Introduction

Somatic mutations play a pivotal role in the development and progression of cancer1. Accurate identification of the somatic variants is essential for understanding tumorigenesis, developing targeted therapies, and advancing precision oncology2,3,4. Typically, the detection of somatic variants relies on the comparison between tumor and matched normal samples, allowing for identifying somatic mutations from germline variants (including rare and de novo germline variants), and technical artifacts[5](https://www.nature.com/articles/s41467-025-64547-z#ref-CR5 “Zheng, Z. et al. ClairS: a deep-learning method for long-read somatic small variant calling. Preprint at bioRxiv, https://doi.org/10.1101/2023.08.17.553778

(2023).“),[6](https://www.nature.com/articles/s41467-025-64547-z#ref-CR6 “Park, J. et al. DeepSomatic: Accurate somatic small variant discovery for multiple sequencing technologies. Preprint at bioRxiv, https://doi.org/10.1101/2024.08.16.608331

(2024).“). However, in many real-world clinical scenarios, matched normal is not always available7,8,9,10,11, necessitating developing tumor-only somatic variant caller.

Tumor-only somatic variant calling is challenging compared with calling in paired samples. Without the matched normal as a reference, distinguishing true somatic variants from germline variants is very challenging. It is difficult to distinguish somatic variants that have a variant allelic fraction (VAF) close to germline variants, especially since the number of germline variants in a sample is often two orders of magnitude more than somatic variants. It is also difficult to distinguish somatic variants with low VAF from background noise (i.e., sequencing errors, alignment artifacts, etc.) without using a paired normal as a reference. The tumor-only calling methods are required to accurately discriminate true somatic variants from germline variants and technical artifacts without a matched normal control. Endeavors have been made to solve the problem for Illumina short-read sequencing data. Several short-read tumor-only somatic variant callers are commonly used in recent studies8,11,[12](#ref-CR12 “Benjamin, D. et al. Calling somatic SNVs and indels with Mutect2. Preprint at bioRxiv, https://doi.org/10.1101/861054

(2019).“),13,14. These callers primarily use statistical methods designed for short-read data, which may not be applicable to long-read sequencing technologies such as Oxford Nanopore Technologies (ONT) and Pacific Biosciences (PacBio). Unlike short-read sequencing, long-read sequencing generates reads spanning thousands of bases, enabling better resolution of complex genomic regions and structural variants (SVs). However, it also exhibits higher sequencing error rates and distinct error profiles across different genomic regions15. Due to these advantages, long-read sequencing is increasingly relevant in cancer research[5](https://www.nature.com/articles/s41467-025-64547-z#ref-CR5 “Zheng, Z. et al. ClairS: a deep-learning method for long-read somatic small variant calling. Preprint at bioRxiv, https://doi.org/10.1101/2023.08.17.553778

(2023).“),[6](https://www.nature.com/articles/s41467-025-64547-z#ref-CR6 “Park, J. et al. DeepSomatic: Accurate somatic small variant discovery for multiple sequencing technologies. Preprint at bioRxiv, https://doi.org/10.1101/2024.08.16.608331

(2024).“),[16](https://www.nature.com/articles/s41467-025-64547-z#ref-CR16 “Simpson, J. T. Detecting somatic mutations without matched normal samples using long reads. Preprint at bioRxiv, https://doi.org/10.1101/2024.02.26.582089

(2024).“),[17](https://www.nature.com/articles/s41467-025-64547-z#ref-CR17 “Keskus A. et al. Severus detects somatic structural variation and complex rearrangements in cancer genomes using long-read sequencing. Nat. Biotechnol. https://doi.org/10.1038/s41587-025-02618-8

(2025).“) and clinical diagnosis18,19, particularly for detecting disease-causing SVs and resolving challenging genomic architectures. Given this trend, there is a pressing need for efficient and accurate long-read somatic variant callers that work with tumor-only samples.

In this work, we present ClairS-TO (Tumor-Only), a long-read tumor-only somatic variant caller. ClairS-TO has its name stemmed from ClairS[5](https://www.nature.com/articles/s41467-025-64547-z#ref-CR5 “Zheng, Z. et al. ClairS: a deep-learning method for long-read somatic small variant calling. Preprint at bioRxiv, https://doi.org/10.1101/2023.08.17.553778

