ClairS: a deep-learning method for long-read tumor–normal pair somatic small variant calling

Abstract

Somatic variant discovery in tumors is crucial for clinical analysis, yet most existing methods are designed for short-read sequencing, with few developed specifically for long reads. Here we present Clair-Somatic (ClairS), a deep-learning-based somatic small-variant caller designed for long-read tumor–normal pairs. Trained on synthetic somatic variants with diverse coverages and variant allele fractions, ClairS accurately detects a wide range of somatic variants. Using the Nanopore Q20+ HCC1395–HCC1395BL dataset at 50/25× tumor/normal coverage, ClairS achieved F1 scores of 89.83% for single-nucleotide variations and 73.38% for indels; augmenting training with real cancer cell lines improved performance to 96.19% and 79.67%, respectively. Our findings indicate that improved read phasing enabled by long-read sequencing is key to accurate single-nucleotide variation detection, especially at low variant allele fractions. Through experiments across varied coverage, purity, contamination levels, multiple platforms and real cancer cell lines, we demonstrate that ClairS is a robust and reliable caller. ClairS is open source and available at https://github.com/HKU-BAL/ClairS.

ClairS accurately detects somatic small variants using long-read sequencing data from tumor–normal pair samples.

Main

Analysis of cancer genomes that identify and characterize somatic variants has enabled a better understanding of tumor progression¹ and led to precision oncology². Identifying somatic variants, however, remains challenging due to intra- and inter-tumor heterogeneity, which often leads to low variant allele fractions (VAFs) and confounding factors, including sequencing artifacts, inadequate sequencing coverage and normal contamination³. Endeavors were made to address these challenges and maximize sensitivity and accuracy in identifying somatic variants using next-generation sequencing (NGS) short reads^4–13; however, constrained by read length, short reads have limited germline variant discovery capability in hard-to-map genomic regions, such as homopolymers and segmental duplications¹⁴. This problem is expected to be alleviated through long-read sequencing¹⁵. Pacific Biosciences (PacBio) and Oxford Nanopore Technologies (ONT) are two widely used long-read sequencing platforms. PacBio HiFi sequencing provides highly accurate circular consensus sequencing reads with native error rates typically below 0.2% and is comparable to short read¹⁶. ONT, another prominent long-read technology, offers miniaturized sequencing devices and fast sample-to-data turnaround, which is a credible step toward democratizing sequencing by drastically reducing the cost of carrying out sequencing experiments. ONT raw reads were reported to have an error rate of 3–15% in the past¹⁷. This was reduced to 1% or lower using ONT’s latest Q20+ chemistry¹⁸. The gap is still notable, however, compared to NGS short reads, which have an average error rate at 0.1% (ref. ¹⁹), making the somatic variant callers once designed for short reads practically unworkable for ONT long reads.

Germline variants are often considered to be easier to correctly identify than somatic variants. An early attempt to call germline small variants using noisy ONT long reads was made by Clairvoyante in 2018 (ref. ²⁰). The work was enabled using (1) a deep neural network, which was also used for variant calling by DeepVariant²¹; and (2) the high-quality known truth variants in GIAB (Genome In A Bottle) reference samples for neural network model training²². Clair²³ and Clair3 (ref. ²⁴)built on Clairvoyante and improved germline variant-calling accuracy by optimizing network input, network output, network architecture and workflow designs for use with noisy ONT data. Both Clair3 and a pipeline named PEPPER-Margin-DeepVariant²⁵ (DeepVariant), which is also designed for ONT long-read germline small-variant calling, have demonstrated better single-nucleotide polymorphism (SNP)-calling performance than using the same coverage of Illumina short reads; however, while solutions are ready for ONT long-read germline small-variant calling, there have been few callers available for ONT long-read somatic small-variant calling. We note that ONT long-read somatic SV (structural variant) callers, including Sniffles2 (ref. ²⁶), Nanomonsv²⁷, SAVANA²⁸ and Severus²⁹, which were developed recently, called for the development of a small-variant caller to complete the ONT long-read somatic variant-calling workflow.

However, some designs critical to ONT long-read germline variant calling are not applicable to somatic variant calling. First, in their network output, both Clair3 and DeepVariant apply a strong diploid genome assumption. Clair3 uses a 21-genotype output, which is a two-combination of A, C, G, T, insertion and deletion²⁴. DeepVariant uses a three-category output that includes hom-ref (homozygous reference), het (heterozygous) and hom-alt (homozygous alternative)²⁵. Both Clair3 and DeepVariant are classification models that use observed allelic fraction of alternative alleles as network input, and output the category that represents the expected allelic fraction of a variant (for example, allelic fraction 0, 0.5, 1 and 0.5/0.5 for genotype 0/0, 0/1, 1/1 and 1/2, respectively); however, somatic variants have VAFs ranging continuously from 0 to 1. Without a certain ploidy, somatic variant candidates have no expected allelic fraction for a model to test against. Thus, a new design is required. As an example, a new design could use a regression model to derive VAF directly, or a classification model to simply determine whether a candidate is a somatic variant or not and infer VAF subsequently. Second, the seven standard GIAB reference samples HG001–HG007 provide approximately 25 million truth germline variants²², which are critical for the model training of state-of-the-art, deep-learning-based germline variant callers; however, in terms of known truth somatic variants, only the HCC1395–HCC1395BL (a human triple-negative breast cancer cell line and a normal cell line derived from the B lymphocytes of the same donor, hereafter referred to as HCC1395/BL) tumor–normal pair was published by the Somatic Mutation Working Group of the SEQC2 (Sequencing Quality Control Phase II) consortium³. It contains only 39,560 single-nucleotide variations (SNVs) and 1,922 indels, which is orders of magnitude fewer than the available truth germline variants, and far from enough for deep-neural-network model training. In late 2024, DeepSomatic³⁰, a somatic variant caller trained on multiple real cancer cell lines, was developed, which performs well on different long-read and short-read cancer samples; however, it requires a diverse set of real cancer samples for training and relies on multiple-technology sequencing data to establish ground truths for model training; a study by its authors showed that models trained on multiple cancer types outperform those trained on a single type³⁰. This reliance poses a limitation, as such diverse and multi-platform-validated cancer datasets are not always accessible, especially for emerging sequencing technologies or latest iterations of existing platforms. Moreover, due to cancer-type-specific mutational signatures, a training set missing a particular cancer type may lead to reduced accuracy or biased variant calling for that type.

In the absence of adequate real tumor–normal samples, synthetically generated tumor data are a potential solution. Bamsurgeon was developed to generate synthetic tumor dataset by spiking somatic variants into existing germline alignments³¹. This was a successful method for short reads, but it faces difficulties for long reads for two reasons: (1) Illumina short reads are read one base at a time, whereas ONT long reads are read as the signals of a sliding five-mer or six-mer window (a spike-in variant also changes the signals of the adjacent bases). Bases can be base-called from signals but cannot be authentically turned back into signals, so the spike-in method cannot be applied to long reads; and (2) Bamsurgeon does not apply the point that a somatic variant is usually found in only one haplotype (either maternal or paternal) but is missing in another. The use of long reads over short reads for somatic variant calling is preferable only if the long-read advantage of haplotyping (also called phasing) is utilized. All things considered, a new method for synthesizing long-read data to provide abundant plausible somatic variants for model training is needed.

In this study, we present Clair-Somatic (ClairS), a tumor–normal pair somatic small-variant caller for ONT long reads, which is named after its germline variant caller predecessors. ClairS draws on the successful experience of the Clair series, and uses a new network output and a new workflow design to address the continuous VAF space of somatic variants. We devised a data synthesis strategy that uses only real long reads from GIAB reference samples with known germline variants, yet can simulate somatic variants at any desired tumor purity, sequencing coverage (up to that of the source data) and level of normal contamination. Our approach treats two different GIAB samples, A and B, as tumor and normal, respectively; germline variants unique to sample A then serve as simulated somatic variants when called against sample B. Unlike the method DeepSomatic³⁰, which is trained on a set of real cancer samples, our strategy can theoretically produce an infinite number of somatic variants for model training. We show results to highlight how phasing improves somatic variant-calling performance on long reads. To leverage more remote alignment information, which is computationally impractical to include in the network input, we devised a post-processing step that searches for ancestral haplotype support for any somatic variant candidates. This step removed a considerable number of false-positive calls in our experiments. For benchmarking, we sequenced in total 75-fold HCC1395 and 45-fold HCC1395BL ONT Q20 + 4 kHz long reads (data deposited to NCBI Sequence Read Archive (SRA)) and used the truth somatic variants provided by the SEQC2 consortium³. With 50/25-fold tumor/normal, ClairS achieved 93.01%p (precision)/86.86%r (recall) for SNVs, and 78.26%p/69.07%r for indels when targeting VAF ≥ 0.05. In late 2024, Park et al.³⁰ released the Illumina, PacBio and ONT reads of five real cancer cell line samples. Among them, except for HCC1395 that has known truths from SEQC2, Park et al. have generated the benchmarking Variant Call Format (VCF) files (that offer confident variants but not as stringent as SEQC2) of HCC1397, HCC1954, H1437 and H2009 using multiple sequencing platforms and multiple somatic variant callers, with ClairS as one of them for ONT long-read somatic variant calling. The VCF file of the four real cancer cell line samples have enabled us to augment ClairS model training to better accommodate the tumor-specific characteristics in real tumors. The augmented model has achieved 97.83%p/94.60%r for SNVs, and 91.57%p/70.50%r for indels with 50/25-fold of HCC1395/BL with the latest ONT Q20 + 5 kHz chemistry. We also show the performance of ClairS at various tumor/normal coverage, tumor purities and normal contamination, as well as on real cancer cell line samples. ClairS outperformed other methods, especially at tumor purities below 20%. ClairS is designed for ONT long reads, but the whole method is also applicable to PacBio platform and Illumina short reads. This versatility allowed us to benchmark ClairS against state-of-the-art short-read somatic variant callers. The results show that ClairS performed comparably with the current heuristic-based and deep-learning-based callers on benchmarked short-read data from cancer cell lines.

Results

The ClairS method

ClairS is a deep-learning-based long-read somatic variant caller. Its novelty lies in both training data synthesis and its workflow and design. GIAB avails seven reference samples with ~25 million truth germline variants. In contrast, only one tumor–normal pair with ~40,000 truth SNVs and ~2,000 truth indels is available, prohibiting training a proper-scale deep-learning model for somatic variant calling generalizable to rich cancer types and sample settings. ClairS solved the problem by generating synthetic tumor and synthetic normal for model training (Fig. 1a), given the observation that a germline variant specific to an individual can be considered a somatic variant of another individual when the genome data of the two individuals are combined. Theoretically, any coverages, tumor purities and normal contaminations can be generated.

