Contrasting Patterns of Genetic Isolation Among Diadromous, Landlocked, and Non‐Migratory Gobiid Fish Populations Authors: Jeffery C. F. Chan, Bi Wei Low, Anthony Lau, Frank E. Rheindt, David Dudgeon, Jia Huan Liew Published: 2026 ABSTRACT ABSTRACTDams can landlock populations of amphidromous fishes that formerly migrated between rivers and seas. However, the genetic consequences of these life‐cycle changes remain unclear. Rhinogobius is a speciose genus of freshwater gobies widely distributed in Asia, including non‐migratory R. duospilus and facultatively amphidromous R. similis . Using genome‐wide SNPs from Hong Kong, southern China, we compared the genetic diversity and structure of R. duospilus (six populations), non‐landlocked R. similis (three populations), and landlocked R. similis (four populations) that had been isolated for 58–106 years. We also used demographic modelling to simulate patterns of gene flow of a panmictic amphidromous population of R. similis fragmented by dam construction 60 years ago. Rhinogobius duospilus populations were deeply divergent, with low genetic diversity; interpopulation differentiation reflected the underlying paleodrainage architecture. In contrast, non‐landlocked R. similis populations were genetically homogenous with high genetic diversity. However, all four landlocked populations of R. similis were highly differentiated, with three populations above tall (> 15 m) dams exhibiting comparable genetic clustering to non‐migratory R. duospilus , although overall genetic diversity remained high. Demographic modelling of a landlocked R. similis population indicated negligible migration into the reservoir and none outward, in accordance with the high levels of genetic divergence among landlocked populations. The effective population size of the original amphidromous population was small (Ne = 81), but the population landlocked above the reservoir for 60 years grew eightfold (Ne = 679). Our modelling suggests that even a limited level of immigration was sufficient to minimize genetic drift and lessen interpopulation divergence. Introduction The global proliferation of dams has severely altered freshwater ecosystems and represents a significant threat to freshwater biodiversity (Barbarossa et al. 2020; Dudgeon and Strayer 2025; Chan et al. 2025). Dams disrupt lateral and longitudinal connectivity of rivers, altering hydrological cycles, flow regimes and sediment transport, with direct and indirect consequences on biodiversity, ecosystem processes and productivity (Magilligan and Nislow 2005; Limburg and Waldman 2009; Barbarossa et al. 2020). Species that migrate between marine and freshwater habitats (i.e., diadromous species) are disproportionately affected as movement into spawning and feeding grounds is impeded (Chan et al. 2026). Many such species are commercially important (e.g., eels), and underpin cultural practices (e.g., stewardship of salmon among Alaskan tribes) and ecosystem services (e.g., inland transfer of marine‐derived nutrients) (Carothers et al. 2021). Diadromous species exhibit one of three modes of life cycles: amphidromy (larvae drift to the sea from fresh water, returning to freshwater as juveniles), anadromy (adults migrate to fresh water from the sea where they reproduce), and catadromy (adults migrate from fresh water to reproduce at sea) (McDowall 2001). Diadromous fishes typically display low levels of population differentiation because the marine portion of their life cycles permits cross‐drainage gene flow (Bracken et al. 2015; Skovrind et al. 2016). Some facultatively diadromous fishes can adapt and survive as landlocked populations above dams; these are genetically isolated from diadromous populations and can display patterns of isolation by distance akin to non‐diadromous fishes (Zarri et al. 2022). A recent meta‐analysis revealed that fragmentation by dams has caused significant declines in genetic diversity and increased genetic differentiation among populations of diadromous fishes (Chan et al. 2025). Outcomes can include changes in behaviour, life history, temperature tolerance and morphology, as well as reductions in phenotypic variability, which can alter adaptive potential (Pearse et al. 2014; Zarri et al. 2022). Reduced genetic diversity in isolated populations can accelerate the erosion of ecosystem services and affect fisheries beyond the level that would be attained by species loss alone (Schindler et al. 2010). Most efforts to better understand genetic responses of landlocked diadromous fishes to fragmentation have focused on anadromous salmonids (e.g., Whiteley et al. 2013; Samarasin et al. 2017). But as dams have proliferated, this phenomenon has become increasingly prevalent in amphidromous fishes, which are at increasing risk