(2023).“), which requires paired tumor-normal samples for somatic variant calling. To maximize the algorithm’s own ability to tell somatic variant from germline and noise, ClairS-TO leverages an ensemble of two disparate neural networks that were trained from the same samples but for two exact opposite tasks—an affirmative neural network that determines how likely a candidate is a somatic variant, and a negational neural network that determines how likely a candidate is not a somatic variant. A posterior probability for each variant candidate is calculated from the outputs of the two networks and prior probabilities derived from the training samples. After the networks, ClairS-TO applies three techniques to further remove non-somatic variants: (1) Nine hard-filters that were found effective for short-read data are applied but with algorithms optimized or parameters tuned for long-read data; (2) Four panels of normals (PoNs), including three built from short-read datasets and one from long-read datasets are applied; and (3) a statistical method is applied to classify each variant into either a germline, somatic, or a subclonal somatic variant, using the estimated tumor sample purity and ploidy, as well as the copy number profile of each variant. To train the models, ClairS-TO depends on synthetic tumor samples and can be augmented using real tumor samples. Synthetic tumor sample combines the real sequencing reads of two samples and regards a germline variant that is specific to one sample as a somatic variant to the other sample. While somatic variants are scarce in real samples, the method can generate a number of training samples comparable to that of germline variants, providing sufficient samples to train a deep neural network robustly. The method also enabled fine-tuning the network using a much smaller number of somatic variants available in the real tumor samples, which is essential to allow the network to learn cancer-specific variant characteristics, such as mutational signatures20. We evaluated the performance of ClairS-TO using COLO829 (metastatic melanoma)21 and HCC1395 (breast cancer)22 cell lines. Both cancer samples are richly studied and with reliable truths for benchmarking. With ONT Q20+ long-read sequencing data, our experiments show that ClairS-TO consistently outperformed DeepSomatic (another long-read tumor-only somatic variant caller released after ClairS-TO) at multiple coverages, tumor purities, and VAF ranges. With PacBio Revio long-read data, ClairS-TO also outperformed DeepSomatic but with a smaller edge. ClairS-TO is optimized for long-read sequencing data but is also applicable to short-read data. Our experiments also show that ClairS-TO has outperformed Mutect2, Octopus, Pisces, and DeepSomatic at 50-fold coverage of Illumina short-read data.

Results

Overview of ClairS-TO

ClairS-TO comprises two disparate neural networks, the affirmative network (AFF) and the negational network (NEG), and is specifically designed for long-read tumor-only somatic variant calling. ClairS-TO offers two pre-trained models for inference: (1) a model trained exclusively on synthetic samples (ClairS-TO SS), and (2) a model initially trained on synthetic samples and further augmented with real samples (ClairS-TO SSRS), as illustrated in Fig. 1b. For the SS model training, synthetic tumors were created by combining variants from two biologically-unrelated individuals, where germline variants unique to one individual were treated as somatic variants to another individual in the mixed synthetic sample. ClairS-TO SSRS was then fine-tuned using the SS model weights with real cancer cell lines, as detailed in the “Methods—Training data preparation” section. Figure 1a provides an overview of the ClairS-TO somatic variant calling workflow. Following the ensembled network prediction, ClairS-TO applies three post-filtering steps: artifact filtering with nine hard-filters, germline variant tagging with four PoNs, and the Verdict module for distinguishing germline and somatic variants. Further details are described in the “Methods—ClairS-TO workflow and design” section.

Fig. 1: Overview of ClairS-TO somatic variant calling and model training workflow.

a The three steps in the somatic variant calling workflow of ClairS-TO. In Step 1, the pileup inputs of variant candidates in the tumor are fed into the affirmative neural network (AFF) and the negational neural network (NEG) to obtain the probabilities of the candidate being a somatic variant and not being a somatic variant, respectively. In Step 2, a joint posterior probability (P(,{y|}({P}_{{\mathrm{AFF}}}\left(y|x\right),{1-P}_{{\mathrm{NEG}}}\left({\neg y}|x\right)))) is calculated using the output probabilities from both networks for inference. (P(y)) is derived from the same samples used for training AFF and NEG. In Step 3, multiple post-processing strategies are applied to tag non-somatic variants. b The training workflow of ClairS-TO. ClairS-TO provides two types of pre-trained models: a model trained exclusively on synthetic tumor samples (ClairS-TO SS) and a model initially trained on synthetic tumor samples and then augmented with real tumor samples (ClairS-TO SSRS). c The data synthesis workflow of ClairS-TO. For example, using 90-fold coverage of GIAB HG002 (Sample A, with GIAB-known truth germline variants as GA) and 80-fold coverage of GIAB HG001 (Sample B, with GIAB-known truth germline variants as GB) from ONT WGS alignments as sources for synthetic tumor and normal, the alignments were split into smaller non-overlapping chunks with an average of 4-fold coverage. These smaller chunks were used to simulate synthetic tumors with varying coverages and different VAFs. For variants in the synthetic tumor (T), “Somatic” refers to GIAB truth germline variants present in A but not in B (T ∩ (GA−GB)); “Germline” refers to GIAB truth germline variants present in both A and B (T ∩ GA ∩ GB); and “Artifact” refers to variants not present in the GIAB truth germline variants of either A or B (T-GA-GB). When using Sample B as the tumor and Sample A as the normal, the definitions remain identical except for switching the subscripts. d The VAF distributions of synthetic somatic SNVs at simulated tumor purities of 100, 75, 50, and 25%, using either HG002 or HG001 as the tumor source. e The VAF distributions of real somatic SNVs in the four real cancer cell lines (i.e., HCC1937, HCC1954, H1437, and H2009).

