A prehistoric East-Asian Yersinia pestis genome and a ~5.3 ka trans-Eurasian expansion of plague Authors: Bing Sun, Yarong Wu, Daxuan Zhang, Kai Mu, Yan Guo, Shuang Fu, Mingru Sui, Chunxiang Li, Wenrui Zhang, Shan Chen, Andrei Poliakov, Igor Lazaretov, Qiankun Quan, Yajun Song, Ruifu Yang, Yinqiu Cui, Yujun Cui Published: 2026 ABSTRACT Plague, caused by Yersinia pestis, has affected human societies for millennia, yet the timing, scale and biological shifts underlying its earliest spread remain unclear. Here, we screened 1,400 ancient human remains from eastern Eurasia (8,000-400 cal. BP) and recovered three prehistoric Y. pestis genomes, including one from East Asia. Integrating these data with 136 published ancient and modern Y. pestis genomes, we resolved a trans‑Eurasian polytomy dated to ~5.3 ka BP that links Scandinavia, Central Europe, and northern China. Low SNP distances and tip-dating indicate splits over century‑scale intervals across > 7,500 km (median lower-bound dispersal rate of 32.9 km yr⁻¹), consistent with rapid long-distance dispersal. Mapping variation onto the deep phylogeny showed that the basal stem of the LNBA⁺ lineage is the only one of the seven deep phases examined to show significant enrichment of disruptive genomic remodeling relative to the genome-wide background distribution of mutation categories, including acquisition of the ymt gene and multiple gene‑disrupting indels and gene losses. This phylogenetic concentration is consistent with an early episode of adaptation associated with later flea-adapted lineages. Our results extend the prehistoric record of plague into East Asia and refine the timing and scale of its earliest trans-Eurasian expansion. Introduction Plague, caused by Yersinia pestis, has shaped human societies through three historically documented pandemics: the Justinianic plague (541-750 CE), the second pandemic initiated by the Black Death (1346-1353 CE) with recurrent waves for more than 500 years, and the third pandemic that began in the late nineteenth century and still persists in wildlife reservoirs1. Y. pestis evolved from the gastrointestinal species Y. pseudotuberculosis after acquiring two plasmids and a suite of virulence loci that enable flea-borne spread and rapidly fatal systemic infection2. Modern strains circulate endemically in rodent-flea cycles on every inhabited continent except Australia and have caused between 250 and 2,700 reported human cases annually in the 21st century3,4; accordingly, Y. pestis is listed among the World Health Organization’s bacterial priority pathogens5. Comparative genomics of present-day isolates reveals a highly clonal population structure (branches 0–4) but, on its own, offers limited power to date and localize the earliest diversification and to quantify the pace of long-distance dispersal of Y. pestis6–8. Ancient-DNA (aDNA) studies have begun to fill this gap by providing time-stamped genomes that anchor molecular clocks and preserve now-extinct diversity9. To date, more than 100 ancient Y. pestis genomes have been recovered, extending the documented history of the pathogen to more than 5,000 years before present (BP)4. The oldest strain, recovered from the Lake Baikal region, occupies a basal position relative to two later-diverging lines: LNBA⁻, which spread westwards to Great Britain and persisted for ~2000 years, and LNBA⁺, the ancestral stem of both historically documented pandemic and modern plague lineages10,11. Additional preLNBA⁻ genomes from Scandinavia and Central Europe bridge the temporal and geographic gap between the Baikal strain and the western LNBA⁻ cluster, underscoring the complexity of early dispersal12,13. Despite this progress, major uncertainties remain. East Asia—today home to the greatest natural foci and genetic diversity of plague—has so far yielded no prehistoric genome, leaving early trans-Eurasian routes unresolved and the full spectrum of ancient diversity obscured. Here we present three prehistoric Y. pestis genomes, spanning the preLNBA⁻ and LNBA⁻ lineages, recovered from a systematic screen of 1,400 human skeletons spanning East Asia and adjacent regions. We integrate these data with 136 published ancient and modern genomes to refine prehistoric phylogeny, geographic spread and early genomic remodeling in Y. pestis. Results Three prehistoric plague genomes, including one from East Asia We screened 1400 human skeletal remains dated between 8000 and 400 calibrated years before present (cal. BP) from sites across eastern Eurasia (Supplementary Fig. 1 and Supplementary Data 1), and detected authentic Y. pestis DNA in three individuals: CY470 (5465–5074 cal. BP) from the Middle-Neolithic cemetery of Zhengjiagou (ZJG) in northern China, and B275 (4,520-4,420 cal. BP) and C276 (4518–4248 cal. BP) from the Early-Bronze-Age sites of Uybat-Charkov and Itkol-II in the Minusinsk Basin, Russia14 (Fig. 1a–c, and Supplementary Fig. 2, and Supplementary Notes 1 and 2).Fig. 1Geographic distribution and phylogenetic reconstruction of ancient Y. pestis genomes.a Geographic distribution of ancient Y. pestis genomes, including newly reported prehistoric samples. Colors denote phylogenetic lineages and genomes with insufficient coverage for lineage assignment are labeled as “unknown”. Source data are provided as a Source Data file. b Enlarged view highlighting the locations of newly identified Y. pestis-positive individuals from this study. Source data are provided as a Source Data file. c Excavation photographs of ancient individuals infected with Y. pestis. Left, individual B275 from the Uybat-Charkov cemetery (~4200–4500 BP; Charkov Village, Republic of Khakassia, Russia; photo courtesy of Andrei Poliakov), indicated by a dashed circle. Middle, individual C276 from the Itkol-II site (~4200–4500 BP; Shirinsky District, Republic of Khakassia, Russia; photo courtesy of Andrei Poliakov). Right, burial CY470 from the Zhengjiagou (ZJG) site (~5000–6000 BP; Hebei Province, China; photo courtesy of Wenrui Zhang). d Maximum-likelihood tree of Y. pestis. Three newly sequenced genomes from this study are labeled in