Fig. 1. Overview of ClairS training data synthesis workflow.

a, The workflow demonstrates how to produce synthetic somatic variants using two biologically unrelated samples with known truth germline variants for ClairS model training. In this study, specifically, we used 80× ONT WGS data of GIAB HG002 as sample A and 50× HG001 as sample B. First, germline variants G_A and G_B were defined as known truth germline variants in samples A and B given by GIAB. G_A and G_B include both homozygous and heterozygous germline variants of a sample. To generate synthetic tumor variants T and synthetic normal variants N_A and N_B for each sample, the alignments were split into smaller chunks with 4× coverage each. Then, the chunks from both samples were combined and the variants called from them were defined as T. With the flexibility of combining any number of chunks from both samples, T effectively covered variants called at different coverages and VAF. Similarly, the chunks from a sample were combined at multiple coverages for calling synthetic normal variants N_A and N_B. With a small number of chunks from another sample combined into a synthetic normal, N_A and N_B effectively covered different contamination levels. The variants G_A, G_B, T, N_A and N_B were then used to generate four categories of variants (‘somatic’, ‘germline’, ‘artifact’ and ‘normal-only’) with different rules. Somatic, germline and artifact match the three categories in the inference task of the ClairS network architecture. The variants of the three categories were used for model training. When using sample B as tumor and A as normal, somatic is defined as ‘(T − N_A) ⋂ (G_B − G_A)’ (variants that were (1) found in synthetic tumor T; (2) not found in synthetic normal N_A; (3) found as a germline variant in G_B; and (4) not found in G_A). Germline is defined as ‘T ⋂ N_A ⋂ G_A ⋂ G_B’ (variants that were found in all T, N_A, G_A and G_B). Artifact is defined as ‘T − N_A − G_A − G_B’, which signifies the variants found only in T and not in the germlines or synthetic normal. When using sample A as tumor and B as normal, the definitions remain the same except for switching the subscripts. b, The VAF distribution of the synthetic somatic SNVs at three different simulated tumor purities (100%, 50% and 25%), using either HG002/HG001 or HG001/HG002 as tumor/normal. As both heterozygous and homozygous variants were used in synthesis, at 100% tumor purity, the variants were gathered at 0.5 and 1.0 VAF. The distribution showed good coverage of typical somatic SNV VAF by the synthetic SNVs. c, The breakdown of the number of synthetic variants for training. The numbers (1) using either HG002/HG001 or HG001/HG002 as tumor/normal and (2) of the three categories of somatic, germline and artifact, as defined in a, are shown. The number of somatic categories is further divided into those synthesized from either homozygous SNPs or heterozygous SNPs. These numbers explain why including heterozygous SNPs in the synthesis is essential to ensure a sufficient number of synthetic somatic variants for model training.

Figure 2a shows an overview of the ClairS workflow. In the first step, ClairS uses Clair3 for germline variant calling, and then LongPhase or WhatsHap for phasing and read haplotagging (Fig. 2b). ClairS, by default, uses the heterozygous SNPs called in the normal sample to phase the tumor sample reads, but it can also be configured to use the high-quality heterozygous SNPs called in the tumor sample. The latter works better when the normal sample coverage is low. In the second step, ClairS performs pileup-based and full-alignment-based variant calling. Unlike Clair3, in which the faster pileup-based calling is used before the more computational-demanding full-alignment-based calling to pull off most variant candidates that are obvious variants, ClairS considers the power of the two neural networks equal. We observed that pileup-based calling is more sensitive at lower VAF, especially when under 0.1, whereas full-alignment-based calling performs better at mid-to-high-range VAFs when the coverage of both samples is good. In the third step, ClairS searches for ancestral haplotype support (Fig. 2d) to reduce false-positive calling mistakes made by the networks by leveraging relatively remote germline variants to find the correct ancestral haplotype for a somatic variant. Variants are filtered if no ancestral haplotype support is found.

Fig. 2. Overview of the ClairS somatic variant-calling workflow.

a, The workflow illustrates the three steps of ClairS. In step 1, ClairS uses Clair3 and LongPhase for germline variant calling, phasing and read haplotagging. The processed alignments are then used for both pileup- and full-alignment-based somatic variant calling in step 2. Step 3 involves post-processing filters that eliminate somatic variant callings if an ancestral haplotype (maternal or paternal) from which the somatic variant could originate cannot be found. The details of steps 1, 2 and 3 are shown in b, c and d, respectively. b, Step 1 details. Clair3 is applied to both tumor and normal samples for germline variant calling. High-quality heterozygous germline variants shared by both samples are selected and used by LongPhase to phase the germline variants found in the tumor sample. Using the phased germline variants, the tumor reads are then haplotagged to belong to either haplotype 1, 2 or unknown. c, Step 2 details. The processed alignments from step 1 are fed into both the pileup-based variant-calling neural network and the full-alignment-based variant-calling neural network. On a single somatic variant candidate, both networks give respective predictions on the probability of three categories: ‘somatic’, ‘germline’ and ‘artifact’. The predictions are then merged according to a set of rules introduced in Methods. d, Step 3 details. The somatic variants called in step 2 are examined to determine if they are supported by an ancestral haplotype. Ancestral haplotypes, which can be either maternal or paternal, are derived using germline variants. A somatic variant is considered supported by an ancestral haplotype if the haplotype containing the somatic variant is believed to originate from one of the ancestral haplotypes.

Depending on the source germline variants’ zygosity and ownership, ClairS derives multiple categories of variants from a synthetic tumor/normal pair, including ‘somatic’, ‘germline’, ‘artifact’ and ‘normal-only’. The data synthesis method complies well with the observation that an authentic somatic variant is usually found in only one haplotype and absent from another (depicted in Extended Data Fig. 1a). Both the training and inference in ClairS fully utilize phasing information for somatic variant calling. To cover the characteristics only in real tumor samples, such as cancer-specific indel length distribution and cancer-type specific mutational signatures, ClairS provided a real sample augmented model fine-tuning the pretrained synthetic model using real cancer cell line (Fig. 4a).

Extended Data Fig. 1. Performance differences between phasable and not-phasable SNVs.

(a) The figure shows that a somatic variant usually originates in a single somatic cell and then spreads to more cells through cell division, resulting in a clonal carrying the same variant. It also shows how the mismatches in the tumor sample and normal sample are different from each other. A somatic variant is more likely to be assigned to a haplotype through phasing, while a variant caused by random sequencing errors is less likely to be successfully phased. (b) A performance comparison of somatic variants where ‘P’: can be phased, and ‘NP’: cannot. The figure shows a higher performance in somatic variants that can be phased, especially at lower VAFs. We used 50/25-fold HCC1395/BL and favor-recall mode.

Fig. 4. Further enhance ClairS with real cancer cell line datasets.

a, Workflow of incorporating real cancer cell line datasets in model training. The cell line datasets were made available by Park et al.³⁰, including two breast cancer cell line pairs (HCC1937/BL and HCC1954/BL) and two lung cancer cell line pairs (H1437/BL and H2009/BL). b, The breakdown of variant categories in the four cancer cell lines. c, The VAF distribution of the somatic SNVs used for training in the four cancer cell lines. d, The SNV and indel performance benchmarks of the ClairS RS and SSRS models using the entire ONT COLO829/BL and HCC1395/BL 5 kHz datasets. High, medium and low tumor coverage settings were benchmarked.

Performance on ONT data

A summary of the ONT data used for model training and benchmarking is shown in Supplementary Table 1. We trained the ClairS ONT model using synthetic data generated from two GIAB samples: HG001 and HG002 (ref. ²²). We used both HG001/HG002 and HG002/HG001 as tumor/normal samples for data synthesis. The HG002 sample has 76.29-fold coverage and was made available by Nanopore through EPI2ME Labs. The HG001 sample has 48.44-fold coverage and was sequenced at HKU, with details given in the ‘ONT library preparation and sequencing’ section in Methods. Following the convention of the state-of-the-art deep-learning-based variant callers^21,24,25, we excluded reads and variants from chromosome 20 from model training.

For benchmarking, we used the HCC1395/BL tumor/normal pair, with known truth variants provided by the SEQC2 consortium³. The HCC1395 sample has 75.97-fold coverage. The HCC1395BL sample has 45.55-fold coverage. Both samples were sequenced at two sequencing centers (HKU and Novogene) for quality control purposes. The yield of the two centers is shown in Supplementary Table 2. Only SEQC2 truth variants labeled ‘HighConf’ (high confidence) and ‘MedConf’ (medium confidence) were used for benchmarking. While there are ~40,000 SNVs but only ~2,000 indels in the truth set, the big difference in the number of truth variants results in different analytical power between the two variant types. So we first benchmarked multiple coverages, tumor purities and normal contaminations with only SNVs, and then benchmarked and discuss somatic indel-calling performance in a separate section. We used Illumina’s Haplotype Comparison Tools³² to generate the performance figures, including F1 score, precision and recall, and cross-validated them with the ‘compare_vcf’ submodule in ClairS. More clarifications and parameters are shown in the ‘Benchmarking’ section in Methods.

For both model training and benchmarking, we used GRCh38, which is the newest reference genome version on which both GIAB and SEQC2 truth variants are based. All ONT sequencing data were base-called using Guppy v.6.1.5 and aligned to GRCh38 using minimap2 v.2.17-r941. The command line used is given in the ‘Command lines used’ section in Supplementary Notes. All data mentioned above, including ONT sequencing data, GIAB truth variants, SEQC2 truth variants and reference genomes, are publicly accessible via links or SRA accession IDs listed in the ‘Data availability’ section in Supplementary Notes.

Performance with different tumor combinations and normal coverage

We assessed the ClairS performance with different combinations of tumors and normal coverage. We tested three tumor coverages: 25-, 50- and 75-fold (hereafter referred to as ‘the three tumor coverages’). We applied 25-fold as the first step as it represents a conservative throughput estimation of an R10.4.1 (also called Q20+) PromethION flowcell. We also tested three normal coverages: 20-, 25- and 30-fold (hereafter referred to as ‘the three normal coverages’). The fivefold step resembles the throughput variance of a single flowcell. Our experiments aimed to imitate a practical setting for clinical cancer diagnosis, with the tumor sample coverage increased one flowcell at a time to seek higher discovery power, but the normal sample coverage is fixed at a single flowcell for cost-effectiveness.

The results are shown as precision–recall curves in Fig. 3c. With normal coverage fixed at 25-fold, ClairS achieved area under precision–recall curve (AUPRC) of 0.9342, 0.9538 and 0.9544 at the three tumor coverages. Users can choose their own variant quality cutoff in ClairS to favor a metric in different usage scenarios. The evaluation metrics at two variant quality cutoffs (8 and 15) are shown in Supplementary Table 3. A quality cutoff of 15 (hereafter referred to as ‘favor-f1 mode’) filters more variants and aims for balanced precision and recall. A quality cutoff of 8 (‘favor-recall mode’) retains more variants and aims for higher recall. In the favor-f1 mode, ClairS achieved 95.03%p (precision)/78.71%r (recall)/86.11%f (F1 score), 93.01%p/86.86%r/89.83%f and 92.94%p/86.92%r/89.83%f at the three tumor coverages, respectively. From 25- to 50-fold, the recall increased from 78.11%r to 86.86%r (+8.75%), with a 2.02% precision drop (from 95.03%p to 93.01%p). From 50- to 75-fold, however, no improvement was observed. In the favor-recall mode, ClairS achieved 82.66%p/91.55%r/86.88%f, 70.21%p/96.10%r/81.14%f and 63.80%p/96.60%r/76.85%f at the three tumor coverages, respectively. Compared to favor-f1, the favor-recall mode reached 91.55%r at 25-fold (against 78.71%r, +12.84%) and 96.10%r at 50-fold (against 86.86%r, +9.24%). The observed drop in precision for higher-coverage samples at a fixed quality threshold is mainly attributed to the inclusion of low-VAF variants in R10.4 chemistry with higher sequencing error rate. Although Clair3 is not designed for somatic variants, to give a reference point, we conducted experiments with it using the same datasets by considering the germline variants found in tumor but not in normal as somatic variants. Clair3 had 19.27%p/72.10%r/30.42%f, 29.82%p/68.32%r/41.52%f and 36.21%p/64.56%r/46.40%f at the three tumor coverages. The results highlight the inappropriateness of using a germline variant caller for somatic variant calling.

