The Taiwan Precision Medicine Initiative provides a cohort for large-scale studies

Abstract

Han Chinese people comprise nearly 20% of the global population but remain under-represented in genetic studies^1,2, so there is an urgent need for large-scale cohorts to advance precision medicine. Here we present the Taiwan Precision Medicine Initiative (TPMI), established by Academia Sinica in collaboration with 16 major medical centres around Taiwan, which has recruited 565,390 participants who consent to provide DNA samples for genetic profiling and grant access to their electronic medical records (EMRs) for research. EMR access is both retrospective and prospective, allowing longitudinal studies. Genetic profiling is done with population-optimized arrays of single-nucleotide polymorphisms for people of Han Chinese ancestry, which enable genome-wide association^3,4, phenome-wide association^5,6 and polygenic risk score^7,8 studies to be performed to evaluate common disease risk and pharmacogenetic response. Participants also agreed to be re-contacted for future research and receive personalized genetic risk profiles with health management recommendations. The TPMI has established the TPMI Data Access Platform, a central database and analysis platform that both safeguards the security of the data and facilitates academic research. As a large cohort of individuals with non-European ancestry that merges genetic profiles with EMR data and enables longitudinal follow-up, TPMI provides a unique resource that could be used to validate genetic risk prediction models, perform clinical trials of risk-based health management and inform health policies. Ultimately, the TPMI cohort will contribute to global genetic research and serve as a model for population-based precision medicine.

Fig. 1. Map of medical centres, their satellite hospitals and sample sizes.

Locations of 16 partner medical centres and 33 affiliated hospitals, along with the numbers of DNA samples, genotyped samples, individuals with EMRs received, individuals with EMRs stored in the TPMI Data Lake and individuals with both genotype and EMR data. Source Data.

Fig. 2. Cohort characteristics.

a, Sex-specific age distribution. b, Top 20 most prevalent ICD-10 codes: E78 (disorders of lipoprotein metabolism and other lipidaemias), I10 (EHT), E11 (type 2 diabetes mellitus), K21 (gastro-oesophageal reflux disease), J30 (vasomotor and allergic rhinitis), G47 (sleep disorders), K05 (gingivitis and periodontal diseases), N39 (other urinary disorders), M47 (spondylosis), K59 (other functional intestinal disorders), M79 (other and unspecified soft tissue disorders), R10 (abdominal and pelvic pain), H10 (conjunctivitis), I25 (chronic ischaemic heart disease), N40 (enlarged prostate), L30 (other and unspecified dermatitis), I11 (hypertensive heart disease), H04 (lacrimal system disorders), N18 (chronic kidney disease) and R07 (pain in throat or chest). c, Age of onset for the top 20 diseases. Onset ages in male individuals (blue) and female individuals (pink) are presented as box plots, ordered by median. Box plots represent minima, first quartile, median, third quartile and maxima. Values and sample sizes are in the Source Data. d, Top 20 most prevalent laboratory tests: creatinine_B (blood creatinine), WBC (white blood cell count), SGPT (serum glutamic pyruvic transaminase or alanine aminotransferase; S-GPT/ALT), HB (haemoglobin), platelet (platelet count), HCT (haematocrit), RBC (red blood cell count), EGFR (estimated glomerular filtration rate), SGOT (serum glutamic–oxaloacetic transaminase or aspartate aminotransferase; S-GOT/AST), TG (triglyceride), cholesterol_T (Total Cholesterol), BUN (blood urea nitrogen), glucose_AC (fasting glucose), LDL_C (low-density lipoprotein cholesterol), HDL_C (high-density lipoprotein cholesterol), uric acid_B (blood uric acid), HbA1c (haemoglobin A1c), bilirubin_T (bilirubin, total value), albumin and TSH (thyroid-stimulating hormone, measured by enzyme immunoassay or luminescence immunoassay). Left, sex-specific distribution of record counts per individual (winsorized at the 95th percentile); middle, proportion of individuals with test data; right, distribution of average follow-up years. Box plots represent minima, first quartile, median, third quartile and maxima. Values and sample sizes are in the Source Data. e, The top pie chart shows the proportions of related and unrelated samples. The bottom pie chart shows relationship categories: duplicate (DUP) or monozygotic twin (MZ), parent–offspring (PO), full sibling (FS), second degree (2nd) and third degree (3rd). Source Data.

