Tsallis-order q fixation statistic Fq is available for use in the snputils Python library docs.snputils.org/api/stats.…

On 865 Oceanian samples spanning 1.8M sites, these profiles help identify which populations drive regional structure and distinguish signals associated with drift, founder effects, admixture, and migration.

In controlled genealogical simulations, Fq also tracks changes through time: OVR measures the level of differentiation, while LOO attributes that structure; capturing isolation and the sharp re-homogenization that follows reconnection.

We propose two complementary modes to use it: • One-vs-Rest (OVR): how distinct is each population from the rest? • Leave-One-Out (LOO): which populations are driving (or homogenizing) regional structure?

Her Tsallis-order q fixation index, Fq, measures the fraction of Tsallis q-entropy lost within subpopulations relative to the pooled population, placing familiar measures inside a single entropy family. • q=2 recovers the classical heterozygosity-based Fst • q=1 gives Shannon differentiation (i.e., normalized mutual information), the absolute q=1 form is mutual information between allele identity and population label Instead of looking at only one statistic, we can now examine a spectrum across q (and its slope). Within 1≤q≤2, lower q gives more weight to rare variants; higher q increasingly emphasizes common variants. This can expose differentiation patterns that a single Fst value may obscure.

Fst is a key genetic differentiation metric, but it emphasizes common variants over rare. In @ieeeICASSP out now @geletavc introduces Fq, which can be tuned across the allele freq spectrum. Fst & Shannon differentiation fall out of it as special cases. arxiv.org/abs/2511.03063

Congrats to first author @d_bonet, who presented ARGformer yesterday in a talk at the European Conference on Computational Biology @ECCBinfo in Geneva!

ARGformer is out in Bioinformatics! doi.org/10.1093/bioinformati… Just as PCA produces a low-dimensional representation of population structure from genotypes, ARGformer can do this now for ARGs...a representation that comes now directly from inferred genealogical topologies & at a haplotype level across each individual's genome. We do this by turning local genealogical paths into tokens & learning dense embeddings, similar to language embeddings, to visualize, cluster & query. Code github.com/AI-sandbox/ARGfor… Team @d_bonet, Cole Shanks, @marcal_cc, @jordiatjhu

Lots of standardized benchmarks. Phase error correction for LAI, the only scalable method with this. Insights...including that Native American dogs 🐕 did not vanish; living on (in part) in the Mexican Xoloitzcuintli. Also, Chihuahuas have a bit of coyote in them. Credit also to contributor, @Rich_Rast And maestro, @_danielmas Use it with our toolkit, snputils.org

Gnomix is published! nature.com/articles/s41467-0… Still the most accurate & fastest local ancestry inference (LAI) method, bests Flare and others (Recomb-Mix & Orchestra) released since our preprint. Easy use github.com/AI-sandbox/gnomix Leads @HelgiHilmars @arvind0422 @BarrabesMiriam

Our repo for exploration, github.com/AI-sandbox/aegen Fin

4. Synthetic genomes Conditioning was crucial for synthetic genomes. An unconditional VAE blurred population differences, while ancestry-specific models generated distributions much closer to real founder populations. The synthetic genomes reproduced real biological structure: correlations between nearby variants decayed with genomic distance, and locus-specific linkage-disequilibrium peaks aligned with real data. Also entropy patterns reflected.

3. Ancestry Discrete global ancestry classification jumped from 74.1% with PCA to 85.7% with a VAE. An ancestry-conditioned VAE pushed test accuracy to 87.1%.