Ancestry imbalance and population structure in openSNP, a public direct-to-consumer genomic resource

Ancestry imbalance and population structure in openSNP, a public direct-to-consumer genomic resource | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Ancestry imbalance and population structure in openSNP, a public direct-to-consumer genomic resource Francesco Montinaro, Vincenzo Angiulli, Francesco Perrone, Raffaella Ribatti This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9305945/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 6 You are reading this latest preprint version Abstract Direct-to-consumer (DTC) genetic datasets shared through public platforms are increasingly used for citizen science and research, yet their population structure and representativeness remain poorly characterised. Here, we analysed 6,118 genotype files from openSNP, harmonised across three major DTC providers (23andMe, AncestryDNA, and FamilyTreeDNA), and compared them with worldwide reference populations from the 1000 Genomes Project (1KGP) and the Human Genome Diversity Project (HGDP). To characterise ancestry patterns, we applied five complementary approaches: principal component analysis, two supervised population-assignment methods based on k-Nearest Neighbours and Random Forest, ADMIXTURE, and PANE/NNLS ancestry decomposition. Random forest showed higher predictive performance than k-nearest neighbours on the independent HGDP test set (96.59% vs 85.23% accuracy), supporting robust assignment of openSNP samples. Across all methods, the dataset showed a marked excess of European ancestry. Supervised classification assigned 91.6–94% of individuals to European populations, ADMIXTURE estimated 91.4–95.1% European ancestry across K values, and PANE identified 85.6% of individuals with European ancestry as the dominant component. African, East Asian, and South Asian ancestries were each represented at roughly 2–3%, while American ancestry remained low, although a subset of individuals displayed substantial admixture. These results confirm and refine previous observations, primarily based on PCA, showing that ancestry imbalance in openSNP persists even when analysed with multiple orthogonal methods and a substantially larger sample size. Public DTC repositories remain valuable resources for methodological and population-genetic studies. Still, their strong ancestry skew should be considered in downstream analyses and in broader discussions of equity and representation in consumer genomics. Biological sciences/Genetics/Population genetics Biological sciences/Genetics/Population genetics/Genetic variation Direct-to-consumer genetics population structure ancestry inference openSNP genomic diversity Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction Following the sharp decline in sequencing and genotyping costs, the advent and rapid proliferation of Direct-to-Consumer (DTC) genetic testing services have revolutionised how individuals access and interact with their genetic information. Several companies offering genetic testing have generated and made available to users genotyping arrays or whole-genome sequencing data. However, the medical information provided in some DTC genetic testing company reports requires scrutiny due to its potential impact on users’ health. Consequently, this area has undergone strict regulation, especially following an initial period of stringent legislation( 1 – 4 ). A less problematic matter, although not free of risks and concerns, is the possibility of tracing users' genetic ancestral patterns, which is often one of the main reasons citizens purchase such services( 5 ). However, for the ancestry reports and interpretations, DTC companies may use different automated frameworks to infer ancestry-related information tailored to the geographic provenance of most of their users. They may overlook some ancestries or populations in favour of others. Moreover, they are often poorly detailed, leaving the user with only broad classifications. For this reason, some users prefer to analyse their own genomes autonomously, also with the help of the involved community ( 6 – 8 ). In this context, new platforms emerged that allow users to share genomic profiles, improving individual self-knowledge and laying the groundwork for citizen science projects, creating valuable resources for population genetics and human health research ( 9 , 10 ). However, these datasets, generated across diverse genotyping platforms and potentially biased by demographics, pose significant challenges for standardisation, harmonisation, and interpretation. Furthermore, analysing these data may provide insights into the accessibility of DTC services across society and citizens' willingness to share their data with the community. While most published studies on DTC data focus on discovering genetic associations and validating ancestry-prediction algorithms, a thorough characterisation of the genetic variation and structure intrinsic to these cohorts is fundamental. Understanding the patterns of genetic variation and the levels of admixture is crucial not only for correctly interpreting the results of association studies but also for evaluating the accuracy of widely used ancestry inference tools. Among others, one of the largest DTC genetic datasets freely available to the community is openSNP( 11 ), which started in 2011 and was definitively closed on 10th April 2025( 12 ) after collecting data from more than 6,000 individuals. An initial analysis of the openSNP genomes conducted in 2017 and reviewed in 2025( 13 ) has highlighted a significant imbalance in continental ancestry representation, with Europeans accounting for almost 95% of the analysed individuals, and the remaining proportions equally distributed between Africa and Asia. This pattern is very similar to the well-known imbalance in translational studies, such as genome-wide association studies, which, despite efforts over the last decade, continue to show under-representation of non-European groups. However, these estimates were obtained solely from Principal components analysis and do not account for subcontinental genetic structure. Moreover, this study includes only 2,280 individuals, a fraction of the final openSNP dataset. The present study aims to address these challenges by analysing a genetic dataset of 6,118 individuals sourced from the openSNP platform, which aggregates data from multiple DTC providers. In doing so, we compared the openSNP dataset with other freely available genomic datasets from worldwide populations. We applied five orthogonal methods to explore and evaluate genetic variation and structure across the openSNP dataset, including two ad hoc-designed K-NN and random forest tools for population assignment. The findings from these objectives will enhance understanding of the complex genetic structure in consumer genetics datasets and shed light on existing sampling biases. Materials and Methods Dataset acquisition Bulk genotype files were downloaded from the openSNP database in February 2024. The dataset comprised 6,967 genotype files, primarily generated by 23andMe (71.2%), AncestryDNA (16.7%), and FamilyTreeDNA (9.6%). As these three providers accounted for 97.5% of all samples, we excluded all other sources to minimise platform‑specific batch effects and ensure data consistency (Figure S1 ). Pre-processing, quality control, and standardisation Data preprocessing was performed separately for each of the three retained DTC platforms to account for platform-specific file formats and genomic build versions. For each platform, genotype files underwent integrity checks using the snps ( 14 ) Python library, which validates file structure, identifies the reference genome build, and detects potential data corruption. Files that failed integrity checks due to corruption, incomplete downloads, or incompatible formats (primarily .pdf and .zip) were logged and excluded from further analysis. All successfully validated files already aligned to the GRCh37/hg19 (build 37) were retained without modification, while files in other builds —GRCh36 or GRCh38— were lifted using the chain file-based remapping function implemented in the snps library. After that, quality control was performed using the GenomePrep ( 15 ) Python library, which applies platform-specific quality filters and converts genotype files to standardised Variant Call Format (VCF). GenomePrep validates chromosome identifiers, genomic positions, and allele designations, then generates VCF files with proper header information and genotype encoding in accordance with VCF v4.2 specifications. VCF files were converted to PLINK binary format using PLINK v1.9 ( 16 ), and datasets were merged iteratively, removing multiallelic variants. The three platform-specific unfiltered datasets were sequentially merged to create a unified openSNP dataset. Standard quality control filters were applied, excluding variants and individuals with call rates below 95% in both platform-specific and openSNP datasets. In doing so, we obtained a unique “DTC dataset” of 6,118 individuals and 73,382 SNPs with a 99.5% genotyping rate. Integration with reference population datasets To contextualise genomic variation within global population structure, we integrated the openSNP dataset with two large-scale reference panels: the 1000 Genomes Project Phase 3 (1KGP) and the Human Genome Diversity Project (HGDP)( 17 ). We downloaded the 1KGP dataset genotyped on Illumina Omni 2.5 array and converted it from VCF to PLINK binary format. We removed SNPs and individuals with more than 5% missing data. For HGDP, we downloaded the dataset genotyped with Illumina HuHap 650k, lifted the genomic coordinates to build 37, and applied the same filtering described above. The three quality-controlled datasets —openSNP, 1KGP, and HGDP— were merged using PLINK's --bmerge function, and multiallelic variants were excluded using the iterative exclusion procedure described above. A final round of quality control was applied to the integrated dataset using identical thresholds ( --geno 0.05 for variant call rate and --mind 0.05 for individual call rate) to remove any variants or individuals that acquired excessive missing data due to differential SNP coverage across platforms. The final merged dataset was sorted by natural sort order of family IDs and within-family individual IDs using the --indiv-sort PLINK command to ensure consistent sample ordering for downstream analyses. After excluding individuals and SNPs with more than 5% missing values, the final dataset contains 9,376 individuals and 66,843 variants, with a genotyping rate of 99.63%. Principal Component Analysis Principal Component Analysis (PCA) was conducted using EigenSoft's smartpca tool( 18 ), calculating the first 50 principal components calculated and no outlier removal. Data manipulation, visualisation, and statistical analyses were carried out using R version 4.4.2( 19 ). To improve the overall interpretation of the results, we pooled populations into African, American, European, East Asian, South-West Asian, and Oceanian groups, following the International Genome Sample Resource (IGSR) classification ( 20 ). The first 3 principal components were used for primary visualisation. Population assignment with k-NN and Random Forest We trained and evaluated two machine learning models for population assignment using genetic data: k-Nearest Neighbors (k-NN) and Random Forest (RF). The reference dataset consisted of samples from the 1000 Genomes Project (1KGP) used as the training set, while samples from the Human Genome Diversity Project (HGDP) served as an independent test set to evaluate model performance. For both models, we implemented a 10-fold cross-validation strategy, repeated 10 times, to assess model performance on the training data. For the k-NN model, we fixed k = 10 neighbours, whereas the Random Forest classifier was implemented using 500 decision trees. To evaluate model performance, we used confusion matrices and overall accuracy metrics. The models were tested on an independent HGDP-derived test subset to assess their generalisability across samples from different sources. It is important to note that the HGDP test subset did not fully overlap with the 1KGP training label space. Specifically, the evaluation set was restricted to HGDP populations with explicit 1KGP counterparts, namely BASQUE, JAPANESE, FRENCH, BANTUKENYA, CHS, HAN, HAN-NCHINA, ITALIAN, TUSCAN, and YORUBA, which were harmonised to the corresponding 1KGP labels IBS, JPT, GBR, LWK, CHB, TSI, and YRI. After prediction, class-level results and confusion-matrix target rows were filtered to retain only classes present in the test set. This same filtering procedure was applied to both the k-NN and random forest models. As a consequence, sensitivity, specificity, and F1-scores were reported only for classes represented in the HGDP test subset, whereas overall accuracy was computed on the complete prediction set and was therefore unaffected by class-level filtering. For the final application to openSNP samples of unknown ancestry, we extracted assignment probabilities and implemented a confidence threshold of 0.5; assignments with probabilities below this threshold were flagged as low confidence. All analyses were performed in R v4.4.2 using the following packages: randomForest for the RF implementation, class ( 21 ) for k-NN, and caret ( 22 ) for