red. Ancient genomes representing five distinct phylogroups are highlighted in different colors, while modern genomes are shown in gray. To enhance visualization of the prehistoric Y. pestis tree topology, phylogroups corresponding to the First Pandemic, Second Pandemic, and modern strains have been collapsed. The full, uncollapsed phylogenetic tree with bootstrap support values for all nodes is provided in Supplementary Fig. 4. DA101 belongs to the same phylogroup as the First Pandemic genomes but predates them and is included here within the First Pandemic group. See Methods for details on the placement of Warburg1 and Warburg2. Geographic distribution and phylogenetic reconstruction of ancient Y. pestis genomes. a Geographic distribution of ancient Y. pestis genomes, including newly reported prehistoric samples. Colors denote phylogenetic lineages and genomes with insufficient coverage for lineage assignment are labeled as “unknown”. Source data are provided as a Source Data file. b Enlarged view highlighting the locations of newly identified Y. pestis-positive individuals from this study. Source data are provided as a Source Data file. c Excavation photographs of ancient individuals infected with Y. pestis. Left, individual B275 from the Uybat-Charkov cemetery (~4200–4500 BP; Charkov Village, Republic of Khakassia, Russia; photo courtesy of Andrei Poliakov), indicated by a dashed circle. Middle, individual C276 from the Itkol-II site (~4200–4500 BP; Shirinsky District, Republic of Khakassia, Russia; photo courtesy of Andrei Poliakov). Right, burial CY470 from the Zhengjiagou (ZJG) site (~5000–6000 BP; Hebei Province, China; photo courtesy of Wenrui Zhang). d Maximum-likelihood tree of Y. pestis. Three newly sequenced genomes from this study are labeled in red. Ancient genomes representing five distinct phylogroups are highlighted in different colors, while modern genomes are shown in gray. To enhance visualization of the prehistoric Y. pestis tree topology, phylogroups corresponding to the First Pandemic, Second Pandemic, and modern strains have been collapsed. The full, uncollapsed phylogenetic tree with bootstrap support values for all nodes is provided in Supplementary Fig. 4. DA101 belongs to the same phylogroup as the First Pandemic genomes but predates them and is included here within the First Pandemic group. See Methods for details on the placement of Warburg1 and Warburg2. Deep sequencing after targeted enrichment produced three genomes with 13- to 128-fold mean coverage and >90% chromosomal length (Supplementary Data 2). All three assemblies contain plasmids pCD1, pMT1 and pPCP1, but lack the flea-adaptation locus ymt (Supplementary Fig. 3). A maximum-likelihood phylogeny places the ZJG strain basally within the preLNBA⁻ lineage, whereas the two Minusinsk Basin strains cluster inside LNBA⁻ (Fig. 1d and Supplementary Fig. 4). These genomes respectively represent the earliest plague evidence from East Asia, and extend the temporal range of LNBA⁻ in the Minusinsk region by ~300 years. Notably, although the ymt-positive LNBA⁺ lineage appears and subsequently diversifies in the current time‑calibrated phylogeny, it remains sparsely represented in the human aDNA record for the ensuing ~3000 years. Across the 1400 screened human individuals from eastern Eurasia (8,000-400 cal. BP), we did not detect additional LNBA⁺ infections, and only two LNBA⁺ genomes (RT5 and I2470) from human remains have been reported to date11,15. Together, these observations delineate a ~3000‑year interval in which LNBA⁺ diversified yet is rarely recovered from human remains (Supplementary Fig. 5). Rapid trans-Eurasian spread in a star-like diversification Earlier studies have documented several star-like radiations in Y. pestis6,15–17. Our expanded dataset reveals an even earlier event, the prehistoric polytomy, dated to 5,281 years ago (95% highest posterior density, HPD: 5,411-5,165) (Fig. 2a and Supplementary Fig. 5). In the maximum-likelihood tree this appears as an unresolved branch owing to the absence of shared derived SNPs (Supplementary Fig. 4), whereas BEAST analysis, under a time-calibrated coalescent model, places the node within the Frälsegården clade (Supplementary Fig. 5). Three branches emerge in close succession: (i) the Frälsegården cluster in western Sweden (southern Scandinavia)12, (ii) a basal genome (Warburg1) from Warburg, Germany (Central Europe)13, and (iii) the stem that later gives rise to LNBA⁻ and LNBA⁺. The ZJG strain (CY470) from northern China differs by only five SNPs from the inferred most recent common ancestor (MRCA) at the base of the prehistoric polytomy (termed Prehistoric Big Bang, PBB node), placing East Asia squarely within the geographic span of this diversification.Fig. 2Prehistoric polytomy and transcontinental dispersal.a Schematic phylogeny highlighting strains around the prehistoric polytomy (PBB node). The most recent common ancestor (MRCA) of the polytomy is marked by a red triangle. Solid branch lengths reflect those in the maximum-likelihood tree, while dashed branches are shortened for clarity. For visualization, only the three strains closest to the polytomy in the LNBA⁻ and LNBA⁺/modern groups are shown. b Geographic spread and inferred dispersal routes. Red text denotes preLNBA⁻ genomes associated with prehistoric polytomy, while black text marks preLNBA⁻ genomes at basal phylogenetic positions, with the dashed circle indicating the oldest samples. The black dashed line indicates regions separated by physical distances of more than 7500 km. Red arrows indicate putative transmission directions, and the green shading marks the potential origin region. The world map was created using the maps package in R. c Distribution of inferred spread speed. One-dimensional probability density of the spread speed V (km yr−1) for prehistoric Y. pestis dispersal, obtained from 2 × 106 Monte-Carlo draws under the distance prior and time constraints. The shaded band marks the 95% confidence interval (CI) and the red dashed line denotes the median. A notched break compresses the extreme right tail for clarity. Source