A pooled sequencing run combines material from multiple biosamples, making the sequencing run—not each constituent biosample—the defensible statistical unit when the measured taxonomic result exists only once per run. In this report, 1,067 of 2,759 runs were pooled from multiple biosamples, so treating each biosample as an independent observation would create pseudo-replication, meaning artificial inflation of the apparent sample size by duplicating one run-level measurement. 1
Evidence from the NMDC analysis
The analysis therefore operated at the workflow_run_id level rather than the biosample level because 1,067 of 2,759 runs were pooled from multiple biosamples. 1 This choice preserved the unit at which the GOTTCHA2 eukaryotic abundance measurement was observed and avoided counting one pooled result multiple times. 1
Pooled runs still required an environmental and study label for downstream association tests. Each pooled run inherited environment and collection metadata from a single representative biosample selected by MIN(biosample_id), the minimum biosample identifier in the joined set. 1 If the pooled biosamples differed in local metadata, this assignment introduced predictor label noise rather than additional independent environmental observations. 1 The report characterizes this as a conservative bias that can weaken associations rather than manufacture them. 1
The metadata bridge linked runs to biosamples and studies through nmdc.metadata.biosample_to_workflow_run, joined on workflow_run_id, and then through biosample_set_associated_studies, whose child tables are keyed by parent_id. 1 This procedure linked 99%+ of classified runs while retaining run-level aggregation. 1 Large read-based taxonomy tables, including the approximately 29M-row kraken2_classification_report, were aggregated to one row per workflow_run_id before metadata joins. 1
Statistical implications
The pooled-run structure separates two problems: pseudoreplication changes the effective number of observations, whereas representative-biosample assignment can mislabel predictors even when the run remains the correct response unit. 1 Consequently, a run-level analysis can avoid false precision without eliminating uncertainty about which biosample-level environment generated the pooled signal. 1
This issue is directly relevant to cross-tenant-data-bridging, because the validity of ecological inference depends on joining workflow results to the correct metadata grain. 1 It also refines cross-tenant-data-bridging: successful key-based joins do not by themselves demonstrate that a pooled run has a uniquely correct environmental label. 1 The problem contributes to environmental-resistome, since metadata recorded at one biosample or representative-biosample level may not describe all material contributing to a pooled run. 1
Practical interpretation
The accompanying analysis used run-level observations for hypothesis testing and treated the inherited metadata as potentially noisy, rather than expanding pooled runs into independent biosample records. 1 This design is preferable to biosample-level pseudo-replication, but it does not replace explicit information about pooling composition, relative biosample contributions, or biosample-specific environmental metadata. 1
Related non-independence and label-noise problems in other projects
The soil–freshwater comparison in the carbon census shows what happens when non-independent samples are counted as independent, which supports the run-level choice above. Each metagenome was treated as an independent observation in a rank test over compositional, zero-inflated relative abundances. With thousands of non-independent samples, the report states that this inflates significance massively: all 83 genera reach q<0.05, many at q~1e-70. Here q is a false-discovery-rate (FDR) adjusted p-value, where FDR is the expected share of false positives among significant calls. These significance values should therefore be read as uncalibrated rather than as strong evidence. 2 The same inflation also bears on study-batch-confounding-of-environmental-associations. 2
A lab–field ecology analysis took the opposite route and aggregated before testing. Its community data included multiple communities per sample, from different filter sizes and replicates, and these were aggregated within samples. 3 This is analogous to collapsing pooled material to one run-level observation. It avoids counting replicates as independent samples, but it gives up any within-sample variation between filter fractions and replicates. 3
The IBD phage-targeting project refines the trade-off by showing that aggregation also has a cost. Its analysis aggregated at the subject level even though each subject contributed multiple serology and metagenomics visits, so longitudinal information was lost. The report notes that mixed-effects regression would have had more power, but unbalanced visit counts complicate interpretation. 4 Choosing the subject as the unit avoids pseudoreplication, while a hierarchical model could keep the repeated visits without treating them as independent. 4
The same project also separates technical replication from biological replication. Patient 6967 is the only multi-timepoint UC Davis Crohn’s-disease (CD) patient with biological-replicate samples. As a result, the state-dependent dosing rule is n=1, and the report says prospective validation requires expanded longitudinal sampling; it should be treated as a single-patient hypothesis rather than an established rule. 4 Patient 1112 instead has 2 reseq replicates of the same biological sample. These gave Spearman ρ = 1.000 (p < 0.001) on 6 Tier-A targets, meaning perfect rank concordance that validates Kaiju reliability and sets the technical-noise floor. The report is explicit that this Spearman ρ=1.000 replicate is not a biological-replicate longitudinal trajectory. It therefore cannot add an independent biological observation to the n=1 longitudinal contrast. 4
Predictor-label noise can also come from text inference rather than pooling, which extends the representative-biosample problem above. In the clay-confined subsurface project, the compartment field was keyword-inferred from isolation_source strings. A small number of bentonite or “rock” entries could plausibly be either porewater or rock-attached. The report states that H3 results were robust to two stricter compartment definitions in sensitivity testing. 5 That sensitivity analysis is a model for checking whether the MIN(biosample_id) labels change pooled-run conclusions. 5,1
Cluster-aware models need correct grouping labels, and the central pitfalls digest shows that recovering them depends on the method. In curatedmetagenomicdata (cMD), the sub-study for IBD samples is encoded in the middle colon-separated token of participant_id, for example CMD:HallAB_2017:SKST006. For healthy-control (HC) samples, the sub-study sits in dim_samples.external_ids as a JSON blob with key "study". Together these cover ≈ 80 % of cMD samples, whereas short-prefix regex on sample IDs covers < 20 % and produces mostly noise categories. 6
Open Directions
- Recover the complete set of biosamples and their contribution weights for the 1,067 pooled runs, then fit weighted or hierarchical models to test whether representative-biosample assignment changes effect estimates. 1
- Compare
MIN(biosample_id)labels with alternative pooled labels, such as majority environment or contribution-weighted environment, and quantify how much metadata-label noise changes association strength. 1 - Reanalyze the eukaryotic-fraction models at the run level with cluster-robust or hierarchical uncertainty, using pooling structure as a grouping variable, to determine whether conclusions remain stable without biosample pseudo-replication. 1
- Validate the run-to-biosample-to-study joins against explicit parent-child key audits and unresolved-record counts, asking whether the reported 99%+ linkage rate hides systematic failures among pooled or unusually structured records. 1
- Rerun the soil–freshwater genus contrasts with study- or site-level blocking, or with cluster-aware compositional models, to test how many of the 83 genera stay significant once non-independence is modelled. 2
- Refit the IBD serology–metagenomics associations with mixed-effects models across visits, and report how unbalanced visit counts affect estimates compared with subject-level aggregation. 4
- Apply a sensitivity analysis like the clay-compartment one to pooled NMDC runs: drop runs whose constituent biosamples have discordant metadata and test whether eukaryotic-fraction associations hold. 5,1