.1_e8.2_400bps_sup@v4.1.0 model was used. As MinKNOW was upgraded to perform sequencing at 5 kHz instead of 4 kHz in between the sequencing of both flow cells, the second flow cell was rebasecalled using the dna_r10.4.1_e8.2_400bps_sup@v5.0.0 model.

From the rebasecalled .fastq files, we selected the reads corresponding to those read_ids enriched by AS. In particular, we filtered the adaptive_sampling.csv file obtained from MinKNOW by only retaining the read_ids with the value “unblock” in the decision column, i.e., retaining the read_ids corresponding to off-target reads. Using grep, we selected all read_ids from the rebasecalled .fastq files that did not correspond to this list of off-target reads. Alignment was performed using minimap2 (v2.28) (Li, 2021) with the GRCh38 reference. For variant calling, we used Clair3 (v1.0.10) (Zheng et al., 2022) with the corresponding pretrained model r1041_e82_400bps_sup_v410 (for HG002 and HG005) and r1041_e82_400bps_sup_v500 (for HG001, HG01190, and NA19785), obtained from Rerio (https://github.com/nanoporetech/rerio). Variant phasing and haplotagging were performed using WhatsHap (v2.4) (Patterson et al., 2015), with flags --include-homozygous --indels --distrust-genotypes --ignore-read-groups.

StarPhase (v1.4.0) (Holt et al., 2024) was used for star-allele calling. The phased variant calls and alignment .bam obtained from WhatsHap were used as input. Chromosome naming in the .bam files was adjusted to include the “chr” prefix. StarPhase was run using pbstarphase diplotype --database --bam {bam} –vcf {vcf} --reference {GRCh38} --include-set {list_of_interest} --output-calls PGx_passport.json --normalize-d6-only --output-debug {debug_dir}. The list of interest includes ABC0G2, CACNA1S, CFTR, CYP1A2, CYP2B6, CYP2C18, CYP2C19, CYP2C8, CYP2C9, CYP2D6, CYP3A4, CYP3A5, CYP4F2, DPYD, G6PD, HLA-DRB1, HLA-B, IFNL3, MT-RNR1, NAT2, NUDT15, RYR1, SLCO1B1,TPMT, UGT1A1, and VKORC1.

We retrieved the alignment .bam files for the Twist Alliance PGx and Dark Genes panel from PacBio (https://downloads.pacbcloud.com/public/dataset/). For the Dark Genes panel, we downloaded the dataset for HG001 and HG002, sequenced on the PacBio Revio. For the PGx panel, we downloaded the dataset for HG002 and HG01190, sequenced on the PacBio Sequel II platform. Similarly, we retrieved the alignment .bam files for the Twist Alliance PGx panel sequenced on ONT of sample HG01190 (data available upon request to ONT). The corresponding .bed files listing the targeted gene regions were obtained from the Twist website. For consistency between variant callers, the retrieved .bam files were used as input for Clair3 and WhatsHap with the same settings as listed above. For Clair3 variant calling, the dedicated Clair3 models for Sequel II and Revio data were used.

QC-statistics N50s were obtained using NanoComp (De Coster and Rademakers, 2023). For calculation of the phasing success, the phased .vcf files were filtered to only include the 44 common genes between the ONT-AS PGx panel and the Twist PGx panel using bcftools view (v1.20) (Danecek et al., 2021) by providing a .bed file listing each gene’s start and stop positions. Subsequently, we used WhatsHap (v2.4) stats to obtain the phasing metrics, which were subsequently summarized using multiqc (v1.28) (Ewels et al., 2016). Mosdepth (v.0.3.10) was used for the sequencing depth calculations (Pedersen and Quinlan, 2018).

Holt et al. recently developed StarPhase, a tool specifically suited to perform diplotyping of PGx genes based on HiFi-sequencing data (Holt et al., 2024). StarPhase allows diplotyping for all genes with CPIC Level A annotation, except IFNL4, and a growing number of additional genes. In particular, for genes harboring complex structural variation such as the HLA-loci and CYP2D6, StarPhase generates consensus haplotypes from the alignment .bam file to call the diplotypes, without supplying a separate .vcf file containing the variant calls. Upon benchmarking the performance of StarPhase to the GeT-RM reference set for 12 common genes, the authors of StarPhase reported 73.8% accuracy. They concluded that the incongruent calls were due to the detection of alleles that were not assayed in the GeT-RM benchmark, updated star-allele nomenclature, or phasing errors in the truth set. These are all known limitations of the GeT-RM reference calls (van der Maas et al., 2025).

Therefore, while developed for PacBio HiFi reads, we tested the performance of StarPhase to call the CYP2D6 star-alleles based on our existing ONT-AS data. We first directly performed StarPhase on the existing alignment .bam files from our previous research (Table 1, “StarPhase not-rebasecalled ONT data”). Although the star-allele provided for NA19785 is correct, the calls for the other samples remained incorrect. We hypothesized that the lower quality of the provided ONT data confused StarPhase, as marked by the detection of two copies of the *68 allele.