data are provided as a Source Data file. d Joint likelihood landscape of dispersal distance and spread speed. Three-dimensional surface showing the joint probability density for geographic distance S (y-axis, km) and spread speed V (x-axis, km yr−1) of prehistoric Y. pestis dispersal. Higher ridges mark combinations of S and V with greater support under the model. Color gradient encodes density. For clarity, the display window is restricted to the region around the 95% CI of V reported in (c), and the extreme high-speed tail is omitted. Source data are provided as a Source Data file. Prehistoric polytomy and transcontinental dispersal. a Schematic phylogeny highlighting strains around the prehistoric polytomy (PBB node). The most recent common ancestor (MRCA) of the polytomy is marked by a red triangle. Solid branch lengths reflect those in the maximum-likelihood tree, while dashed branches are shortened for clarity. For visualization, only the three strains closest to the polytomy in the LNBA⁻ and LNBA⁺/modern groups are shown. b Geographic spread and inferred dispersal routes. Red text denotes preLNBA⁻ genomes associated with prehistoric polytomy, while black text marks preLNBA⁻ genomes at basal phylogenetic positions, with the dashed circle indicating the oldest samples. The black dashed line indicates regions separated by physical distances of more than 7500 km. Red arrows indicate putative transmission directions, and the green shading marks the potential origin region. The world map was created using the maps package in R. c Distribution of inferred spread speed. One-dimensional probability density of the spread speed V (km yr−1) for prehistoric Y. pestis dispersal, obtained from 2 × 106 Monte-Carlo draws under the distance prior and time constraints. The shaded band marks the 95% confidence interval (CI) and the red dashed line denotes the median. A notched break compresses the extreme right tail for clarity. Source data are provided as a Source Data file. d Joint likelihood landscape of dispersal distance and spread speed. Three-dimensional surface showing the joint probability density for geographic distance S (y-axis, km) and spread speed V (x-axis, km yr−1) of prehistoric Y. pestis dispersal. Higher ridges mark combinations of S and V with greater support under the model. Color gradient encodes density. For clarity, the display window is restricted to the region around the 95% CI of V reported in (c), and the extreme high-speed tail is omitted. Source data are provided as a Source Data file. Bayesian tip-dating indicates that the European ancient genomes and CY470 diverged within ~200 years (Supplementary Fig. 5). Given the >7500 km great‑circle distance separating the discovery sites (Fig. 2b), we propagated uncertainty from both the BEAST time to the MRCA (tMRCA) estimate (95% HPD) and the radiocarbon calibration (95% confidence interval, CI) to estimate trans-Eurasian dispersal speed. The median speed is 32.9 km yr⁻¹, with the full credible interval shown in Fig. 2c, d. Because prehistoric transmission would not have followed straight‑line routes and likely involved staged transfers, these values should be interpreted as lower‑bound estimates (details in Methods). Even the lower-bound exceeds the continuous dispersal rates documented for rodent hosts (typically ≤3.5 km yr⁻¹)18–27. Across northern and central Europe, close affinities between the Warburg genomes (Warburg1/2) and the Frälsegården cluster indicate repeated short-range transfers superimposed on the continental-scale jump (Fig. 2a, b). In addition, the presence of basal preLNBA⁻ genomes in the Lake Baikal region (Russia)10, at Riņņukalns (Latvia)12, and in northern China (CY470) supports the Eurasian steppe as the most parsimonious source for these early lineages in the currently sampled record. Only the basal LNBA⁺ stem shows significant enrichment of disruptive genomic remodeling To scan for signatures of selection during early Y. pestis evolution, we mapped variation onto the maximum-likelihood phylogeny and partitioned the deep backbone around the LNBA⁺/LNBA⁻ split node into seven phases: five pre‑split stem phases (P1-P5), and the two immediate post‑split stems to the LNBA⁺ and LNBA⁻ MRCAs (P5⁺ and P5⁻). Among the seven phases examined, only P5+ on the basal LNBA⁺ stem showed significant enrichment of disruptive genomic remodeling relative to the genome-wide background distribution of mutation categories (Fig. 3 and Supplementary Data 3). This branch accumulated 14 indels, compared with 0–5 on other prehistoric branches, including 11 in loci annotated as pseudogenes (Supplementary Data 4). Together with one frameshift indel and one SNP located in a pseudogene, the P5⁺ branch carried 13 gene-disrupting mutations in total, representing a significant enrichment of disruptive variants relative to the genome-wide background distribution (Fisher’s exact test, P = 4.1 × 10⁻⁶; Supplementary Data 3–5). Concurrently, the clade acquired a 19-kb segment on plasmid pMT1 carrying ymt and lost three chromosomal loci including the ypm cassette (Supplementary Fig. 6 and Supplementary Data 6).Fig. 3Genomic changes across early evolutionary phases of Y. pestis.Genomic variants mapped onto the Y. pestis phylogeny, highlighting seven evolutionary phases (P1-P5, P5⁺, and P5⁻) marked by red inverted triangles. Three adjacent blocks per phase indicate the number of SNPs, indels, and large genomic gains/losses. The red filled star marks the phase showing significant enrichment of disruptive genomic remodeling, whereas the red unfilled star marks a weaker deviation limited to nonsynonymous SNPs. Genes with mutations linked to flea transmission, immune evasion, and pla in the P5+ phase are annotated. Source data are provided as a Source Data file. Genomic changes across early evolutionary phases of Y. pestis. Genomic variants mapped onto the Y. pestis phylogeny, highlighting seven evolutionary phases (P1-P5, P5⁺, and P5⁻) marked by red inverted triangles. Three adjacent blocks per phase indicate the number of SNPs, indels, and large genomic gains/losses. The red