model training and evaluation. Parallel processing was utilised to optimise computational efficiency during the training step. ADMIXTURE To infer ancestral components of the openSNP dataset, we performed admixture ( 23 ) analysis with K ranging from 2 to 10 (S3). In detail, we have run ten iterations for each K value using the -s time and --cv options. To avoid bias from sample-size imbalance, we projected the openSNP samples onto the ancestral components inferred from all other individuals. Therefore, we first estimated the component proportions (K) and allele frequencies using the 1KGP and HGDP populations exclusively and subsequently projected the openSNP dataset using the -P option. PANE We inferred fine-scale ancestry proportions using PANE ( 24 ). This approach estimates ancestry proportions from multiple reference populations by performing non-negative least-squares analysis on principal component vectors. For the analysis, we first computed the mean PC coordinates for each of the 72 reference populations across the first 50 principal components. These population-specific mean vectors served as "donors" in the PANE regression framework. We then projected each of the 6,118 openSNP samples onto this reference space using the same 50 PCs as "recipients". The algorithm decomposed each openSNP sample's PC coordinates as a non-negative linear combination of the reference population mean vectors, yielding proportion estimates for all 72 populations. To facilitate interpretation and account for fine-scale population structure, we aggregated the population-level proportions to seven continental groups: Africa (including sub-Saharan African populations), Europe (including European populations), Middle East (including North African and Middle Eastern populations), South Asia (including Pakistani and Indian populations), East Asia (including East and Southeast Asian populations), Americas (including Native American and admixed American populations), and Oceania (including Papuan and Melanesian populations). Results Characterisation of the analysed openSNP dataset Following data curation and standardisation, our analysis included genotype data from 6,118 individuals within the openSNP dataset, sourced from three major Direct-to-Consumer (DTC) companies. The data processing pipeline yielded the following final filtered datasets: for 23andMe, starting with 4,961 initial files, the dataset was reduced to 4,430 individuals and 100,727 SNPs, compared to the 4,643 individuals and 3,394,458 SNPs retained before the final filtering step. The AncestryDNA dataset began with 1,161 files, yielding genetic data for 1,121 individuals and 389,000 SNPs after filtering, compared with the initial acquisition of 1,137 individuals and 1,531,219 SNPs. Finally, for FamilyTreeDNA, of the 672 initial files, the final dataset comprised 474 individuals and 177,520 SNPs, after an intermediate step that had retained 511 individuals and 1,724,813 SNPs. To provide an overall description of genetic variation across the whole sample, we merged the three company-specific unfiltered datasets, yielding an initial dataset of 6,291 individuals and 3,396,244 SNPs. After filtering, the final cleaned dataset comprised 6,118 individuals and more than 73,000 SNPs, with 4,488, 1,131, and 499 samples from 23andMe, AncestryDNA, and FamilyTreeDNA, respectively (Figure S1 ). We merged the DTC dataset with two available resources genotyped on two different Illumina arrays, the Human Genome Diversity Project and the 1000 Genome Project, resulting in a “Full Dataset” of 9,376 samples and 66,843 SNPs. All the subsequent analyses were performed on this dataset. Principal Component Analysis We assessed the genetic variation in the DTC samples using a PC Analysis, projecting the genetic variation of openSNP individuals into the PC space inferred from the HGDP and 1KGP populations, as shown in Fig. 1 . When the first two PCs are considered, the vast majority of samples fall near European samples, suggesting a high proportion of individuals of European origin. The remaining samples are scattered following two main clines, evident in PC2 (Fig. 1 A). A substantial proportion of individuals are distributed toward African groups, possibly due to the high African ancestry among African Americans ( 25 , 26 ). Similarly, a relatively high number of individuals stretches between European and East Asian, capturing both the native American Ancestry and the XX-century migration from East Asia to the Americas. When PC3 is taken into account, the impact of these two ancestries becomes discernible, revealing their presence. In all cases, some individuals are placed outside of this principal clines, suggesting the presence of more than two main ancestries and revealing a complex admixture scenario (Fig. 1 B). Model performance evaluation The k-NN model with k = 10 achieved an overall accuracy of 85.51% (SD ± 1.74%) and Cohen's Kappa of 0.846 (SD ± 0.018) across 100 resampling iterations. The RF model demonstrated superior performance with 97.42% overall accuracy (SD ± 0.96%) and Cohen's Kappa of 0.973 (SD ± 0.010) ( Supplementary Tables 1–4 ). Analysis of per-class performance during cross-validation revealed heterogeneous F1-scores across populations for the k-NN model. Four populations achieved perfect classification (F1 = 1.00): Finnish (FIN), Gujarati Indian (GIH), Luhya (LWK), and Peruvian (PEL). Several populations showed high performance with F1-scores exceeding 0.90, and for the worst classes, the second most frequent assignment is to a close population. For example, for CHS (F1 = 0.59), 56.9% of the individuals are assigned to CHB (Figure S2 A). The RF model demonstrated more consistent per-class performance, achieving perfect F1 Scores (1.00) across five populations: Finnish (FIN), British (GBR), Gujarati Indian (GIH), Luhya (LWK), and Peruvian (PEL). Moreover additional seven populations showed near-perfect classification with F1-scores exceeding 0.99 (TSI, CLM, IBS, YRI, JPT, PUR, MXL). The lowest F1-score was observed for Han Chinese in Beijing (CHB, F1 = 0.869), while all other populations exceeded 0.90 (Figure S2 C). When applied to the independent HGDP test dataset, the k-NN model achieved 85.23% overall accuracy (Figure S2 B). Three populations were perfectly classified: Iberian, Japanese, and Luhya, each with 100% sensitivity, 100% precision, and F1 = 1.00. Chinese samples showed 79.5% sensitivity and 100% precision (F1 = 0.886), while British samples showed 78.6% sensitivity and 100% precision (F1 = 0.880). Tuscan samples showed 100% sensitivity but 76.9% precision (F1 = 0.870). Yoruba samples showed 47.6% sensitivity with 100% precision, resulting in the lowest F1-score of 0.645. The RF model achieved 96.59% overall accuracy on the HGDP test set (Figure S2 D). Four populations demonstrated perfect classification with 100% sensitivity, 100% precision, and F1 = 1.00: Japanese, Luhya, and Yoruba. Chinese samples showed the highest performance among imperfectly classified groups with 97.7% sensitivity and 100% precision (F1 = 0.989). Tuscan samples demonstrated 100% sensitivity with 90.9% precision (F1 = 0.952), while Iberian samples showed 100% sensitivity with 88.9% precision (F1 = 0.941). British samples exhibited 82.1% sensitivity with 100% precision (F1 = 0.902). openSNPs population assignment using unsupervised models Application of both trained models to 6,118 openSNP user samples revealed predominant European ancestry across the cohort. For both models, we first considered raw assignments regardless of their probability, then filtered out those with a probability below 50%, obtaining high-confidence assignments. Looking at raw assignments, the k-NN model assigned 5,580 samples (91.2%) to European populations, 193 samples (3.2%) to East Asian populations, 173 samples (2.8%) to African populations, 127 samples (2.1%) to South Asian populations, and 45 samples (0.7%) to American populations. Considering only high-confidence assignments, the k-NN model assigned a total of 5,949 samples (97.2% of all samples). Among these, 91.6% (n = 5,452) were to European populations, 3.0% (n = 178) to East Asian populations, 2.7% (n = 158) to African populations, 2.1% (n = 127) to South Asian populations, and 0.6% (n = 34) to American populations. Instead, for the RF model, raw assignments show 5,485 samples (89.7%) assigned to European ancestry, 207 samples (3.4%) to African ancestry, 154 samples (2.5%) to East Asian ancestry, 137 samples (2.2%) to American ancestry, and 135 samples (2.2%) to South Asian ancestry. The confidence distribution differed markedly from k-NN, with only 4,077 samples (66.6% of all samples) assigned with high confidence. Of these, 94.0% (n = 3,833) were assigned to European ancestry, 2.8% (n = 115) to South Asian ancestry, 1.9% (n = 76) to East Asian ancestry, 0.8% (n = 32) to African ancestry, and 0.5% (n = 21) to American ancestry. The per-population assignment distribution analysis revealed a considerable imbalance in representation across populations, as shown in Fig. 2 . British ancestry (GBR) was the dominant assignment in both models, with the k-NN model assigning 4,266 samples (71.7% of 5,949 high-confident assignments, Fig. 2 A) and the RF model assigning 3,487 samples (85.5% of 4,077 Fig. 2 B). Tuscan (TSI) received the second-highest assignments: 1,144 samples (19.2%) from k-NN and 308 samples (7.6%) from RF. Together, these two populations accounted for over 90% of all assignments from k-NN and 93% from RF. The remaining populations showed substantially lower frequencies, with most receiving fewer than 100 assignments. Notable differences between models included RF's higher detection of admixed American populations (137 total assignments, including 98 MXL, 26 PUR, 11 CLM, 2 PEL) compared to k-NN (44 total assignments: 40 MXL, 5 PUR), and RF's more distributed European assignments (including 66 IBS and 43 FIN) compared to k-NN (1 IBS, 42 FIN). ADMIXTURE To infer the ancestry composition of the analysed individuals and assess the presence of multiple ancestries within individuals, we performed ADMIXTURE analysis with K ranging from 2 to 10 (Figure S3 ). At K = 4, the main components of the DTC dataset are differentiated. Most individuals show a very high proportion of European ancestry (mean = 0.831, SD = 0.072, Fig. 3 A), with minor contributions from Asian, African, and Native American components. A substantial group shows a high proportion of non-European ancestry. In fact, 152 individuals have a predominant African ancestry, while 152 have a modal Asian ancestry. For Native American Ancestry, 9 individuals have a predominant ancestry. At K = 7, which is the most supported according to the performed cross-validation analysis, we additionally identify a component modal in Papuan and Melanesian populations. Moreover, West Eurasian populations are mostly composed of two components, one modal in Northern Europe (yellow) and the other in Northern Africa and the Middle East (blue), but are present in high proportions in many Southern European populations. Moreover, South Asians show a high proportion of a component which is less present further East (Fig. 3 B). DTC individuals are mainly composed of admixture profiles very similar to those of Eurasian groups, although many individuals suggest different origins. Among the others, we identified 44 individuals characterised by a large Native American ancestral component, exceeding 40%, although only a single individual exceeded 60%. Furthermore, 135 individuals have more than 60% of East Asian component. Similarly, a subset of samples (n = 133) is characterised by an ancestral component which is modal in South Asia. We also identified 172 individuals with an ancestral profile consistent with that observed in the ASW group and are therefore likely to be African American. We did not identify any individuals with more than 90% African ancestry. PANE To characterise fine-scale admixture patterns across the openSNP cohort, we estimated ancestral proportions using PANE, which implements Non-Negative Least Squares (NNLS) regression on principal component coordinates (Fig. 4 ). Considering proportions estimated from the first 50 principal components, the analysis confirmed predominant European ancestry across the cohort, with a median proportion of 83.5% (mean = 73.5%, SD = 24.4%). The remaining continental components showed substantially lower contributions: American ancestry had a median of 4.44% (mean = 6.44%, SD = 8.42%), Middle Eastern ancestry showed a median of 2.82% (mean = 5.72%, SD = 9.05%), South Asian ancestry had a median of 2.36% (mean = 5.08%, SD = 12.6%), African ancestry showed a median of 2.05% (mean = 4.95%, SD = 12.7%), East Asian ancestry had a median of 0.78% (mean = 3.93%, SD = 14%), and Oceanian ancestry was minimal with a median of 0.31% (mean = 0.52%, SD = 1.37%). The distribution of ancestry proportions revealed considerable heterogeneity within the cohort. While the majority of samples showed high European ancestry (> 66%), a substantial subset exhibited significant non-European components, consistent with the population assignment results from k-NN and RF models. The high standard deviations observed for African (SD = 12.7%), East Asian (SD = 14%), and South Asian (SD = 12.6%) components indicate the presence of both predominantly European individuals and those with substantial ancestry from these regions. Using a