Fig. 3. The ClairS neural network architecture and benchmark result on ONT HCC1395/BL 4 kHz dataset.

a,b, The pileup network (a) and full-alignment network (b) use alignments of both the tumor and normal samples as input. Tensors are created from both samples using methods detailed in Methods and are then concatenated. The tensors are then processed by their respective neural network for inference. Both networks output the probability of three categories: ‘germline’, ‘somatic’ and ‘artifact’. The sequence of layers and layer configurations are shown. The letters c, s and k represent channel, stride and kernel, respectively. c, The precision–recall curve of different combinations of tumor and normal coverage. The dot on each dashed line shows where the best F1 score was achieved. The values in the legend are the area under precision–recall curve at each coverage. d, The precision–recall curve of different tumor–normal purity combinations with tumor coverage fixed at 50× and normal coverage fixed at 25×.

ClairS achieved 0.9387, 0.9538 and 0.9579 AUPRC at the three normal coverages. In both favor-f1 and favor-recall modes, raising the normal coverage consistently increased somatic variant-calling performance. With tumor coverage fixed at 50-fold and using the favor-f1 mode, ClairS achieved 91.86%p/83.77%r/87.63%f, 93.01%p/86.86%r/89.83%f and 92.06%p/88.51%r/90.25%f at the three normal coverages. Using the favor-recall mode, the numbers were 68.36%p/95.69%r/79.75%f, 70.21%p/96.10%r/81.14%f and 70.56%p/96.43%r/81.49%f.

Supplementary Fig. 1 shows the performance of ClairS in favor-recall mode, broken down to four VAF ranges: 0.5–1, 0.2–0.5, 0.1–0.2 and 0.05–0.1. At different coverages, the performance of ClairS at range 0.2–0.5 (low-mid) was found to be as good as 0.5–1 (mid-high). For example, at 50/25-fold tumor/normal coverage, ClairS achieved 94.7%p/99.3%r/96.9%f at 0.5–1 and 95.2%p/98.1%r/96.7%f at 0.2–0.5. At range 0.1–0.2, precision was reduced to ~60%p, while recalls plateaued at ~90%r. At range 0.05–0.1, precision was further reduced to below 10%p (11.8%p, 5.7%p and 4.6%p at the three tumor coverages), whereas recalls were raised with increasing tumor coverage (16.5%r, 32.8%r and 46.1%r). The reason for the drop in precision was that the higher coverage led to a drastic increase in the number of variant candidates at very low VAF. At range 0.05–0.1, the number of candidates was about 131,000, 310,000 and 419,000 at the three tumor coverages.

Performance at different tumor purities and normal contamination

We assessed the performance of ClairS at different combinations of tumor purity (1.0, 0.8, 0.6, 0.4 and 0.2) and normal purity (1.0, 0.95 and 0.90). All the experiments in this section used 50-fold tumor and 25-fold normal coverage. The results are shown in Fig. 3d and Supplementary Table 4. The two modes (favor-f1 and favor-recall) behaved differently with varying purity. With normal purity fixed at 1.0, in favor-f1 mode, precision remained above 90%p (93.01%p, 96.25%p, 97.79%p, 98.68%p and 98.99%p at tumor purity 1.0, 0.8, 0.6, 0.4 and 0.2), whereas recall dropped (86.86%r, 81.63%r, 71.08%r, 52.94%r and 22.43%r) with decreasing tumor purity. In the favor-recall model, precision varied (70.21%p, 80.11%p, 88.14%p, 93.81%p and 97.54%p), while recall was boosted, especially at lower tumor purity (96.10%r, 94.45%r, 90.51%r, 80.81%r and 53.06%r). According to these results, generally, we suggest using favor-f1 mode at higher tumor purity and favor-recall mode at lower tumor purity.

Our results also showed that lower normal purity harmed somatic variant-calling performance, especially on recall. With tumor purity fixed at 1.0, in favor-f1 mode, the recalls were 86.86%r, 70.35%r and 51.74%r at normal purity 1.0, 0.95 and 0.90, respectively. In favor-recall mode, the recalls were 96.10%r, 85.36%r and 69.42%r. Accordingly, we suggest using a high-purity normal sample with ClairS for somatic variant discovery. Last, we showed the results using 0.8/0.95 tumor/normal purity. In favor-recall mode, ClairS achieved 79.68%p/81.13%r/80.40%f, demonstrating the reliability of ClairS in challenging sample conditions.

Analysis of false-positive and false-negative calls

Using 50-fold tumor and 25-fold normal coverage, we manually analyzed 300 false-positive and 300 false-negative calls randomly picked from all variant calls. Each false positive and false negative was assigned with the most obvious limitation as the reason why the call was false. The reasons for the 300 false-positive calls and 300 false-negative calls are listed in Supplementary Table 5a,b, respectively. A distribution of the reasons is given in Extended Data Fig. 5 as a pie chart.

Extended Data Fig. 5. Performance comparison with DeepSomatic on the PacBio platform.

(a) The breakdown of variant categories and the VAF distribution of synthetic variants used for training the ClairS SS model on PacBio data. (b) The breakdown of variant categories and the VAF distribution of real cancer cell line variants used for training the ClairS SSRS model on PacBio data. (c) show the performance of SNV and Indel at different tumor/normal coverages between ClairS SSRS and DeepSomatic. (c) and (d) The precision-recall result of ClairS SSRS with DeepSomatic on five PacBio real cancer cell lines (HCC1395/BL, HC1937/BL, HCC1954/BL, H1437/BL, H2009/BL). Benchmarking was restricted to chromosome 1. The orthogonal-tools benchmark truth set from Park et al. was used for evaluation. (e) The SNV and Indel F1 score comparison between ClairS SSRS and DeepSomatic in six PacBio cancer cell lines. WGS were used for COLO829/BL, and chromosome 1 was used for other datasets.

Among the false-positive calls, 39% had no matching truth but were with tumor 0.05 ≤ VAF < 0.1, 22% were with tumor 0.1 ≤ VAF < 0.15 and 9% were with tumor VAF ≥ 0.15. As elaborated in Methods, these calls are with tumor and normal coverage ≥4, and normal VAF below 0.05. One possible explanation is that ClairS was unable to pick up more hints and tell these calls from the true ones. Some of these cases might be correctly called again with higher tumor coverage, which reduces the statistical bias in VAF. It is also possible that as the SEQC2 truth set is still under active development, its incompleteness caused a few variants that are actually true to be misclassified as false positives. Another major category of false-positive calls is likely to be caused by alignment artifacts because of a repetitive or imperfect genome reference sequence. It includes 6% ‘one or more deletions in flanking 50 bp’, 5% ‘in low complexity region’, 4% ‘excessive mismatches in alignment’, 3% ‘one or more insertions in flanking 50 bp’ and in total, another 10% in repetitive regions of different types. Also, we found 1% false-positive calls, possibly because of insufficient normal coverage.

Among the false-negative calls, 40% truth variants that were not called were with tumor VAF < 0.1, 10% were with ‘normal VAF ≥ 0.05, but tumor VAF < 6 times larger than normal VAF’ and 3% were with <3 reads supporting the variant allele in a tumor. These missed variants might be called again if higher tumor coverage is given. False-negative calls are more likely to be caused by alignment artifacts, as we observed 10% false-negative calls in the homopolymer region, 10% in the low complexity region, 7% in the tandem repeat region and in total, another 9% that were also likely to have been caused by alignment. We also observed 4% false-negative calls caused by extreme strand bias (reads observed in only one strand) and 1% with base quality of all supporting bases <20.

Somatic indel-calling performance

The SEQC2 truth set provides ~40,000 SNVs but only ~2,000 indels. Owing to the scarcity of truth somatic indels that could reduce statistical biases, we benchmarked somatic indel calling separately. The results of different combinations of tumor coverage (25-, 50- and 75-fold) and normal coverage (20-, 25- and 30-fold) are shown in Supplementary Table 6. With normal coverage fixed at 25-fold, in favor-f1 mode (variant quality cutoff at 12), ClairS achieved 73.94%p/61.79%r, 78.26%p/69.07% and 80.03%p/70.28%r, at 25-, 50- and 75-fold tumor coverage. In favor-recall mode (cutoff at 8), ClairS achieved 63.10%p/68.32%r, 54.48%p/77.35%r and 55.12%p/78.93%r. As in our conclusion for SNV calling, generally, we suggest using favor-recall mode for calling somatic indels at lower tumor coverage for higher recall. Furthermore, we benchmarked the indel-calling performance when excluding complex genomic regions, as shown in Supplementary Fig. 2 and Supplementary Table 7, the performance improves an average 13.54%f in favor-f1 mode and reached 88.91%f when with 75/30-fold tumor/normal coverage. These results indicate that complex regions pose greater challenges for accurate indel detection.

Analysis on other cancer cell lines

In addition to HCC1395/BL (which in 2023, when the first preprint of ClairS was released, was the only available tumor–normal pair with known truth variants for benchmarking), a few new tumor–normal pairs were further made available to the public for us to benchmark the performance ClairS. In early 2024, the Nanopore EPI2ME Labs released a melanoma cancer cell line dataset, COLO829/BL, sequenced using both Q20+ 4 kHz (ref. ³³) and Q20+ 5 kHz (ref. ³⁴) chemistries. As shown in Extended Data Fig. 7a, the datasets sequenced with Q20+ 5 kHz chemistry and base-called with Dorado exhibit lower error rates (2.9 versus 3.7% in HCC1395/BL) and higher read quality (peak at 22 versus 20) compared to those processed with Q20+ 4 kHz chemistry and Guppy basecalling. In addition, in late 2024, Park et al.³⁰ from UCSC and Google released the datasets of four pairs of cell lines, including two breast cancer cell lines (HCC1937/BL and HCC1954/BL) and two lung cancer cell lines (H1437/BL and H2009/BL), which were sequenced with Q20+ 5 kHz chemistry. The truth set of COLO829/BL was generated by the New York Genome Center³⁵ and the truth sets of the other four pairs were generated by Park et al. through the use of multiple sequencing platforms and multiple somatic variant callers, including ClairS. These new truth sets enable more comprehensive benchmarks of ClairS across different sequencing chemistries and cancer types, but with the caveat that they are not as reliable as the truth set of HCC1395/BL, which was intensively verified in a wet laboratory. As shown in Supplementary Tables 8 and 14, ClairS synthetic model achieved 92.95%f, 92.93%f, 87.98%f, 91.99%f, 93.33%f and 98.58%f for SNV in 50/25-fold T/N COLO829/BL 4 kHz, COLO829/BL 5 kHz, HCC1937/BL 5 kHz, HCC1954/BL 5 kHz, H1437/BL 5 kHz and H2009/BL 5 kHz datasets, respectively. The performance numbers of the lung cancer cell lines were higher than those of the breast cancer cell lines, which agrees with the studies^35,36 using short read and highlights the influence of different cancer types in somatic variant calling.

