
In a clinical genetics lab, a technician is staring at a screen, scanning fresh sequencing data—tens of millions of reads, a line of colored peaks representing one of the four DNA nucleotides—searching for a faint somatic variant hidden in the heterogeneous mess that is a cancer genome. Occasionally, there’s a small bump overlapping the peak, indicating the detection of a different nucleotide in that same position. The lab tech squints, trying to assess if that little bump is indeed a real mutation or just a sequencing artifact. If incorrect, the answer could influence a child’s diagnosis and treatment, which could mean the difference between life and death.
That single uncertainty, distinguishing a true mutation from a technical artifact amid the chaos of cancer genomes, defines much of somatic variant calling. It’s a deceptively simple goal—find the mutations that occur only in tumor cells—but one that’s proven to be remarkably elusive, particularly for long-read technologies that produce different types of errors than traditional short-read methods.
Until recently, even the best tools struggled with that distinction, especially when it comes to long-read sequencing data. But a collaboration between researchers at the UC Santa Cruz (UCSC) Genomics Institute, Google Research, and Children’s Mercy Hospital may have finally given those clinical genomics labs a way to confidently see through the noise. Their new AI-powered algorithm, DeepSomatic, described in a Nature Biotechnology research paper, pushes somatic variant detection to new levels of precision and robustness.
Benedict Paten, who directs UCSC’s Computational Genomics group, told Inside Precision Medicine, “When we observe recurrent mutations in specific genes that are known to be frequently mutated, DeepSomatic gives us significant confidence that our findings are valid rather than coincidental. Certain mutations proliferate and undergo micro-evolutionary selection in patients and in all of us. So, we see them popping up over and over again. Without some oracle, you can never be a 100% percent sure there isn’t some freak statistical oddity that’s going on. But the proof is in the pudding, and we have substantial evidence to support this conclusion.”
Solving the somatic variant problem
Since his time as a graduate student in Paten’s lab, the paper’s senior author and research scientist working on Health AI at Google Research, Keshwar Shafin, PhD, has been thinking about how to distinguish germline variants from “truth sets”—genomes that have been meticulously curated to serve as references. Trained in both computer science and biomolecular engineering in Paten’s lab at UCSC, Shafin helped develop the variant calling tool DeepVariant and the long-read assembler Shasta during his doctoral training.
But how to pin down real low-frequency somatic variants, which by definition are not present in every cell, is a whole different problem. Compared to the straightforwardness of germline genomes, cancer genomes with enormous heterogeneity of variants are as clear and interpretable as a Jackson Pollock painting. “If you sequence our genome, about 50% of the reads should match a germline variant,” Shafin told Inside Precision Medicine. “But in cancer, many somatic variants have such low allele frequencies that your model must learn the subtle differences between what’s a true mutation and what’s noise.”
To complicate matters further, tumor purity varies greatly between samples, and subclonal variants and clonal heterogeneity make interpretation more difficult. In hematologic malignancies, for example, the neoplastic fraction can range from less than 20% to nearly 100% of nucleated cells, so sensitivity thresholds must be adjusted. The problem, Shafin explained, is that somatic variant calling has not had access to truth sets like those created by the National Institute of Standards and Technology (NIST) used to train machine learning algorithms to call germline variants. Previously published somatic benchmarking efforts were limited in scale or used only short reads. That scarcity makes training and validating neural networks difficult: big models need large, high-quality truth sets, and simulated data can’t replicate the full complexity of real sequencing error distributions.
To address that gap, the UCSC–Google team designed DeepSomatic to learn from experimental reality. To do so, Jimin Park, a graduate student in Paten’s lab and co-first author of the Nature Biotechnology paper, built training sets from six previously characterized tumor–normal cell line pairs that were then sequenced each across three platforms, using Illumina for short-reads typically between 50 and 300 base pairs long and PacBio HiFi and Oxford Nanopore Technologies (ONT) for long reads, which can range from thousands to over a million base pairs in a single read.
“A big part of this project is the fact that we trained this tool with cell line data instead of simulated data, which is what sets it apart from a lot of the existing tools,” Park said to Inside Precision Medicine. “We incorporated long-read technologies as well, developing training sets from those six cell lines and looking at the intersection between different sequencing technologies. There were many iterations back and forth between us and the Google team, sharing results and refining how the current models worked.”
The tri-platform strategy served two purposes. First, long reads expose regions of the genome that short reads cannot resolve—repeats, paralogues, and complex structural contexts. Secondly, the distinct error modes of various platforms serve as a useful consensus filter: when a candidate variant emerges in ONT, PacBio, and Illumina data for the same sample, the likelihood of it being a coincidental error across all three instruments is significantly reduced. “If three platforms all agree that this is a somatic variant at this site, then we have very high confidence that it’s real,” Shafin explained. That cross-validation enabled the team to assemble a high-fidelity somatic “truth set” that could teach a neural network to recognize real low-frequency signals.
Toward all-in-one clinical testing
The UCSC and Google Research team’s benchmarking work is as important as the model itself. Shafin and Park were integral to assembling a reproducible framework—open datasets, multi-platform comparisons, and explicit performance metrics—that other groups can use to test their callers. That standardization matters in translation from bench to clinic. “From a clinical perspective, we want benchmarks,” said Lisa Lansdon, PhD, assistant director of the Molecular Genetics Laboratory at Children’s Mercy Hospital. “We want to be sure we’re calling things robustly.”
Today, many institutions use a patchwork approach: short reads for single-nucleotide variants and small indels, targeted assays or cytogenetics for structural rearrangements, and orthogonal tests when key calls are ambiguous. That multiplicity adds cost, time, and logistical complexity.