because of the limited mitigation afforded by the current design of fish passes (Chan et al. 2025). The strong climbing ability, generally shorter generation time, and obligate pelagic phase of many amphidromous fishes likely results in responses to fragmentation that differ from anadromous or catadromous species. For instance, gobiids, which are the most speciose group of amphidromous fishes, can scale dams with their fused pelvic fins (Cooney and Kwak 2013). However, their pelagic drifting larval stage makes them vulnerable to entrainment due to weak swimming ability, and can result in high mortality rates during downstream migration (Jarvis and Closs 2019). This study used two species of Rhinogobius as model organisms. They are a speciose genus of small (~3–10 cm) benthic freshwater gobies widely distributed in East and Southeast Asia, including non‐migratory Rhinogobius duospilus (lacking a marine larval phase, with direct development into juveniles soon after hatching; Li et al. 2008) and R. similis that can establish both amphidromous and landlocked populations (Ishizaki et al. 2016; Oto et al. 2025). The latter is widespread in East and Southeast Asia, while R. duospilus occurs throughout southern China. Both species are sympatric in coastal streams and reservoirs in the Hong Kong Special Administrative Region (henceforth Hong Kong) where our study was conducted (Chan et al. 2024). Hong Kong has a dense drainage network of coastal streams but connectivity between them has been greatly influenced by sea‐level oscillations. Sea levels have dropped by up to ~150 m at various times during Quaternary glacial periods, connecting now‐isolated streams within larger paleodrainages. Whenever sea levels rise during inter‐glacial periods (such as the present time), the larger paleodrainages become fragmented into smaller drainage networks isolated by salt‐water barriers (Fyfe et al. 2000). Contemporary R. duospilus populations show marked genetic divergence reflecting historic paleodrainage structure (Wu et al. 2016). In contrast, R. similis occurs in the form of amphidromous populations in coastal streams that have not been dammed, but there are some landlocked populations in reservoirs (Chan et al. 2024). None of these populations have been studied genetically, but we anticipated that genetic differences in populations of Rhinogobius would reflect their life cycles. We predicted that non‐landlocked populations would be panmictic because of a presumptively amphidromous life cycle with larval mixing in Hong Kong's coastal waters, whereas land‐locked populations isolated within reservoirs would be relatively divergent, although less so than the non‐migratory stream dwelling R. duospilus populations that have been isolated from each other since the Last Glacial Maximum ~18,000 years ago. Materials and Methods Site Description Sampling took place across four unobstructed streams and four reservoirs in Hong Kong (Figure 1; Table 1). The reservoirs are situated along coastal streams that were impounded between 58 and 106 years ago (as of sample collection dates) by dams at least 15 m tall (Table 1). As R. similis can establish both amphidromous and landlocked populations in streams (Ishizaki et al. 2016), we refer to those in unobstructed streams as non‐landlocked populations and those in reservoirs as landlocked populations. Populations were selected to maximize the spatial representation across the three main regions (Lantau Island, New Territories, Hong Kong Island; total extent: 1114 km2) and paleodrainages (Figure S1) of Hong Kong. In addition, two populations of R. similis in close proximity on Lantau Island were selected for demographic modelling to determine whether there had been any gene flow between them. One was a landlocked population within Shek Pik Reservoir (completed in 1963; hereafter SPR), and the other was a non‐landlocked population in nearby Shui Hau Stream (SHS). A catchwater runs along the ~100 m contour parallel to the South Lantau coastline and collects water from various coastal streams, including SHS, before emptying into SPR. Migrations into the catchwater, and subsequently into SPR, may occur occasionally, likely during the wet season when R. similis ascends streams in Hong Kong (Chan et al. 2024). No R. similis occurred downstream of the SPR Dam where stream flows have been reduced to a trickle and were subject to periodic interruption. (A) Sites sampled across Hong Kong and inset showing demographic modelling locations (SHS, Shui Hau Stream; SPR, Shek Pik Reservoir); (B) SHS; (C) SPR; site abbreviations given in Table 1. TABLE 1Number of specimens per species and per site used for genetic analyses.SiteAbbreviationYear collectedReservoir commissioned (years isolated)Dam height (m) Rhinogobius