Extended Data Fig. 7. Performance analysis of ClairS with different sequencing platforms and settings.

(a) The read length and read quality distributions of the HCC1395/BL sample are depicted across two sequencing chemistry and basecaller settings: R10.4 4 kHz Guppy and R10.4.1 5 kHz Dorado. (b) Variant-calling performance with and without low-VAF filtering across different datasets and the total number of truth variants in each dataset before/after applying the low-VAF filter are shown. 50x/25x HCC1395/BL and COLO829/BL datasets in different platforms were used for benchmarking. (c) Venn diagrams show the true positive (TPs) SNVs and Indels identified by ClairS (using the SSRS model) and DeepSomatic for each cancer cell line. The analysis was performed on WGS data for the COLO829/BL cancer cell line, and chromosome 1 for others.

Analysis of excluding complex genomic regions

We also benchmarked the performance of different models with and without difficult genomic regions, as defined by GIAB genomic stratifications³⁷. These regions include repetitive sequences, low-mappability, difficult-to-map segments and segmental duplications regions. As shown in Supplementary Fig. 2 and Supplementary Table 7, excluding difficult regions improved variant detection by an average of 1.43%f for SNVs and 16.58%f for indels across multiple coverages in HCC1395/BL and COLO829/BL. These results underscore the challenges in accurately detecting variants, particularly somatic indels, within complex genomic regions.

Performance after adding phasing information to the input in step 2

The reconstruction of haplotypes, also known as haplotype-resolved assembly or phasing, greatly improved the performance of long-read germline variant calling in previous practice^24,25. In Methods, we elaborate how phasing information can be used to enhance somatic variant-calling performance. Using 50/25-fold HCC1395/BL and favor-recall mode for benchmarking, we show in Extended Data Fig. 1b that phasable somatic variants performed better than unphasable ones, especially at lower VAF. At VAF 0.1–0.15, phasable somatic variants had 43.8%p/85.6%r/57.9%f, but unphasable somatic variants had only 16.4%p/81.2%r/27.3%f. At VAF 0.05–0.1, phasable somatic variants had 5.8%p/46.6%r/10.3%f, but unphasable somatic variants had only 2.3%p/27.2%r/4.3%f. We tried disabling phasing in ClairS and fed no phasing information to the calling networks. As shown in Extended Data Fig. 1b, the overall F1 score dropped from 81.14%f to 78.66%f (−2.48%).

Performance of the two respective networks in step 2

In contrast to Clair3, in which the pileup network handles all the variant candidates, and the full-alignment network processes only the undecided candidates using the pileup network, ClairS uses both networks equally to make collective decisions. The rationale and details are elaborated in Methods. We tested the performance using both the pileup network only and the full-alignment network only, with 50/25-fold of HCC1395/BL and favor-recall mode. The results are shown in Supplementary Table 9. When we used only the pileup network, the F1 score dropped from 81.14%f to 73.27%f (−7.87%). When we used only the full-alignment network, the F1 score dropped from 81.14%f to 79.14%f (−2.00%).

Performance of step 3: searching for ancestral haplotype support

As elaborated in Methods, step 3 utilizes remote alignment signals that could not be included in the network inputs due to computational limitations to improve the precision of the called somatic variants. As shown in Supplementary Table 9a, without this step, the precision dropped from 70.21%p to 67.14%p (−3.07%, using 50/25-fold HCC1395/BL and favor-recall mode). Furthermore, the performance effect of applying the ancestral haplotype support post-filter is shown in Supplementary Fig. 3 and Supplementary Table 9b across different datasets, basecallers and sequencing chemistries, demonstrating the consistent benefit of this strategy.

Further enhancing ClairS with augmented model training using real cancer cell lines

Synthetic tumor samples used in ClairS model training have no limit in providing a broad coverage of tumor purity, normal contamination and allelic fraction spectrum for a robust model training. Nevertheless, they fail to cover characteristics that only exist in real cancer samples, such as cancer-specific indel length distribution and cancer-type specific mutational signatures³⁸. Intuitively, there are two ways of using real cancer samples for model training, use them only, or use them together with the synthetic tumor samples. Because of the existence of cancer-type-specific mutational signatures, using only real cancer samples for model training risks having a lower or distorted variant-calling performance because of missing a cancer type in model training. The problem gets even worse when considering the cost of continuously generating new data of reference cancer samples of tens of cancer types³⁶ on new sequencing technologies and chemistries, and we believe that the problem can be mitigated if the cancer-type-agnostic synthetic samples are used together with real samples for model training. To prove this, and enabled by the real cancer cell line datasets publicly released by Park et al.³⁰, we trained two new types of ClairS model using (1) only real cancer samples (ClairS RS) or (2) using synthetic samples for a base model to be augmented with real cancer samples (ClairS SSRS). The model training details for both ClairS RS and SSRS are described in the ‘Real cancer cell lines for augmented model training’ section in Methods. Figure 4a–c shows an overview of the ClairS RS and SSRS training workflow, the count of each variant type used in training, and the VAF distribution of real cancer cell lines used in training. Following the SS model protocol, SSRS benchmarking excluded variants with VAF < 0.05. The benchmarkable variants are shown in Extended Data Fig. 7b and details are described in the ‘Benchmarking’ section in Methods.

The results of the two models across different tumor coverages are shown in Fig. 4d and Supplementary Table 10. HCC1395/BL and COLO829/BL were excluded from model training and used completely for benchmarking (dataset details shown in Supplementary Table 1b). ClairS SSRS has unanimously outperformed ClairS RS. In 50/25-fold HCC1395/BL, ClairS RS and SSRS achieved 95.53% and 96.19% F1 scores for SNVs and 73.96% and 79.67% for indels. In 50/25-fold COLO829/BL, the F1 scores were 94.79% and 95.70% for SNVs and 65.41% and 67.93% for indels. Notably, in the COLO829/BL dataset, the precision values were generally higher than recall, whereas the opposite trend was observed in HCC1395/BL, reflecting intrinsic differences between the two cancer samples.

We also compared ClairS against DeepSomatic³⁰, which was trained exclusively on real cancer cell lines. The whole genome of COLO829/BL (including 42,993 SNVs and 985 indels) was used for main benchmarking due to the following reasons: (1) COLO829/BL is a melanoma, a cancer type not included in both ClairS and DeepSomatic’s model training; (2) HCC1395/BL has the best somatic variant quality for benchmarking purposes, but it cannot be used for whole-genome sequencing (WGS) benchmarking because DeepSomatic has used it in model training; and (3) although DeepSomatic has excluded chromosome 1 of HCC1395/BL from model training, chromosome 1 contains only 3,457 SNVs and 147 indels, leading to insufficient statistical power for benchmarking in various coverage and purity settings. Although the higher quality of HCC1395/BL somatic variants will benefit model training and further improve the performance of ClairS, considering long-read somatic variant calling is still in its early stage of development, we have decided to exclude it from model training and reserve it for comprehensive benchmarks across sequencing platforms and new somatic variant callers. The results are shown in Extended Data Fig. 5a,b and Supplementary Table 12. ClairS SSRS outperformed DeepSomatic at most coverages and tumor purity ranges. At high, medium and low coverages (75/25, 50/25 and 25/25-fold), the SNV F1 scores of ClairS SSRS over DeepSomatic were 0.97%, 3.16% and 4.22% and the indel F1 score differences were 2.14%, 3.38% and 5.25%. Across a range of tumor purities (1.0, 0.8, 0.6, 0.4 and 0.2), ClairS achieved a higher mean F1 score than DeepSomatic by an average of 2.20% (93.71% versus 91.51%). The performance difference was driven by a trade-off: DeepSomatic exhibited higher recall at tumor purities above 60%, whereas ClairS demonstrated higher precision. The advantage of ClairS was most pronounced at 20% tumor purity, where it attained 87.27% F1 score, compared to 83.34% for DeepSomatic. These results underscore the importance of combining synthetic samples and real tumor data for augmented training to enhance somatic variant calling performance.

In addition to WGS benchmarking in COLO829/BL, we performed a chromosome 1-specific evaluation on five cancer cell line datasets (HCC1395/BL, HCC1937/BL, HCC1954/BL, H1437/BL and H2009/BL), which were excluded from DeepSomatic’s model training. For this assessment, we employed the orthogonal tool benchmarking methodology proposed by Park et al.³⁰, which integrates calls from multiple tools to generate a high-confidence truth set. As shown in Fig. 5c and Supplementary Table 11a, ClairS achieved better SNV calling performance to DeepSomatic in four of the five cell lines (HCC1395/BL, HCC1937/BL, HCC1954/BL and H2009/BL), while DeepSomatic performed better in the H1437/BL cell line. To further validate these findings, we conducted a second orthogonal technology benchmark using a truth set derived from DeepSomatic’s own calls on Illumina–PacBio hybrid data to evaluate performance on ONT platform. As shown in Supplementary Fig. 4, the overall performance of ClairS is comparable between the two orthogonal benchmarks (averaging 96.04% versus 95.71% in F1 score), demonstrating its consistent accuracy regardless of tool-specific or technology-specific truth set definition. This finding aligns with the DeepSomatic team’s observation that different validation methods yield complementary performance metrics with similar trends. While absolute performance may vary slightly between benchmarks, the performance trends using two benchmarks are similar. Furthermore, Venn diagrams of true-positive calls (Extended Data Fig. 7c) reveal that while ClairS and DeepSomatic share major true positives, each method uniquely identifies a substantial fraction (1–6%, mean 3.24%) of somatic variants across different cell lines. A higher proportion of true somatic variants was found in indels. Specifically, between 13% and 31% (mean 19.34%) of these true indels were called exclusively by DeepSomatic and ClairS across different cancer cell lines. These method-specific calls are crucial for constructing a more comprehensive benchmark for real tumor samples.

Fig. 5. Performance comparison with DeepSomatic on different cancer cell lines.

a, The SNV and indel performance comparison between ClairS SSRS and DeepSomatic on the COLO829/BL 5 kHz dataset at different tumor–normal coverages. Performance is evaluated under high, medium and low tumor coverage settings, each with two different normal coverage levels. b, Performance comparison of ClairS SSRS and DeepSomatic with different tumor purities on 50×/25× COLO829/BL 5 kHz dataset. c, The precision–recall result of ClairS SSRS with DeepSomatic on five real cancer cell lines (HCC1395/BL, HCC1937/BL, HCC1954/BL, H1437/BL and H2009/BL). Benchmarking was restricted to chromosome 1 to avoid bias, as DeepSomatic’s training data included other chromosomes. The orthogonal-tools benchmark truth set from Park et al.³⁰ was used for evaluation. d, The SNV and indel F1 score comparison between ClairS SSRS and DeepSomatic in six ONT cancer cell lines. WGS was used for COLO829/BL and chromosome 1 was used for other datasets.