threshold of 50% for dominant ancestry assignment, we identified 5,237 individuals (85.6%) with predominant European ancestry, 160 individuals (2.62%) with African ancestry, 149 individuals (2.44%) with East Asian ancestry, 129 individuals (2.11%) with South Asian ancestry, 67 individuals (1.1%) with American ancestry, 49 individuals (0.8%) with Middle Eastern ancestry, one individual (0.02%) with Oceanian ancestry, and 326 individuals (5.33%) with uncertain/admixed profiles showing no single dominant component exceeding the threshold. Discussion Here, we evaluated genetic variation in the openSNP dataset, comprising 6,118 individuals from three major direct-to-consumer genetic testing companies, providing essential insights into both the genetic ancestry of DTC testing users and the effectiveness of different computational approaches for population assignment. Our analysis revealed remarkable variability in the genomic ancestry of openSNP users, while still highlighting a strong imbalance in continental ancestries. In fact, in principal component analysis, most of the analysed samples fall within the European variation, with much fewer individuals within the African and Asian variation. However, the observation of distinct genetic clines extending toward African and East Asian populations highlights complex admixture patterns, particularly evident in individuals with African-American and Asian-American ancestry. However, from PCA alone, it is often challenging to estimate each individual's exact ancestry composition, especially when substantial recent admixture has occurred, as is the case in modern human populations. Therefore, to provide an extensive picture of genomic ancestry variation, we performed three ancestry inference and population assignment methods. This allowed us to perform a more reliable and detailed population assignment, beyond the classification into three primary continents reported in previous surveys (Supplementary Table 5). Based on cross-validation analysis, our population assignment methods using k-NN and RF have shown high accuracy and sensitivity, with the latter achieving slightly better predictive performance, suggesting a higher capability of ensemble methods than distance-based methods. When applied to the openSNP dataset, these methods confirmed the overwhelming European predominance in the 6,118 openSNP samples (91.6–94% for k-NN and RF, respectively). The minimal representation of East Asian (3-1.9%), African (2.7–0.8%), South Asian (2.1–2.8%), and American (0.6 − 0.5%) ancestries highlights the limited global reach and persistent demographic imbalances in open genomic data repositories. ADMIXTURE further supported these findings, revealing subtle ancestry components that discrete population assignments might miss. The high proportion of European ancestry in most individuals, coupled with varying contributions from Asian, African, and Native American components, demonstrates the complexity of genealogy in modern populations. These results align with the PCA findings and provide a more nuanced view of population structure. The PANE analysis, incorporating non-negative least squares with PCA, offered additional resolution by quantifying ancestry proportions, showing an average European ancestry of 73% (median = 83.1%, sd = 24.4%) across the dataset. The relatively low proportion of American (mean = 6.4%, median = 4.4%, sd = 8.4%), Asian and African ancestry reflects both historical migration patterns and sampling biases in DTC testing demographics. Taken together, our results have several important implications. First, we highlighted a wide variation in the openSNP DTC dataset. Nevertheless, we confirmed a strong imbalance in ancestry composition across the whole dataset, even when more refined methods than PCA are used. When compared with a recent survey of a much more limited set of individuals and methods performed in 2017, we uncovered some differences. In fact, we observed a slight increase in the Asian proportion, which shows a higher proportion than previously reported (5% vs 2.9% in Corpas et al 2017). On the other hand, the African ancestry proportion increased only slightly from 2.2% to 2.8%, demonstrating that this market is unequally distributed across the global population. These results reveal that, despite the DTC's recent initiatives to expand customers' sociocultural and ancestral backgrounds, these efforts had only a limited effect. However, even though access to DTC services for individuals of non-European ancestry has increased, their users are more reluctant to share their data. In this context, further studies documenting both access to and willingness to share genetic data would be needed to provide a better overview of the observed disparities, which have a similar magnitude to those observed in translational studies Declarations Data Availability The assembled analysed genotype data from openSNP cohort ( DTC dataset ) and the Full Dataset are available at https://doi.org/10.5281/zenodo.18493898. Acknowledgments Vincenzo Angiulli is a PhD student within the European School of Molecular Medicine (SEMM). Author Contributions V.A. and F.M. conceived and designed the study. V.A. performed data processing, statistical analyses, and ancestry inference. F.P. and R.M.R. contributed to data interpretation and methodological validation. V.A. wrote the original draft of the manuscript. F.M. supervised the study and provided critical revisions. All authors reviewed and approved the final version. Fundings The authors declare that no funds, grants, or other support were received during the preparation and development of this manuscript. Ethical Approval All data in this study were obtained from publicly available resources. No new human subjects were recruited, and no individually identifiable information was used. Ethical approval and informed consent were obtained by the original primary studies. No additional ethical approval was required for this secondary analysis. Competing Interests The authors declare no competing interests. References Commissioner O of the. FDA [Internet]. FDA; 2024 [cited 2026 Mar 17]. FDA authorizes first direct-to-consumer test for detecting genetic variants that may be associated with medication metabolism. Available from: https://www.fda.gov/news-events/press-announcements/fda-authorizes-first-direct-consumer-test-detecting-genetic-variants-may-be-associated-medication Yim SH, Chung YJ. Reflections on the US FDA’s Warning on Direct-to-Consumer Genetic Testing. Genomics Inform. 2014;12(4):151–5. doi: 10.5808/GI.2014.12.4.151 Annas GJ, Elias S. 23andMe and the FDA. New England Journal of Medicine. 2014;370(11):985–8. doi: 10.1056/NEJMp1316367 Su P. Direct-to-Consumer Genetic Testing: A Comprehensive View. Yale J Biol Med. 2013;86(3):359–65. PubMed PMID: 24058310; PubMed Central PMCID: PMC3767220. Roth WD, Lyon KA. Genetic Ancestry Tests and Race: Who Takes Them, Why, and How Do They Affect Racial Identities? In: Gates J Henry Louis, Suzuki K, Von Vacano DA, editors. Reconsidering Race: Social Science Perspectives on Racial Categories in the Age of Genomics [Internet]. Oxford University Press; 2018 [cited 2026 Mar 23]. p. 0. Available from: https://doi.org/10.1093/oso/9780190465285.003.0008 doi:10.1093/oso/9780190465285.003.0008 Majumder MA, Guerrini CJ, McGuire AL. Direct-to-Consumer Genetic Testing: Value and Risk. Annual Review of Medicine. 2021;72(Volume 72, 2021):151–66. doi: 10.1146/annurev-med-070119-114727 Tandy-Connor S, Guiltinan J, Krempely K, LaDuca H, Reineke P, Gutierrez S, et al. False-positive results released by direct-to-consumer genetic tests highlight the importance of clinical confirmation testing for appropriate patient care. Genetics in Medicine. 2018;20(12):1515–21. doi: 10.1038/gim.2018.38 Nelson SC, Bowen DJ, Fullerton SM. Third-Party Genetic Interpretation Tools: A Mixed-Methods Study of Consumer Motivation and Behavior. The American Journal of Human Genetics. 2019;105(1):122–31. doi: 10.1016/j.ajhg.2019.05.014 Wiggins A, Wilbanks J. The Rise of Citizen Science in Health and Biomedical Research. The American Journal of Bioethics. 2019;19(8):3–14. doi: 10.1080/15265161.2019.1619859 PubMed PMID: 31339831. Prainsack B. The Powers of Participatory Medicine. PLOS Biology. 2014;12(4):e1001837. doi: 10.1371/journal.pbio.1001837 Greshake B, Bayer PE, Rausch H, Reda J. openSNP–A Crowdsourced Web Resource for Personal Genomics. PLOS ONE. 2014;9(3):e89204. doi: 10.1371/journal.pone.0089204 openSNP [Internet]. [cited 2026 Mar 17]. openSNP. Available from: https://opensnp.github.io/ Bridging genomics’ greatest challenge: The diversity gap. Cell Genomics. 2025;5(1):100724. doi: 10.1016/j.xgen.2024.100724 Riha A. apriha/snps [Python] [Internet]. 2026 [cited 2026 Mar 17]. Available from: https://github.com/apriha/snps A survey of direct-to-consumer genotype data, and quality control tool (GenomePrep) for research. Computational and Structural Biotechnology Journal. 2021;19:3747–54. doi: 10.1016/j.csbj.2021.06.040 Chang CC, Chow CC, Tellier LC, Vattikuti S, Purcell SM, Lee JJ. Second-generation PLINK: rising to the challenge of larger and richer datasets. Gigascience. 2015;4:7. doi: 10.1186/s13742-015-0047-8 PubMed PMID: 25722852; PubMed Central PMCID: PMC4342193. A global reference for human genetic variation. Nature. 2015;526(7571):68–74. doi: 10.1038/nature15393 Price AL, Patterson NJ, Plenge RM, Weinblatt ME, Shadick NA, Reich D. Principal components analysis corrects for stratification in genome-wide association studies. Nat Genet. 2006;38(8):904–9. doi: 10.1038/ng1847 R: The R Project for Statistical Computing [Internet]. [cited 2026 Mar 18]. Available from: https://www.r-project.org/ Fairley S, Lowy-Gallego E, Perry E, Flicek P. The International Genome Sample Resource (IGSR) collection of open human genomic variation resources. Nucleic Acids Res. 2020;48(D1):D941–7. doi: 10.1093/nar/gkz836 Modern Applied Statistics with S, 4th ed [Internet]. [cited 2026 Mar 18]. Available from: https://www.stats.ox.ac.uk/pub/MASS4/ Kuhn M. Building Predictive Models in R Using the caret Package. Journal of Statistical Software. 2008;28:1–26. doi: 10.18637/jss.v028.i05 Alexander DH, Novembre J, Lange K. Fast model-based estimation of ancestry in unrelated individuals. Genome Res. 2009;19(9):1655–64. doi: 10.1101/gr .094052.109 PubMed PMID: 19648217; PubMed Central PMCID: PMC2752134. de Gennaro L, Molinaro L, Raveane A, Santonastaso F, Saponetti SS, Massi MC, et al. PANE: fast and reliable ancestral reconstruction on ancient genotype data with non-negative least square and principal component analysis. Genome Biol. 2025;26(1):29. doi: 10.1186/s13059-025-03491-z PubMed PMID: 39934833; PubMed Central PMCID: PMC11818073. Micheletti SJ, Bryc K, Ancona Esselmann SG, Freyman WA, Moreno ME, Poznik GD, et al. Genetic Consequences of the Transatlantic Slave Trade in the Americas. The American Journal of Human Genetics. 2020;107(2):265–77. doi: 10.1016/j.ajhg.2020.06.012 Ongaro L, Scliar MO, Flores R, Raveane A, Marnetto D, Sarno S, et al. The Genomic Impact of European Colonization of the Americas. Current Biology. 2019;29(23):3974–3986.e4. doi: 10.1016/j.cub.2019.09.076 Additional Declarations There is no duality of interest Supplementary Files FigureS1.png Figure S1 Sample distribution in the openSNP dataset by genotyping provider. Bar plot showing the number of samples from each DTC genetic testing company after quality control and filtering. The final dataset includes 6,118 individuals: 4,488 from 23andMe, 1,131 from AncestryDNA, and 499 from FamilyTreeDNA. FIgureS2.png Figure S2 Confusion matrices for k-NN and Random Forest population assignment models. A. k-Nearest Neighbors (k=10) performance on 10-fold cross-validation (repeated 10 times) using 1000 Genomes Project training data across all 18 populations (overall accuracy: 85.51%, Kappa: 0.844). B. k-NN performance on the HGDP test dataset comprising seven population pairs mapped to corresponding 1000 Genomes populations (overall accuracy: 85.23%). C.Random Forest (mtry=2, 500 trees) performance on 10-fold cross-validation (repeated 10 times) across all 18 populations (overall accuracy: 97.42%, Kappa: 0.972). D. RF performance on the HGDP test dataset using the same population mapping as panel B (overall accuracy: 96.59%). Colour intensity represents the number of samples assigned to each population pair. FigureS3.png Figure S3 Detailed ADMIXTURE inference of the openSNP dataset. ADMIXTURE analysis performed for each K ranging from 2 to 10. On the left column, population average ancestral components are shown. On the right column, the corresponding individual ancestries are shown with colors indicating the proportion of each K-dominant continental ancestry. SupplementaryTable1.xlsx Supplementary Table S1 KNN per-class performance metrics in internal cross-validation. Per-class classification performance of the k-Nearest Neighbors model (k = 10) on the 1KGP training set, estimated from repeated 10-fold cross-validation. Reported metrics include Sensitivity, Specificity, Positive Predictive Value, Balanced Accuracy, and F1 score. SupplementaryTable2.xlsx Supplementary Table S2 KNN per-class performance metrics in external HGDP test set. Per-class classification performance of the k-Nearest Neighbors model (k = 10) on the independent HGDP-derived test set after population harmonisation to 1KGP labels. Reported metrics include