filled star marks the phase showing significant enrichment of disruptive genomic remodeling, whereas the red unfilled star marks a weaker deviation limited to nonsynonymous SNPs. Genes with mutations linked to flea transmission, immune evasion, and pla in the P5+ phase are annotated. Source data are provided as a Source Data file. Previous work defined six key changes that enable efficient flea-blockage transmission: acquisition of ymt, loss-of-function in pde2, rcsA and ureD, and two inactivating substitutions in pde328–30. All but the two pde3 variants arose on this early LNBA+ lineage, indicating that most previously described flea-associated changes accumulated along the same deep branch (Fig. 3). Bayesian tip-dating places the MRCA of all LNBA+ genomes at 4356 years ago (95% HPD: 4,790-4,033), which is 786 years after the LNBA⁺/LNBA⁻ split (5,142 years ago; 95% HPD: 5285–5009) (Supplementary Fig. 5), implying that eleven gene-inactivating indels, acquisition of ymt and the first pla substitution (V51L) accumulated within roughly eight centuries along the ancestral P5+ branch (Fig. 3). Another early branch (P4) on the stem leading to the PBB showed a weaker shift limited to SNP composition, with five substitutions, all nonsynonymous (Fisher’s exact test, P = 0.024; Fig. 3 and Supplementary Data 3). Clusters of Orthologous Groups (COG) analysis31 assigned four of these loci to categories spanning cell motility, secretion, amino-acid/carbohydrate metabolism, and cell-envelope biogenesis, representing broad metabolism and cellular-process modules (Supplementary Data 5). No indels or large segment gains or losses were observed in P4, indicating subtle genomic tuning rather than extensive remodeling seen on P5+. No other deep phase showed significant enrichment of disruptive variants relative to genome-wide background patterns under this framework (Supplementary Data 3). Discussion Our study extends the palaeogenomic record of Y. pestis geographically and temporally by adding one prehistoric genome from East Asia. By integrating these new data with published ancient and modern genomes, we better constrain the timing and geographic scale of early prehistoric trans-regional transmission. We note that East Asia is currently represented by a single prehistoric genome, and geographic inferences will benefit from additional sampling. We identified a prehistoric “big bang” polytomy dated to ~5,000 years ago that gave rise to three closely contemporaneous offshoots. Two branches, represented by the Scandinavian family cluster and the single Warburg1 genome, are not detected beyond their initial occurrences in the currently available ancient genome record. The third generated the “twins” stem, which subsequently split into the ymt-negative LNBA⁻ clade and its ymt-positive sister, LNBA⁺. The finding that Scandinavian, central-European and northern-Chinese genomes differing by only a few SNPs diverged within ~200 years implies dispersal on the order of tens of kilometre per year. Interpreted as lower‑bound estimates (because routes were non‑linear and likely involved staged transfers), these rates exceed continuous dispersal documented for rodent hosts18 and also outpace demographic diffusion estimates for LNBA human populations (~4–11 km yr⁻¹)32,33. This contrast is consistent with human mediation beyond background migration, plausibly via relay trade and logistical movements along emerging wheeled‑vehicle corridors34,35. Taken together, the rate contrast and the >7500 km span make exclusive rodent‑borne spread unlikely. The rapid trans-regional spread inferred here requires transmission opportunities compatible with long-distance mobility and repeated inter-group contacts, but human-mediated dispersal could reflect transport of infectious sources (people, materials or vectors) rather than sustained stepwise spread in rodent hosts alone. Given sparse and uneven sampling, genomic and archaeological evidence cannot yet resolve disease severity or the dominant transmission mode in prehistoric settings. Nonetheless, repeated recovery of closely related genomes within specific burial contexts (e.g., multi-generational occurrence at Frälsegården19) is consistent with recurrent introductions within connected communities. Early preLNBA⁻ infections are not necessarily synonymous with deadly flea‑borne plague and may have involved alternative clinical and transmission presentations, which is compatible with scenarios involving lower case fatality and/or longer infectious periods36, but these inferences remain indirect and require experimental validation. Efficient interhuman transmission mediated by human ectoparasites, including body lice (Pediculus humanus), is also a plausible mechanism based on historical-pandemic patterns and modeling37,38. Alternative scenarios, including predominantly zoonotic circulation with occasional spillover into humans, remain plausible. Our results also highlight an enigmatic ~3000-year interval in which the LNBA+ lineage diversified, yet only sparse evidence is recovered from human remains (Supplementary Fig. 5). Rather than indicating true absence, this pattern is consistent with a combination of sampling and ecological factors. First, ancient pathogen recovery is strongly biased toward human skeletal material, while infections circulating primarily in animal hosts or specific environments would be under-detected. In line with this, recent reports of prehistoric Y. pestis DNA from non-human contexts—including a Bronze Age Y. pestis genome recovered from an Ovis (sheep) tooth and falling within the LNBA⁻ lineage39, and a Neolithic dog associated with the preLNBA⁻ lineage13—support the presence of contemporaneous infections outside humans and argue for a broader host/environmental footprint. Second, if early LNBA+ circulation was dominated by enzootic dynamics with only intermittent spillover into humans, human cases could have remained rare and geographically patchy, leaving few detectable traces in the archaeological record. Our study has several limitations. Geographic sampling in East Asia remains sparse, and