Benchmarking datasets

We used two cancer cell lines COLO829 and HCC1395 for benchmarking. The truth somatic variants of COLO829 are provided by NYGC21, with a total of 42,993 SNVs and 985 Indels. The truths of HCC1395 are provided by the SEQC2 consortium22. We only used those labeled as “HighConf” (high confidence) and “MedConf” (medium confidence) in the high-confidence regions, with a total of 39,447 SNVs and 1602 Indels. To reflect the real performance of the callers and remove the false negative (FN) calls that are actually caused by no coverage or no alternative allele support at all in the input, we required a truth to be included for benchmarking if it has (1) coverage (\ge)4; (2) reads supporting an alternative allele (\ge)3; and (3) VAF (\ge)0.05.

Performance analysis on ONT

A summary of the ONT datasets used for model training and benchmarking is provided in Supplementary Table 1a. ClairS-TO SS was trained on synthetic datasets derived from GIAB HG002 and HG001 from the ONT EPI2ME Labs[23](https://www.nature.com/articles/s41467-025-64547-z#ref-CR23 “Nanopore EPI2ME Labs, https://github.com/epi2me-labs/wf-somatic-variation

(2023).“). ClairS-TO SSRS was augmented from the SS model using four real cancer cell lines—HCC1937, HCC1954, H1437, and H2009—from Park et al.[6](https://www.nature.com/articles/s41467-025-64547-z#ref-CR6 “Park, J. et al. DeepSomatic: Accurate somatic small variant discovery for multiple sequencing technologies. Preprint at bioRxiv, https://doi.org/10.1101/2024.08.16.608331

(2024).“). For performance evaluation, two well-established cancer cell lines, COLO829 and HCC1395, were used. Our evaluation involved comparing results with other callers, including Clair324, DeepSomatic[6](https://www.nature.com/articles/s41467-025-64547-z#ref-CR6 “Park, J. et al. DeepSomatic: Accurate somatic small variant discovery for multiple sequencing technologies. Preprint at bioRxiv, https://doi.org/10.1101/2024.08.16.608331

(2024).“), and smrest[16](https://www.nature.com/articles/s41467-025-64547-z#ref-CR16 “Simpson, J. T. Detecting somatic mutations without matched normal samples using long reads. Preprint at bioRxiv, https://doi.org/10.1101/2024.02.26.582089

(2024).“). Since Clair3 is a germline variant caller, we treated variants called by Clair3 as somatic, excluding those tagged by the PoNs used in ClairS-TO. DeepSomatic is a deep-learning-based somatic variant caller that supports tumor-only calling and is trained exclusively on real cancer cell lines. For benchmarking, we used the best-performing “multi-cancer” model, following the authors’ recommended practices. Smrest is a haplotype-resolved statistical method designed for tumor-only somatic SNV calling using long-read data. As reported in its publication, smrest is mainly designed for low tumor-purity data[16](https://www.nature.com/articles/s41467-025-64547-z#ref-CR16 “Simpson, J. T. Detecting somatic mutations without matched normal samples using long reads. Preprint at bioRxiv, https://doi.org/10.1101/2024.02.26.582089

(2024).“). For all benchmarking methods, we followed the authors’ recommended pipeline and parameter settings.

Performance with different sequencing coverages

We evaluated ClairS-TO and other callers across three sequencing coverages: 25-, 50-, and 75-fold, considering that 25-fold coverage typically represents the throughput of one ONT R10.4.1 PromethION flow cell. We incrementally increased the coverage by one flow cell throughput, matching real-world clinical sequencing approaches to enhance variant discovery capacity.