similis specimens Rhinogobius duospilus specimensHa Pak Nai StreamHPNS2022——1111Pui O StreamPOS2022——1515Shui Hau StreamSHS2022——1010Sheung Sze Wan StreamSSWS2022———10Lower Shing Mun ReservoirLSMR20231965 (58)551010Shek Pik ReservoirSPR20231963 (60)5415—Shap Long ReservoirSLR20231955 (68)1515—Tai Tam Tuk ReservoirTTR20231917 (106)181111Total specimens8767 Number of specimens per species and per site used for genetic analyses. Sample Collection A total of 87 R. similis (seven sites) and 67 R. duospilus (six sites) were collected for genetic analysis (10–15 per population; Figure 1; Table 1) between April 2022 and August 2023 using D‐framed hand‐nets (30 × 30 cm, 2‐mm mesh‐size). A mixed size class of specimens with a standard length range of 20–100 mm ( R. similis ) and 20–60 mm ( R. duospilus ) was collected to reduce sampling bias that may arise from sampling the same cohort. Specimens were euthanized in the field with MS‐222 following Underwood and Anthony (2020). Entire caudal fin clips were preserved in 96% ethanol to prevent DNA degradation prior to DNA extraction. DNA Extraction and Sequencing Entire caudal fin clips were taken for DNA extraction (Table 1). Total genomic DNA was extracted using the QIAGEN Blood & Tissue Kit following the manufacturer's protocol and quantified using a Qubit 2.0 Broad Range DNA Assay. Extracted DNA samples were then sent to NovogeneAIT Genomics Shanghai for library preparation and sequencing, following a protocol adapted from Elshire et al. (2011). DNA was digested with two restriction enzymes, MseI and NIaIII, purified by Agencourt AMPure XP beads, and then underwent PCR amplification. Amplified samples were pooled, and size selection was performed to select fragments that were between 350 and 400 base pairs (with indexes and adaptors) in length. Fragments were subsequently purified by beads again and sequenced on two lanes on the NovaSeq PE150 platform to obtain 150 bp paired‐end reads. SNP Filtering, Identification, and Alignment Sequence data for the eight goby populations were filtered to maximize high‐quality reads for optimized downstream analyses. The following reads were filtered out: (1) reads with ≥ 10% unidentified nucleotides; (2) reads with > 50% bases having low quality scores (phred quality < 5); (3) reads with > 10 nucleotides aligned to the adapter, allowing up to 10% mismatches. In the absence of a reference genome for R. duospilus , sequence data of both species were aligned to the R. similis genome (ASM1945343v1; Hu et al. 2022), using the Burrows‐Wheeler Aligner version 0.7.8 (BWA) software (Li and Durbin 2009). Single‐nucleotide polymorphisms (SNPs) were then called using the ‘ref_map.pl’ program in STACKS version 2.61 (Catchen et al. 2013; Rochette et al. 2019). The ‘populations’ program in STACKS was then run to generate population summary statistics for downstream analyses. To maximize repeatability, we retained the first SNP per locus in loci that were present in a minimum of 80% of individuals across all 7 populations, removing rare variants (minimum minor allele count of 3; Paris et al. 2017). Further filter criteria were applied in PLINK version 1.90 (Purcell et al. 2007) to remove SNPs with more than 10% missing data, individuals with > 50% missing data, and variants which had Hardy–Weinberg equilibrium exact test values of 1e‐50 with mid‐p adjustment (Graffelman and Moreno 2013). Population Genetic Analyses To assess population‐level genetic diversity of the six ( R. duospilus ) and seven ( R. similis ) goby populations, expected heterozygosity, observed heterozygosity, and inbreeding coefficients were calculated using GenoDive version 2.0b27 (Meirmans and Van Tienderen 2004). Next, genetic clusters were identified through principal component analyses (PCA) using the R package ‘SNPRelate’ (Zheng et al. 2012). Population structure was further scrutinized with a Bayesian clustering approach using STRUCTURE version 2.3.4 (Pritchard et al. 2000) where admixture proportions of individuals in each population are estimated. The admixture‐locprior model was used, with a burn‐in value of 100,000 and 500,000 Markov chain Monte Carlo (MCMC) iterations to test for predicted levels of genetic structuring, using genetic clusters or K values ranging from 1 to 6 for each species and five replicates per K value. Structure runs were averaged and summarized using CLUMPAK (Kopelman et al. 2015), and the most likely K value identified using the Evanno ∆K method (Evanno et al. 2005). Pairwise genetic differentiation between populations of each species was then quantified with the pairwise fixation index (FST) using GenoDive. Isolation by distance (IBD) analyses were conducted in the R package ‘dartR’ (Gruber et al. 2018) for each species using Mantel tests