Performance on PacBio data

Although our method was initially designed for the ONT platform, we evaluated its performance on PacBio sequencing to assess its generalizability. We applied the same synthetic training data synthesis strategy to the PacBio platform. A summary of the PacBio data used for training and benchmarking is provided in Supplementary Table 1. We trained the ClairS model (ClairS SS) on PacBio data using synthetic variants generated from two samples: HG003 and HG004. We used both HG003–HG004 and HG004–HG003 as tumor–normal synthetic samples for data synthesis. The HG003 and HG004 samples were acquired from the PacBio research team³⁹. The datasets were sequenced with PacBio Revio system. Then we trained the ClairS SSRS using the synthetic SS model on the two breast cancer cell lines (HCC1937/BL and HCC1954/BL) and two lung cancer cell lines (H1437/BL and H2009/BL) acquired from Park et al.³⁰. Similar to ONT, we excluded the chromosome 1 and 20 in SSRS model training for all cell lines. The distribution of different training categories and VAF is shown in Extended Data Fig. 5a,b.

For benchmarking, we used the HCC1395/BL and COLO829/BL cell lines in WGS with additional four cancer cell lines in chromosome 1. We benchmarked the performance with DeepSomatic in COLO829/BL in WGS and other samples in chromosome 1, as DeepSomatic has included these cell lines in model training. The SNV results (Extended Data Fig. 5c and Supplementary Table 10) show that ClairS achieved performance comparable to DeepSomatic on the COLO829/BL sample across multiple coverages, with an average 96.16% F1 score versus DeepSomatic’s 95.77%. DeepSomatic outperformed ClairS on a 25/30-fold tumor–normal pair, whereas ClairS outperformed DeepSomatic on other coverage combinations. On the other cell lines, as shown in Extended Data Fig. 5d and Supplementary Table 11b, their performance was variable: ClairS performed better on some (for example, HCC1937/BL and H2009/BL), whereas DeepSomatic performed better on others. A similar trend was observed for indel detection. DeepSomatic performed better on HCC1395/BL cancer cell lines, whereas ClairS performed better on others. Notably, as shown in Extended Data Fig. 7b and Supplementary Table 14, ClairS achieved better performance on PacBio data compared to ONT long read. This demonstrates the advantage of our method when using PacBio data, which has a lower error rate.

Performance on Illumina data

Short-read somatic small-variant calling has been intensively studied. A nonexhaustive list of state-of-the-art methods include Strelka2¹⁰, Mutect2⁸, Lancet⁶, NeuSomatic⁵, Octopus¹², SomaticSniper⁹ and VarNet⁴. ClairS was designed primarily for long-read somatic small-variant calling; however, it poses no limitations to small reads, and we expect a variant-calling method that works for long reads to perform as well as or even better than existing short-read methods. Also, benchmarking against other short-read somatic small-variant callers provides insights into how ClairS performs against existing methods, as no other long-read callers are available for comparison.

A summary of the Illumina data used for model training and benchmarking is shown in Supplementary Table 1. We used both HG003–HG004 and HG004–HG003 as tumor–normal samples for data synthesis. The HG003 and HG004 samples had 91.15- and 88.49-fold coverage, and were publicly shared by Google Health Center. For benchmarking, we also used the HCC1395/BL tumor–normal pair, but with data from six sequencing centers, made available by the SEQC2 consortium (NS, NovaSeq at Illumina; NC, HiSeq at the National Cancer Institute; IL, HiSeq at Illumina; EA, HiSeq at European Infrastructure for Translational Medicine; FD, HiSeq at Fudan University; and NV, HiSeq at Novartis) with coverage ranging from 37.93- to 87.54-fold. The six multicenter replicates enabled us to verify ClairS’ performance consistency. If not specifically mentioned, other training and benchmarking details are the same as those of the ONT data experiments, and we benchmarked only SNVs for all callers. Like the ONT data experiments, the command lines and links to the data used for both model training and benchmarking are listed in the ‘Data availability’ section in Supplementary Notes.

The evaluation metrics of the eight callers on the six datasets are shown in Supplementary Table 13. Figure 6a shows the precision–recall curves and Fig. 6b shows a histogram of the F1 scores. ClairS consistently performed comparably or slightly better than the two top-performing callers, Strelka2 and Mutect2 on benchmarked cancer cell lines. On the six datasets, ClairS achieved a 97.88%f, 97.82%f, 97.46%f, 97.53%f, 97.03%f and 96.41%f. Strelka2 achieved a 96.16%f, 96.18%f, 97.03%f, 97.35%f, 96.32%f and 95.47%f. Mutect2 achieved 95.21%f, 95.75%f, 95.33%f, 96.24%f, 96.46%f and 94.35%f. Broken down into different VAF ranges, the ClairS performance was also consistently comparable to or better than that of other callers, as shown in Fig. 6c. Figure 6d shows Venn diagrams of the overlaps of false-positive calls between Strelka2, Mutect2 and ClairS. The diagrams show that although rare, there are 30 to 41 false-positive calls in the six datasets that were called by all three callers. Owing to the possible incompleteness of the truth set, a false-positive variant can be either a wrong call or a true variant missing from the truth set. These false-positive variant calls called by all three callers are worth conducting further verification on and might further contribute to the completeness of the truth set.

Fig. 6. Illumina HCC1395/BL dataset benchmarking results.

a, The precision–recall curve of HCC1395/BL short-read datasets from six SEQC2 sources (NS, NC, IL, EA, FD and NV) using eight tools (Strelka2, Lancet, Mutect2, NeuSomatic, Octopus, SomaticSniper, VarNet and ClairS). Variants were ranked by Strelka2 (SomaticEVS), Mutect2 (TLOD), VarNet (Score), SomaticSniper (SSC) and other callers (QUAL). The dot on each line shows where the best F1 score was. b, The overall F1 score of the experiments shown in a. c, The F1 score at different VAFs of the experiments shown in a. d, Venn diagrams showing the overlap of false-positive variant calls between Strelka2, Mutect2 and ClairS.

Differences in somatic variants called with long and short reads in different genomic stratifications

Although short-read sequencing underpins most current methods for somatic variant detection, its performance is fundamentally limited in repetitive or complex genomic regions. Long-read sequencing overcomes this limitation by generating reads that span tens of kilobases, enabling unique mappability and comprehensive haplotype resolution⁴⁰. The capacity of long-read sequencing suggests its potential to identify somatic variants that are missed by short-read approaches, which are worth further investigation. Extended Data Fig. 6a,b illustrates that ONT and PacBio long read consistently calls more variants than Illumina short read in the GIAB difficult genomic regions. Previous studies on long-read germline variant calling have shown a better performance on SNPs compared to short read, particularly in challenging genomic regions^24,25. In our benchmarks, even when using HCC1395/BL that offers the highest quality of truth somatic variants up to date, the truth set was generated primarily based on short reads, thus risking misclassifying some variants that only long read has revealed as false positives. We found 3,601 and 671 of these ‘false-positive’ variants in the HCC1395/BL and COLO829/BL, which were detected by both ONT and PacBio long read but missed by Illumina short read, and have the potential to be further experimentally verified as a truth. Extended Data Fig. 6c shows that these variants have a broad coverage of VAF and generally have good base qualities.

Extended Data Fig. 6. Performance analysis of ClairS by genomic context.

(a) Comparison of the number of somatic small variants called with ONT, PacBio, and Illumina data at different GIAB difficult genomic regions. 50x/25x HCC1395/BL or COLO829/BL dataset was used. (b) The count of somatic variants detected by Nanopore and Illumina, stratified by GIAB genomic context. 50x/25x HCC1395/BL or COLO829/BL dataset was used. (c) The quality score and VAF distribution of the variants detected by ONT but missed by Illumina stratified by GIAB genomic context. 50x/25x HCC1395/BL or COLO829/BL dataset was used.

Performance analysis on copy-number variation-prone regions

Somatic copy-number variations (CNVs) are a hallmark of cancer genomes, with profiles varying across cancer types⁴¹. These alterations can confound the accurate detection of somatic SNVs by altering the expected allelic balance and coverage. To evaluate the performance of long-read somatic variant callers in these challenging regions, we conducted a targeted analysis in CNV-prone areas.

As a high-confidence, fully validated CNV call set was unavailable for all tested cell lines, we first identified CNV regions using a third-party tool, SAVANA²⁸, run on the HCC1395/BL and COLO829/BL 5 kHz datasets (Extended Data Fig. 8a) acquired from Park et al.³⁰. We then categorized the genome into seven categories based on the log₂ copy-number ratio: ‘high gain’, ‘medium gain’, ‘low gain’, ‘normal’, ‘low loss’, ‘low loss’, ‘medium loss’ and ‘high loss’ (Extended Data Fig. 8b). This stratification allowed us to assess somatic variant-calling performance across a spectrum of copy-number states, from high-level amplifications to deep deletions. Our analysis revealed that copy-number profiles were highly heterogeneous across different cell lines and chromosomes (Extended Data Fig. 8c,d), and the accuracy of somatic variant calling was more adversely affected in regions of copy-number loss compared to regions of gain (Extended Data Fig. 8e). This indicates that CNVs have a substantial impact on somatic calling performance and allelic losses present a greater challenge for accurate somatic variant detection than copy gains.

Extended Data Fig. 8. Performance analysis on CNV-prone regions in different samples.

(a) Distribution of copy number variation (CNV) segments across all chromosomes for the HCC1395/BL and COLO829/BL 5 kHz datasets. (b) Demonstration of the seven copy number categories used to classify genomic regions, shown for chromosome 1. (c) Distribution of CNV segments per chromosome. In box plots, the center line is the median (50th percentile), the box bounds are the 25th and 75th percentiles, and whiskers span to the minima and maxima within 1.5× the interquartile range. Exact number of CNV segments for each chromosome are listed in Supplementary Table 17. (d) Distribution of the CNV profiles across datasets. (e) Performance of ClairS SSRS on HCC1395/BL and COLO829/BL 5 kHz datasets, stratified by the degree of copy number gain or loss.

Discussion

In this study, we present ClairS, a somatic small-variant caller for ONT long reads. In our benchmarks, we showed that it is reliable at different sample coverages, tumor purities and normal contaminations. With the training data synthesis method that we devised, ClairS can be trained for somatic small-variant calling for different sequencing platforms. ClairS draws on its germline variant caller predecessors’ experience, while using a redesigned workflow, network architecture, network output and post-processing procedure for the more challenging somatic variant-calling tasks. Compared with DeepSomatic, which was trained on real cancer samples and relied on multiple-technologies sequencing data to establish ground truths for model training, ClairS relies on carefully designed synthetic samples, reliable training labels, and incorporating phasing for enhanced generalizability. Unlike ClairS-TO⁴², which is tailored for tumor-only somatic calling and uses a dual network architecture to identify artifacts, along with multiple post-processing strategies and germline resources to discriminate germline from somatic variants, ClairS benefits from a matched normal control. This allows ClairS to concentrate on distinguishing challenging somatic variants from sequencing artifacts and achieved outperformed accuracy in tumor–normal pair calling.