our conclusions about trans-Eurasian connections rely on a small number of key genomes, including a single prehistoric genome from northern China. Ancient pathogen detection is also strongly biased toward human skeletal remains, limiting direct inference about animal reservoirs and enzootic dynamics. In addition, quantitative estimates of dispersal speed depend on assumptions about plausible travel routes and distances, and functional interpretations of gene changes remain hypotheses in the absence of experimental validation. Despite these limitations, the newly reported prehistoric East Asian Y. pestis genome, together with two new Siberian genomes, strengthens the evidence that plague lineages had already achieved a continent-scale distribution across Eurasia by ~5.3 ka, with near-synchronous splits occurring across continental distances within a few centuries. Together, the phylogenetic, chronological, and geographic evidence are consistent with rapid long-distance spread at this early stage of plague history, while lineage-specific genomic remodeling accompanied the emergence of later flea-adapted clades. Methods Ethical statement and sample permissions This study relies on previously excavated archaeological remains and does not involve new excavations or research on living humans or animals. We affirm that all archaeological materials were originally collected, handled, and sampled in a responsible manner, in strict accordance with relevant local laws, regulations, and ethical guidelines. All newly reported ancient samples were analyzed with explicit permissions from the sample managers—who are also co-authors of this study—ensuring that destructive ancient-DNA analysis was scientifically and ethically appropriate. DNA extraction and library preparation Sample pre-treatment was conducted within the clean room facilities of the ancient-DNA laboratories at Jilin University, Changchun. To prevent modern DNA contamination, we first abraded the exterior surface of each tooth with a handheld rotary tool before drilling to collect dental pulp powder from the pulp cavity. DNA was extracted from approximately 50 mg of dental pulp powder, following the protocol described by Meyer et al. 40, using a silica-column-based extraction procedure optimized for short and degraded ancient-DNA fragments. Briefly, the dental pulp powder from each sample was transferred into a 1.5 mL tube and incubated in 1 mL extraction buffer containing 0.45 M EDTA, 0.25 mg/mL proteinase K, and 0.05% Tween-20. The mixture was vortexed thoroughly and incubated with rotation at 50 °C for 16-24 h to digest the mineralized tissue matrix and release DNA into solution. After incubation, the lysate was centrifuged at maximum speed for 2 min to pellet residual solid material. The supernatant was transferred into a 50 mL tube containing approximately 10 mL binding buffer composed of 5 M guanidine hydrochloride, 40% isopropanol, and 0.05% Tween-20, together with 400 μL of 3 M sodium acetate. The mixture was gently mixed and loaded onto silica spin columns by centrifugation to allow DNA binding to the silica membrane. The columns were then washed twice with 750 μL PE buffer and subjected to a dry spin to remove residual wash buffer. DNA was eluted from the silica membrane in 50 μL TET buffer containing 10 mM Tris-HCl, 1 mM EDTA, and 0.05% Tween-20. The eluate was applied to the membrane twice to maximize DNA recovery, resulting in a final extract volume of 50 μL. We then prepared non-UDG-treated, double-stranded DNA libraries from the extracts. Negative controls (no-template reactions with RNase-free water) were processed in parallel to monitor contamination. The indexed libraries were then shotgun sequenced on the Illumina HiSeq X10 platform, generating 150 bp paired-end reads. Screening with HOPS For each of the 1,400 samples, we assessed raw-read quality with FastQC (v0.11.8, https://www.bioinformatics.babraham.ac.uk/projects/fastqc/). Subsequently, we used AdapterRemoval (v2.3.1)41 to clip adaptors and merge paired-end reads, retaining only sequences with a minimum length of 30 bp and a minimum quality score of 20. To deplete host DNA, these processed reads were first aligned to the human reference genome (hs37d5) using the BWA-aln algorithm (v0.7.12)42 with parameters “-n 0.01 -l 32”, as implemented in the EAGER1 pipeline (v1.92.37)43. Reads unmapped to the human genome, identified using SAMtools (v1.14)44 with the “view” command and the “-f 4” flag, were extracted for downstream analysis. These putative non-human reads were then converted back to FASTQ format using BEDTools’ bamtofastq utility (v2.25.0)45. Taxonomic assignment of non-human reads was performed using HOPS (v0.34)46 against a custom database comprising reference or representative genome assemblies of 138 selected pathogens, including genomes representing all modern Y. pestis diversity, as well as other bacteria, viruses, fungi, and parasites. Genome assemblies were downloaded from NCBI GenBank (2–100 genomes per species). Further details are provided in Supplementary Data 7. Semi-global alignment was applied with a minimum identity threshold of 90%. HOPS generates a diagnostic plot only when the edit-distance distribution shows a decreasing trend accompanied by characteristic aDNA damage. Three samples, B275, CY470, and C276, were identified as candidate Y. pestis-positive based on the HOPS results. To further assess negative cases, we applied the competitive alignment method of Andrades Valtueña et al30. to all 1,400 samples and calculated a competitive alignment score. Samples with scores >0 were considered candidate Y. pestis-positive cases. For these, raw reads were mapped to Yersinia pseudotuberculosis IP32953 (GCA_000047365.1) and Y. pestis CO92 reference genomes (GCF_000009065.1) using BWA-aln, and edit-distance profiles were computed. Twenty-one samples exhibited scores >0 (Supplementary Fig. 7 and Supplementary Data 8). Only those showing a decreasing edit-distance trend were classified as confirmed Y. pestis-positive cases and retained for downstream analyses, leaving B275, CY470, and