Panagram Compares 605 Vertebrate Genomes Without Alignment

Genome collections now span crop varieties and dog breeds as well as distant animal species. Looking for DNA associated with a trait across these collections creates a scaling problem for whole-genome alignment. ^[1]

The finding

Profiling genomes without full alignment

The standard tool for aligning many genomes at once, Progressive Cactus, needed months of cloud computing and about 2.1 million CPU hours to align a set of 605 vertebrate genomes. ^[2] Panagram built a comparable profile of the same 605 genomes, anchored to the CHM13 human reference genome, in about one day on a single computer with 40 processor cores, using fewer than 2,000 CPU hours. That one-day figure covers building the reference-anchored profile itself, not the entire downstream trait-association workflow. Under the same test conditions, two alignment- and graph-based tools, minigraph and pggb, did not finish, and neither did Mumemto, a separate alignment-free method the authors also benchmarked. ^[1] That gap is the paper's headline claim, and it is a compute-time comparison, not a claim that Panagram finds everything those tools would find. The work has not yet been peer reviewed.

A first pass with sequence trade-offs

Pangenome collections, crop varieties, dog breeds, conservation genomics panels, are growing faster than alignment-based tools can keep up with. Panagram doesn't produce the same output as a full alignment; it trades that letter-by-letter record for a much faster read on which short DNA snippets each genome shares or lacks, a trade the authors note works less well on repetitive or highly divergent sequence and can miss copy-number changes from recent duplications that lack distinguishing sequence differences. Even with those limits, it lowers the cost of taking a first pass at hundreds of genomes at once. The dog result shows the approach can recover a real, previously confirmed gene; the larger plant and vertebrate screens show it can also generate many more candidates than anyone has had the chance to check. Sorting the real signals from the noise in those catalogs is the work still to come.

The evidence

From DNA snippets to trait associations

A k-mer is simply a short, fixed-length stretch of DNA letters. Panagram breaks every genome in a collection into overlapping k-mers and records, for each one, which genomes contain it. That record is compressed into what the authors call a pan-k-mer bitvector, then mapped back onto each genome's own sequence to form a pan-k-mer bitmap. Because each genome keeps its own coordinates, there is no need to force every genome into one shared alignment grid, which is the step that becomes slow as genomes pile up.

From these bitmaps, Panagram computes two numbers at every position. Multiplicity is how many of the genomes in the set share a given k-mer there, which separates DNA that is common to nearly everything (core sequence) from DNA found in only a few genomes (variable sequence). Containment estimates how similar two genomes are at a given stretch, based on what fraction of one genome's k-mers also appear in the other. Scanning these two numbers across a genome, and comparing them against which individuals have which trait, is how the authors looked for trait-linked DNA without ever building a full alignment.

Known dog traits and untested gene links

The authors tested Panagram on four pangenomes. In 51 canine haplotypes from 37 animals, including related wild canids, the strongest signal, by a wide margin, involved three exons in a gene called RSPO2, already known to cause the wiry, mustached look called facial furnishings; the screen turned up 283 additional statistically significant exon-trait associations beyond that top hit. ^[1], ^[3] Finding a known gene first is a useful internal check on the method. Alongside it, the screen flagged newer candidates: two genes involved in mitochondrial metabolism, IVD and ACO2, linked to body mass index, and EMC2 and CCDC150 linked to coat length and furnishings. A separate window-based version of the analysis also recovered three previously known height genes, IGF1, HMGA2, and a signal downstream of SMAD2, though the exon-level analysis missed six previously reported weight genes. The authors flag two other hits as uncertain: GHR, an upstream signal near a coat-length gene, and TNR, an intronic signal near a furnishings gene, noting that neither is itself an established coat-length gene and that each signal may reflect linkage to a nearby causal variant rather than a direct effect. ^[1]

In 160 Arabidopsis thaliana genomes, the method found 1,361 statistically significant links between genes and local environment. Among the strongest was AINTEGUMENTA (ANT), a gene already known to control flower and organ size, linked to elevation. The authors are careful to note that elevation correlates with temperature, frost exposure, and growing-season length, and that this kind of association does not by itself separate cause from coincidence. ^[1]

In a panel of 20 genomes, 19 cultivated Solanum aethiopicum (the eggplant relative known as African eggplant) plus one wild Solanum anguivi, including eight genomes the team newly sequenced and assembled with Oxford Nanopore long reads, the authors used Panagram mainly to look for introgressed stress-response genes moving between the wild and cultivated lines, and separately found that genomes sequenced with PacBio HiFi gave more reliable similarity scores than the Oxford Nanopore assemblies.

Across 605 vertebrate genomes, the screen produced 7,724 significantly associated exons spanning 6,421 genes across seven traits, including body mass, longevity, and metabolism. Among the strongest exon-level signals was FXR2 (adjusted p = 2.5 x 10^-17); among gene-level signals, the strongest association with body mass specifically involved a microRNA gene, MIR1224 (adjusted p = 1.8 x 10^-10). As with the dog and plant results, the authors caution in their Discussion that these are statistical associations, not confirmed causes, and call for fine-mapping and functional validation before treating any of them as settled; the vertebrate screen is best read as an early-stage catalog of leads.

Future directions

The dog coat-length and furnishing results leave the roles of GHR and TNR unresolved. The authors caution that these signals could track a nearby causal gene. Fine-mapping or functional experiments could test whether GHR affects coat length or TNR affects furnishings, keeping association separate from mechanism. ^[1]

The trade-offs described for repetitive or highly divergent sequence suggest another starting point: compare Panagram's presence-and-absence calls with alignment-based results in those regions. In those regions, a benchmark could examine which copy-number changes are missed when recent duplications lack distinguishing sequence differences. ^[1]