The use of long reads for somatic SV discovery has unraveled complex somatic SVs that could not be detected using short reads alone^28,29,43. With the unprecedented power of long reads to cover repetitive genome regions, we expect the use of long reads for somatic small-variant calling to reveal more somatic variants that were previously inaccessible by short reads, and lead to a better understanding of the mutational processes and functional consequences of the somatic variants in different cancer types. To allow more researchers to achieve these goals, ClairS has been included as the small-variant caller in ONT’s somatic variant-calling workflow⁴⁴.

Somatic VAF is influenced by several factors in real tumors, including intratumor heterogeneity, contamination from normal cells, intracell variation (especially in polyploid tumor cells) and sequencing errors. ClairS does not model these factors individually and instead simplifies the VAF as a value between 0 and 1. A key direction for future research is to quantify each factor’s impact to improve accuracy.

Our study has several limitations. First, our model training and analyses were conducted on cancer cell lines, which provide high-quality DNA but lack the heterogeneity and contamination typical of real tumor samples. The performance of our pipeline on clinically relevant specimens, such as formalin-fixed paraffin-embedded tissues with degraded DNA, remains to be evaluated. Second, the current lower sequencing coverage of long-read technologies compared to short-read sequencing at a similar cost may constrain sensitivity for detecting subclonal mutations in low-purity samples. Future work will focus on optimizing method for various real clinical cancer samples, as well as exploring the cost–benefit trade-offs between coverage and the unique variant detection capabilities of long-read sequencing.

Methods

Training data synthesis

Generating synthetic tumors and synthetic normal

A tumor comprises normal cells and tumor cells; the latter are regarded as foreign. Similarly, the normal cells of an individual are considered foreign to the normal cells of another individual, and vice versa. We observed that the germline variants unique to an individual can mimic a somatic variant when mixed with another individual. It is noticed that cancer mutation signatures that have arisen from specific repair pathway dysfunction would not be effectively covered by synthetic tumors, but with insufficient known truth somatic variants and standard tumor–normal sample pairs available, this observation lays the foundation for generating ample synthetic somatic variants from known truth germline variants in the GIAB reference samples with real sequencing data for deep-learning model training. The detailed training data generation workflow is shown in Fig. 1a and Supplementary Fig. 4. To illustrate our method, we used 80-fold GIAB HG002 ONT WGS alignments as the tumor source (hereafter referred to as A) and 50-fold HG001 as the normal source (hereafter referred to as B). We first calculated the average sequencing coverages for both samples using the mosdepth⁴⁵ tool. Based on a minimum sampled coverage (for example, fourfold), we then determined the necessary number of data ‘chunks’ (for example, 20 chunks for A and 12 chunks for B), by dividing the sample coverage with the sampled coverage. Reads from the source alignment were distributed into these chunks sequentially. Assuming a relatively uniform read distribution, this process ensures each chunk approximates the desired sampled coverage. These chunks serve as the basic building units for generating synthetic tumor and normal datasets. First, they can be used to simulate different sequencing coverages and allelic fractions with fine granularity. For example, to simulate a synthetic tumor sample with an ideal allelic fraction of 67%, we can combine ten chunks from A (representing a total coverage of 40-fold) with five chunks from B (representing 20-fold). Second, the chunks can be used to model varying levels of normal sample contamination. For instance, instead of using 100% of A as the normal sample, one or more chunks from B can be added to simulate contamination scenarios.

Coverage advice for the source data

We applied two restrictions to tumor and normal synthesis. First, to avoid any biases caused by reusing individual reads, we used sampling without replacement in tumor and normal synthesis (a chunk that was used could not be used again). Second, we required the synthetic tumor to have equal or higher coverage than the synthetic normal to align with common practice. The two restrictions together imply that the raw sample with higher coverage is more favorable to be used for supplying the coverage for both the synthetic normal and the normal portion in the synthetic tumor, whereas the other raw sample with lover coverage is more favorable to be used for the tumor portion in the synthetic tumor. To maximize the diversity of synthetic tumor–normal admixtures, we preferentially used the higher-coverage sample to generate a larger number of data chunks. This practice is for maximizing both the synthetic normal coverage and the tumor purity spectrum. If we do the opposite, using the previous example (80-fold A and 50-fold B), if 20-fold B is reserved to mix with A for the tumor, then 30-fold B is left for normal, but we can achieve tumor purity only between 33% (tenfold A + 20-fold B) and 100% (80-fold A + no B); however, if 30-fold B is reserved to mix with A for the tumor, tumor purity between 0% (no A + 30-fold B) and 100% (80-fold A + no B) can be achieved, but only 20-fold B is left for normal.

Deriving multiple categories of variants from a synthetic tumor/normal pair

Four categories of variants, namely ‘somatic’, ‘germline’, ‘artifact’ and ‘normal-only’, were derived from a synthetic tumor–normal pair, as explained in detail in Fig. 1a. We used GIAB truth variants in the synthetic tumor and synthetic normal, and we defined any truth variants (including both homozygous and heterozygous) as ‘callable variants’. Basically, somatic variants are callable variants in the synthetic tumor but not in the synthetic normal, which are also known as truth germline variants in the tumor source, but not in the normal source. Germline variants are callable variants in both the synthetic tumor and synthetic normal, which are also known as truth germline variants in both the tumor source and normal source. Artifact variants are callable variants only in the synthetic tumor. They are not known as truth in either the tumor source or the normal source. Normal-only variants are callable variants found only in the synthetic normal. Three of the four categories, ‘somatic’, ‘germline’ and ‘artifact’, are used for model training. Only the variants located in the overlapping GIAB-defined high-confidence regions of both the tumor and normal sources were used for training to ensure the quality of the training variants. Because of the subsampling process, some variants might have few supporting reads in the synthetic tumor, especially for low VAF somatic variants. These variants were excluded from model training to avoid confusing the neural network. More exclusion details are given in the following paragraph. Our sample synthesis method supports generating synthetic tumors at any purity level, so we can use as many purities as possible to achieve fine coverage of VAFs from 0 to 1, but for practicality, we used three tumor purities (25%, 50% and 100%), and applied subsampling to all variants from the three purities to achieve acceptable VAF distribution. This is feasible because (1) the innate variance of the VAF of the germline variants from the tumor and normal sources enables a pool of somatic variants fully covering VAFs from 0 to 1, even with just three purities, and (2) applying subsampling to the pool enables us to enrich difficult somatic variants and reduce the number of less common somatic variants for model training. In terms of subsampling, the VAFs of chosen somatic variants were randomly selected from a β distribution with shape parameters α = 2 and β = 5. The same distribution was used by the SEQC2 consortium for spike-in somatic variants in Sahraeian et al.⁴⁶. Our experiment showed that subsampling itself resulted in a ~1.8% increase in the F1 score using 50×/25× of HCC1395/BL. The resulting VAF distribution of SNVs is shown in Fig. 1b. We tried adding one more tumor purity at 12.5%, but apart from longer model training time, no performance gain was observed.

Other details about the variants selected for model training

For the somatic variants used for model training, a minimum coverage of four, and a minimum of three reads supporting the somatic variant allele are required. Somatic variants with (1) synthetic normal VAF > 0.03 and (2) synthetic tumor VAF ≤ 6 times of synthetic normal VAF, were excluded from training. With this rule, it is worth pointing out that those with synthetic normal VAF > 0.03 and synthetic tumor VAF six times or more higher than the synthetic normal VAF are included for training, and ClairS has learned from the rule to give a lower variant quality score to those somatic variants with a relatively high normal VAF but a tumor VAF not plausibly higher than the normal VAF. A detailed normal VAF distribution is shown in Supplementary Fig. 6. For the artifacts, the non-reference VAF was capped at 0.05 to avoid using a large number of obvious artifacts for training. For germline variants, a minimum coverage of four reads and a minimum of three reads supporting the germline variant allele, were required in both the synthetic tumor and normal. Germline variants with a difference in VAF larger than 0.1 between the synthetic tumor and normal were excluded from training. To prevent the model from inferring a somatic variant from its coexistence with two or more adjacent germline variants, which is a confounding factor that can be easily learned by the model, germline variants that were less than 33 bp (the window size of our model design) from each other were excluded from training. Our experiment showed that this exclusion alone increased somatic variant-calling precision by ~1%. With three tumor purities and the exclusions explained above, 12,489,342 training samples were left. The breakdown is shown in Fig. 1c.

Phasing information for somatic variant-calling

An authentic somatic variant usually originates from either the maternal or paternal haplotypes, whereas a random error usually has a fair chance happening in both (Extended Data Fig. 1a). Thus, somatic variants that have a single ancestral haplotype (either maternal or paternal) should be considered more reliable than those with two ancestral haplotypes, except for somatic variants with high VAF that might be a result of copy-number alteration or clonal duplication⁴⁷. ClairS uses phasing information for both model training and inference. Clair3 and LongPhase are used for phasing and read haplotagging. More details are given in the ‘ClairS workflow and design’ section. ClairS uses phasing information during full-alignment-based variant calling, in which a channel named ‘Tumor/Normal/Phasing Info’ is used. In this channel, the alignments are grouped into haplotype-unknown, haplotype 1 and haplotype 2, each using the read order of the alignments. Although long-read sequencing enables outstanding phasing performance, some somatic variants in difficult genomic regions or without a heterozygous germline in their vicinity still cannot be covered by any phased reads. Thus, during model training, for each variant that has a heterozygous origin from the tumor source, if one or more reads can be phased, both a version of input with reads after phasing and a version before phasing were used.

ClairS workflow and design

Overview

Figure 2 shows an overview of the ClairS somatic variant-calling workflow. Starting from the alignments in the BAM/CRAM format of a tumor–normal sample pair, ClairS follows three steps to derive the somatic variants in a tumor and outputs them to a VCF file. In step 1, ClairS uses Clair3 and LongPhase for germline variant calling, phasing and read haplotagging. The processed alignments are then used for both pileup and full-alignment-based somatic variant calling in step 2. Step 3 involves post-processing filters that eliminate somatic variant calling if an ancestral haplotype (either maternal or paternal) from which the somatic variant could originate cannot be found.

Step 1: germline variant calling, phasing and read haplotagging

Step 1 is depicted in Fig. 2b. Clair3 (ref. ²⁴) is integrated into ClairS for calling high-quality heterozygous germline variants in both tumor and normal to maximize the performance of the subsequent phasing task. Unlike Clair3’s default, VAF ≥ 0.2 and coverage ≥ 10 were applied to ensure the quality of the called variants and reduce computational overhead. Only the heterozygous germline variants found in both tumor and normal were chosen for phasing. For phasing and haplotagging the tumor alignments, both LongPhase⁴⁸ and WhatsHap⁴⁹ are allowed in ClairS. We chose LongPhase over WhatsHap as the default because LongPhase runs ~15 times faster while delivering similar or longer phase sets on human samples. Notably, ClairS does not phase and haplotag the normal alignments. Our experiment showed phasing the normal alignments doubled the processing time but did not result in any improvement in calling performance.