Sensitivity, Specificity, Positive Predictive Value, and F1 score. SupplementaryTable3.xlsx Supplementary Table S3 Random Forest per-class performance metrics in internal cross-validation. Per-class classification performance of the Random Forest model on the 1KGP training set, estimated from repeated 10-fold cross-validation. Reported metrics include Sensitivity, Specificity, Positive Predictive Value, Balanced Accuracy, and F1 score. SupplementaryTable4.xlsx Supplementary Table S4 Random Forest per-class performance metrics in external HGDP test set. Per-class classification performance of the Random Forest model on the independent HGDP-derived test set after population harmonisation to 1KGP labels. Reported metrics include Sensitivity, Specificity, Positive Predictive Value, and F1 score. SupplementaryTable5.xlsx Supplementary Table S5 Percentage of Continental ancestry in the openSNP dataset inferred using four different methods. Continental ancestry percentages in the openSNP dataset, estimated by each method (kNN, RF, Admixture with K=4, K=7, and PANE) and reported across Africa, East Asia, South Asia, Europe, America, and Others. Cite Share Download PDF Status: Under Review Version 1 posted Reviewer # 2 agreed at journal 08 May, 2026 Reviewer # 1 agreed at journal 08 May, 2026 Reviewers invited by journal 28 Apr, 2026 Submission checks completed at journal 02 Apr, 2026 Editor assigned by journal 02 Apr, 2026 First submitted to journal 02 Apr, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9305945","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":617116338,"identity":"187444e4-14d2-47cd-aa58-32c50d9ab129","order_by":0,"name":"Francesco Montinaro","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABCUlEQVRIiWNgGAWjYBAC9gZUvgSIYHwAZCQA6QZ05SDAcwCFmwDWwmwA1dKITQ+6FjDJJgFlYrWGR/rwswc/Kurk+Rm4Ex8X/rDI4599+Fg17w4go4G5/QE2LXxp5oY9Zw4bzmzg3Ww8I0GiWOJcWtpt3jNAxgHsDrPnYTCTZmw7kGBwgHebNE+CRGLDGR6z27xtQAYOLTw87N+kGf/VIbTMB2opBmmZj1MLD9CWBmaElg1ALcwgLRtwaymT7DkG9Esz0C88aRKJG8+wJUvObZMoNjzM2DgDu8O2SfyoAYYYe+/Gxzw2dYnzzjAf/PC2rS5P7nj7gw9YtCAAMxEio2AUjIJRMAqIBADWhlxuiFrLNgAAAABJRU5ErkJggg==","orcid":"","institution":"Università degli Studi di Bari \"Aldo Moro\"","correspondingAuthor":true,"prefix":"","firstName":"Francesco","middleName":"","lastName":"Montinaro","suffix":""},{"id":617116339,"identity":"60a079c7-c0e6-4ae2-994b-6224bbaf0bdd","order_by":1,"name":"Vincenzo Angiulli","email":"","orcid":"https://orcid.org/0009-0000-2580-0467","institution":"University of Bari \"Aldo Moro\"","correspondingAuthor":false,"prefix":"","firstName":"Vincenzo","middleName":"","lastName":"Angiulli","suffix":""},{"id":617116340,"identity":"f4750581-540a-4858-a606-d5cc81219521","order_by":2,"name":"Francesco Perrone","email":"","orcid":"","institution":"University of Bari \"Aldo Moro\"","correspondingAuthor":false,"prefix":"","firstName":"Francesco","middleName":"","lastName":"Perrone","suffix":""},{"id":617116341,"identity":"278d9458-321f-458b-ac79-d21e138ca5e5","order_by":3,"name":"Raffaella Ribatti","email":"","orcid":"https://orcid.org/0000-0001-5534-0123","institution":"University of Bari \"Aldo Moro\"","correspondingAuthor":false,"prefix":"","firstName":"Raffaella","middleName":"","lastName":"Ribatti","suffix":""}],"badges":[],"createdAt":"2026-04-02 17:50:23","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9305945/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9305945/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":109068060,"identity":"d8aa7f0e-511c-4bd9-a11d-bcaa7434f7b1","added_by":"auto","created_at":"2026-05-12 10:03:21","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":2332811,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003e\u003cstrong\u003eGenetic ancestry of openSNP users revealed by projection onto reference population structure.\u003c/strong\u003e\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003ePrincipal component analysis of 9,376 individuals based on 66,843 SNPs. Principal components were computed on reference populations (1KGP and HGDP), and openSNP samples (n=6,118, gray) were projected onto the reference PC space. Samples are coloured by continental ancestry. A. PC1 vs PC2. B. PC1 vs PC3.\u003c/p\u003e","description":"","filename":"Figure1.png","url":"https://assets-eu.researchsquare.com/files/rs-9305945/v1/adb1e4f103d97db855cec0ac.png"},{"id":108607174,"identity":"243c5b71-0219-4b30-ad8b-4c633335e8af","added_by":"auto","created_at":"2026-05-06 12:28:19","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":623470,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003e\u003cstrong\u003ePopulation assignment distribution of openSNP samples using unsupervised models.\u003c/strong\u003e\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003ek-Nearest Neighbors (A) and Random Forest (B) classification results showing log₁₀-transformed sample counts per assigned population. Bar colours indicate continental ancestry. Opacity represents prediction confidence: solid bars show \u0026gt;50% probability assignments, and transparent bars show ≤50% probability. Percentages indicate the proportions of intra-population confidence predictions.\u003c/p\u003e","description":"","filename":"Figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-9305945/v1/85574a6f854777fe152500f7.png"},{"id":108805919,"identity":"e87710af-7d75-47b7-94a3-84c514361acb","added_by":"auto","created_at":"2026-05-08 15:27:11","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":641783,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003e\u003cstrong\u003eContinental ancestry composition of openSNP samples estimated by ADMIXTURE.\u003c/strong\u003e\u003c/em\u003e\u003cbr\u003e\nOn the top panel, population average ancestral components are shown. On the bottom panel, individual ancestries are shown with colors indicating the proportion of K-dominant continental ancestry. A) ADMIXTURE for K=4 – Europe (orange), Africa (blue), Asia (green), America (red). B) ADMIXTURE for K=7 – Europe (yellow), Africa (purple), South Asia (pink), East Asia (red), America (orange), Middle East (blue), Oceania (green).\u003c/p\u003e","description":"","filename":"Figure3.png","url":"https://assets-eu.researchsquare.com/files/rs-9305945/v1/690a287f26ecb0082bf22d0e.png"},{"id":108805439,"identity":"1a87311e-12fb-45ff-801d-e49628098325","added_by":"auto","created_at":"2026-05-08 15:25:58","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":381744,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003e\u003cstrong\u003eContinental ancestry composition of openSNP samples estimated by PANE.\u003c/strong\u003e\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eEach vertical bar represents an individual, with colours indicating the proportion of ancestry from seven continental groups: Africa (purple), Europe (yellow), Middle East (blue), South Asia (pink), East Asia (red), Americas (orange), and Oceania (green).\u003c/p\u003e","description":"","filename":"Figure4.png","url":"https://assets-eu.researchsquare.com/files/rs-9305945/v1/5b9f9a5a2ff52ff4478f9ba5.png"},{"id":109069236,"identity":"c5302f4c-6ca7-4f80-b034-5778279a04ed","added_by":"auto","created_at":"2026-05-12 10:21:40","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":6892053,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9305945/v1/0238fd62-a1c7-4c03-9733-9c82e981b2b1.pdf"},{"id":108805559,"identity":"926c42ce-fdb7-412a-bebd-cdc63bb2071a","added_by":"auto","created_at":"2026-05-08 15:26:15","extension":"png","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":223624,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eFigure S1\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u003cstrong\u003eSample distribution in the openSNP dataset by genotyping provider.\u003c/strong\u003e\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eBar plot showing the number of samples from each DTC genetic testing company after quality control and filtering. The final dataset includes 6,118 individuals: 4,488 from 23andMe, 1,131 from AncestryDNA, and 499 from FamilyTreeDNA.\u003c/p\u003e","description":"","filename":"FigureS1.png","url":"https://assets-eu.researchsquare.com/files/rs-9305945/v1/064a22b21cad42f976547204.png"},{"id":108805452,"identity":"63f3e7c4-057c-4f7c-a0e0-8bd2bf817fe7","added_by":"auto","created_at":"2026-05-08 15:26:01","extension":"png","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":9187576,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eFigure S2\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u003cstrong\u003eConfusion matrices for k-NN and Random Forest population assignment models.\u003c/strong\u003e\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eA.\u003c/strong\u003e k-Nearest Neighbors (k=10) performance on 10-fold cross-validation (repeated 10 times) using 1000 Genomes Project training data across all 18 populations (overall accuracy: 85.51%, Kappa: 0.844). \u003cstrong\u003eB.\u003c/strong\u003e k-NN performance on the HGDP test dataset comprising seven population pairs mapped to corresponding 1000 Genomes populations (overall accuracy: 85.23%). \u003cstrong\u003eC.\u003c/strong\u003eRandom Forest (mtry=2, 500 trees) performance on 10-fold cross-validation (repeated 10 times) across all 18 populations (overall accuracy: 97.42%, Kappa: 0.972). \u003cstrong\u003eD. \u003c/strong\u003eRF performance on the HGDP test dataset using the same population mapping as panel B (overall accuracy: 96.59%).\u003c/p\u003e\n\u003cp\u003eColour intensity represents the number of samples assigned to each population pair.\u003c/p\u003e","description":"","filename":"FIgureS2.png","url":"https://assets-eu.researchsquare.com/files/rs-9305945/v1/1927dacb378bcb602e2d3215.png"},{"id":108607175,"identity":"44d6353c-7705-4de9-afb6-16c552149c9d","added_by":"auto","created_at":"2026-05-06 12:28:19","extension":"png","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":1265107,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003e\u003cstrong\u003eFigure S3\u003c/strong\u003e\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u003cstrong\u003eDetailed ADMIXTURE inference of the openSNP dataset.\u003c/strong\u003e\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eADMIXTURE analysis performed for each K ranging from 2 to 10. On the left column, population average ancestral components are shown. On the right column, the corresponding individual ancestries are shown with colors indicating the proportion of each K-dominant continental ancestry.\u003c/p\u003e","description":"","filename":"FigureS3.png","url":"https://assets-eu.researchsquare.com/files/rs-9305945/v1/5ef77519ed45010beaa4e7f2.png"},{"id":108805024,"identity":"6e519ca3-b517-4040-8239-f576f6ea2c0d","added_by":"auto","created_at":"2026-05-08 15:24:32","extension":"xlsx","order_by":4,"title":"","display":"","copyAsset":false,"role":"supplement","size":29234,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSupplementary Table S1\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u003cstrong\u003eKNN per-class performance metrics in internal cross-validation.\u003c/strong\u003e\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003ePer-class classification performance of the k-Nearest Neighbors model (k = 10) on the 1KGP training set, estimated from repeated 10-fold cross-validation. Reported metrics include Sensitivity, Specificity, Positive Predictive Value, Balanced Accuracy, and F1 score.\u003c/p\u003e","description":"","filename":"SupplementaryTable1.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-9305945/v1/042edf79eec17725c40d182b.xlsx"},{"id":108607182,"identity":"6d7e83a7-e6cd-45f9-b39e-076a3fad7702","added_by":"auto","created_at":"2026-05-06 12:28:19","extension":"xlsx","order_by":5,"title":"","display":"","copyAsset":false,"role":"supplement","size":13706,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSupplementary Table S2\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u003cstrong\u003eKNN per-class performance metrics in external HGDP test set.\u003c/strong\u003e\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003ePer-class classification performance of the k-Nearest Neighbors model (k = 10) on the independent HGDP-derived test set after population harmonisation to 1KGP labels. Reported metrics include Sensitivity, Specificity, Positive Predictive Value, and F1 score.\u003c/p\u003e","description":"","filename":"SupplementaryTable2.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-9305945/v1/d36053d873107c58c530b93e.xlsx"},{"id":108805191,"identity":"914284ec-0d02-4cc1-8432-24659c163fd7","added_by":"auto","created_at":"2026-05-08 15:25:09","extension":"xlsx","order_by":6,"title":"","display":"","copyAsset":false,"role":"supplement","size":14917,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSupplementary Table S3\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u003cstrong\u003eRandom Forest per-class performance metrics in internal cross-validation.\u003c/strong\u003e\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003ePer-class classification performance of the Random Forest model on the 1KGP training set, estimated from repeated 10-fold cross-validation. Reported metrics include Sensitivity, Specificity, Positive Predictive Value, Balanced Accuracy, and F1 score.\u003c/p\u003e","description":"","filename":"SupplementaryTable3.