Instead of a patchwork of tools tuned for particular platforms, labs may increasingly adopt AI components that handle basecalling, error correction, variant detection, phasing, and even annotation in coordinated, learnable ways. “AI isn’t just about calling variants,” Shafin said. “It’s about building models that understand sequencing error, sample complexity, and biological context. That’s how we move from research to clinical impact.”
“If long reads can give us both small-variant precision and structural-variant resolution in one assay, that has real operational and clinical value,” Lansdon said. She notes that Children’s Mercy has already used long-read sequencing for non-oncology cases; the question has been whether long reads can be trusted for cancer diagnostics, where the stakes and sample complexity are higher. “This study suggests we’re getting there,” she said, but she stresses that broader validation is required.
Lansdon also highlights operational considerations that go beyond raw accuracy: throughput, turnaround time, sample type, and sample preservation. Tumor specimens often come as FFPE (formalin-fixed, paraffin-embedded) blocks, which introduce artifacts and degrade DNA. The team explicitly trained and tested models to generalize across preservation protocols and showed that multi-platform training helps the model tolerate those real-world variances. “That upstream filtering—figuring out what calls exist at all—is what DeepSomatic accomplished,” Lansdon said. “Clinical interpretation remains our job: applying gene filters, mapping findings to diagnosis and prognosis, and integrating them into the medical record.”
To test the clinical applicability of DeepSomatic, Lansdon’s lab contributed eight pediatric tumor samples from the hospital’s research biorepository, enabling an independent, clinically oriented validation of DeepSomatic’s outputs. “What this study did was say, ‘Here’s proof of principle,’” Lansdon said. “Across these cell lines, across these technologies, you can now be very certain about what to expect if your caller performs well. That’s a big step for clinical confidence.”
The Children’s Mercy team looked for two things: did DeepSomatic recover the variants that the hospital’s clinical workflows had already identified, and did it miss anything important? The answer was reassuring.
A crucial clinical example came in acute myeloid leukemia. Dual mutations in CEBPA can define a specific AML subtype and directly affect prognosis and treatment choice. DeepSomatic replicated clinically relevant calls—including complex small variants in tumor CEBPA—that some callers miss and flagged no unexpected false negatives in the tested cohort. “If the caller misses it, we’re blind,” Lansdon said. “Being able to call those mutations confidently changes what we do for the patient.”
Importantly, DeepSomatic’s output is meant to feed into existing clinical workflows rather than replace them. Variant annotation tools, such as ClinVar, and other knowledge bases remain essential to determine whether an identified variant is truly actionable or clinically reportable. Operationally, labs will need to consider instrument capacity, data storage, and bioinformatics support. Long reads produce large volumes of data; the computational footprint to run DeepSomatic at scale is nontrivial. Lansdon’s team is working through these questions as they plan expanded clinical validation.
Additionally, there are guardrails. Lansdon stresses transparency and validation. “We can’t deploy a black box,” Lansdon said. For clinical use, models must be auditable, interpretable to a degree, and validated across diverse cohorts. Regulatory pathways will demand clinical-grade evidence of sensitivity, specificity, reproducibility, and equivalence or superiority to existing tests.
At Children’s Mercy, the next step is deliberate scale-up. Lansdon’s group plans to sequence additional pediatric cancer samples—tumor and matched germline where possible—to validate DeepSomatic across a wider range of tumor purities, histologies, and preservation states. “If the model continues to perform as it has on these early cohorts, we’ll consider rolling it into a clinical test,” Langsdon said. “But scaling is the only way to know if it’s robust across the messy reality of clinical practice.”
Moving beyond cancer
On the research side, Paten, Shafin, and Park emphasize continued dataset expansion—more tumor types, more sample conditions, and greater representation of rare variant classes. The team also wants to encourage community adoption of the benchmark so other groups can test and improve alternative callers. Open datasets and cross-lab comparisons are essential to build trust.
Although the primary focus is cancer, the approach has broader translational potential. Mosaic variants—mutations present in only some tissues or cell populations—play a role in developmental disorders and have been implicated in neurodegeneration. Detecting low-frequency, tissue-specific mutations could help solve “missing heritability” problems and inform new diagnostic categories. “It wasn’t done in this study,” Lansdon noted, “but theoretically DeepSomatic could be used to pick up mosaic variants that are meaningful but often missed by existing tools.”
Benedict Paten sees the same conceptual arc: mapping somatic variation across tissues has implications for understanding, diagnosing, and treating countless conditions. “You are a tree of cells, with every cell in your body tracing back to that original fertilized egg,” said Paten. “Over time, these mutations contribute to the formation of the body, and as a result, we are discovering various mutational signatures in different tissues.”
Paten continued, “For example, UV radiation heavily affects skin, while environmental toxins influence bronchial epithelium and lungs. Different mutagens are affecting our body and causing disruptions to DNA. Understanding how that works is crucial for determining whether something is normal or not. This work will help us move towards a greater understanding of how these factors play into a broader context. Jimin has already begun exploring applications in brain tissue to study neurodegeneration and its potential links to somatic mutation.”
The DeepSomatic project doesn’t promise instant clinical revolution; it offers an incremental, pragmatic pathway. It replaces hand-tuned heuristics with learned models grounded in multi-platform experimental truth sets. It reduces the ambiguity that costs time and money in the clinic. It suggests that the long-read promise—resolving structural complexity and offering comprehensive variant profiles—may soon be practical for oncology and beyond. For clinicians, patients, and the tech sitting in a clinical genomics lab, DeepSomatic provides a path to see what was once invisible—and act on it.