Fig. 3. Population structure.

a, PCA analysis. The TPMI cohort was compared with TWB, SGDP and 1KGP samples. The top-left inset compares TPMI participants born before 1950 with those born after 1950; the top-right inset compares the TPMI and the 1KGP. The main figure shows TPMI, TWB and two Taiwan Indigenous tribes (SGDP). Admixture fraction plots show ancestry fractions from ten ancestral populations (K = 10), with principal component (PC) 1 on the bottom axis and PC2 on the right. b, Coancestry and fine-scale structure. The coancestry heat map shows individuals (rows, columns) clustered by shared haplotypes, with colour intensity indicating haplotype copying. Darker blue or red indicates higher coancestry; yellow or light orange indicates lower. Diagonal blocks mark within-group sharing: K1–K6 show strong within-group haplotype sharing; K1–K2 (Han Chinese-enriched) exhibit strong coancestry with each other but less with K3–K6 (Indigenous-enriched), reflecting genetic differentiation; K3–K6 form distinct blocks, with some asymmetric sharing suggesting admixture or shared ancestry. The dendrogram shows clustering consistent with subgroup distinctions. c, Admixture graph depicting relationships and gene flow among K1–K6. Solid arrows represent drift edges (genetic drift from ancestral populations); dotted arrows represent admixture, with percentages indicating fractions. Edge numbers denote drift lengths (f2 units). K1 derives around 90% of ancestry from a lineage that also contributes to K2, plus 10% admixture from a lineage related to K6, indicating close K1–K2 affinity with minor Indigenous input. K4 shows around 49% of ancestry from a K5-related lineage (shared with K3) and 51% from a branch that also contributes to K2, reflecting Han–Indigenous admixture. K6 is mostly unadmixed with a long drift branch (f2 = 70), consistent with a highly diverged Indigenous lineage. K5 seems to be ancestral to other Indigenous groups (K3, K4 and possibly indirectly K6), with considerable early divergence (drift = 36 on both edges).

Fig. 4. Comparison of T2D GWAS results in the TPMI with those from four biobanks.

a, Comparison of GWAS results from the TPMI and the PheWebs of the Biobank Japan (BBJ), China Kadoorie Biobank (CKB), the Korean Genome and Epidemiology Study (KoGES) and the UK Biobank (UKB). A Firth logistic regression was applied for the T2D GWAS in the TPMI. All statistical tests were two-sided. Multiple-testing adjustment was applied using a genome-wide significance threshold of P < 1 × 10⁻⁸. Novel T2D-associated SNPs identified in our GWAS but absent in biobanks at the genome-wide significance level are shown. b, Pairwise comparison with each biobank. Different statistical methods were applied across cohorts. For BBJ, CKB and KoGES, a generalized linear mixed model was implemented using SAIGE; for UKB: linear regression was used. All tests were two-sided. Bonferroni correction was applied for multiple-testing adjustment across loci. The four graphs (from top to bottom) show T2D-associated SNPs identified in the TPMI but not in BBJ, CKB, KoGES and UKB, respectively. Source Data.

Fig. 5. PRS analysis for T2D.

a, AUC of PRS (red curve) and of PRS, age, sex and BMI (blue curve). b, Dose–response effect of PRS levels on the odds ratio (OR) of T2D. Dose–response effect of PRS (red line) and of a combination of PRS, age, sex and BMI (blue curve) with n = 205,779 independent samples. Error bars represent the 95% confidence interval, calculated as exp (β ± Z_0.025 × s.e.), based on maximum likelihood estimation from a logistic regression model with different decile intervals of PRS values included as covariates. The estimated coefficient (β) is provided in column B (‘Estimate’) and the standard error (s.e.) is provided in column C (‘Std. Error’) in the Source Data for b. Source Data.