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-9305945/v1/2c7da48eaf6c61f119a61272.xlsx"},{"id":108607179,"identity":"4a91d266-9e0a-489b-83fe-5c1ea3d1f33d","added_by":"auto","created_at":"2026-05-06 12:28:19","extension":"xlsx","order_by":7,"title":"","display":"","copyAsset":false,"role":"supplement","size":13781,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSupplementary Table S4\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u003cstrong\u003eRandom Forest per-class performance metrics in external HGDP test set.\u003c/strong\u003e\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003ePer-class classification performance of the Random Forest model on the independent HGDP-derived test set after population harmonisation to 1KGP labels. Reported metrics include Sensitivity, Specificity, Positive Predictive Value, and F1 score.\u003c/p\u003e","description":"","filename":"SupplementaryTable4.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-9305945/v1/5abb3507551eb24d97ee6fb0.xlsx"},{"id":108805431,"identity":"19bd911b-be10-4d7e-9851-ad29a48dfe8e","added_by":"auto","created_at":"2026-05-08 15:25:57","extension":"xlsx","order_by":8,"title":"","display":"","copyAsset":false,"role":"supplement","size":78794,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSupplementary Table S5\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u003cstrong\u003ePercentage of Continental ancestry in the openSNP dataset inferred using four different methods.\u003c/strong\u003e\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eContinental ancestry percentages in the openSNP dataset, estimated by each method (kNN, RF, Admixture with K=4, K=7, and PANE) and reported across Africa, East Asia, South Asia, Europe, America, and Others.\u003c/p\u003e","description":"","filename":"SupplementaryTable5.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-9305945/v1/6986684c19b843da834cceec.xlsx"}],"financialInterests":"There is no duality of interest","formattedTitle":"Ancestry imbalance and population structure in openSNP, a public direct-to-consumer genomic resource","fulltext":[{"header":"Introduction","content":"\u003cp\u003eFollowing the sharp decline in sequencing and genotyping costs, the advent and rapid proliferation of Direct-to-Consumer (DTC) genetic testing services have revolutionised how individuals access and interact with their genetic information. Several companies offering genetic testing have generated and made available to users genotyping arrays or whole-genome sequencing data. However, the medical information provided in some DTC genetic testing company reports requires scrutiny due to its potential impact on users\u0026rsquo; health. Consequently, this area has undergone strict regulation, especially following an initial period of stringent legislation(\u003cspan additionalcitationids=\"CR2 CR3\" citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eA less problematic matter, although not free of risks and concerns, is the possibility of tracing users' genetic ancestral patterns, which is often one of the main reasons citizens purchase such services(\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e). However, for the ancestry reports and interpretations, DTC companies may use different automated frameworks to infer ancestry-related information tailored to the geographic provenance of most of their users. They may overlook some ancestries or populations in favour of others. Moreover, they are often poorly detailed, leaving the user with only broad classifications. For this reason, some users prefer to analyse their own genomes autonomously, also with the help of the involved community\u003c/p\u003e \u003cp\u003e(\u003cspan additionalcitationids=\"CR7\" citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eIn this context, new platforms emerged that allow users to share genomic profiles, improving individual self-knowledge and laying the groundwork for citizen science projects, creating valuable resources for population genetics and human health research (\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e). However, these datasets, generated across diverse genotyping platforms and potentially biased by demographics, pose significant challenges for standardisation, harmonisation, and interpretation. Furthermore, analysing these data may provide insights into the accessibility of DTC services across society and citizens' willingness to share their data with the community.\u003c/p\u003e \u003cp\u003eWhile most published studies on DTC data focus on discovering genetic associations and validating ancestry-prediction algorithms, a thorough characterisation of the genetic variation and structure intrinsic to these cohorts is fundamental. Understanding the patterns of genetic variation and the levels of admixture is crucial not only for correctly interpreting the results of association studies but also for evaluating the accuracy of widely used ancestry inference tools. Among others, one of the largest DTC genetic datasets freely available to the community is openSNP(\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e), which started in 2011 and was definitively closed on 10th April 2025(\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e) after collecting data from more than 6,000 individuals.\u003c/p\u003e \u003cp\u003eAn initial analysis of the openSNP genomes conducted in 2017 and reviewed in 2025(\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e) has highlighted a significant imbalance in continental ancestry representation, with Europeans accounting for almost 95% of the analysed individuals, and the remaining proportions equally distributed between Africa and Asia. This pattern is very similar to the well-known imbalance in translational studies, such as genome-wide association studies, which, despite efforts over the last decade, continue to show under-representation of non-European groups. However, these estimates were obtained solely from Principal components analysis and do not account for subcontinental genetic structure. Moreover, this study includes only 2,280 individuals, a fraction of the final openSNP dataset.\u003c/p\u003e \u003cp\u003eThe present study aims to address these challenges by analysing a genetic dataset of 6,118 individuals sourced from the openSNP platform, which aggregates data from multiple DTC providers.\u003c/p\u003e \u003cp\u003eIn doing so, we compared the openSNP dataset with other freely available genomic datasets from worldwide populations. We applied five orthogonal methods to explore and evaluate genetic variation and structure across the openSNP dataset, including two ad hoc-designed K-NN and random forest tools for population assignment.\u003c/p\u003e \u003cp\u003eThe findings from these objectives will enhance understanding of the complex genetic structure in consumer genetics datasets and shed light on existing sampling biases.\u003c/p\u003e"},{"header":"Materials and Methods","content":"\u003cp\u003eDataset acquisition\u003c/p\u003e \u003cp\u003eBulk genotype files were downloaded from the openSNP database in February 2024. The dataset comprised 6,967 genotype files, primarily generated by 23andMe (71.2%), AncestryDNA (16.7%), and FamilyTreeDNA (9.6%). As these three providers accounted for 97.5% of all samples, we excluded all other sources to minimise platform‑specific batch effects and ensure data consistency (Figure \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003ePre-processing, quality control, and standardisation\u003c/p\u003e \u003cp\u003eData preprocessing was performed separately for each of the three retained DTC platforms to account for platform-specific file formats and genomic build versions. For each platform, genotype files underwent integrity checks using the \u003cb\u003esnps\u003c/b\u003e(\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e) Python library, which validates file structure, identifies the reference genome build, and detects potential data corruption. Files that failed integrity checks due to corruption, incomplete downloads, or incompatible formats (primarily .pdf and .zip) were logged and excluded from further analysis. All successfully validated files already aligned to the GRCh37/hg19 (build 37) were retained without modification, while files in other builds \u0026mdash;GRCh36 or GRCh38\u0026mdash; were lifted using the chain file-based remapping function implemented in the \u003cem\u003esnps\u003c/em\u003e library.\u003c/p\u003e \u003cp\u003eAfter that, quality control was performed using the \u003cb\u003eGenomePrep\u003c/b\u003e(\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e) Python library, which applies platform-specific quality filters and converts genotype files to standardised Variant Call Format (VCF). GenomePrep validates chromosome identifiers, genomic positions, and allele designations, then generates VCF files with proper header information and genotype encoding in accordance with VCF v4.2 specifications.\u003c/p\u003e \u003cp\u003eVCF files were converted to PLINK binary format using PLINK v1.9 (\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e), and datasets were merged iteratively, removing multiallelic variants. The three platform-specific unfiltered datasets were sequentially merged to create a unified openSNP dataset.\u003c/p\u003e \u003cp\u003eStandard quality control filters were applied, excluding variants and individuals with call rates below 95% in both platform-specific and openSNP datasets.\u003c/p\u003e \u003cp\u003eIn doing so, we obtained a unique \u003cem\u003e\u0026ldquo;DTC dataset\u0026rdquo;\u003c/em\u003e of 6,118 individuals and 73,382 SNPs with a 99.5% genotyping rate.\u003c/p\u003e \u003cp\u003eIntegration with reference population datasets\u003c/p\u003e \u003cp\u003eTo contextualise genomic variation within global population structure, we integrated the openSNP dataset with two large-scale reference panels: the 1000 Genomes Project Phase 3 (1KGP) and the Human Genome Diversity Project (HGDP)(\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e). We downloaded the 1KGP dataset genotyped on Illumina Omni 2.5 array and converted it from VCF to PLINK binary format. We removed SNPs and individuals with more than 5% missing data. For HGDP, we downloaded the dataset genotyped with Illumina HuHap 650k, lifted the genomic coordinates to build 37, and applied the same filtering described above.\u003c/p\u003e \u003cp\u003eThe three quality-controlled datasets \u0026mdash;openSNP, 1KGP, and HGDP\u0026mdash; were merged using PLINK's \u003cem\u003e--bmerge\u003c/em\u003e function, and multiallelic variants were excluded using the iterative exclusion procedure described above. A final round of quality control was applied to the integrated dataset using identical thresholds (\u003cem\u003e--geno 0.05\u003c/em\u003e for variant call rate and \u003cem\u003e--mind 0.05\u003c/em\u003e for individual call rate) to remove any variants or individuals that acquired excessive missing data due to differential SNP coverage across platforms.\u003c/p\u003e \u003cp\u003eThe final merged dataset was sorted by natural sort order of family IDs and within-family individual IDs using the \u003cem\u003e--indiv-sort\u003c/em\u003e PLINK command to ensure consistent sample ordering for downstream analyses.\u003c/p\u003e \u003cp\u003eAfter excluding individuals and SNPs with more than 5% missing values, the \u003cem\u003efinal dataset\u003c/em\u003e contains 9,376 individuals and 66,843 variants, with a genotyping rate of 99.63%.\u003c/p\u003e \u003cp\u003ePrincipal Component Analysis\u003c/p\u003e \u003cp\u003ePrincipal Component Analysis (PCA) was conducted using EigenSoft's \u003cem\u003esmartpca\u003c/em\u003e tool(\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e), calculating the first 50 principal components calculated and no outlier removal.\u003c/p\u003e \u003cp\u003eData manipulation, visualisation, and statistical analyses were carried out using \u003cem\u003eR\u003c/em\u003e version 4.4.2(\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e). To improve the overall interpretation of the results, we pooled populations into African, American, European, East Asian, South-West Asian, and Oceanian groups, following the International Genome Sample Resource (IGSR) classification (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e). The first 3 principal components were used for primary visualisation.\u003c/p\u003e \u003cp\u003ePopulation assignment with k-NN and Random Forest\u003c/p\u003e \u003cp\u003eWe trained and evaluated two machine learning models for population assignment using genetic data: k-Nearest Neighbors (k-NN) and Random Forest (RF). The reference dataset consisted of samples from the 1000 Genomes Project (1KGP) used as the training set, while samples from the Human Genome Diversity Project (HGDP) served as an independent test set to evaluate model performance.\u003c/p\u003e \u003cp\u003eFor both models, we implemented a 10-fold cross-validation strategy, repeated 10 times, to assess model performance on the training data. For the k-NN model, we fixed k\u0026thinsp;=\u0026thinsp;10 neighbours, whereas the Random Forest classifier was implemented using 500 decision trees.