As the performance of ONT basecalling has improved in the last 2 years, the raw squiggle data were rebasecalled using the latest Dorado algorithm (Methods). Crucially, no new sequencing was performed. Rebasecalling shifted the median read quality Phred score significantly, as illustrated for the NA12878 data, resulting in a shift from 15.5 to 20.4 (Supplementary Figure S1). Supplying the newly basecalled alignment .bam files to StarPhase resulted in correct star-alleles of CYP2D6 in all samples (Table 1, bottom row). These detected CYP2D6 star-allele calls are identical to the Get-RM diplotypes and were further documented in the literature using orthogonal LR-sequencing workflows such as CRISPR–Cas9 target enrichment (Bettinotti et al., 2018; Gaedigk et al., 2019; Pratt et al., 2016; Pratt et al., 2010; Rubben et al., 2022). Therefore, together with the results of our previous research using Aldy (Hari et al., 2023) for the other PGx genes, the ONT-AS PGx panel results in an accurate complete long-read PGx passport for nearly all CPIC Level A genes. Only for HLA-A (currently not a part of the target file for AS, but can easily be added in our .bed file) and IFNL4 (currently not covered by StarPhase), calls are not yet included.

The Twist Alliance PGx Panel targets 49 genes, of which 44 are shared with our ONT-AS PGx panel. For future work, HLA-A can flexibly be included in our AS .bed file. The other four genes (CTBP2P2, NAGS, APOL1, and GBA) are absent in the clinical annotations from PharmGKB. Some genes are only partially covered in the Twist PGx panel. For ADD1, CACNA1S, CFTR, DPYD, F2, F5, GRIK4, HTR2C, POLG, and YEATS4, the Twist PGx capture probes only target specific parts of the gene, which limits the identification of haplotypes defined by SNVs outside these regions and the detection of novel variants. Focusing on DPYD, the Twist PGx panel covers 354 of the 434 annotated polymorphisms included in PharmVar 6.2.5 and, for example, does not include the rs773499329 (MAF 0.1% in the South Asian population) and rs762198241 variants (Pratt et al., 2024; Shrestha et al., 2018). Notably, the mitochondrial encoded MT-RNR1 gene, included within the CPIC Level A annotations in view of aminoglycoside-induced hearing loss, is not ordinarily included in the Twist PGx panel but may be spiked in, as was the case for the PacBio Twist PGx benchmark discussed below (McDermott et al., 2022). Similarly, it is not included in our ONT AS PGx panel, although it can be flexibly added by appending it to the enrichment .bed file for future experiments. Therefore, and because of the possibility of nuclear mitochondrial DNA segments being falsely aligned to the mitochondrial reference sequence without dedicated enrichment, we did not retain this gene in our analysis (Supplementary Note S1). Twist also supplies the Dark Genes panel, which captures fragments of 389 unique genes, of which 46 are shared with the AS panel list. The three panels share five genes (HLA-B, HLA-DRB1, IFNL3, CYP2D6, and VKORC1) (Supplementary Figure S2).

Of the common genes between the Twist Alliance PGx and ONT PGx panels, 25 (26 including MT-RNR1) can currently be diplotyped using StarPhase. The genes include all CPIC Level A genes (except for HLA-A and IFNL4 as discussed above) and CYP1A2, CYP2C8, CYP3A4, HLA-DRB1, HLA-DQA1, and NAT2. We therefore performed a direct comparison between both strategies using the GIAB reference sample HG01190. Additionally, for this sample, both ONT and PacBio data using the Twist PGx panel are available. A perfect star-allele match was observed across the three datasets, with consistent calls compared to the Get-RM truth set (Table 2). For CYP2B6, CYP2C9, and UGT1A1, the reference calls were improved, as discussed before (Deserranno et al., 2023). One exception seems to be HLA-DRB1, for which both strategies do not correspond to the reference call found by others (Thuesen et al., 2022). Of note, StarPhase was not able to call a diplotype for HLA-DRB1 based on the ONT sequencing data of the Twist PGx panel, probably due to the lower coverage of the first part of the gene (Supplementary Figure S3). According to the literature, due to the high complexity of the HLA locus, high sequencing depth is required (Thuesen et al., 2022). The sequencing depth for each of the genes profiled can be found in Supplementary Table S1. The median sequencing depths along HLA-DRB1 were 15.79 X and 19.62 X, for ONT-AS PGx and Twist PacBio, respectively, most probably explaining the discording call. Of note, for HLA-DQA1, a minor discrepancy regarding a difference in the non-coding region is found between ONT-AS PGX and PacBio Twist on one side and ONT Twist on the other side. As historical methods do not allow eight-digit HLA discrimination, we could not retrieve a comprehensive reference call for this sample; hence, we cannot accurately determine which one is correct.