Figure 2a and Supplementary Table 2 showed the AUPRC (area under precision-recall curve) and the best achievable F1-score results of each caller. For the COLO829 dataset, ClairS-TO SSRS achieved 0.6489, 0.6634, and 0.6685 AUPRC for SNV at 25-, 50-, and 75-fold coverage, respectively. The improvement is more pronounced from 25- to 50-fold (+0.0145 AUPRC) compared with 50- to 75-fold (+0.0051 AUPRC). The SSRS model’s consistent outperformance against the SS model demonstrates that incorporating real samples during training enhances the ability to learn variant characteristics inherent in real tumors.

Fig. 2: Performance at different sequencing coverages using ONT COLO829.

a Precision-Recall curves at 25-, 50-, and 75-fold tumor coverages of ONT COLO829. The dot on each dashed line indicates where the best F1-score was achieved. The area under the Precision-Recall curve (AUPRC) is shown in parentheses in the legend. b The performance of the best achievable F1-score, along with the corresponding precision and recall, at 25-, 50-, and 75-fold tumor coverages of ONT COLO829. Source data are provided as a Source data file.

Among other callers, Clair3 achieved an average AUPRC value of 0.5605 at the three coverages, underscoring the limitations of using a germline variant caller for somatic variant calling. DeepSomatic achieved AUPRC values 0.0459, 0.0300, and 0.0320 lower than those of ClairS-TO SSRS, respectively. In terms of the best achievable F1-score, as shown in Fig. 2b, ClairS-TO SSRS surpasses DeepSomatic by 6.38, 4.38, and 4.91% at the three coverages, respectively. This improvement highlights the advantage of augmenting training with both synthetic and real tumor samples.

For the HCC1395 dataset, DeepSomatic was excluded from benchmarking because HCC1395 was part of its training data. The rationale for ClairS-TO’s decision to exclude HCC1395 entirely, and how this decision benefits the community by enabling more comprehensive performance comparisons between algorithms and tools in the long term, is explained in the “Methods—Training data preparation” section. As shown in Fig. 3a, b, ClairS-TO SSRS performed the best, achieving an average AUPRC of 0.5929 and an F1-score of 78.69% across the three coverages. The results across different datasets demonstrate that ClairS-TO consistently outperforms other callers under various coverage settings. Notably, ClairS-TO achieved ~92% recall across different coverages, comparable to ClairS, a tumor-normal-pair caller, making ClairS-TO a reliable option for identifying somatic variants of interest with clinical sequencing panels.

Fig. 3: Performance at different sequencing coverages using ONT HCC1395.

a Precision-Recall curves at 25-, 50-, and 75-fold tumor coverages of ONT HCC1395. The dot on each dashed line indicates where the best F1-score was achieved. The area under the Precision-Recall curve (AUPRC) is shown in parentheses in the legend. DeepSomatic was excluded from benchmarking with the HCC1395 WGS dataset because the sample was included in its model training. b The performance of the best achievable F1-score, along with the corresponding precision and recall, at 25-, 50-, and 75-fold tumor coverages of ONT HCC1395. Source data are provided as a Source data file.

Performance at different VAF ranges

We also benchmarked ClairS-TO across different VAF ranges using 50-fold coverage of COLO829, as shown in Fig. 4 and Supplementary Table 3. In the low VAF range (0.05–0.2), low-mid VAF ranges (0.2–0.3, 0.3–0.4, and 0.4–0.5), and mid-high VAF range (0.5–1.0), ClairS-TO SSRS achieved F1-scores of 32.85, 75.35, 79.66, 80.90, and 77.60%. Across all VAF ranges, ClairS-TO consistently outperformed Clair3 and DeepSomatic, as detailed in Supplementary Table 3. Notably, ClairS-TO performed similarly in low-to-mid and mid-to-high VAF ranges. While higher VAFs typically yield better results, precision dropped in the mid-to-high range due to misclassified germline variants.

Fig. 4: Performance in different VAF ranges using ONT COLO829.

Performance on 50-fold coverage of ONT COLO829 across different VAF ranges, with VAF divided into five ranges: 0.05–0.2, 0.2–0.3, 0.3–0.4, 0.4–0.5, and 0.5–1.0. Source data are provided as a Source data file.

Performance at different tumor purities

In real-world clinical scenarios, tumor samples are often impure and sequenced with varying tumor purities. We evaluated the performance of ClairS-TO at different tumor purities—1.0, 0.8, 0.6, 0.4, and 0.2—by mixing normal samples into tumors in silico. The results for 50-fold coverage of COLO829 are shown in Fig. 5a and Supplementary Table 4a. ClairS-TO SSRS achieved AUPRC values from 0.6634 to 0.4797 across the five tumor purities, outperforming DeepSomatic by 0.0300 to 0.1919. ClairS-TO SSRS also outperformed smrest by 0.4419 and 0.2929 in 0.4 and 0.2 tumor purities, respectively. As shown in Fig. 5b, ClairS-TO SSRS achieved F1-scores from 76.83% to 66.99% across the five tumor purities, respectively. The performance decline was most pronounced when tumor purity dropped from 0.4 to 0.2, primarily due to a reduction in recall.