Step 2: pileup-based and full-alignment-based variant calling

Step 2 is depicted in Fig. 2c. For a variant candidate (explained in ‘Selecting variant candidates’), a pileup input and a full-alignment input are generated (explained in ‘The design of pileup input and full-alignment input’). Then the inputs are sent to a Bi-GRU-based pileup-calling neural network and a ResNet-based full-alignment-calling neural network (explained in ‘The design of neural networks’) for inference. Both networks have the same output of a single task with three categories, ‘somatic’, ‘germline’ and ‘artifact’, which match exactly the three categories defined in the synthetic training data. In contrast to Clair3, in which the faster pileup-based calling cleans up most variant candidates that are obvious variants, and the more computational-demanding full-alignment-based calling handles the tricky and less obvious candidates, ClairS considers the power of the two neural networks equal. We observed that full-alignment-based calling is performant at mid-range VAFs; however, pileup-based calling requires less evidence than full-alignment calling to draw the same conclusion. When VAF goes under 0.1, pileup-based calling becomes increasingly more sensitive and usually outperforms full-alignment-based calling. This observation makes pileup-based calling more important for somatic variant calling than its role in Clair3 for germline variant calling, especially in multiple clinical usage scenarios when sensitivity is emphasized. In ClairS, a somatic variant is called when both networks give somatic the highest probability. The variant quality (QUAL) is Phred-like and is calculated as $\max \left(-10{\log }_{10}\left(\frac{1-p}{p}\right)+\mathrm{2,0}\right)$, where $p=\frac{p_{\mathrm{Somatic}}^{\mathrm{Pileup}}+p_{\mathrm{Somatic}}^{\mathrm{Full}-\mathrm{alignment}}}{2}$. Also, in contrast to Clair3, which uses the same network for both SNP and indel calling, ClairS uses two different networks respectively trained for SNV and indel calling. This means that ClairS runs on four networks in total: pileup for SNVs, pileup for indels, full-alignment for SNVs and full-alignment for indels. The rationale behind the new design is that unlike germline variants that are commonly diploid, somatic variants have no ploidy assumption, meaning that the existence of SNVs and indels in the same position are independent events. Our tests found that using separated networks led to a 1.5% increase in SNV recall. The use of separated networks also allowed the use of different variant quality cutoffs for SNVs and indels, which is useful for somatic variant calling, especially when the sample condition is not ideal.

Selecting variant candidates

Sending every genome position as a variant candidate to the neural networks guarantees maximum sensitivity; however, it is not only computationally infeasible, but also unreasonable to work on nonstarter positions, such as those without any non-reference allele support. A good variant candidate selection strategy is essential to achieve a balance between sensitivity and running time. In ClairS, the selection criteria are as follows. Let r ∈ K = (A, C, G, T) be the reference base of a genome position, and m ∈ K-r be the alternative bases. D^X_m denotes the coverage of m at the position in sample X ∈ {T, N}, where T and N represent the tumor and normal sample. An alternative base in m is selected as a variant candidate if it matches either criterion below:

$$\left\{\begin{array}{c}{D}_m^T\ge 3\mathrm{and}\frac{D_m^T}{{\sum }_i^K{D}_i^T}\ge \alpha \mathrm{and}\frac{D_m^N}{{\sum }_i^K{D}_i^N}<\alpha \hfill \\ D_m^T\ge 3\mathrm{and}\frac{D_m^T}{{\sum }_i^K{D}_i^T}\ge \alpha \mathrm{and}\frac{D_m^N}{{\sum }_i^K{D}_i^N}\ge \alpha \mathrm{and}\frac{\frac{D_m^T}{{\sum }_i^K{D}_i^T}}{\frac{D_m^N}{{\sum }_i^K{D}_i^N}}\ge \beta \hfill \end{array}\right)$$

where α sets the minimum VAF, and β sets the minimum tumor VAF to normal VAF ratio for a candidate to be selected. Intuitively, $D_m^T\ge 3$ means disregarding variant candidates with <3 reads in tumor supporting the variant allele. $\frac{D_m^T}{{\sum }_i^K{D}_i^T}\ge \alpha$ means disregarding those with tumor VAF < α. The first criterion equation means selecting a variant candidate if its VAF is <α in normal, but ≥α in tumor. The second equation means selecting a variant candidate, even if its VAF in normal is ≥α, the VAF in tumor is ≥β-times larger than the VAF in normal. In ClairS, α and β are configurable and default to 0.05 and 6, respectively. Typically, increasing α reduces the number of candidates, decreases recall rate at low VAF, but might increase precision rate because low-VAF somatic variants are harder to be called correctly. Increasing β also reduces the number of candidates, leading to lower recall but higher precision in all VAF ranges. Like model training data preparation, coverage ≥4 is required in both tumor and normal for a candidate to be selected for variant calling.

The design of pileup input and full-alignment input

ClairS pileup input comprises 1,122 integers, which is 33 positions wide with 34 features at each position (an example is given in Extended Data Fig. 2a). A detailed explanation of each feature is given under the ‘Description of pileup input features’ section in Supplementary Methods. ClairS has 18 pileup features in common with Clair3, and 16 additional features. The 16 new features are read counts N_LMQ+, N_LMQ−, N_LBQ+ and N_LBQ−, where N is either of the nucleotides A, C, G and T, LMQ subscript means a mapping quality (MQ) < 20, LBQ means base quality (BQ) < 30 and + and − mean the forward and reverse strand, respectively. The rationale behind the new features is that in ClairS, the results of pileup-based calling and full-alignment-based calling are trusted equally, so the mapping quality and base quality information that used to be exclusive to full-alignment calling need to be added to pileup-based calling. Our experiment showed that removing the 16 new features reduced precision by ~2% using 50×/25× of HCC1395/BL. ClairS full-alignment input comprises 30,030 integers (seven channels, each with 33 positions and 130 rows) to support, at most, 76 tumor reads, 52 normal reads and two empty rows as space between tumor and normal (an example is given in Extended Data Fig. 2b). Like Clair3, random subsampling down to the maximum supported coverage is used at excessive coverages. A detailed explanation of each channel is given under the ‘Description of full-alignment input channels’ section in Supplementary Methods. In both inputs, the candidate variant is centered at the 16th position. Positions uncovered by any base in full-alignment input are filled with zero. Moreover, for the full-alignment input, we represent the reference and alternative bases as scalar integers instead of one-hot encoding to improve computational efficiency. Further details are provided under the ‘Design of full-alignment input representation’ section in Supplementary Methods.

Extended Data Fig. 2. Visualization of neural network inputs.

(a) Pileup-based calling input visualization. The candidate site is centered and marked by two dashed lines. (b) Full-alignment-based calling input visualization. In b, the top and bottom are padded with zero when the total coverage of tumor and normal samples does not reach the input limit. The normal read alignments and tumor read alignments in all channels are separated by two rows filled with zeros. The two demonstrations involved truth variants randomly picked from the HCC1395/BL dataset.

Design of neural networks

The pileup and full-alignment network architecture and important parameters are shown in Fig. 3a,b. The pileup network uses two bidirectional gate recurrent unit (Bi-GRU) layers, each with 128 and 192 units. Compared to the Clair3 pileup network, the use of Bi-GRU instead of bidirectional long short-term memory (Bi-LSTM) architecture reduced trainable parameters from 2,532,995 to 2,309,507 and matrix computations from 3.11 billion to 2.38 billion, but improved performance in our experiment. The full-alignment network is a residual neural network (ResNet) comprising three standard residual blocks (detailed in the Supplementary Fig. 5) and a total of 2,468,131 parameters. A convolutional layer is added immediately before each residual block to expand the number of channels. More details of the model architecture and input design is described in the ‘Selection of flanking window size and network architecture’ section in Supplementary Methods. In both networks, a dropout rate at 0.3 was set for the flattened layer and dense layer to prevent overfitting.

Model training hyperparameters

ClairS trains both the pileup and full-alignment models for 30 epochs. Initially, the learning rate was 5 × 10⁻⁴, which was reduced by 10% after every epoch. Although the performance in the final few epochs is generally similar, we select the best-performing epoch for benchmarking. For full-alignment model training, we augmented the data with a mixture of phased and unphased alignments. This included generating unphased data by sorting reads by their start positions to handle unphasable variants. During model training, we employed ‘focal loss’⁵⁰, a technique that applies a modulating term to the cross-entropy loss to focus training on challenging, misclassified variants. This was critical because our strategy of including abundant artifact candidates to enhance model robustness across diverse sequencing scenarios resulted in severe class imbalance. By downweighting the loss contribution from well-classified majority classes, focal loss ensured that the model dedicated greater capacity to learning complex cases and underrepresented variant types, ultimately improving performance and providing more flexible handling of imbalanced data. A detailed description is provided under the ‘Design of the Focal loss in model training’ section in Supplementary Methods. Additionally, we applied dropout and L2 regularization to prevent overfitting. We utilized the Adam optimizer for optimization. The batch size was set to 800, which fits into 8 GB of graphic memory. The entire model training process took approximately a week on a NVIDIA GTX 3090 or equivalent GPU. The training time scales linearly with the increase in training samples.

Step 3: search for ancestral haplotype support

Step 3 is depicted in Fig. 2d. The neural networks exhibited good power in distinguishing real variants from false-positive candidates; however, useful signals remote to a variant candidate are not covered by the current neural network designs in ClairS, which considers only the flanking 16 bp of a candidate. Notably, even if the flanking window is extended to 50 bp, it is still too short for an accurate inference of which haplotype a variant candidate belongs to using only the networks, but the networks would already be computationally infeasible for somatic variant calling. In ClairS, post-processing step 3 is designed to reduce false-positive calling mistakes made by the networks by leveraging relatively remote germline variants to find the correct ancestral haplotype for a somatic variant. Any somatic variant calls that cannot be found with ancestral haplotype support are switched to an artifact and are excluded from the output. The haplotagged reads produced in step 1 are used in this step. For a somatic variant that covers any haplotagged reads, we required the somatic variant to coexist with the heterozygous germline variants less than 100 bp away on its left and right in the reads in the haplotype group the somatic variant supporting reads were in. An example of a false-positive somatic variant filtered by this rule is given in Extended Data Fig. 3a. A somatic variant at chr4:38,012,942 was called by the two networks. A phased heterozygous germline variant was found 61 bp left of the somatic call. Three reads in haplotype 2 that supported the somatic variant were found not to have the heterozygous germline variant. Thus, the somatic variant was considered unsupported by an ancestral haplotype. For a somatic variant that covers no haplotagged read, it is probably because there are no germline variants or only homozygous variants in the vicinity. In this case, we required the somatic variant to coexist with the homozygous germline variants less than 100 bp away on its left and right in all somatic variant supporting reads. An example of this is given in Extended Data Fig. 3b. A somatic variant at chr1:100,632,158 was called by the two networks. A homozygous germline variant was found 39 bp left of the somatic call. Multiple reads that support the somatic variant were found not to have the homozygous germline variant. Thus, the somatic variant was not considered to be supported by an ancestral haplotype. Somatic variants that do not have any germline variants less than 100 bp away on their left or right are not applicable in this step.

