Size dependent loss of structural variant allele counts in pangenome derived VCFs

Every human carries a unique set of instructions written in the language of DNA. While much of this code is identical from person to person, the differences are what make us individuals and drive our health. Scientists have long focused on tiny, single-letter changes in the code, which are easy to spot and record. However, the most dramatic differences often involve large chunks of DNA—sometimes thousands of letters long—that are deleted, duplicated, or flipped around. These are called structural variants. Because they are so large and complex, they are much harder to map than single-letter changes. In the past, scientists tried to force these large, messy differences into a simple, linear list, much like trying to describe a winding mountain road by listing only the straight highway segments it crosses. This approach often lost crucial details about how many people actually carried a specific change.

To solve this, researchers recently developed a new way to store genetic data called a pangenome graph. Instead of a single straight line, this graph weaves together the DNA sequences of many different people into a single, branching structure. In this map, every person's DNA is a path that winds through the same landscape, diverging and rejoining at different points. This allows scientists to see large structural variants as natural features of the terrain rather than forced annotations. However, almost all the software used to study disease and evolution still expects data in the old, simple list format. To use the new graph, scientists must convert it back into a list, a process that acts like a translator between two very different languages. The critical question was whether this translation process preserves the true count of how many people carry a specific large variant, or if the act of converting the data breaks the information apart.

A team of researchers set out to test exactly this. They created a controlled experiment where they knew the answer before they began. Using real human DNA sequences as a foundation, they simulated panels of twenty individuals, carefully placing large structural variants at specific locations and assigning them to known carriers. Because they designed the experiment themselves, they knew the exact number of people carrying each variant. They then built a pangenome graph from this simulated data and ran it through the standard conversion pipeline used by major research consortia to turn the graph back into a list. By comparing the final list against their original truth, they could see exactly what information survived the journey and what was lost.

The results revealed a clear and systematic problem. For small variants, the conversion worked almost perfectly, preserving the count of carriers. But as the size of the variant increased, the accuracy of the count dropped sharply. For variants that were 100 letters long, the conversion correctly identified the carrier count in nearly every case. However, for variants that were 30,000 letters long, the correct count was recovered in only about half of the cases. The larger the variant, the more likely it was that the conversion process would fragment the data, splitting a single group of carriers into many smaller, unrelated groups in the final list. This fragmentation was even worse in regions of the genome that are already difficult to read, known as segmental duplications, where the correct count was recovered in less than 40 percent of the largest variants.

The researchers traced the source of this error to the very first step of the conversion process. When the graph is broken down into a list, the software looks for "bubbles"—sections where paths diverge and rejoin. It treats every unique path through a bubble as a distinct variant. Because large bubbles often contain many tiny, unrelated differences between the carriers, the software mistakenly sees a single large deletion as dozens of slightly different events. The subsequent steps in the pipeline, which are designed to clean up and simplify the list, could not fix this fundamental misunderstanding. They could merge some of the fragments, but they could not reconstruct the original, unified group of carriers because the initial breakdown had already lost the connection.

To prove that this was a flaw in the translation method rather than a flaw in the graph itself, the team tried a different approach. Instead of converting the graph to a list first, they wrote a tool that counted the carriers directly by walking through the graph paths. This method bypassed the translation entirely. When they used this direct counting method, they recovered the correct carrier count for 99 percent of the variants, regardless of their size. This demonstrated that the information was still present in the graph, but the standard method of converting it to a list was failing to read it correctly. The issue was not with the map, but with the way the map was being read.

The team then looked at real-world data from the Human Pangenome Reference Consortium to see if this problem existed outside of their simulations. They examined a specific large deletion found in the human population, carried by 216 out of 461 individuals. In the raw graph data, this single event appeared as 84 different alleles, with the largest group containing only 25 carriers. When they looked at the final published list, the event was split into four separate records. While the total number of carriers was roughly correct when you added them all up, the event was no longer recognized as a single unit. Any analysis that counted how many distinct structural variants existed in the genome would mistakenly count this one event as four, inflating the diversity and distorting the understanding of how common the variant truly was.

The study concludes that while pangenome graphs offer a more complete picture of human genetic variation, the standard tools used to share and analyze this data are currently breaking the most important information for large variants. The loss of accurate counts is not a rare glitch but a predictable consequence of how the data is translated. The researchers suggest that for studies relying on the frequency of large structural variants, scientists should count carriers directly from the graph paths rather than relying on the converted lists. This approach is faster and far more accurate, ensuring that the true scale of genetic diversity is not lost in the translation from a complex map to a simple list.