Pansoma, a machine learning tool for identifying somatic variants using pangenome graphs

Breaking the Linear Bias in Cancer Mutation Detection

Identifying somatic mutations—the genetic changes acquired during a lifetime that drive diseases like cancer—is a cornerstone of precision oncology. Traditionally, this is done by comparing a tumor sample to a "normal" healthy sample from the same patient. This process filters out inherited traits. However, in many clinical scenarios, a matched healthy sample is unavailable due to cost, ethics, or sample contamination. This forces researchers to rely on "tumor-only" analysis. This is a much harder task. The computer must guess which mutations are unique to the cancer and which are simply part of the patient's natural genetic diversity.

Current methods struggle because they rely on a single, linear reference genome. This is a one-dimensional string of DNA that acts as a universal template. This template fails to capture the immense complexity of human genetic variation. This leads to "reference bias," where the computer misinterprets natural diversity as a mutation. A new study introduces Pansoma, an AI tool that moves away from this linear constraint by using a pangenome graph. Instead of a single line, the pangenome is a complex web of possible DNA paths. This web incorporates diverse genetic variations from across the human population. By learning to recognize mutation patterns directly on this graph, Pansoma aims to distinguish true cancer drivers from the background noise of human diversity with much higher precision.

The failure of the single-string reference

Most existing somatic variant callers operate on the assumption that there is one "correct" version of the human genome. When sequencing data does not match this linear reference, the software assumes a mutation has occurred. In tumor-only analysis, this creates a massive problem. If a patient has a natural, inherited variation that is not in the reference, the tool might flag it as a cancer mutation. This leads to a high rate of false positives, which are incorrect detections that clutter clinical reports with irrelevant data.

The authors note that while advanced deep-learning models like DeepSomatic have attempted to improve this, they remain tethered to the linear reference. Furthermore, traditional "gold standard" workflows rely on comparing tumor DNA to matched normal tissue to subtract the germline (inherited) background. Without that subtraction step, the "noise" of inherited variation becomes overwhelming. As shown in the study's investigation of the pangenome's utility [Figure 2e], a significant portion of the variants found in healthy individuals are already encoded as alternative paths within a pangenome graph. Current linear tools simply cannot "see" these paths. They are forced to treat natural variation as errors or novel mutations.

From linear strings to graph-native tensors

Pansoma solves this by treating the genome as a navigation problem on a graph rather than a search on a string. The authors implement a three-stage pipeline designed to be "pangenome-native" [Figure 1a].

First, the tool performs graph-based data preparation. Sequencing reads are aligned to the pangenome graph using a graph-aware mapper. This mapper identifies which specific path through the web the DNA fragment belongs to. To make this data usable for machine learning, the authors developed a new "node pileup" (NPU) format. Think of a node as a specific intersection in a city's road network. The NPU format allows the software to instantly look up every "car" (sequencing read) passing through that specific intersection. This avoids scanning the entire map.

Second, the tool converts these intersections into "image-like" tensors [Figure 1b]. A tensor is a multidimensional array of numbers. Instead of looking at a simple list of letters, the convolutional neural network (CNN) sees a multi-channel 3D block of data. Each "pixel" in this block encodes several pieces of information. It includes the DNA base, the confidence of that base call (quality), the certainty of the alignment (mapping quality), and the specific type of alignment event (CIGAR operations). This allows the model to learn the visual "texture" of a true mutation versus a sequencing error or a natural graph path.

Finally, the model reconstructs the findings. Most medical software still expects linear coordinates. Therefore, Pansoma provides a dual-output system [Figure 1d]. It projects variants back onto the standard linear reference for compatibility. However, it also retains "graph-reference" records. This ensures that mutations occurring on unique, non-reference paths are preserved. These mutations would otherwise be invisible to older tools.

Leading precision in tumor-only settings

The researchers benchmarked Pansoma against industry standards like Mutect2, ClairS-TO, and DeepSomatic-TO. They used highly characterized datasets, including the SMaHT COLO829 melanoma line and the GIAB HG008 pancreatic tumor sample.

The results show that Pansoma achieves leading precision and F1-scores in tumor-only somatic variant calling .

Precision refers to the accuracy of the positive calls made. In the COLO829T benchmark, the authors report that Pansoma maintained strong performance across a wide range of tumor purity levels, from 25% to 100% [Figure 3a]. A critical driver of this success was the reduction of the "background burden." By using the pangenome, the tool significantly reduced the number of false candidate sites that the AI had to process. For instance, on Illumina data, graph-based candidate generation reduced the number of suspected Single Nucleotide Variant (SNV) sites to just 63.8% of the count produced by linear-reference methods [Supplementary Figure 5]. This reduction means the AI spends less time analyzing false alarms.

The impact of the graph structure was even more pronounced in long-read sequencing data. The authors report that graph alignment increased the number of perfectly aligned reads by more than four-fold compared to linear alignment. This capability allows Pansoma to excel across different sequencing platforms, including Illumina, PacBio, and Oxford Nanopore (ONT) [Figure 3b].

Navigating the graph-based frontier

Despite these gains, the transition to pangenomics introduces new complexities. One major limitation is the difficulty of benchmarking. Almost all "truth sets" (gold-standard lists of known mutations) are defined using linear coordinates. This makes it technically challenging to verify the accuracy of the variants Pansoma finds on non-reference graph paths. While the authors argue these variants carry valid biological signals, such as UV-associated mutational signatures in melanoma [Figure 4c], they exist in a scientific "blind spot." Current validation tools struggle to reach them.

Furthermore, the tool's performance is inherently tied to the quality of the pangenome itself. If a patient's unique genetic background is not represented in the underlying graph, the advantages of pangenome-native alignment diminish. The authors also acknowledge that Indel (insertion/deletion) calling remains a difficult, platform-dependent task. This is especially true with ONT long-read data. Finally, while Pansoma bridges the gap by projecting results back to linear coordinates, this projection can cause a loss of information. This might complicate clinical interpretation in existing diagnostic pipelines.

The verdict: A necessary shift for precision oncology

Pansoma is a highly effective step toward a post-linear era of genomics. By moving the heavy lifting of variant discrimination to the initial alignment stage, the authors have successfully lowered the "noise floor" of tumor-only sequencing. The ability to achieve high precision without relying heavily on external "panel-of-normals" filters [Figure 4d] suggests that pangenome-native alignment is a robust way to handle human diversity.

For practitioners looking to implement this, the tool is ready for research environments. The authors have made the code and pretrained models available via GitHub and Zenodo. However, for widespread clinical deployment, the ecosystem must catch up. Until downstream annotation tools and clinical truth sets are equally "graph-aware," the full potential of Pansoma's non-reference discoveries will remain partially locked behind the necessity of linear projection.

Figures from the paper