Fig. 5: Performance at different tumor purities, genomic contexts, and program settings.

a Precision-Recall curves on 50-fold coverage of ONT COLO829 at five tumor purities: 1.0, 0.8, 0.6, 0.4, and 0.2. The dot on each dashed line indicates where the best F1-score was achieved. AUPRC is shown in parentheses in the legend. b The best achievable F1-score on 50-fold coverage of ONT COLO829 at five tumor purities: 1.0, 0.8, 0.6, 0.4, and 0.2. c Performance with or without the Verdict module. Notably, Verdict significantly improved performance, particularly when tumor purity is low, and is recommended to be enabled when the estimated purity is below 0.6. d The best achievable F1-score on 50-fold coverage of ONT COLO829 across different challenging genomic contexts, categorized into five groups: “Low complexity”, “Segmental duplications”, “Low mappability”, “Functional regions”, and “Other difficult regions”, as defined by GIAB Stratifications v3.3. e The best achievable F1-score on 50-fold coverage of ONT COLO829 using different PoNs or combinations for filtering. Source data are provided as a Source data file.

As shown in Fig. 5c and Supplementary Table 4b, ClairS-TO SSRS achieved F1-scores of 75.14, 74.95, and 66.99% at tumor purities of 0.6, 0.4, and 0.2, respectively, when using the Verdict module. Disabling Verdict resulted in F1-score decreases by 1.99, 4.38, and 7.81%. Notably, when tumor purity exceeds 0.6, the improvements are negligible. The results indicate that the Verdict module significantly reduces the misclassification of germline variants, particularly in tumor samples with low purity, even without a matched normal sample as a reference.

Performance across various genomic contexts

We benchmarked ClairS-TO using 50-fold coverage of COLO829 and categorized challenging genomic contexts into five categories: “Low complexity”, “Segmental duplications”, “Low mappability”, “Functional regions”, and “Other difficult regions”, as defined by GIAB Stratifications v3.325. The best achievable F1-scores are shown in Fig. 5d and Supplementary Table 5. ClairS-TO consistently outperformed DeepSomatic across all genomic contexts. Although ClairS-TO outperformed DeepSomatic by 11.11% and 13.69% in complex homopolymer and tandem repeat regions, the best F1-scores achieved by ClairS-TO SSRS were only 65.63% and 69.09%, respectively. The performance remains substantially below WGS benchmarks (67.36% versus 76.83%), highlighting the persistent challenges in somatic variant detection and the critical need for enhanced performance in complex genomic regions.

Benchmarking different network and training configurations

We evaluated the performance of various network configurations on 50-fold coverage of COLO829 using: (1) the default ensemble (AFF & NEG) in ClairS-TO, (2) the AFF network only (AFF only), and (3) the NEG network only (NEG only). As shown in Supplementary Fig. 1b and Supplementary Table 6b, the ClairS-TO SSRS model achieved an AUPRC of 0.6634 with the AFF & NEG ensemble, while AFF only achieved 0.6363 AUPRC, and NEG only achieved 0.6562 AUPRC. These results highlight the benefits of network ensembling for enhanced performance.

We also compared the performance of three ClairS-TO models: ClairS-TO SS, ClairS-TO RS, and ClairS-TO SSRS, using 50-fold coverage of COLO829. Specifically, ClairS-TO RS was trained from scratch using four real cancer cell lines (HCC1937, HCC1954, H1437, and H2009). As shown in Supplementary Fig. 1a and Supplementary Table 6a, ClairS-TO SS, ClairS-TO RS, and ClairS-TO SSRS achieved AUPRC values of 0.6176, 0.6582, and 0.6634, respectively. The superior performance of ClairS-TO SSRS demonstrates the importance of combining synthetic samples and real tumors for augmented training.