on geographic distances, calculated from coordinates, and pairwise FST with Rousset's correction (Rousset 1997). Demographic Scenario Modelling We conducted demographic modelling to determine whether there had been any gene flow between land‐locked R. similis in SPR and the non‐landlocked population in nearby SHS. All demographic models were based on coalescent simulations and were initialized on the assumption that both contemporary populations represented a single historical amphidromous panmictic population with high rates of migration (0.1; Table S1; following Ruzich et al. 2019) until construction of the SPR dam in 1963 (60 years prior) isolated the SPR population from the panmictic SHS population. We compared four different demographic scenarios after impoundment: no gene flow; migration from reservoirs to free‐flowing streams; migration from the stream into the reservoir; and bidirectional migration. Site‐frequency spectra (SFS) were constructed for each population using the ‘easySFS’ tool (https://github.com/isaacovercast/easySFS) from a PLINK association file generated from the filtered SNPs. Observed SFS were then input into FastSimCoal2 version 2.7 for simulation of each scenario using the composite likelihood method (Excoffier et al. 2021). Each independent run involved 100,000 coalescent simulations, 40 expectation–maximization cycles, stopping criteria with 0.001 relative difference in parameter values, a minimum of 10 SFS entry counts for parameter estimation, and a 100–100,000 log‐uniform distribution search range (following Meier et al. 2017). For each of the four scenarios, 100 independent FastSimCoal2 runs were conducted to calculate parameter estimates for maximum likelihood, with the best model determined using Akaike information criterion (AIC) values (Excoffier et al. 2013). As there are no reliable estimates of the mutation rate of Rhinogobius, parameter estimates were derived from the time that the SPR dam was completed, that is, 60 generations ago, assuming a 1‐year generation time (Oto et al. 2025). To account for uncertainty, we also examined the likelihood distributions for each population model scenario by rerunning each model with the best parameter values (100 expected SFS with 1,000,000 simulations per model). Overlapping likelihood distributions between the original and rerun models would indicate an equally good fit to the observed data, with any differences in AIC reflecting stochasticity of the SFS approximation (Meier et al. 2017). We conducted block‐bootstrapping to obtain 95% confidence intervals (CIs) for the best‐supported demographic model for the SPR population of R. similis . The original dataset was then split into 100 equal‐sized blocks to create 50 bootstraps by sampling with replacement, 100 blocks for each bootstrap. Each bootstrap then underwent 100 independent runs in FastSimCoal2 with the same parameters as above; however, fewer expectation conditional maximization (ECM) cycles (20) were applied per run to reduce computational intensity because the SNPs were already filtered. The best run determined by AIC in each of the bootstraps was then used to compute 95% CIs using the R‐package ‘rcompanion’ (Mangiafico 2015; James et al. 2021). Results A total of 40.16GB of data were generated from 154 individuals (Table 1) after filtering, with mapping rates of each sample ranging from 62.76% to 98.02%. The average sequence coverage was 5.4× for R. duospilus and 4.3× for R. similis . A total of 9416 SNPs were retained from the filtering process for R. duospilus and 17,725 for R. similis . We were able to retain data from all individuals sequenced. Population Genetic Diversity The six populations of R. duospilus had mostly comparable levels of observed heterozygosity (H o = 0.019–0.031) and expected heterozygosity (H e = 0.023–0.037) across sites; the exception was LSMR (H o = 0.084, H e = 0.112; Table 2), which had three to four times higher genetic diversity than other populations. The inbreeding coefficients (Fis) for R. duospilus varied widely between −0.007 and 0.256. There were comparatively higher measures of heterozygosity in R. similis (H o = 0.031–0.043; H e = 0.036–0.051) than R. duospilus , which likewise exhibited minimal variation across populations. Measures of heterozygosity barely differed among the three non‐landlocked and four land‐locked R. similis populations (both respectively averaged H o = 0.037 and H e = 0.044; Table 2). Fis in R. similis was always greater than zero (Table 2). TABLE 2Population genetic diversity summary statistics for Rhinogobius similis and R. duospilus .Site Rhinogobius duospilus Rhinogobius similis Ho He Fis Ho He Fis