Extended Data Fig. 3. Two examples of haplotype inconsistency that signifies a false somatic call.

(a) Example of a false somatic call with a haplotype inconsistent with the haplotype derived from a heterozygous germline variant nearby. (b) Example of a false somatic call with a haplotype inconsistent with the haplotypes derived from a homozygous germline variant nearby. The bases A, C, G, and T are depicted in green, blue, yellow, and red, respectively. The background in gray, purple, and pink represents an unknown haplotype, haplotype 1, and haplotype 2, respectively. GT is an abbreviation of genotype.

Output

ClairS supports VCF format output. Somatic variants are marked ‘PASS’ or ‘LowQual’ if the variant quality is low (QUAL < 8, configurable by option) or they are filtered in step 3. For each variant, the allelic fraction and supporting coverage of the reference allele and all alternative alleles are shown. The options ‘–print_germline_calls’ and ‘–print_ref_calls’ enable outputting germline variants and artifacts, respectively.

Real cancer cell lines for augmented model training

In addition to HCC1395/BL, which was first released by SEQC2, Park et al.³⁰ and Keskus et al.²⁹ have further released four pairs of real cancer cell line datasets, including two breast cancer cell lines, HCC1937/BL, HCC1954/BL, and two lung cancer cell lines H1437/BL and H2009/BL. These datasets have also enabled ClairS to fine-tune its synthetic samples-based model with real tumor–normal pair cancer data.

Following the practices of Park et al.³⁰, we excluded chromosome 1 of all datasets from training and trained our model on only four real tumor–normal pairs: HCC1937/BL, HCC1954/BL, H1437/BL and H2009/BL, reserving the entire HCC1395/BL dataset for benchmarking. To the best of our knowledge, HCC1395/BL and COLO829/BL are, so far, the two best-studied tumor samples with the most confident somatic variants called and reported. HCC1395 particularly, has its truth variants curated by the SEQC2 consortium that involves multiple institutes. In fact, ClairS will benefit from adding HCC1395 to model training, but we have chosen to exclude it to allow future comparisons between ClairS and other works based on the whole HCC1395/BL. Furthermore, because our training set already includes two breast cancer lines (HCC1937/BL and HCC1954/BL), excluding HCC1395/BL carries minimal risk of losing essential breast cancer-specific mutational signatures. For real-data training, the ‘somatic’ and ‘germline’ variants used in training were generated by Park et al.³⁰ through cross-validation with multiple platforms and callers. ‘Artifact’ variants were defined as those located in high-confidence regions but not classified as ‘somatic’ or ‘germline’. Similar to using synthetic samples for training, we required a minimum of four read coverage for tumor and normal, and two alternative allele read support in tumor. A summary of the real cancer samples, allelic fraction distributions and variant categories is provided in Fig. 5c. The ClairS RS model maintained the exact hyperparameters used for synthetic training. For the ClairS SSRS model, we fine-tuned the synthetic-trained model using a reduced learning rate (1 × 10⁻⁴) and early stopping. We reserved 20% of the training data as a validation set and selected the epoch with the best validation performance for inference.

ONT library preparation and sequencing

Genomic DNA (gDNA) of a triple-negative breast cancer cell line (HCC1395) and a B lymphocyte-derived normal cell line (HCC1395BL) from the same donor were purchased from the American Type Culture Collection. The gDNA of HG001 was purchased from the Coriell Institute. The high-molecular-weight gDNA was examined by Nanodrop, Qubit and 0.35% agarose electrophoresis for its concentration, purity and integrity. The gDNA was then fragmented with gTube to generate DNA fragments approximately 20 kb in length. These fragments were then being sequenced at two sequencing centers: HKU and Novogene. At HKU, the fragments of HCC1395, HCC1395BL and HG001 were prepared and ligated with a sequencing adaptor using ONT’s ligation sequencing kit V14 SQK-LSK114. The ligated samples were sequenced on Q20+ PromethION flowcells using a PromethION 2 Solo device and MinKNOW software v.1.18.02 for 96 h. At Novogene, the fragments of HCC1395 and HCC1395BL were prepared and ligated with a sequencing adaptor using ONT’s ligation sequencing kit V12 SQK-LSK112. The ligated samples were sequenced on R10.4 PromethION flowcells using PromethION 48, for 96 h.

Benchmarking

For the HCC1395/BL dataset, we used the truth set of somatic variants generated and maintained by the SEQC2 consortium³. The truth set was orthogonally validated with multiple sequencing replicates from multiple sequencing centers that comprise over 1,500-fold sequencing data in total. We used only the somatic variants labeled ‘HighConf’ (high confidence) or ‘MedConf’ (medium confidence) as truth. Somatic variants labeled ‘LowConf’ (low confidence, VAF ≤ 0.05, not a part of the truth set as defined by SEQC2) were not used for benchmarking. In total, there were 39,560 truth SNVs and 1,922 truth indels; 39,447 of the SNVs and 1,602 of the indels were within the high-confidence regions defined in a BED file provided by SEQC2. For the COLO829/BL dataset, we used the truth set from New York Genome Center³⁵, which included 42,993 truth SNVs and 985 truth indels. A variant call was considered correct only if it matched both the genome position and variant allele of the truth. For both the ONT and Illumina benchmarks, some truth variants were excluded for the following reasons. First, although the mean sequencing coverage for the ONT HCC1395 4 kHz benchmarks is 75.97x, per-site coverage varies substantially. An extremely small number of truth sites (around ten) show no reads, and a slightly larger fraction have only one or two reads supporting the alternate allele (Supplementary Table 15). These low-support variants are due to truth set origin, higher error rate in ONT, and challenging genomic context, with the detailed discussion can be found under the ‘Handling truth variants with low alternate read support’ section in Supplementary Methods. Because such sites would fail all benchmarks and confuse model training, we exclude them from analysis. Second, some benchmarks tested multiple sequencing coverages and required sequencing read subsampling from the full dataset. The subsampling process might remove reads supporting a truth variant to an extent that few or no supporting reads are left. This affects especially the somatic variants that already have a low VAF. For example, a VAF 0.05 somatic variant with 20-fold coverage and one read supporting the variant allele can be reduced to VAF 0 by removing just one read during subsampling. This reduces the quality of the benchmarking results, especially for low-VAF truth variants when subsampled datasets are used. To alleviate the problem, any truth variants that have very low VAF (<0.05) observed in the full dataset before subsampling should be excluded. Summing up the two reasons above, for each of the full datasets we used in both the ONT and Illumina benchmarks, we excluded truth variants that matched any of the following criteria from benchmarking: (1) VAF < 0.05; (2) reads supporting the variant allele <3; (3) tumor coverage <4; and (4) normal coverage <4. Extended Data Fig. 7b and Supplementary Table 16 summarize the benchmarking results and the composition of the truth sets, comparing the inclusion versus exclusion of low-VAF variants across different sequencing platforms. For standardization, we used som.py, provided in Illumina’s Haplotype Comparison Tools³² (v.0.3.12) to generate evaluation metrics, including F1 score, precision and recall against the truth variants. The ‘compare_vcf’ submodule in ClairS produces identical results to som.py, but automates the exclusion of unqualifying truth variants. The truth set materials are publicly available to the community. All tools, their version and command lines used are given in the ‘Command lines used’ section in Supplementary Notes.

Computational performance

ClairS was written in Python and C++. The Python parts leveraged PyPy for speed up. The neural network implementations used PyTorch. Training ClairS neural networks requires a high-end GPU, but using ClairS for somatic variant calling requires only a CPU. For the 50×/25× HCC1395/BL pair, ClairS finished running in ~5 h for ONT and PacBio data and ~2 h for Illumina data (30% slower than Strelka2, but faster than all other short-read somatic variant callers), using two 12-core Intel Xeon Silver 4116 processors. The memory footprint is low and is controlled at lower than 1 GB per CPU. For model training, we tested Nvidia GeForce RTX 2080 Ti, 3090 and 4090, and found that each new model provided a ~35% speed increase from the previous generation.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Online content

Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at 10.1038/s41592-026-03152-4.

Extended data

Extended Data Fig. 4. Categorizing the FPs and FNs in ClairS.

The pie charts show the distribution of reasons for the FPs and FNs in ClairS. A 50x/25x ONT HCC1395/BL dataset was used for calling. 300 FPs and 300 FNs were randomly chosen from the results and analyzed. The tandem repeat, low complexity, homopolymer, and segmental duplication regions were defined using GIAB v3.0 Genome Stratification. Among the categories, ‘excessive mismatches in alignment’ and ‘insufficient normal coverage’ were decided manually, that is, without certain cutoffs. ‘Excessive mismatches in alignment’ was given if an eye check of the alignments revealed excessive inconsistent mismatches than usual alignments with a true somatic variant. ‘Insufficient normal coverage’ was given when a germline variant signal existed in both tumor and normal, but the coverage of normal was low, so the germline variant signal in normal was obviously weaker than in tumor.

Author contributions

R.L. conceived of the study. Z.Z. and R.L. designed the algorithms and implemented ClairS. Z.Z., R.L. and L.C. wrote the paper. L.C., J.S., X.Y., M.H, Y.-L.L. and T.-W.L. evaluated the benchmarking results. All authors revised the paper.

Peer review

Peer review information

Nature Methods thanks Kai Wang and other, anonymous, reviewer(s) for their contribution to the peer review of this work. Peer reviewer reports are available. Primary Handling Editor: Lin Tang, in collaboration with the Nature Methods team.

Funding

R.L. was supported by Hong Kong Research Grants Council grants CRF (C7003-24Y), GRF (17113721) and TRS (T21-705/20-N), the Shenzhen Municipal Government General Program (JCYJ20210324134405015), the URC fund from HKU and ONT.

Data availability

The links to the reference genomes, truth somatic variants, benchmarking materials, ONT, PacBio and Illumina data are given in the ‘Data availability’ section in Supplementary Notes. All analysis output, including the VCFs and running logs, is available at http://www.bio8.cs.hku.hk/clairs/analysis_result and via Zenodo at 10.5281/zenodo.18974512 (ref. ³⁸). The HCC1395/BL 4 kHz sequencing data generated in this study were deposited in the NCBI SRA with accession ID PRJNA986292.

Code availability

ClairS is open source and available at https://github.com/HKU-BAL/ClairS under the BSD 3-Clause license. The results in this paper were based on the ClairS initial release (v.0.4.4). Multiple installation options are available for ClairS, including Docker and Singularity. ClairS has also been included as the small-variant caller in ONT’s somatic variant calling workflow⁴⁴ since v.0.1.0.

Competing interests

R.L. receives research funding from ONT. The other authors declare no competing interests.

Extended data

is available for this paper at 10.1038/s41592-026-03152-4.

Supplementary information

The online version contains supplementary material available at 10.1038/s41592-026-03152-4.