Analysis of introducing different PoNs for germline tagging

ClairS-TO utilizes four PoNs—gnomAD26, dbSNP27, 1000 G PoN28, and CoLoRSdb[29](https://www.nature.com/articles/s41467-025-64547-z#ref-CR29 “Lake, J. A. Consortium of Long Read Sequencing database (CoLoRSdb). Zenodo, https://doi.org/10.5281/zenodo.13145123

(2024).“)—to tag non-somatic variants. As shown in Fig. 5e and Supplementary Table 6c, using 50-fold coverage of COLO829, ClairS-TO SSRS achieved F1-scores of 2.31, 39.56, 44.80, 3.87, 71.19, and 76.49% when using no PoNs, gnomAD only, dbSNP only, 1000 G PoN only, CoLoRSdb only, and all four PoNs, respectively. Precision improved by 62.86% when transitioning from no PoNs to using all four PoNs, with a total of ~3.37 million germline variants tagged. These results underscore the benefits of incorporating multiple PoNs, particularly CoLoRSdb, a PoN specifically designed for long-read data[29](https://www.nature.com/articles/s41467-025-64547-z#ref-CR29 “Lake, J. A. Consortium of Long Read Sequencing database (CoLoRSdb). Zenodo, https://doi.org/10.5281/zenodo.13145123

(2024).“). While the use of PoNs effectively identifies common germline variants, it may also inadvertently remove clonal hematopoiesis mutations if they are present in the PoNs, which is a limitation shared by most tumor-only methods relying on PoNs filtering.

Performance evaluation of different hard-filters

We further assessed the impact of nine hard-filters, the details of which are described in the “Methods—ClairS-TO workflow and design” section. Using 50-fold coverage of COLO829, the filters “MultiHap”, “NoAncestry”, “LowAltBQ”, “LowAltMQ”, “VariantCluster”, “ReadStartEnd”, and “StrandBias” removed 125, 26, 30, 1, 6541, 87, and 237 non-somatic variants, respectively. With these hard-filters applied, precision improved by 1.85%, demonstrating their effectiveness in eliminating germline variants and sequencing artifacts.

Performance analysis on PacBio

To validate the effectiveness of our training workflow and proposed post-filters, we also trained models on PacBio data, with a summary of the datasets used provided in Supplementary Table 1a. The ClairS-TO SS PacBio model was trained on a synthetic dataset derived from GIAB HG003 and HG004, sequenced on the PacBio Revio system. The ClairS-TO SSRS model was augmented from the SS model using four real cancer cell lines (HCC1937, HCC1954, H1437, and H2009) provided by Park et al.[6](https://www.nature.com/articles/s41467-025-64547-z#ref-CR6 “Park, J. et al. DeepSomatic: Accurate somatic small variant discovery for multiple sequencing technologies. Preprint at bioRxiv, https://doi.org/10.1101/2024.08.16.608331

(2024).“). For benchmarking, we used 50-fold coverage of the PacBio COLO829. As shown in Fig. 6a and Supplementary Table 7a, ClairS-TO SSRS performed the best, achieving a 0.6667 AUPRC and a 78.64% F1-score, and outperformed DeepSomatic by 3.99% in F1-score.

Fig. 6: Performance on PacBio and Illumina data.

a The Precision-Recall curves and the best achievable F1-score of SNV and Indel on coverage PacBio Revio COLO829 using ClairS-TO SS, ClairS-TO SSRS, Clair3, and DeepSomatic. The dot on each dashed line represents the best F1-score achieved. AUPRC is shown in parentheses in the legend. b The Precision-Recall curves and the best achievable F1-score of SNV and Indel on 50-fold coverage of Illumina COLO829 using ClairS-TO SS, ClairS-TO SSRS, DeepSomatic, Mutect2, Octopus, and Pisces. The dot on each dashed line represents the best F1-score achieved. AUPRC is shown in parentheses in the legend. Source data are provided as a Source data file.

Performance analysis on Illumina

Short-read tumor-only somatic variant calling has been extensively studied7,8,9,10,11,[12](#ref-CR12 “Benjamin, D. et al. Calling somatic SNVs and indels with Mutect2. Preprint at bioRxiv, https://doi.org/10.1101/861054

(2019).“),13,14,30,31,32, with numerous notable methods developed, such as Mutect2[12](https://www.nature.com/articles/s41467-025-64547-z#ref-CR12 “Benjamin, D. et al. Calling somatic SNVs and indels with Mutect2. Preprint at bioRxiv, https://doi.org/10.1101/861054

(2019).“), Octopus13, and Pisces14. ClairS-TO was primarily designed for long-read sequencing data but can be adapted to short-read data by excluding long-read-specific post-filters, such as long-read PoNs, variant phasing, and the “MultiHap” and “NoAncestry” filters. For training, we used GIAB HG003 and HG004 for the SS model and augmented the SSRS model with four real cancer cell lines (HCC1937, HCC1954, H1437, and H2009). All datasets were sequenced on Illumina’s latest HiSeq X or NovaSeq 6000 platforms, with more details provided in Supplementary Table 1a.