HPNS0.0260.0270.0260.0390.0460.149SSWS0.0270.027−0.007———SHS0.0310.0370.1730.0370.0430.151POS0.0190.0230.190.0350.0430.177SLR———0.0310.0360.143SPR———0.0380.0450.138LSMR0.0840.1120.2560.0370.0440.156TTR0.0280.0330.1620.0430.0510.152 Note: Site abbreviations given in Table 1.Abbreviations: Fis, inbreeding coefficient; He, expected heterozygosity; Ho, observed heterozygosity. Population genetic diversity summary statistics for Rhinogobius similis and R. duospilus . Note: Site abbreviations given in Table 1. Abbreviations: Fis, inbreeding coefficient; He, expected heterozygosity; Ho, observed heterozygosity. Population Structure Four out of six R. duospilus populations formed distinct genetic clusters, although the two adjacent Lantau Island populations (SHS and POS) overlapped on the PCA biplot (accounting for 32.8% of overall variability; Figure 2A). STRUCTURE identified the genetic clusters K = 2 (ΔK = 23967.096; Figure 2B) and K = 6 (ΔK = 2682.315) to best represent genetic populations (K = 3 to K = 5 ranged from ΔK = 0.450 to 5.782). Populations were grouped by regions and paleodrainages, with K = 2 reflecting the separation of Lantau Island (SHS, POS) from Hong Kong Island (TTR) and the New Territories (HPNS, SSWS, LSMR), and K = 6 reflecting the separation by distinct paleodrainages (Figure S1). Pairwise FST values between R. duospilus populations were all significantly different (p ≤ 0.001; Table 3). We also observed a mild but significant (p = 0.02, R 2 = 0.22; Figure S2) pattern of isolation by distance. (A) Principal component analysis (PCA) ordination biplot and (B) STRUCTURE (best K = 2) barplots showing the genetic structure of Rhinogobius duospilus ; (C) PCA ordination biplot (inset in the top right corner is an enlargement of the data points in the bottom left) and (D) STRUCTURE (best K = 2) bar plots showing the genetic structure of R. similis ; abbreviations as given in Table 1. TABLE 3Pairwise genetic divergence (FST) of Rhinogobius duospilus populations. All pairwise comparisons were significantly different (p ≤ 0.001).HPNSSSWSSHSPOSLSMRTTRHPNS—0.8910.8550.8860.6390.855SSWS0.891—0.8590.8910.6790.865SHS0.8550.859—0.6180.6210.838POS0.8860.8910.618—0.6880.871LSMR0.6390.6790.6210.688—0.628TTR0.8550.8650.8380.8710.628— Note: Site abbreviations given in Table 1. Pairwise genetic divergence (FST) of Rhinogobius duospilus populations. All pairwise comparisons were significantly different (p ≤ 0.001). Note: Site abbreviations given in Table 1. PCA suggested that non‐landlocked populations of R. similis are genetically distinct from land‐locked populations (Figure 2C), although individuals from SLR were situated close to, while still distinct from, the non‐landlocked populations in ordinal space (see inset in Figure 2C). The spread in data points along both axes within landlocked populations was larger than the spread among non‐landlocked populations, which may explain the low variability explained in the PCA plot (7.1%). STRUCTURE analysis suggested highest support for K = 2 (∆K = 728.831; Figure 2D), whereas K = 3 (∆K = 2.133) and K = 4 (∆K = 6.789) were not well supported. At K = 2, non‐landlocked (blue; Figure 2D) and land‐locked (orange) populations were generally distinct, although one of the reservoir populations (SLR) was genetically mixed with the non‐landlocked populations (HPNS, SHS, POS). The three remaining large reservoirs (SPR, LSMR, TTR) with tall dams (≥ 18 m) exhibited varying levels of admixture with the non‐landlocked populations (Figure 2D), while non‐landlocked populations did not exhibit admixture with landlocked populations, suggesting a one‐way migration into reservoirs. Pairwise FST values of all four reservoir populations were significantly different (p ≤ 0.001; Table 4) across all comparisons, while non‐landlocked populations significantly differed from all landlocked populations. There was no evidence of genetic isolation by distance (p = 0.09; Figure S3). TABLE 4Pairwise genetic divergence (FST) of Rhinogobius similis populations.HPNSSHSPOSSLRSPRLSMRTTRHPNS—0.0020.000 0.065 0.133 0.323 0.147 SHS0.002—0.000 0.054 0.127 0.326 0.144 POS0.0000.000— 0.061 0.133 0.327 0.150 SLR 0.065 0.054 0.061 — 0.154 0.368 0.176 SPR 0.133 0.127 0.133 0.154 — 0.293 0.150 LSMR 0.323 0.326 0.327 0.368 0.293 — 0.132 TTR 0.147 0.144 0.150 0.176 0.150 0.132 — Note: Significantly different comparisons (p ≤ 0.001) are in bold.Abbreviations: HPNS, Ha Pak Nai Stream; LSMR, Lower Shing Mun Reservoir; POS, Pui O Stream; SHS, Shui Hau Stream; SLR, Shap Long Reservoir; SPR, Shek Pik Reservoir; TTR, Tai Tam Tuk Reservoir. Pairwise genetic divergence (FST) of Rhinogobius similis populations. Note: Significantly different comparisons (p ≤ 0.001) are in bold. Abbreviations: HPNS, Ha Pak Nai Stream; LSMR, Lower Shing Mun Reservoir; POS, Pui O Stream; SHS, Shui Hau Stream; SLR, Shap Long Reservoir; SPR, Shek Pik