\u003c/p\u003e \u003cp\u003eTo evaluate model performance, we used confusion matrices and overall accuracy metrics. The models were tested on an independent HGDP-derived test subset to assess their generalisability across samples from different sources. It is important to note that the HGDP test subset did not fully overlap with the 1KGP training label space. Specifically, the evaluation set was restricted to HGDP populations with explicit 1KGP counterparts, namely BASQUE, JAPANESE, FRENCH, BANTUKENYA, CHS, HAN, HAN-NCHINA, ITALIAN, TUSCAN, and YORUBA, which were harmonised to the corresponding 1KGP labels IBS, JPT, GBR, LWK, CHB, TSI, and YRI. After prediction, class-level results and confusion-matrix target rows were filtered to retain only classes present in the test set. This same filtering procedure was applied to both the k-NN and random forest models. As a consequence, sensitivity, specificity, and F1-scores were reported only for classes represented in the HGDP test subset, whereas overall accuracy was computed on the complete prediction set and was therefore unaffected by class-level filtering.\u003c/p\u003e \u003cp\u003eFor the final application to openSNP samples of unknown ancestry, we extracted assignment probabilities and implemented a confidence threshold of 0.5; assignments with probabilities below this threshold were flagged as low confidence.\u003c/p\u003e \u003cp\u003eAll analyses were performed in R v4.4.2 using the following packages: \u003cb\u003erandomForest\u003c/b\u003e for the RF implementation, \u003cb\u003eclass\u003c/b\u003e(\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e) for k-NN, and \u003cb\u003ecaret\u003c/b\u003e(\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e) for model training and evaluation. Parallel processing was utilised to optimise computational efficiency during the training step.\u003c/p\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eADMIXTURE\u003c/h2\u003e \u003cp\u003eTo infer ancestral components of the openSNP dataset, we performed \u003cb\u003eadmixture\u003c/b\u003e(\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e) analysis with K ranging from 2 to 10 (S3). In detail, we have run ten iterations for each K value using the \u003cem\u003e-s\u003c/em\u003e time and \u003cem\u003e--cv\u003c/em\u003e options. To avoid bias from sample-size imbalance, we projected the openSNP samples onto the ancestral components inferred from all other individuals. Therefore, we first estimated the component proportions (K) and allele frequencies using the 1KGP and HGDP populations exclusively and subsequently projected the openSNP dataset using the \u003cem\u003e-P\u003c/em\u003e option.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003ePANE\u003c/h3\u003e\n\u003cp\u003eWe inferred fine-scale ancestry proportions using \u003cb\u003ePANE\u003c/b\u003e(\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e). This approach estimates ancestry proportions from multiple reference populations by performing non-negative least-squares analysis on principal component vectors.\u003c/p\u003e \u003cp\u003eFor the analysis, we first computed the mean PC coordinates for each of the 72 reference populations across the first 50 principal components. These population-specific mean vectors served as \"donors\" in the PANE regression framework. We then projected each of the 6,118 openSNP samples onto this reference space using the same 50 PCs as \"recipients\". The algorithm decomposed each openSNP sample's PC coordinates as a non-negative linear combination of the reference population mean vectors, yielding proportion estimates for all 72 populations.\u003c/p\u003e \u003cp\u003eTo facilitate interpretation and account for fine-scale population structure, we aggregated the population-level proportions to seven continental groups: Africa (including sub-Saharan African populations), Europe (including European populations), Middle East (including North African and Middle Eastern populations), South Asia (including Pakistani and Indian populations), East Asia (including East and Southeast Asian populations), Americas (including Native American and admixed American populations), and Oceania (including Papuan and Melanesian populations).\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eCharacterisation of the analysed openSNP dataset\u003c/p\u003e \u003cp\u003eFollowing data curation and standardisation, our analysis included genotype data from 6,118 individuals within the openSNP dataset, sourced from three major Direct-to-Consumer (DTC) companies. The data processing pipeline yielded the following final filtered datasets: for 23andMe, starting with 4,961 initial files, the dataset was reduced to 4,430 individuals and 100,727 SNPs, compared to the 4,643 individuals and 3,394,458 SNPs retained before the final filtering step. The AncestryDNA dataset began with 1,161 files, yielding genetic data for 1,121 individuals and 389,000 SNPs after filtering, compared with the initial acquisition of 1,137 individuals and 1,531,219 SNPs. Finally, for FamilyTreeDNA, of the 672 initial files, the final dataset comprised 474 individuals and 177,520 SNPs, after an intermediate step that had retained 511 individuals and 1,724,813 SNPs.\u003c/p\u003e \u003cp\u003eTo provide an overall description of genetic variation across the whole sample, we merged the three company-specific unfiltered datasets, yielding an initial dataset of 6,291 individuals and 3,396,244 SNPs. After filtering, the final cleaned dataset comprised 6,118 individuals and more than 73,000 SNPs, with 4,488, 1,131, and 499 samples from 23andMe, AncestryDNA, and FamilyTreeDNA, respectively (Figure \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eWe merged the DTC dataset with two available resources genotyped on two different Illumina arrays, the Human Genome Diversity Project and the 1000 Genome Project, resulting in a \u003cem\u003e\u0026ldquo;Full Dataset\u0026rdquo;\u003c/em\u003e of 9,376 samples and 66,843 SNPs. All the subsequent analyses were performed on this dataset.\u003c/p\u003e \u003cp\u003ePrincipal Component Analysis\u003c/p\u003e \u003cp\u003eWe assessed the genetic variation in the DTC samples using a PC Analysis, projecting the genetic variation of openSNP individuals into the PC space inferred from the HGDP and 1KGP populations, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e1\u003c/span\u003e. When the first two PCs are considered, the vast majority of samples fall near European samples, suggesting a high proportion of individuals of European origin. The remaining samples are scattered following two main clines, evident in PC2 (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e1\u003c/span\u003eA). A substantial proportion of individuals are distributed toward African groups, possibly due to the high African ancestry among African Americans (\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e, \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e). Similarly, a relatively high number of individuals stretches between European and East Asian, capturing both the native American Ancestry and the XX-century migration from East Asia to the Americas. When PC3 is taken into account, the impact of these two ancestries becomes discernible, revealing their presence.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eIn all cases, some individuals are placed outside of this principal clines, suggesting the presence of more than two main ancestries and revealing a complex admixture scenario (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e1\u003c/span\u003eB).\u003c/p\u003e \u003cp\u003eModel performance evaluation\u003c/p\u003e \u003cp\u003eThe k-NN model with k\u0026thinsp;=\u0026thinsp;10 achieved an overall accuracy of 85.51% (SD\u0026thinsp;\u0026plusmn;\u0026thinsp;1.74%) and Cohen's Kappa of 0.846 (SD\u0026thinsp;\u0026plusmn;\u0026thinsp;0.018) across 100 resampling iterations. The RF model demonstrated superior performance with 97.42% overall accuracy (SD\u0026thinsp;\u0026plusmn;\u0026thinsp;0.96%) and Cohen's Kappa of 0.973 (SD\u0026thinsp;\u0026plusmn;\u0026thinsp;0.010) (\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003eSupplementary Tables\u0026nbsp;1\u0026ndash;4\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eAnalysis of per-class performance during cross-validation revealed heterogeneous F1-scores across populations for the k-NN model. Four populations achieved perfect classification (F1\u0026thinsp;=\u0026thinsp;1.00): Finnish (FIN), Gujarati Indian (GIH), Luhya (LWK), and Peruvian (PEL). Several populations showed high performance with F1-scores exceeding 0.90, and for the worst classes, the second most frequent assignment is to a close population. For example, for CHS (F1\u0026thinsp;=\u0026thinsp;0.59), 56.9% of the individuals are assigned to CHB (Figure \u003cspan refid=\"MOESM2\" class=\"InternalRef\"\u003eS2\u003c/span\u003eA).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe RF model demonstrated more consistent per-class performance, achieving perfect F1 Scores (1.00) across five populations: Finnish (FIN), British (GBR), Gujarati Indian (GIH), Luhya (LWK), and Peruvian (PEL). Moreover additional seven populations showed near-perfect classification with F1-scores exceeding 0.99 (TSI, CLM, IBS, YRI, JPT, PUR, MXL). The lowest F1-score was observed for Han Chinese in Beijing (CHB, F1\u0026thinsp;=\u0026thinsp;0.869), while all other populations exceeded 0.90 (Figure \u003cspan refid=\"MOESM2\" class=\"InternalRef\"\u003eS2\u003c/span\u003eC).\u003c/p\u003e \u003cp\u003eWhen applied to the independent HGDP test dataset, the k-NN model achieved 85.23% overall accuracy (Figure \u003cspan refid=\"MOESM2\" class=\"InternalRef\"\u003eS2\u003c/span\u003eB). Three populations were perfectly classified: Iberian, Japanese, and Luhya, each with 100% sensitivity, 100% precision, and F1\u0026thinsp;=\u0026thinsp;1.00. Chinese samples showed 79.5% sensitivity and 100% precision (F1\u0026thinsp;=\u0026thinsp;0.886), while British samples showed 78.6% sensitivity and 100% precision (F1\u0026thinsp;=\u0026thinsp;0.880). Tuscan samples showed 100% sensitivity but 76.9% precision (F1\u0026thinsp;=\u0026thinsp;0.870). Yoruba samples showed 47.6% sensitivity with 100% precision, resulting in the lowest F1-score of 0.645.\u003c/p\u003e \u003cp\u003eThe RF model achieved 96.59% overall accuracy on the HGDP test set (Figure \u003cspan refid=\"MOESM2\" class=\"InternalRef\"\u003eS2\u003c/span\u003eD). Four populations demonstrated perfect classification with 100% sensitivity, 100% precision, and F1\u0026thinsp;=\u0026thinsp;1.00: Japanese, Luhya, and Yoruba. Chinese samples showed the highest performance among imperfectly classified groups with 97.7% sensitivity and 100% precision (F1\u0026thinsp;=\u0026thinsp;0.989). Tuscan samples demonstrated 100% sensitivity with 90.9% precision (F1\u0026thinsp;=\u0026thinsp;0.952), while Iberian samples showed 100% sensitivity with 88.9% precision (F1\u0026thinsp;=\u0026thinsp;0.941). British samples exhibited 82.1% sensitivity with 100% precision (F1\u0026thinsp;=\u0026thinsp;0.902).\u003c/p\u003e \u003cp\u003eopenSNPs population assignment using unsupervised models\u003c/p\u003e \u003cp\u003eApplication of both trained models to 6,118 openSNP user samples revealed predominant European ancestry across the cohort. For both models, we first considered raw assignments regardless of their probability, then filtered out those with a probability below 50%, obtaining high-confidence assignments.\u003c/p\u003e \u003cp\u003eLooking at raw assignments, the k-NN model assigned 5,580 samples (91.2%) to European populations, 193 samples (3.2%) to East Asian populations, 173 samples (2.8%) to African populations, 127 samples (2.1%) to South Asian populations, and 45 samples (0.7%) to American populations. Considering only high-confidence assignments, the k-NN model assigned a total of 5,949 samples (97.2% of all samples). Among these, 91.6% (n\u0026thinsp;=\u0026thinsp;5,452) were to European populations, 3.0% (n\u0026thinsp;=\u0026thinsp;178) to East Asian populations, 2.7% (n\u0026thinsp;=\u0026thinsp;158) to African populations, 2.1% (n\u0026thinsp;=\u0026thinsp;127) to South Asian populations, and 0.6% (n\u0026thinsp;=\u0026thinsp;34) to American populations. Instead, for the RF model, raw assignments show 5,485 samples (89.7%) assigned to European ancestry, 207 samples (3.4%) to African ancestry, 154 samples (2.5%) to East Asian ancestry, 137 samples (2.2%) to American ancestry, and 135 samples (2.2%) to South Asian ancestry. The confidence distribution differed markedly from k-NN, with only 4,077 samples (66.6% of all samples) assigned with high confidence. Of these, 94.0% (n\u0026thinsp;=\u0026thinsp;3,833) were assigned to European ancestry, 2.8% (n\u0026thinsp;=\u0026thinsp;115) to South Asian ancestry, 1.9% (n\u0026thinsp;=\u0026thinsp;76) to East Asian ancestry, 0.8% (n\u0026thinsp;=\u0026thinsp;32) to African ancestry, and 0.5% (n\u0026thinsp;=\u0026thinsp;21) to American ancestry.\u003c/p\u003e \u003cp\u003eThe per-population assignment distribution analysis revealed a considerable imbalance in representation across populations, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e2\u003c/span\u003e. British ancestry (GBR) was the dominant assignment in both models, with the k-NN model assigning 4,266 samples (71.7% of 5,949 high-confident assignments, Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e2\u003c/span\u003eA) and the RF model assigning 3,487 samples (85.5% of 4,077 Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e2\u003c/span\u003eB). Tuscan (TSI) received the second-highest assignments: 1,144 samples (19.2%) from k-NN and 308 samples (7.6%) from RF. Together, these two populations accounted for over 90% of all assignments from k-NN and 93% from RF. The remaining populations showed substantially lower frequencies, with most receiving fewer than 100 assignments. Notable differences between models included RF's higher detection of admixed American populations (137 total assignments, including 98 MXL, 26 PUR, 11 CLM, 2 PEL) compared to k-NN (44 total assignments: 40 MXL, 5 PUR), and RF's more distributed European assignments (including 66 IBS and 43 FIN) compared to k-NN (1 IBS, 42 FIN).