Results

Of the 50 bats captured during this study, 39 were males, five of which were sub-adults, and only 11 were females, four of which were reproductive (lactating). All females and sub-adults were caught exclusively in Juniperus procera or Erica arborea forests. Only adult males were caught at the higher altitude Afroalpine moorlands (Supplementary Table S1).

Ecological niche models

Models had very high discrimination ability (AUCtrain=0.997; AUCtest=0.995 ±0.01). The main variables affecting the probability of occurrence of P. balensis were maximum temperatures (highest contribution to model and highest effect on gain when removed), Ethiopian Afroalpine moorland and Ethiopian montane grassland and woodland ecoregions, high topographic ruggedness and forest land cover (Supplementary Table S2). All other variables, including anthropogenic impact (human footprint), contributed very little to the model (<0.5%), and were therefore removed. The model projected suitable areas for P. balensis (i.e. falling above the threshold that maximises model sensitivity and specificity) across the Ethiopian Highlands, split into mountain ranges to the north and south of the Rift Valley, as well as mountain ranges in Eritrea and Somaliland (Fig. 2; Supplementary Fig. S2). However, only less than 1% of the Horn of Africa was predicted to be suitable for P. balensis . Climatic models projected across temporal scales predicted that the climatically suitable range was 4.5 times larger during the LGM and 3.8 times larger during the mid-Holocene. Only a quarter of the current range was projected to remain suitable by the end of the century, equating to 5.6% of the LGM suitable range (based on the more severe RCP 8.5 scenario; Fig. 2; Supplementary Fig. S2; Table 1).

Genetic composition

We identified 32 unique cytb haplotypes and 34 unique HVI haplotypes (Genbank Accession numbers to be added after acceptance). Each mountain range had unique haplotypes and no haplotypes were shared between mountain ranges. Overall, the Bayesian phylogenetic tree and haplotype network showed a strong effect of mountain ranges on genetic population structure and supported a split between haplotypes to the north-west and south-east of the Rift Valley. The main divergence was identified between haplotypes from Bale [Harenna Forest and Sanetti Plateau (Bale-S)] and the remaining haplotypes, followed by the divergence of Simien haplotypes (Fig. 3b). Population structure analysis of the microsatellite dataset split individuals into two clusters (K=2), north and south of the Rift Valley (Supplementary Fig. S3-S4). There were no further splits within each cluster. There was some evidence of allele sharing between both sides of the Rift Valley (Fig. 3c). FST and Jost’s D values were strongly correlated (MRDM: R2=0.684, P=0.018). In both cases, values confirmed the separation between populations north and south of the Rift Valley, with low values between the two Bale populations south of the valley and highest values between populations on either side of the Rift Valley (Supplementary Table S5).
Based on the mtDNA dataset, the Bale-S population had highest nucleotide diversity, substantially higher than the rest of the populations. The Simien population had lowest nucleotide diversity, but the highest haplotype diversity, though differences in haplotype diversity between populations were negligible (Table 2). Based on the microsatellite dataset, all populations had relatively high and similar levels of genetic diversity, with a particularly high number of private alleles identified in the Guassa population. Abune Yosef and Bale-S had the highest levels of inbreeding, while Simien and Guassa had the lowest, though all values were low (Table 2).
Allelic richness corrected for sample size decreased as the proportion of arable land in 5 km radius around the population capture location increased (F=10.41, df=1,3, P=0.048, R2=0.717; Supplementary Fig. S5). Levels of inbreeding did not relate to any of the land-use variables (Supplementary Table S6).

Landscape barriers to gene flow

The topographic hypothesis (effect of altitude) had the strongest support (AICcmin=0.501, BICew=0.475), followed by the ecoregions hypothesis (AICcmin=0.412, BICew=0.303). Confidence intervals of both variables did not overlap zero, supporting their effect on gene flow (Table 3). The exact same hypotheses received the strongest support based on Jost’s D measure of genetic differentiation, though the ecoregions hypothesis received the highest support (Supplementary Table S7). Projected movement density maps based on the effect of these two landscape variables highlight the strong effect of the Rift Valley on genetic connectivity in P. balensis and the split between populations to the north and south of the valley (Fig. 4). The remaining hypotheses had very low support. Hypotheses that included multiple variables had little support, likely due to the small number of nodes (Table 3).

Demographic history

The best-supported scenario was of a recent approximately six-fold decline of the south-eastern population but no change in the north-western population (scenario 4; probability of scenario 0.988, overall model error < 0.001, type 1 error = 0.095, type 2 error averaged over scenarios (± standard deviation) = 0.046 ±0.04; Supplementary Fig. S6; Table S8). Model checking for the most probably scenario (scenario 4) indicated a good fit between observed and simulated datasets, whereby the observed dataset fell within the cloud of simulated points (Supplementary Fig. S7). Rift valley split time was estimated at a median of 89,600 years ago (95% CI: 35,000-155,400), based on a generation time of two years, while the decline of the south-eastern population was estimated to have occurred at a median of 150 years ago (95% CI: 30-1000). Bias in parameter estimation was low (mean relative bias 0.18-0.26 for population size parameters, and 0.29 for time of decline of southern population; Table 4).