Additionally, we repeated the star-allele calling benchmark for HG002 and found near-perfect agreement between both strategies (no Get-RM truth set available), except for the suballele of CYP2D6, the non-coding variant in HLA-DRB1 and one HLA-DQA1 genotype (Supplementary Table S2). CYP2D6 suballeles do not convey any additional functional impact compared to the major star-allele. The CYP2D6*4.014 suballele detected in the ONT-AS PGx panel is different from the CYP2D6*4.015 suballele detected using the Twist panel in the presence of rs1473203326 in the *4.014 allele. The presence of this SNV could not be deduced from manual inspection of the ONT-AS PGx reads, pointing to a possible artifact in StarPhase (Supplementary Table S2 and Supplementary Figure S4). The disagreeing HLA allele calls are discussed below.

Furthermore, we assessed the usability of the Twist Dark Genes panel for PGx as it shares five genes with the ONT-AS PGx panel and the Twist PGx Panel (Table 3, Supplementary Table S3). Strikingly, the star-allele call for CYP2D6 and IFNL3 for HG001 is wrong using the Twist Dark Genes panel. Upon closer examination, the capture region of CYP2D6 in this panel is too limited to reliably call the diplotype. The limited capture region is especially problematic for samples harboring complex CYP2D6 gene configurations, involving non-identical gene duplications and hybrids with neighboring pseudogenes, as is the case for HG001. Comparing the capture region of CYP2D6 between the Twist Dark Genes and Twist PGx panel, the PGx panel also captures the neighboring regions of CYP2D6, which might explain the wrong call with the Dark Genes panel (Supplementary Figure S5). For samples with non-complex CYP2D6 gene configurations such as HG002, our results indeed confirm that this limited capture region is less problematic (Table 3). Similarly, for IFNL3, the capture region included in the Dark Genes panel is too limited to reliably detect all possible star alleles (Supplementary Figure S6).

The results for HLA typing across the different panels are more complex (Table 3). HLA-B and HLA-DRB1 reference calls for HG001 and HG002 were obtained from the literature (Lai et al., 2024; Chin et al., 2020). For HG001, the ONT-AS PGx panel demonstrates concordance with the reference, while HLA-DRB1 is most probably false for the Twist Dark Genes panel. Manual curation (Supplementary Figure S7) indicated that the increased aligned read length in the ONT-AS PGx panel compared to the Dark Genes panel might have contributed to the correct HLA-DRB1 gene assignment. Despite the full inclusion of HLA-B in the Dark Genes panel, the HLA-B call for HG002 with this panel is most likely wrong, probably due to the limited sequencing depth (Supplementary Figure S8). *35:636 N differs from *35:08:01:01 only by a C→T substitution in gDNA position 331 of exon 2 (Loginova et al., 2025). For HLA-DRB1, strikingly, only the Dark Genes panel seems to be concordant with the literature reference benchmark reported by Lai et al. (2024). Of note, one of the analysis tools used in the construction of their benchmark, i.e., HLA-VBseq, did output the HLA-DRB1*10:01:01:03/*10:01:01:03 diplotype retrieved by StarPhase for the ONT-AS PGx dataset. For HLA-DRB1, the Twist Dark Genes panel had a median sequencing depth of 30 X compared to 21 X both for the ONT-AS PGx and Twist PGx panels. Conversely, the Dark Genes panel had the lowest sequencing depth for HLA-B (median sequencing depth: 19 X) compared to the ONT-AS PGx panel (44 X) and Twist PGx panel (143 X), therefore most probably explaining the assumed wrong calls. Our results confirm that high coverage is crucial for HLA-typing, more than for other PGx genes, and longer read lengths might help in calling the correct allele.

Finally, we benchmarked the overall read length and variant phasing capabilities between the ONT-AS PGx and the Twist PGx panel data. Due to the nature of the hybridization capture library preparation, the aligned N50 read length for the hybridization panels is significantly smaller (ONT Twist PGx: 4,040 bp; PacBio Twist PGx: 5,370 bp) than the native ONT-AS PGx reads (8,047 bp), as profiled for HG01190. Corroborating the findings of previous research, this also impacted the level of the target region that can be phased. For the 44 common genes, the ONT-AS PGx panel had significantly larger proportions of the target gene loci phased (Figure 1A) than the Twist PGx panel. The percentage of the targeted region phased for the Twist PGx panel was similar to previously reported results (van der Lee et al., 2022). Additionally, the number of individual phasing blocks for the ONT-AS PGx panel was approximately half of the number of blocks in the Twist PGx data, and each phasing block included more variants (Figure 1B). For the ONT-AS PGx panel, the maximum read length that can be obtained theoretically only depends on the length of the input DNA molecules. For this research, no shearing was performed. However, longer DNA fragments tend to cause increased blocking of the nanopores, and DNA shearing to, for example, 10 kb has been proposed to address this. Although shearing might boast yield during ONT AS and allow more samples to be multiplexed on a single-flow cell, it limits the number of unique variants contained within a single read available for phasing.