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e\n\u003ch3\u003eADMIXTURE\u003c/h3\u003e\n\u003cp\u003eTo infer the ancestry composition of the analysed individuals and assess the presence of multiple ancestries within individuals, we performed ADMIXTURE analysis with K ranging from 2 to 10 (Figure \u003cspan refid=\"MOESM3\" class=\"InternalRef\"\u003eS3\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eAt K\u0026thinsp;=\u0026thinsp;4, the main components of the DTC dataset are differentiated. Most individuals show a very high proportion of European ancestry (mean\u0026thinsp;=\u0026thinsp;0.831, SD\u0026thinsp;=\u0026thinsp;0.072, Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e3\u003c/span\u003eA), with minor contributions from Asian, African, and Native American components. A substantial group shows a high proportion of non-European ancestry. In fact, 152 individuals have a predominant African ancestry, while 152 have a modal Asian ancestry. For Native American Ancestry, 9 individuals have a predominant ancestry.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eAt K\u0026thinsp;=\u0026thinsp;7, which is the most supported according to the performed cross-validation analysis, we additionally identify a component modal in Papuan and Melanesian populations. Moreover, West Eurasian populations are mostly composed of two components, one modal in Northern Europe (yellow) and the other in Northern Africa and the Middle East (blue), but are present in high proportions in many Southern European populations. Moreover, South Asians show a high proportion of a component which is less present further East (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e3\u003c/span\u003eB).\u003c/p\u003e \u003cp\u003eDTC individuals are mainly composed of admixture profiles very similar to those of Eurasian groups, although many individuals suggest different origins. Among the others, we identified 44 individuals characterised by a large Native American ancestral component, exceeding 40%, although only a single individual exceeded 60%. Furthermore, 135 individuals have more than 60% of East Asian component. Similarly, a subset of samples (n\u0026thinsp;=\u0026thinsp;133) is characterised by an ancestral component which is modal in South Asia. We also identified 172 individuals with an ancestral profile consistent with that observed in the ASW group and are therefore likely to be African American. We did not identify any individuals with more than 90% African ancestry.\u003c/p\u003e\n\u003ch3\u003ePANE\u003c/h3\u003e\n\u003cp\u003eTo characterise fine-scale admixture patterns across the openSNP cohort, we estimated ancestral proportions using PANE, which implements Non-Negative Least Squares (NNLS) regression on principal component coordinates (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e4\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eConsidering proportions estimated from the first 50 principal components, the analysis confirmed predominant European ancestry across the cohort, with a median proportion of 83.5% (mean\u0026thinsp;=\u0026thinsp;73.5%, SD\u0026thinsp;=\u0026thinsp;24.4%). The remaining continental components showed substantially lower contributions: American ancestry had a median of 4.44% (mean\u0026thinsp;=\u0026thinsp;6.44%, SD\u0026thinsp;=\u0026thinsp;8.42%), Middle Eastern ancestry showed a median of 2.82% (mean\u0026thinsp;=\u0026thinsp;5.72%, SD\u0026thinsp;=\u0026thinsp;9.05%), South Asian ancestry had a median of 2.36% (mean\u0026thinsp;=\u0026thinsp;5.08%, SD\u0026thinsp;=\u0026thinsp;12.6%), African ancestry showed a median of 2.05% (mean\u0026thinsp;=\u0026thinsp;4.95%, SD\u0026thinsp;=\u0026thinsp;12.7%), East Asian ancestry had a median of 0.78% (mean\u0026thinsp;=\u0026thinsp;3.93%, SD\u0026thinsp;=\u0026thinsp;14%), and Oceanian ancestry was minimal with a median of 0.31% (mean\u0026thinsp;=\u0026thinsp;0.52%, SD\u0026thinsp;=\u0026thinsp;1.37%).\u003c/p\u003e \u003cp\u003eThe distribution of ancestry proportions revealed considerable heterogeneity within the cohort. While the majority of samples showed high European ancestry (\u0026gt;\u0026thinsp;66%), a substantial subset exhibited significant non-European components, consistent with the population assignment results from k-NN and RF models. The high standard deviations observed for African (SD\u0026thinsp;=\u0026thinsp;12.7%), East Asian (SD\u0026thinsp;=\u0026thinsp;14%), and South Asian (SD\u0026thinsp;=\u0026thinsp;12.6%) components indicate the presence of both predominantly European individuals and those with substantial ancestry from these regions. Using a threshold of 50% for dominant ancestry assignment, we identified 5,237 individuals (85.6%) with predominant European ancestry, 160 individuals (2.62%) with African ancestry, 149 individuals (2.44%) with East Asian ancestry, 129 individuals (2.11%) with South Asian ancestry, 67 individuals (1.1%) with American ancestry, 49 individuals (0.8%) with Middle Eastern ancestry, one individual (0.02%) with Oceanian ancestry, and 326 individuals (5.33%) with uncertain/admixed profiles showing no single dominant component exceeding the threshold.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eHere, we evaluated genetic variation in the openSNP dataset, comprising 6,118 individuals from three major direct-to-consumer genetic testing companies, providing essential insights into both the genetic ancestry of DTC testing users and the effectiveness of different computational approaches for population assignment.\u003c/p\u003e \u003cp\u003eOur analysis revealed remarkable variability in the genomic ancestry of openSNP users, while still highlighting a strong imbalance in continental ancestries.\u003c/p\u003e \u003cp\u003eIn fact, in principal component analysis, most of the analysed samples fall within the European variation, with much fewer individuals within the African and Asian variation. However, the observation of distinct genetic clines extending toward African and East Asian populations highlights complex admixture patterns, particularly evident in individuals with African-American and Asian-American ancestry. However, from PCA alone, it is often challenging to estimate each individual's exact ancestry composition, especially when substantial recent admixture has occurred, as is the case in modern human populations. Therefore, to provide an extensive picture of genomic ancestry variation, we performed three ancestry inference and population assignment methods. This allowed us to perform a more reliable and detailed population assignment, beyond the classification into three primary continents reported in previous surveys (Supplementary Table\u0026nbsp;5).\u003c/p\u003e \u003cp\u003eBased on cross-validation analysis, our population assignment methods using k-NN and RF have shown high accuracy and sensitivity, with the latter achieving slightly better predictive performance, suggesting a higher capability of ensemble methods than distance-based methods. When applied to the openSNP dataset, these methods confirmed the overwhelming European predominance in the 6,118 openSNP samples (91.6\u0026ndash;94% for k-NN and RF, respectively). The minimal representation of East Asian (3-1.9%), African (2.7\u0026ndash;0.8%), South Asian (2.1\u0026ndash;2.8%), and American (0.6\u0026thinsp;\u0026minus;\u0026thinsp;0.5%) ancestries highlights the limited global reach and persistent demographic imbalances in open genomic data repositories.\u003c/p\u003e \u003cp\u003eADMIXTURE further supported these findings, revealing subtle ancestry components that discrete population assignments might miss. The high proportion of European ancestry in most individuals, coupled with varying contributions from Asian, African, and Native American components, demonstrates the complexity of genealogy in modern populations. These results align with the PCA findings and provide a more nuanced view of population structure.\u003c/p\u003e \u003cp\u003eThe PANE analysis, incorporating non-negative least squares with PCA, offered additional resolution by quantifying ancestry proportions, showing an average European ancestry of 73% (median\u0026thinsp;=\u0026thinsp;83.1%, sd\u0026thinsp;=\u0026thinsp;24.4%) across the dataset. The relatively low proportion of American (mean\u0026thinsp;=\u0026thinsp;6.4%, median\u0026thinsp;=\u0026thinsp;4.4%, sd\u0026thinsp;=\u0026thinsp;8.4%), Asian and African ancestry reflects both historical migration patterns and sampling biases in DTC testing demographics.\u003c/p\u003e \u003cp\u003eTaken together, our results have several important implications. First, we highlighted a wide variation in the openSNP DTC dataset. Nevertheless, we confirmed a strong imbalance in ancestry composition across the whole dataset, even when more refined methods than PCA are used. When compared with a recent survey of a much more limited set of individuals and methods performed in 2017, we uncovered some differences. In fact, we observed a slight increase in the Asian proportion, which shows a higher proportion than previously reported (5% vs 2.9% in Corpas et al 2017). On the other hand, the African ancestry proportion increased only slightly from 2.2% to 2.8%, demonstrating that this market is unequally distributed across the global population.\u003c/p\u003e \u003cp\u003eThese results reveal that, despite the DTC's recent initiatives to expand customers' sociocultural and ancestral backgrounds, these efforts had only a limited effect. However, even though access to DTC services for individuals of non-European ancestry has increased, their users are more reluctant to share their data. In this context, further studies documenting both access to and willingness to share genetic data would be needed to provide a better overview of the observed disparities, which have a similar magnitude to those observed in translational studies\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eData Availability\u003c/h2\u003e\n\u003cp\u003eThe assembled analysed genotype data from openSNP cohort (\u003cem\u003eDTC dataset\u003c/em\u003e) and the \u003cem\u003eFull Dataset\u003c/em\u003e are available at https://doi.org/10.5281/zenodo.18493898.\u003c/p\u003e\n\u003ch2\u003eAcknowledgments\u0026nbsp;\u003c/h2\u003e\n\u003cp\u003eVincenzo Angiulli is a PhD student within the European School of Molecular Medicine (SEMM).\u0026nbsp;\u003c/p\u003e\n\u003ch2\u003eAuthor Contributions\u003c/h2\u003e\n\u003cp\u003eV.A. and F.M. conceived and designed the study. V.A. performed data processing, statistical analyses, and ancestry inference. F.P. and R.M.R. contributed to data interpretation and methodological validation. V.A. wrote the original draft of the manuscript. F.M. supervised the study and provided critical revisions. All authors reviewed and approved the final version.\u003c/p\u003e\n\u003ch2\u003eFundings\u003c/h2\u003e\n\u003cp\u003eThe authors declare that no funds, grants, or other support were received during the preparation and development of this manuscript.\u003c/p\u003e\n\u003ch2\u003eEthical Approval\u003c/h2\u003e\n\u003cp\u003eAll data in this study were obtained from publicly available resources. No new human subjects were recruited, and no individually identifiable information was used. Ethical approval and informed consent were obtained by the original primary studies. No additional ethical approval was required for this secondary analysis.\u003c/p\u003e\n\u003ch2\u003eCompeting Interests\u003c/h2\u003e\n\u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eCommissioner O of the. FDA [Internet]. FDA; 2024 [cited 2026 Mar 17]. FDA authorizes first direct-to-consumer test for detecting genetic variants that may be associated with medication metabolism. Available from: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.fda.gov/news-events/press-announcements/fda-authorizes-first-direct-consumer-test-detecting-genetic-variants-may-be-associated-medication\u003c/span\u003e\u003cspan address=\"https://www.fda.gov/news-events/press-announcements/fda-authorizes-first-direct-consumer-test-detecting-genetic-variants-may-be-associated-medication\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYim SH, Chung YJ. Reflections on the US FDA\u0026rsquo;s Warning on Direct-to-Consumer Genetic Testing. Genomics Inform. 2014;12(4):151\u0026ndash;5. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.5808/GI.2014.12.4.151\u003c/span\u003e\u003cspan address=\"10.5808/GI.2014.12.4.151\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAnnas GJ, Elias S. 23andMe and the FDA. New England Journal of Medicine. 2014;370(11):985\u0026ndash;8. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1056/NEJMp1316367\u003c/span\u003e\u003cspan address=\"10.1056/NEJMp1316367\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSu P. Direct-to-Consumer Genetic Testing: A Comprehensive View. Yale J Biol Med. 2013;86(3):359\u0026ndash;65. PubMed PMID: 24058310; PubMed Central PMCID: PMC3767220.