Figure 6b and Supplementary Table 7b showed the results of ClairS-TO compared to other short-read tumor-only somatic variant callers. ClairS-TO SSRS achieved AUPRC values of 0.6500 and F1-scores of 76.99% at 50-fold coverage of COLO829, surpassing DeepSomatic, Mutect2, Octopus, and Pisces by 15.10, 18.83, 36.69, and 54.59% in F1-score. These results demonstrate that ClairS-TO achieves competitive performance on short-read data and could serve as a supplementary tool for ensemble analysis in short-read workflows.

Indel performance analysis

Indel variant detection is more challenging compared to SNVs, primarily due to two reasons: (1) Long-read technologies exhibit higher error-rate in complex regions, such as homopolymer and tandem repeat regions, leading to substantial Indel artifacts that hinder the detection of true somatic Indels with low VAF; and (2) The number of somatic Indels is inherently limited. For instance, only around 2000 somatic Indels are reported in HCC1395 or COLO829 cell lines, resulting in an insufficient number of Indels from real samples for comprehensive model training and benchmarking currently. While tumor-only somatic Indel detection performance remains limited, we provide the following results as a baseline for future improvements.

As shown in Figs. 2a, 6a, b, ClairS-TO SSRS achieved AUPRC values of 0.2019, 0.1972, and 0.2334 on ONT, PacBio, and Illumina 50-fold coverage of the COLO829 dataset, respectively. ClairS-TO consistently outperformed DeepSomatic and other short-read callers, as detailed in Supplementary Table 2, with similar trends observed in benchmarking at different coverages or using the HCC1395 cell line. Despite ClairS-TO performing the best among the evaluated tools, recall remains capped at ~50%, highlighting the need for further improvements to achieve reliable somatic Indel detection.

Regarding the performance at different Indel lengths and in different sequence contexts, as shown in Supplementary Fig. 2b and Supplementary Table 8a, ClairS-TO has demonstrated better accuracy for longer Indels ((\ge !{5,{\rm{bp}}})) compared to shorter Indels (1–4 bp), with an overall F1-score 35.35% versus 26.80%. This is primarily due to the sequencing artifacts that predominantly affect 1–3 bp gaps, making shorter somatic Indels more challenging to detect. Moreover, we also analyzed the Indel performance in different challenging genomic regions. As shown in Supplementary Fig. 2a and Supplementary Table 5, ClairS-TO’s Indel recall rises to ~80% when excluding all difficult regions, indicating that the difficult regions are the major source of Indel errors. We further analyzed the distribution of Indel FPs, as shown in Supplementary Table 8b, the majority are high-confidence germline Indels, and excluding them significantly improves precision to ~90% in non-difficult regions.

Analysis of false positive and false negative calls

We manually analyzed 300 false positive (FP) calls and 300 FN calls randomly selected from 50-fold coverage of ONT COLO829. Each FP and FN was assigned with a primary and a secondary reason, as shown in Fig. 7 and Supplementary Data 1.

Fig. 7: Categorizing the FPs and FNs.

Pie charts illustrate the distributions of primary reasons for FPs and FNs in ClairS-TO. A total of 300 FPs and 300 FNs were randomly selected from the calling results on 50-fold coverage of ONT COLO829. Regarding FPs, 61% were heterozygous germline variants, and 14% were homozygous germline variants. The distribution of the remaining FPs (excluding germline variants) is also shown. A major category among these FPs is caused by alignment artifacts in complex genomic contexts, including 18% in segmental duplication regions, 12% in tandem repeat regions, 7% in low-mappability regions, and 5% in homopolymer regions, as defined by GIAB Stratifications v3.3. Regarding FNs, 35% were “Filtered by PoNs” and 18% were “Filtered by VariantCluster”, both resulting from post-processing steps. A major category among the remaining FNs (excluding those incorrectly filtered) is also located in complex regions, including 21% in tandem repeat regions, 21% with one or more deletions in the flanking 50 bp, 14% with excessive mismatches in alignment, 10% with one or more insertions in the flanking 50 bp, 10% in segmental duplication regions, 9% in homopolymer regions, and 7% in low-mappability regions. Source data are provided as a Source data file.