Reservoir; TTR, Tai Tam Tuk Reservoir. Demographic Modelling Among the four demographic scenarios of R. similis movement between SHS and SPR, the highest log‐likelihood and lowest AIC values were obtained for the model representing unilateral gene flow from SHS into SPR (Figure 3), although this was only marginally better than the bidirectional model (Table S2). This model suggests a low probability (0.6%) of gene flow from SHS into SPR for each generation, and no migration the other way around, reflecting results from STRUCTURE analysis (Figure 2D). The common ancestral effective population size was small (Ne = 81 [95% CI: 79–82]) but grew exponentially in SPR (Ne = 679 [95% CI: 667–690]) after 60 years of isolation, while the amphidromous non‐landlocked SHS population had a much smaller change (Ne = 105 [95% CI: 104–105]). Likelihood distributions overlapped between models, indicating a good fit (Figure S4). (A) Model comparisons based on 100 FastSimCoal2 runs of demographic scenarios of Rhinogobius similis after a previously non‐landlocked population became landlocked above a dam for 60 generations; each run consisted of 40 expectation–maximization cycles with 100,000 coalescent simulations per cycle. (B) Best model (upstream) based on lowest Akaike information criterion scores and highest log‐likelihood; Ne = effective population size; 95% confidence intervals are shown in brackets based on 100 bootstrap replicates; Shui Hau Stream (SHS)—non‐landlocked population; Shek Pik Reservoir (SPR)—landlocked population; bidirectional—two‐way migration between SHS and SPR. Discussion We found three distinct genetic patterns among our focal species that appear to be specific to their respective life cycles. Populations of non‐migratory and salt‐intolerant Rhinogobius duospilus were deeply divergent among one another, each with a low diversity. This pattern was in stark contrast to the genetically homogenous non‐landlocked R. similis , which had a relatively higher diversity. Earlier marking experiments have demonstrated limited movement of R. duospilus across streams in Hong Kong (~3 m displacement after 35 weeks; Ho and Dudgeon 2016) and consequently, populations can be genetically differentiated even if they are less than 5 km apart (Wu et al. 2016). Our results of genetic structure were best explained by ancient paleodrainage basins during the Pleistocene (0.796–1.395 myr), with unique and distinct lineages observed in each of five Pleistocene drainages (Wu et al. 2016; Figure S1). These patterns of deep genetic divergence have been observed among other primary freshwater loaches in Hong Kong including Schistura fasciolata, Pseudogastromyzon myersi (Wong et al. 2017) and, most notably, Oreonectes platycephalus (Wang et al. 2024), which is confined to headwater tributaries. In contrast to R. duospilus , non‐landlocked R. similis populations were more or less panmictic and genetically variable, typical of an amphidromous life cycle with genetic connectivity maintained along the ~730 km coastline of Hong Kong. These findings are consistent with the absence of genetic structure among amphidromous R. similis populations across East and Southeast Asia (Ju et al. 2016), as well as among populations of other amphidromous goby species, such as Sicyopterus lagocephalus (Berrebi et al. 2005) and Rhinogobius gigas (Liao et al. 2021). This high level of connectivity is expected as sea currents around Hong Kong alternate directionally (north–south) between the wet and dry seasons when the Hainan and Kuroshio‐Taiwan currents prevail respectively (Morton and Morton 1983), allowing amphidromous fishes of different origins to mix (Chan et al. 2024). The lack of genetic structure is also consistent with the absence of evidence for natal homing behaviour among amphidromous fishes (R. McDowall 2010). On the other hand, the overall genetic structure of landlocked R. similis populations was more similar to non‐migratory R. duospilus than to their non‐landlocked conspecifics. After 58–106 years of isolation by impoundments, the genetic structure of the three landlocked R. similis populations above large dams aligned with three different paleodrainage regions. This alignment shows that isolation by human‐induced fragmentation, even over a relatively short evolutionary timescale (58–106 years), can amplify genetic differentiation and structuring of amphidromous metapopulations that typically have limited structuring. Similar patterns of genetic differentiation of fragmented populations have been reported in salmonids, although these usually have longer generation times (~3 years for Oncorhynchus; Froese and Pauly 2025) than R. similis (mostly ~1 year; Oto et al. 2025). In Oncorhynchus, genetic differentiation was discernible after ~40–60 