Profile explorer methods

Source-specific analysis

NMR analyses use the UK Biobank reference source and are run separately for incident (249 × 723), prevalent (249 × 578) and mortality (249 × 127) endpoints. The EPIC-Norfolk MS analysis includes 948 measurements with complete estimates across all 27 incident endpoints. The remaining 65 measurements and 595 missing cells are retained in a completeness ledger and are not imputed.

Study 11 is excluded from clustering because only P values are available and source-key conflicts remain unresolved. Its source records are retained. Studies are analysed separately and are not pooled as independent observations.

Profiles and group selection

For disease-system profiles, source log coefficients are averaged within each registered system and each measurement profile is then standardized across systems. Individual-endpoint profiles are standardized across endpoints.

Groups are derived by Ward hierarchical clustering with Euclidean distance. Candidate solutions with k=2–10 are considered if every group contains at least 10 measurements. The eligible solution with the highest silhouette is reported as the best candidate; ties are resolved in favour of the smaller k.

NMR system labels follow the source ICD chapters. The nine MS systems are project-defined analytical groupings of author-reported endpoints and do not imply ontology equivalence. Heatmap values are standardized relative profile scores, not HRs or ORs.

Descriptive group summaries

Group composition is based on existing source-dictionary measurement types, with percentages calculated using all members of the group as the denominator. The NMR concentration category includes proteins and pooled measurements as well as small molecules. MS measurements are not further subdivided because validated measurement-form annotations are not available. No enrichment test is applied.

Representative measurements must have membership frequency ≥0.90 in both resampling analyses and an individual silhouette >0. Up to five qualifying measurements are shown per group in alphabetical order. No additional measurements are added when fewer than five meet these criteria. Glucose is shown only when it belongs to the currently selected partition and is used as an example, not as validation of the group.

For each group, the three highest and three lowest heatmap scores are shown as higher and lower profile coordinates. These are relative standardized scores and should not be interpreted as effect sizes or risk estimates. Group summaries are calculated from the full selected partition and are unaffected by filters applied to the membership table.

Resampling stability

Each analysis includes 200 endpoint bootstrap replicates and 200 system-block bootstrap replicates. Endpoints are resampled within systems, whereas block resampling draws complete systems with replacement. Each replicate is re-standardized and re-clustered across all candidate k values. Cluster labels are aligned to the reference partition by maximum overlap before membership frequencies are calculated.

A measurement is marked as uncertain if its membership frequency is <0.70 in either resampling analysis or if its silhouette is <0. The prespecified strict stability gate requires both ARI 10th percentiles to be ≥0.60, all membership frequencies to be ≥0.70 in both resampling analyses, and every group to contain at least 10 measurements. These are internal stability criteria and do not constitute external validation.

Sensitivity and structure checks

Fixed-k sensitivity analyses examine the effects of excluding NMR ratios and percentages, equal weighting of disease systems, splitting endpoints into two subsets, restricting the analysis to information-rich endpoints and leaving out one system at a time. Information-rich endpoints are defined as those for which at least 50% of measurements have q ≤ 0.05. NMR display q values and this sensitivity use BH correction over the complete original Atlas family for the corresponding age group, sex, endpoint context and model (283,211 incident, 234,721 prevalent and 43,235 mortality tests). The original frozen Feature q values are retained separately. The MS analysis applies BH to all 26,756 observed estimates across the 27 incident endpoints, including measurements excluded from clustering for incomplete coverage. Its source Bonferroni adjustment (1,014 measurements) is shown separately. Primary clustering uses all complete log coefficients, without a P or FDR filter; primary partitions and resampling results are unchanged.

For the permutation test, endpoint coordinates are independently shuffled within each measurement 100 times. The test statistic is the maximum eligible silhouette across k=2–10. The Monte Carlo P value is calculated as

A common permutation of columns leaves pairwise distances unchanged and is therefore used only as an invariance check.

PCA is used only as a descriptive projection. Directional concordance across related endpoints is also descriptive and is not treated as independent replication. Groups are left unnamed and are not assigned biological interpretations.