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRoth WD, Lyon KA. Genetic Ancestry Tests and Race: Who Takes Them, Why, and How Do They Affect Racial Identities? In: Gates J Henry Louis, Suzuki K, Von Vacano DA, editors. Reconsidering Race: Social Science Perspectives on Racial Categories in the Age of Genomics [Internet]. Oxford University Press; 2018 [cited 2026 Mar 23]. p. 0. Available from: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1093/oso/9780190465285.003.0008\u003c/span\u003e\u003cspan address=\"10.1093/oso/9780190465285.003.0008\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e doi:10.1093/oso/9780190465285.003.0008\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMajumder MA, Guerrini CJ, McGuire AL. Direct-to-Consumer Genetic Testing: Value and Risk. Annual Review of Medicine. 2021;72(Volume 72, 2021):151\u0026ndash;66. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1146/annurev-med-070119-114727\u003c/span\u003e\u003cspan address=\"10.1146/annurev-med-070119-114727\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTandy-Connor S, Guiltinan J, Krempely K, LaDuca H, Reineke P, Gutierrez S, et al. False-positive results released by direct-to-consumer genetic tests highlight the importance of clinical confirmation testing for appropriate patient care. Genetics in Medicine. 2018;20(12):1515\u0026ndash;21. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/gim.2018.38\u003c/span\u003e\u003cspan address=\"10.1038/gim.2018.38\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNelson SC, Bowen DJ, Fullerton SM. Third-Party Genetic Interpretation Tools: A Mixed-Methods Study of Consumer Motivation and Behavior. The American Journal of Human Genetics. 2019;105(1):122\u0026ndash;31. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.ajhg.2019.05.014\u003c/span\u003e\u003cspan address=\"10.1016/j.ajhg.2019.05.014\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWiggins A, Wilbanks J. The Rise of Citizen Science in Health and Biomedical Research. The American Journal of Bioethics. 2019;19(8):3\u0026ndash;14. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1080/15265161.2019.1619859\u003c/span\u003e\u003cspan address=\"10.1080/15265161.2019.1619859\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e PubMed PMID: 31339831.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePrainsack B. The Powers of Participatory Medicine. PLOS Biology. 2014;12(4):e1001837. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1371/journal.pbio.1001837\u003c/span\u003e\u003cspan address=\"10.1371/journal.pbio.1001837\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGreshake B, Bayer PE, Rausch H, Reda J. openSNP\u0026ndash;A Crowdsourced Web Resource for Personal Genomics. PLOS ONE. 2014;9(3):e89204. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1371/journal.pone.0089204\u003c/span\u003e\u003cspan address=\"10.1371/journal.pone.0089204\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eopenSNP [Internet]. [cited 2026 Mar 17]. openSNP. Available from: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://opensnp.github.io/\u003c/span\u003e\u003cspan address=\"https://opensnp.github.io/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBridging genomics\u0026rsquo; greatest challenge: The diversity gap. Cell Genomics. 2025;5(1):100724. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.xgen.2024.100724\u003c/span\u003e\u003cspan address=\"10.1016/j.xgen.2024.100724\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRiha A. apriha/snps [Python] [Internet]. 2026 [cited 2026 Mar 17]. Available from: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/apriha/snps\u003c/span\u003e\u003cspan address=\"https://github.com/apriha/snps\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eA survey of direct-to-consumer genotype data, and quality control tool (GenomePrep) for research. Computational and Structural Biotechnology Journal. 2021;19:3747\u0026ndash;54. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.csbj.2021.06.040\u003c/span\u003e\u003cspan address=\"10.1016/j.csbj.2021.06.040\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChang CC, Chow CC, Tellier LC, Vattikuti S, Purcell SM, Lee JJ. Second-generation PLINK: rising to the challenge of larger and richer datasets. Gigascience. 2015;4:7. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/s13742-015-0047-8\u003c/span\u003e\u003cspan address=\"10.1186/s13742-015-0047-8\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e PubMed PMID: 25722852; PubMed Central PMCID: PMC4342193.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eA global reference for human genetic variation. Nature. 2015;526(7571):68\u0026ndash;74. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/nature15393\u003c/span\u003e\u003cspan address=\"10.1038/nature15393\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePrice AL, Patterson NJ, Plenge RM, Weinblatt ME, Shadick NA, Reich D. Principal components analysis corrects for stratification in genome-wide association studies. Nat Genet. 2006;38(8):904\u0026ndash;9. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/ng1847\u003c/span\u003e\u003cspan address=\"10.1038/ng1847\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eR: The R Project for Statistical Computing [Internet]. [cited 2026 Mar 18]. Available from: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.r-project.org/\u003c/span\u003e\u003cspan address=\"https://www.r-project.org/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFairley S, Lowy-Gallego E, Perry E, Flicek P. The International Genome Sample Resource (IGSR) collection of open human genomic variation resources. Nucleic Acids Res. 2020;48(D1):D941\u0026ndash;7. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/nar/gkz836\u003c/span\u003e\u003cspan address=\"10.1093/nar/gkz836\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eModern Applied Statistics with S, 4th ed [Internet]. [cited 2026 Mar 18]. Available from: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.stats.ox.ac.uk/pub/MASS4/\u003c/span\u003e\u003cspan address=\"https://www.stats.ox.ac.uk/pub/MASS4/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKuhn M. Building Predictive Models in R Using the caret Package. Journal of Statistical Software. 2008;28:1\u0026ndash;26. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.18637/jss.v028.i05\u003c/span\u003e\u003cspan address=\"10.18637/jss.v028.i05\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAlexander DH, Novembre J, Lange K. Fast model-based estimation of ancestry in unrelated individuals. Genome Res. 2009;19(9):1655\u0026ndash;64. doi:\u003cdiv class=\"ExternalRefDOI\"\u003e10.1101/gr\u003c/div\u003e.094052.109 PubMed PMID: 19648217; PubMed Central PMCID: PMC2752134.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ede Gennaro L, Molinaro L, Raveane A, Santonastaso F, Saponetti SS, Massi MC, et al. PANE: fast and reliable ancestral reconstruction on ancient genotype data with non-negative least square and principal component analysis. Genome Biol. 2025;26(1):29. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/s13059-025-03491-z\u003c/span\u003e\u003cspan address=\"10.1186/s13059-025-03491-z\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e PubMed PMID: 39934833; PubMed Central PMCID: PMC11818073.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMicheletti SJ, Bryc K, Ancona Esselmann SG, Freyman WA, Moreno ME, Poznik GD, et al. Genetic Consequences of the Transatlantic Slave Trade in the Americas. The American Journal of Human Genetics. 2020;107(2):265\u0026ndash;77. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.ajhg.2020.06.012\u003c/span\u003e\u003cspan address=\"10.1016/j.ajhg.2020.06.012\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOngaro L, Scliar MO, Flores R, Raveane A, Marnetto D, Sarno S, et al. The Genomic Impact of European Colonization of the Americas. Current Biology. 2019;29(23):3974\u0026ndash;3986.e4. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.cub.2019.09.076\u003c/span\u003e\u003cspan address=\"10.1016/j.cub.2019.09.076\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"european-journal-of-human-genetics","isNatureJournal":false,"hasQc":false,"allowDirectSubmit":false,"externalIdentity":"ejhg","sideBox":"Learn more about [European Journal of Human Genetics](http://www.nature.com/ejhg/)","snPcode":"41431","submissionUrl":"https://mts-ejhg.nature.com/cgi-bin/main.plex","title":"European Journal of Human Genetics","twitterHandle":"@ejhg_journal","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"ejp","reportingPortfolio":"Nature AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"Direct-to-consumer genetics, population structure, ancestry inference, openSNP, genomic diversity","lastPublishedDoi":"10.21203/rs.3.rs-9305945/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9305945/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eDirect-to-consumer (DTC) genetic datasets shared through public platforms are increasingly used for citizen science and research, yet their population structure and representativeness remain poorly characterised. Here, we analysed 6,118 genotype files from openSNP, harmonised across three major DTC providers (23andMe, AncestryDNA, and FamilyTreeDNA), and compared them with worldwide reference populations from the 1000 Genomes Project (1KGP) and the Human Genome Diversity Project (HGDP). To characterise ancestry patterns, we applied five complementary approaches: principal component analysis, two supervised population-assignment methods based on k-Nearest Neighbours and Random Forest, ADMIXTURE, and PANE/NNLS ancestry decomposition. Random forest showed higher predictive performance than k-nearest neighbours on the independent HGDP test set (96.59% vs 85.23% accuracy), supporting robust assignment of openSNP samples. Across all methods, the dataset showed a marked excess of European ancestry. Supervised classification assigned 91.6\u0026ndash;94% of individuals to European populations, ADMIXTURE estimated 91.4\u0026ndash;95.1% European ancestry across K values, and PANE identified 85.6% of individuals with European ancestry as the dominant component. African, East Asian, and South Asian ancestries were each represented at roughly 2\u0026ndash;3%, while American ancestry remained low, although a subset of individuals displayed substantial admixture. These results confirm and refine previous observations, primarily based on PCA, showing that ancestry imbalance in openSNP persists even when analysed with multiple orthogonal methods and a substantially larger sample size. Public DTC repositories remain valuable resources for methodological and population-genetic studies. Still, their strong ancestry skew should be considered in downstream analyses and in broader discussions of equity and representation in consumer genomics.\u003c/p\u003e","manuscriptTitle":"Ancestry imbalance and population structure in openSNP, a public direct-to-consumer genomic resource","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-05-06 12:28:13","doi":"10.21203/rs.3.rs-9305945/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"reviewerAgreed","content":"This content is not available.","date":"2026-05-08T12:53:13+00:00","index":2,"fulltext":"This content is not available."},{"type":"reviewerAgreed","content":"This content is not available.","date":"2026-05-08T11:21:50+00:00","index":1,"fulltext":"This content is not available."},{"type":"reviewersInvited","content":"","date":"2026-04-28T10:24:08+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-04-02T18:18:52+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-04-02T17:46:04+00:00","index":"","fulltext":""},{"type":"submitted","content":"European Journal of Human Genetics","date":"2026-04-02T17:46:03+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"european-journal-of-human-genetics","isNatureJournal":false,"hasQc":false,"allowDirectSubmit":false,"externalIdentity":"ejhg","sideBox":"Learn more about [European Journal of Human Genetics](http://www.nature.com/ejhg/)","snPcode":"41431","submissionUrl":"https://mts-ejhg.nature.com/cgi-bin/main.plex","title":"European Journal of Human Genetics","twitterHandle":"@ejhg_journal","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"ejp","reportingPortfolio":"Nature AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"73d058fe-e777-4e28-8427-22613bab13e8","owner":[],"postedDate":"May 6th, 2026","published":true,"recentEditorialEvents":[{"type":"reviewerAgreed","content":"This content is not available.","date":"2026-05-08T12:53:13+00:00","index":2,"fulltext":"This content is not available."},{"type":"reviewerAgreed","content":"This content is not available.","date":"2026-05-08T11:21:50+00:00","index":1,"fulltext":"This content is not available."}],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[{"id":65676166,"name":"Biological sciences/Genetics/Population genetics"},{"id":65676167,"name":"Biological sciences/Genetics/Population genetics/Genetic variation"}],"tags":[],"updatedAt":"2026-05-06T12:28:14+00:00","versionOfRecord":[],"versionCreatedAt":"2026-05-06 12:28:13","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9305945","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9305945","identity":"rs-9305945","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}