C276, consistent with the HOPS results. Enrichment experiment and re-screening with HOPS To enrich for Y. pestis DNA, we applied an in-solution capture strategy to all libraries that screened positive. For this purpose, 100 bp long RNA baits with a 25 bp tiling density and 99% identity were designed based on a selection of representative genomes (NC003143.1, NC009381.1, NZ_CP009712.1, NC003131.1, NC003134.1). We hybridized 750 ng of an equimolar library pool with these baits at 60 °C for 24 h, following the iGeneTech protocol. The enriched libraries were then prepared for shotgun sequencing on the Illumina HiSeq X10 platform, generating 2 × 150 bp paired-end reads. To eliminate potential false positives, after obtaining post-capture sequencing data, a re-screening step with HOPS was performed, followed by authenticity re-validation. Retrieval and selection of Y. pestis genomic data We retrieved all publicly available, high-quality ancient Y. pestis genomes from the NCBI Sequence Read Archive (SRA), as reported in previous publications (Supplementary Data 9). As our study focuses on prehistoric genomes, we included only representative genomes from the Second Pandemic owing to its broader sampling. When multiple genomes originated from the same archaeological context (i.e., similar radiocarbon date, geographic location, and phylogenetic position), we selected the sample with the highest average sequencing depth and read coverage for Y. pestis. For modern genomes, 57 representative complete assemblies covering major Y. pestis phylogroups were selected, with 2-3 strains per group where available, or a single strain when only one genome was accessible. These assemblies were sourced from BioProject PRJNA119051147 and other published datasets (Supplementary Data 9). Corresponding Illumina short-read data for all selected modern strains were also retrieved for downstream analyses. In total, 139 Y. pestis genomes were included in downstream analyses, comprising 42 prehistoric strains (9 from the preLNBA⁻ phylogroup, 31 from the LNBA⁻ phylogroup, and 2 from the LNBA⁺ phylogroup), 14 genomes assigned to phylogroup 0.ANT4 (including the DA101 genome from the Tian Shan region, dated to the 2nd–3rd century48, and 13 from the First Pandemic), 26 from the 1. PRE phylogroup (Second Pandemic), and 57 modern strains (Supplementary Data 9). Although the DA101 genome predates the First Pandemic, it was treated as part of this group in this study. Y. pseudotuberculosis IP32953 was used as the outgroup. Variant calling and phylogenetic reconstruction Adapter and quality trimming were performed using Trimmomatic (v0.38)49, removing reads shorter than 30 bp or with an average quality score below Q20. Filtered reads were aligned to the Y. pestis CO92 reference genome using BWA-aln with parameters “-n 0.01 -l 32”. PCR duplicates were removed using Picard MarkDuplicates (v2.21.2)50. Variant calling for both chromosomal and plasmid sequences was performed using GATK HaplotypeCaller (v3.8)51, applying a minimum phred-scaled confidence threshold of 30 (-stand_call_conf 30). SNPs were retained if they exhibited allele frequencies ≥0.9 and sequencing depths of at least 10× for modern genomes and 3× for ancient genomes. Variants located in repetitive regions, identified using BLASTN (v2.13.0+) and Tandem Repeats Finder (v4.07b)52, were excluded. For indels, sites with allele frequencies ≥0.8 and the same depth thresholds (≥10× for modern, ≥3× for ancient genomes) were retained. High-confidence chromosomal SNPs were concatenated to construct a maximum-likelihood tree using IQ-TREE (v1.6.5)53 under the GTR + G model with 1,000 ultrafast bootstrap replicates. Trees were visualized using FigTree (v1.4.4; http://tree.bio.ed.ac.uk/software/figtree/), and a schematic highlighting the prehistoric polytomy was generated using GrapeTree (v1.5.0)54. The phylogenetic positions of Warburg1 and Warburg2 were opposite to those reported by Susat et al. 13. Discrepancies in data quality and sample naming between the NCBI SRA database and the publication suggest a potential mix-up during data submission or processing. As we received no response from the original authors, we retained the SRA sample names for consistency with the raw data, but reassigned their dates based on phylogenetic position and the original report. Dating of ancient samples As the radiocarbon (14C) dating data for individual B275 were previously published, we utilized the dating results from that publication55. For the other two individuals, CY470 and C276, pretreatment for dating was carried out at the AMS-14C Graphitization Lab, Guangzhou Carbonage Technology Co., Ltd. The 14C measurements were subsequently performed using the accelerator mass spectrometry (AMS) facility (1.5SDH-1, 0.5MV, NEC, USA) at the Guangzhou Institute of Geochemistry, Chinese Academy of Sciences (GIGCAS). Radiocarbon dates for individuals CY470 and C276 were determined from the same samples used for DNA extraction. These 14C dates were calibrated using OxCal (v4.4.4)56 with the IntCal20 atmospheric curve57. Supplementary Data 2 lists the calibrated 14C dates for each sample, including the mean and range. Temporal signal assessment and time-calibrated phylogenetic analysis TempEst (v1.5.3)58 was used to assess the temporal signal. A time-calibrated phylogeny was inferred using BEAST (v2.6.7)59 with a tip-dating approach, in which modern tips were assigned their sampling years and ancient tips their median radiocarbon or archaeological dates (Supplementary Data 9). All dates were then converted to years before present (BP), with “present” defined as the isolation year of the most recent modern Y. pestis strain (2017). A GTR substitution model with four gamma rate categories was applied. For the full dataset, a relaxed log-normal molecular clock was used. Markov chain Monte Carlo (MCMC) analyses were run for 500 million steps for the full dataset, with the first 10% discarded as burn-in. Convergence (effective sample size [ESS] > 200 for all parameters) was confirmed in Tracer (v1.7.2; http://tree.bio.ed.ac.uk/software/tracer/). Maximum