years for O. mykiss (Clemento et al. 2009), ~70 years for O. masou (Kitanishi et al. 2012), and ~100 years for O. nerka (Samarasin et al. 2017). We could find only one study of amphidromous fishes, in which Japanese Rhinogobius fluviatilis populations exhibited interpopulation genetic differentiation ~35 years after landlocking (Takagi et al. 2012). However, differentiation among R. fluviatilis populations was less consistent than what we found in R. similis where all landlocked populations were significantly differentiated, perhaps because of the longer period of isolation (at least 58 years) in Hong Kong. The short coastal streams of Hong Kong (typically 2–4 km in length) are unlikely to support large populations of R. similis , which is consistent with our demographic model estimates of a small effective size for the ancestral population (Ne = 81) in SPR prior to impoundment. Although fragmentation often reduces the effective population size of amphidromous migrants (Maekawa et al. 2001; Whiteley et al. 2013), we recorded an eight‐fold increase in the effective population size above the SPR dam (Ne = 679) following impoundment. Formation of the reservoir created a substantially larger extent of habitat (~7.8 km2) supporting an expanded population that would, in turn, facilitate genetic exchange among more individuals resulting in higher genetic diversity (Kitanishi et al. 2012). This expansion in population size is likely aided by the fact that R. similis are lotic generalists, as exemplified by their successful invasion of freshwater habitats in Singapore (Larson et al. 2008), Armenia (Pipoyan et al. 2024) and Japan (Tsuji et al. 2025). Landlocked R. similis may also benefit from an abundance of zooplankton prey, which would enhance recruitment success (Iguchi and Mizuno 1999; Sumizaki et al. 2019). The combined effects of short generation time and an increase in effective population size from a small population after impoundment may have expedited genetic differentiation of the landlocked R. similis (Hoffman et al. 2017). Evidence of a shorter pelagic larval duration (PLD) in a landlocked population of R. formosanus (~20 days less than in an amphidromous population) has been reported in Taiwan (Liao et al. 2020), which could have implications for generation times and facilitate the accumulation of genetic mutations. The PLD of R. similis is not known but Liao et al. (2020) also reported lower genetic diversity in the landlocked than in the amphidromous population of R. formosanus , in contrast to our findings for R. similis . This may reflect a founder effect, as the Taiwan population had only been landlocked in the reservoir for 33 years, compared to our shortest period of 58 years, where the founder effect may no longer be detectable. Our demographic modelling results suggest that for each generation, any allele from the SPR population has a 0.6% chance of having been transferred from the non‐landlocked SHS population. The intermittent connectivity between SPR and SHS populations could have increased genetic diversity within SPR as reported for other landlocked fishes (Frankham 1996; Alcala and Vuilleumier 2014). While our results estimate gene flow in terms of allele transfer probability rather than absolute number of migrants, they are consistent with the ‘one‐migrant‐per‐generation’ rule, which suggests a minimum of 1–10 migrants per generation is sufficient to minimize loss of genetic diversity in isolated populations (Mills and Allendorf 1996). Implications and Limitations Three key lessons can be inferred from our study. First, our observations of R. duospilus are consistent with a pattern of restricted gene flow in non‐migratory fishes across Pleistocene paleodrainages, which can lead to higher inbreeding and vulnerability to genetic drift (Allendorf et al. 2012). The lower population genetic variability may increase susceptibility to inbreeding depression, diseases and climatic fluctuations, which increases extinction risk (Frankham 2005). Second, we show that genetic differentiation in R. similis can be relatively strong after 58–106 years of isolation, though we highlight that other amphidromous species may respond differently to land locking. More information is needed on how other threatened landlocked species such as Plecoglossus altivelis ryukyuensis (Kawakami and Tachihara 2011), Rhinogobius sp. aobara‐yosinobori (Nagata et al. 2012), R. formosanus (Liao et al. 2020), and Gobiopterus macrolepis (Chan et al. 2023) in Japan, Taiwan, and Hong Kong have been affected. Third, our findings suggest that low, potentially overlooked levels of migration can maintain genetic diversity of landlocked R. similis populations. However, this may not be possible for other amphidromous fishes that have poor