Contributions
NMDC Context Audit audited KBase Data Lakehouse resources whose names contain nmdc and found that 20 database names resolve to 7 real, maintained resources across three tenants and six provenance classes, so the label does not reliably mark a dataset boundary 1. The audit found that two of the four databases in the nmdc tenant are external re-hosts: an NCBI BioSample harvest with 51,711,888 biosamples, and 27,481 Pfam terms 1. It also found that kbase.nmdc_neon represents NSF’s NEON rather than NMDC 1. It found that the freshest NMDC resource, kbase.nmdc_mags with 62,346 MAGs (metagenome-assembled genomes), sits in the kbase tenant with nothing cross-linking it to the nmdc tenant 1. It identified a ~3,000× scale trap between the 16,640 genuine NMDC samples and the co-hosted NCBI mirror 1. It found that Iceberg snapshot ages span approximately four months across NMDC-labeled resources and that this currency signal is not exposed at discovery 1. The audit delivered a 15-file Open-Knowledge-Format directory and proposed, but did not apply, seven fixes 1. The audit did not directly observe any user choosing the wrong resource, so its link between label overload and sub-optimal selection is inferred 1. Its provenance classes are inferred without an ingestion manifest 1. These findings feed provenance-aware-resource-discovery 1.
Metadata Correlates of Eukaryotic Contamination in NMDC Prokaryote-Targeted Metagenomes quantified eukaryotic read fractions across 2,759 NMDC ReadbasedAnalysis runs from 9 studies using GOTTCHA2 2. It found eukaryotic reads in 77% of runs, a median fraction of 2.7% and a mean of 13.3%, and that plastid sequences made up a median 100% of detected eukaryotic signal 2. The project found that Kraken2 and Centrifuge produced approximately 0 Eukaryota signal because their NMDC reference databases were prokaryote-restricted 2. GOTTCHA2 was therefore the only usable estimator, and its values are relative rather than calibrated absolute contamination 2. It found that eukaryotic fraction differed by matrix (Kruskal–Wallis H=77.8, p=1.3×10⁻¹⁷) 2. However, each biome was approximately 80–100% nested within a single study, and an environment model scored R²=−0.30 under GroupKFold out-of-study validation (cross-validation that holds out complete studies as groups), with an out-of-study detection AUC (area under the ROC curve) of 0.56, approximately chance 2. The report’s statement that environment explained no more variance than study_id alone conflicts with its own random cross-validation (CV) table, which gives environment R²=0.35 against study_id-only R²=0.24 2. This inconsistency was left unresolved 2. Within the NEON soil study of 1,186 runs, the project found that local vegetation (H=119.1, p=7.6×10⁻²¹) and geography across 47 sites (H=310.4, p=2.4×10⁻⁴⁶) were predictive, with within-study R²=+0.17 ± 0.06 2. These within-study signals remain subject to possible sub-batch confounding and were shown for one soil study only 2.
Carbon Source Utilization Predicts Ecology and Lifestyle in Pseudomonas analyzed GapMind carbon-pathway predictions for 12,732 genomes across 433 Pseudomonas species clades (GTDB r214) 3. Comparing 7 Pseudomonas s.s. species and 189 Pseudomonas_E species with at least 5 genomes, it found that 43 of 62 pathways differed significantly (q < 0.05), including xylose completeness of 0.0% versus 74.4% and arabinose of 0.0% versus 62.6% 3. Among 54 free-living and plant-associated species with at least 5 genomes and at least 60% majority-environment agreement, the project found a significant but modest ecological signal, with a permutation test giving p = 0.006 3. A Random Forest classifier trained on 51 species in four environment classes reached a balanced accuracy of 0.408 +/- 0.169 against a 0.250 chance baseline, and the report judged ecology prediction only partially supported 3. It found that the primary axis of a principal component analysis (PCA) separated the two subgenera rather than lifestyles and noted that the analyses do not explicitly control for phylogenetic non-independence 3. The report also flagged two further limitations: P. aeruginosa comprised 53% of genomes (6,760/12,732), and GapMind omits aromatic degradation pathways 3. It treated the link between pathway loss and a transition to animal hosts as an untested hypothesis 3.