clade credibility (MCC) trees were generated with TreeAnnotator (v2.6.7)59 and visualized in FigTree (v1.4.4; http://tree.bio.ed.ac.uk/software/figtree/), with branch lengths scaled to age. Spread speed inference of Y. pestis in transcontinental dispersal To estimate the minimum plausible spread speed vv of prehistoric Y. pestis in transcontinental dispersals, we combine geographic distance (ss) and the transmission time interval (Δ tΔt), calculated from the most recent common ancestor (MRCA) and two descendant ancient genomes (CY470 and Warburg1). The effective elapsed time is:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\Delta} t=({t}_{{{\rm{MRCA}}}}-{t}_{{{\rm{CY}}}470})+({t}_{{{\rm{MRCA}}}}-{t}_{{{\rm{Warburg}}}1})=2{t}_{{{\rm{MRCA}}}}-{t}_{{{\rm{Warburg}}}1}-{t}_{{{\rm{CY}}}470}\,$$\end{document}Δt=(tMRCA−tCY470)+(tMRCA−tWarburg1)=2tMRCA−tWarburg1−tCY470 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${t}_{{{\rm{MRCA}}}}$$\end{document}tMRCA is the BEAST-inferred time to the MRCA. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${t}_{{{\rm{CY}}}470}$$\end{document}tCY470 and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${t}_{{{\rm{Warburg}}}1}$$\end{document}tWarburg1 are calibrated 14C dates of the CY470 and Warburg1. Corresponding spread speed (v)(v) is:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$v=\frac{s}{\Delta t}=\frac{s}{2{t}_{{{\rm{MRCA}}}}-{t}_{{{\rm{Warburg}}}1}-{t}_{{{\rm{CY}}}470}}\,$$\end{document}v=sΔt=s2tMRCA−tWarburg1−tCY470 Dispersal distance model We define ss as the sum of the geographic distances from the MRCA location to the sampling sites of each descendant strain: the Chinese genome (CY470) and the Central-European genome (Warburg1), i.e., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$s={s}_{{{\rm{CY}}}470}+{s}_{{{\rm{Warburg}}}}$$\end{document}s=sCY470+sWarburg. For any hypothesized MRCA location, the minimum distance sum equals the great circle distance between CY470 and Warburg1 sampling sites (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${s}_{\min }\approx {\mathrm{7,500}}\,{{\rm{km}}}$$\end{document}smin≈7, 500km), representing the shortest possible path. To incorporate real-world topographic constraints, including mountain ranges, river systems, and other ecological barriers, that result in nonlinear and circuitous dispersal routes, we apply a factor of three to the minimum distance. This yields an upper bound of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${s}_{\max }=3\times {s}_{\min }={\mathrm{22,500}}\,{{\rm{km}}}$$\end{document}smax=3×smin=22, 500km. Thus, the plausible geographic spread distance lies within the range \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$s\in \left[7500,\,22500\right]{{\rm{km}}}$$\end{document}s∈7500,22500km. Time information Radiocarbon ages for CY470 and Warburg1 (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${t}_{{{\rm{CY}}}470}$$\end{document}tCY470, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${t}_{{{\rm{Warburg}}}1}$$\end{document}tWarburg1) are provided as 5-year-binned histograms calibrated with OxCal (https://c14.arch.ox.ac.uk/). For \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${t}_{{{\rm{MRCA}}}}$$\end{document}tMRCA, we build a histogram-based density from the observed \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${t}_{{{\rm{MRCA}}}}$$\end{document}tMRCA inferred from 500 million Bayesian MCMC trees after discarding a 10% burn-in. Critically, the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${t}_{{{\rm{MRCA}}}}$$\end{document}tMRCA histogram bin width is chosen by K-fold cross-validation (K=5K=5) to maximize the average log-likelihood on validation folds. Monte Carlo inference We perform two million-scale Monte Carlo sampling to propagate uncertainty from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${t}_{{{\rm{MRCA}}}}$$\end{document}tMRCA, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${t}_{{{\rm{Warburg}}}1}$$\end{document}tWarburg1, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${t}_{{{\rm{CY}}}470}$$\end{document}tCY470 and ss into vv. Specifically, We (i) sample \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${t}_{{{\rm{Warburg}}}1}$$\end{document}tWarburg1 and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${t}_{{{\rm{CY}}}470}$$\end{document}tCY470 from their re-normalized empirical histograms by bin-mass selection plus uniform within-bin jitter, (ii) sample \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${t}_{{{\rm{MRCA}}}}$$\end{document}tMRCA from the histogram density, and (iii) sample ss from a Uniform prior model on \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left[7500,\,22500\right]{{\rm{km}}}$$\end{document}7500,22500km. We enforce genealogical constraints \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${t}_{{{\rm{MRCA}}}} > {t}_{{{\rm{Warburg}}}1},\,{t}_{{{\rm{CY}}}470}$$\end{document}tMRCA>tWarburg1,tCY470 by rejection-resampling (up to 50 attempts). For each accepted triple we compute \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$v=\frac{s}{\Delta t}$$\end{document}v=sΔt. Construction of one-dimensional (1D) and two-dimensional (2D) densities The 1D probability density of vv is computed by histogramming the Monte Carlo draws of vv into bins and normalizing the density (Fig. 2c). For readability, we use a broken vv-axis. To visualize the joint probability density distribution of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left(s,{v}\right)$$\end{document}s,v, we compute a 2D histogram with density normalization and render a 3D surface \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left(v,s,H\right)$$\end{document}v,s,H (Fig. 2d). All analyses were implemented in Python (v3.10.3) using NumPy (v2.3.1, probability density mapping), pandas (v2.3.0, data handling), and matplotlib (v3.10.3, 3D visualization). Note on real-world delays Real-world dispersal rarely consists of a single successful transmission event or of uninterrupted travel along a direct route, especially in prehistoric contexts (>5 ka). Accordingly, only a fraction of the