Regarding FPs, 61% were heterozygous germline variants, and 14% were homozygous germline variants called as germline in Clair3 but as somatic in ClairS-TO. Although most germline variants were tagged by PoNs, the remaining individual germline variants accounted for the majority of FPs. For example, a germline variant with a VAF of 0.6167 at chr3:145,978,865 was called as somatic with a QUAL score of 74.86 in ClairS-TO but as a heterozygous germline with a QUAL score of 28.77 in Clair3. This FP was called due to that ClairS-TO has insufficient evidence to classify it as a germline variant. Among the remaining FPs, 17% and 12% were not reported in the COLO829 truth set, with tumor VAF in the ranges of 0.05 (\le) VAF ( < ) 0.1 and 0.1 (\le) VAF ( < ) 0.15, respectively. These variants, some of which are called only with long-read data, might have a chance to be correct and be added to the truth set in the future with additional sequencing coverage and orthogonal validations. Another major category of FPs is caused by alignment artifacts in complex genomic contexts, including 18% in segmental duplication regions, 12% in tandem repeat regions, 7% in low-mappability regions, and 5% in homopolymer regions. These issues might be mitigated as sequencing errors in complex regions are reduced.

Regarding FNs, 35% were filtered by PoNs, and 18% were filtered by the “VariantCluster” hard-filter. While these post-filters result in the loss of some true variants, they provide a more substantial improvement in precision. Among the remaining FNs, 8% of the truth variants were missed due to low tumor VAF (VAF ( < ) 0.1) or insufficient supporting reads (variant allele depth ( < ) 3), which could potentially be detected at a higher coverage. Other categories of FNs were identified in complex genomic regions, including 21% in tandem repeat regions, 21% with one or more deletions in the flanking 50 bp, 14% with excessive mismatches in alignment, 10% with one or more insertions in the flanking 50 bp, 10% in segmental duplication regions, 9% in homopolymer regions, and 7% in low-mappability regions.

Model interpretability analysis

ClairS-TO uses the ensemble of two networks (AFF and NEG, more details are given in the “Methods” section) for a more comprehensive variant calling performance. As shown in Supplementary Fig. 5a, b, we randomly selected two cases to demonstrate the effectiveness of ClairS-TO’s model ensembling. In the ONT COLO829 dataset, for the C->A variant at chr21:10,345,275, the AFF network favored the reference base (58.07% vs 43.30% confidence), while the NEG network rejects the reference with 65.65% confidence. Similarly, for another T->A variant at chr12:61,056,821, the AFF network predicts the variant correctly (62.10% vs 42.43%) despite NEG’s 58.39% rejection confidence. In both cases, Bayesian integration of these complementary network prediction yielded correct predictions (58.32% and 54.61% confidence, respectively). These two cases demonstrate the advantages of ensembling AFF & NEG networks over single-network prediction.

Moreover, to visualize feature importance patterns in our networks, we performed the saliency-map-alike integrated gradients (IG) analysis33, a method to compute input feature contributions by analyzing gradients of model outputs relative to input tensor. As visualized in Supplementary Fig. 6a, b, the IG features reveal that the AFF and NEG networks have learned to focus on different positions and features in the inputs. According to the IG feature maps, the central genomic position contributes most significantly to the model outputs, especially for the A/C/G/T features in forward and reverse strands. The flanking bases also demonstrate measurable influence on prediction confidence.

Discussion

In this study, we present ClairS-TO, a deep-learning-based method for long-read tumor-only somatic variant calling. Our experiments have shown that ClairS-TO is the best-performing caller on ONT and PacBio long-read data. When trained with short-read sequencing data, ClairS-TO has also shown an edge over well-established short-read tumor-only somatic variant callers. More experiments at multiple coverages, tumor purities, and VAF ranges show that ClairS-TO has maintained the lead. Ablation studies have demonstrated the effectiveness of each component and the solution as a whole.

ClairS-TO uses synthetic tumor samples for model training and can use any number of real tumor samples with verified somatic variant truths for fine-tuning. The practically unlimited number of somatic variants that can be generated from synthesizing tumor samples provide a foundation of using deep-learning for somatic variant calling before more real tumor samples of different cancer types with verified truths are available, considering (1) a model trained only on real tumor samples might risk of overfitting when applied to a cancer type that was not included in training—it is impractical to expect for a model training only on real tumor samples to have a full coverage of all cancer types, and (2) the number of somatic variants in a real sample only ranged from 10 to 100k, which is two orders of magnitude fewer than the germline variants a real sample could offer for model training—it takes more efforts and longer to reach the same number of true variants for training as what we had for deep-learning-based germline variant calling if real tumor sample is the only source. Without a paired normal sample for reference, a tumor-only somatic variant caller is required to not only be capable of extracting variant signals, but also telling somatic variants apart from germline variants and background noise. ClairS-TO uses an ensemble of two neural networks of completely different ar