climbing ability, and may therefore require human intervention (e.g., translocation of individuals) to facilitate connectivity. This may represent a possible management strategy for preserving fragmented populations of migratory fishes in instances where the installation of fish passes or the removal of dams and other barriers are not feasible. A potential limitation of the present study is the lack of otolith microchemistry analyses, which could help reveal the potential coexistence of amphidromous and river‐resident R. similis (see Ishizaki et al. 2016). However, we note that none of the non‐landlocked R. similis populations showed any signs of isolation. Future genetic studies supplemented by otolith microchemistry will not only verify amphidromy but also contextualize these results, for example by establishing the pelagic larval duration. Categorizing deeply divergent populations as evolutionarily significant units (ESU) may be necessary in cases where current taxonomy does not adequately reflect genetic diversity (Fraser and Bernatchez 2001). For R. duospilus , the Lantau Island population may be an appropriate ESU, and it has previously been suggested to represent a cryptic species (Wu et al. 2016). Strategic planning for future dam sites, dam removal, or retrofitting with fish passes would be ideal for conserving populations of migratory fish species (Barbarossa and Schmitt 2024; Chan et al. 2025). Where restoring longitudinal connectivity is not feasible (e.g., dam removal of a reservoir) and an insufficient number of individuals can naturally traverse barriers, translocation is a possible management tool to enhance evolutionary potential of fitness characters and genetically rescue inbred populations (Frankham 2015; Pregler et al. 2023). Notably, trial translocations of the salmonid Salvelinus fontinalis resulted in increased genetic diversity and decreased genetic differentiation among recipient populations isolated by in‐stream barriers (Robinson et al. 2017). We advocate the use of long‐term genetic monitoring frameworks (e.g., Hogg et al. 2025) to detect early signs of genetic erosion of fragmented populations (Leroy et al. 2018; McLennan et al. 2025), so that translocations can be undertaken to increase the resilience and persistence of isolated populations. Author Contributions J.C.F.C., D.D., and J.H.L. developed the research concept and study design. J.C.F.C., B.W.L., A.L., and F.E.R. developed the molecular analysis. B.W.L. supervised lab work. J.C.F.C. performed field collection, lab work, molecular analysis, and drafted the initial manuscript. D.D. and J.H.L. supervised the study. All authors contributed critically to editing and drafting the manuscript. Funding This study was supported by the Lantau Conservation Fund (grant number: LCF‐RE‐2021‐02). Full funding for publishing this article for open access was provided by the University of California Libraries under a transformative open agreement with the publisher. Conflicts of Interest The authors declare no conflicts of interest. Supporting information Table S1: Demographic model parameters for the Rhinogobius similis population in Shek Pik Reservoir (SPR) and Shui Hau Stream (SHS) for input into FastSimCoal. Table S2: FastSimCoal output of demographic models. SPR, Shek Pik Reservoir; SHS, Shui Hau Stream. Figure S1: Genetic clusters of Rhinogobius duospilus according to PCA and STRUCTURE plots (K = 6) overlain on the five main paleodrainages of Hong Kong (redrawn from Fyfe et al. 2000); HPNS, Ha Pak Nai Stream; LSMR, Lower Shing Mun Reservoir; POS, Pui O Stream; SHS, Shui Hau Stream; SSWS, Sheung Sze Wan Stream; TTR, Tai Tam Tuk Reservoir. Figure S2: Isolation by distance plot of Rhinogobius duospilus based on Mantel tests on pairwise FST with Rousset's correction (Dgen) and geographic distances calculated from coordinates (Dgeo). Each point represents a pairwise comparison between two populations for a total of 15 comparisons among six populations. A correlation between geographic distance and genetic differentiation suggests isolation by distance. Figure S3: Isolation by distance plot of Rhinogobius similis based on Mantel tests on pairwise FST with Rousset's correction (Dgen) and geographic distances calculated from coordinates (Dgeo). Each point represents a pairwise comparison between two populations for a total of 21 comparisons among seven populations. A correlation between geographic distance and genetic differentiation suggests isolation by distance. Figure S4: Log‐likelihood distributions of models rerun with best parameters based on 100 expected SFS with 1,000,000 simulations per model; downstream—towards Shui Hau Stream (SHS); upstream—towards Shek Pik Reservoir (SPR); bidirectional—two‐way migration between SHS and SPR.