elapsed interval is usable for geographic displacement, namely \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta t-{t}^{{\prime} }$$\end{document}Δt−t′ with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${t}^{{\prime} }\ge 0$$\end{document}t′≥0, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${t}^{{\prime} }$$\end{document}t′ aggregates unavoidable delays (e.g., adverse weather, multiple sequential transmission steps and their success probabilities, illness, and socio-cultural constraints). In our baseline analyses we set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${t}^{{\prime} }=0$$\end{document}t′=0 as a conservative assumption. The delay-adjusted speed is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}^{*}=\frac{s}{\Delta t-t{\prime} }$$\end{document}v*=sΔt−t′, which implies \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}^{*}\ge v$$\end{document}v*≥v, with strict inequality whenever \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${t}^{{\prime} } > 0$$\end{document}t′>0. For interpretability we parameterize downtime as a duty-cycle: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${t}^{{\prime} }=k\Delta t,{k}\in [{\mathrm{0,1}})$$\end{document}t′=kΔt,k∈[0, 1). We evaluated the effect of delays by fixing \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k\in \{0,\,0.2,\,0.4,\,0.6,\,0.8\}$$\end{document}k∈{0,0.2,0.4,0.6,0.8} and recomputing the inferred speeds from the same Monte Carlo pipeline used in the baseline analysis (two million draws; a Uniform prior distance model and genealogical/time constraints). As kk increases, the distribution shifts to higher speeds because a smaller fraction of the elapsed time is available for movement (Supplementary Fig. 8). Sensitivity analysis of the dispersal distance prior Because dispersal speed scales linearly with the assumed geographic distance, we assessed robustness to alternative spatial priors for ss while keeping the same plausible support \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left[7500,\,22500\right]$$\end{document}7500,22500 km. Besides the baseline Uniform prior \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$s\sim {{\rm{Uniform}}}({\mathrm{7500,22500}})$$\end{document}s~Uniform(7500, 22500), we considered three non-uniform priors: (i) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$s\sim {{\mathscr{N}}}\left({\mathrm{15000,2500}}\right)\,$$\end{document}s~N15000, 2500 truncated to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left[7500,\,22500\right]$$\end{document}7500,22500; (ii) a shifted Rayleigh model \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$s=7500+{X}$$\end{document}s=7500+X with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X\sim {{\rm{Rayleigh}}}\left(\sigma=3500\right)\,$$\end{document}X~Rayleighσ=3500 truncated at 22500; and (iii) a truncated Inverse Gaussian prior \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$s\sim {{\rm{InvGauss}}}\left(\mu=12000,\lambda=3000\right)\,$$\end{document}s~InvGaussμ=12000,λ=3000 on \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\,\left[7500,\,22500\right]$$\end{document}7500,22500. For each prior, we reran the same two-million-draw Monte Carlo procedure as in the former analysis. We then recomputed the 1D density of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}$$\end{document}v and the joint density of (s,v)(s,v). While alternative priors shift the absolute scale of inferred speeds as expected, inferred transcontinental spread speeds remain higher than documented rates for rodent hosts (Supplementary Fig. 9). Pangenome analysis and phase-specific gene gains and losses A total of 83 complete genome assemblies representing the Y. pseudotuberculosis complex (57 Y. pestis, 25 Y. pseudotuberculosis, and one Y. similis), obtained from NCBI, were used to construct the pangenome (Supplementary Data 9). Gene annotation and coding sequence prediction were performed using Prokka (v1.14.6)60. The pangenome was inferred using Panaroo (v1.2.8)61 in sensitive mode with default parameters. Shotgun sequencing reads were mapped to the constructed pangenome using BWA-mem, and per-gene coverage was calculated using SAMtools coverage. To identify phase-specific gene gain or loss events, we analyzed the presence/absence profiles of pangenomes across samples assigned to the seven defined phylogenetic phases. Given the relatively low sequencing depth and genome coverage of some ancient samples, genes present in at least two samples within a given phase (coverage ≥80%) and absent from all others were initially selected as candidates. These candidates were then manually inspected to confirm phase specificity. Verified genes were subsequently mapped to the reference genomes, and adjacent genes were merged into larger genomic fragments where appropriate. Selection signal analysis across seven evolutionary phases We identified clade-fixed mutations for seven evolutionary phases (P1-P5, P5+, and P5-) based on the presence of ancestral and derived alleles before and after the corresponding internal nodes in the phylogeny. Mutation types (e.g., synonymous, nonsynonymous, frameshift, and in-frame) were annotated using the CO92 reference genome annotation. For each phase, mutation categories were quantified and compared with the genome-wide background distribution (Supplementary Data 3) using Fisher’s exact test, with P < 0.05 indicating a significant deviation from genome-wide background mutation patterns. World map plotting The world maps shown in Fig. 1a, b, and Supplementary Fig. 1 were generated using publicly available basemaps from Natural Earth (https://www.naturalearthdata.com/downloads/10m-raster-data). The map in Fig. 2b was plotted using the maps package (v3.4.1) in R (v4.2.3). Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Description of Additional Supplementary Files Supplementary Data 1-9 Reporting Summary Transparent Peer Review file Supplementary Information Description of Additional Supplementary Files Supplementary Data 1-9 Reporting Summary Transparent Peer Review file Source data Source data Source data