What an Omics Database Contains, and Which Ones to Use for Your Own Data

An omics database is a searchable store of molecular measurements or their annotations: variants, transcript abundances, protein levels, metabolite concentrations, and the mappings between them. Two kinds exist and they get conflated constantly. Reference/annotation databases tell you what a feature means (gnomAD allele frequencies, ClinVar assertions, UniProt protein entries and Reactome pathways). Cohort databases hold the raw or processed measurements from many individuals (GTEx, TCGA, UK Biobank, GEO and PRIDE). When you have your own molecular data, you query the first kind constantly and the second kind occasionally, for distribution context. The most common beginner error is downloading a cohort dataset expecting annotation and finding only matrices of numbers with no interpretation attached.

The omics data types, and what each file looks like

“Omics” means comprehensive measurement of one molecular layer. In practice five layers matter for a personal profile.

Genomics. DNA sequence. Deliverables: a CRAM or BAM of aligned reads (~30–60 GB for a 30x human genome as CRAM) and a VCF of small variants. The VCF holds ~4–5 million variants for one person against GRCh38, of which ~4–5 thousand are protein-altering rare variants. A separate structural variant VCF usually comes with them. Static across your lifetime except in somatic tissue.

Transcriptomics. RNA abundance, usually from blood in a personal context. Deliverables: FASTQ, a BAM, and a quantification table. We prefer salmon quant or kallisto against a GENCODE transcriptome plus decoy sequences, producing quant.sf with TPM and estimated counts per transcript. Around 12,000–15,000 genes detected above 1 TPM in whole blood, heavily dominated by hemoglobin genes unless globin-depleted.

Proteomics. Protein abundance in plasma. Either affinity-based (Olink, SomaScan) giving NPX or RFU values for a fixed panel of a few thousand proteins, or mass-spec giving peptide-spectrum matches plus inferred protein groups. These are not interchangeable: affinity platforms measure epitope binding, and a variant in the epitope region can shift the reading without changing protein concentration.

Metabolomics and clinical biomarkers. Metabolomics is the messiest layer, with large unannotated feature fractions and heavy batch structure ¹. Standard clinical chemistry (lipid panel, CMP, CBC, HbA1c and hs-CRP) is the well-calibrated subset, delivered as a handful of numbers in defined units.

Continuous glucose. Time series, typically 5-minute or 15-minute intervals, 288 or 96 points per day. Delivered as CSV with timestamp and mg/dL or mmol/L. This is the only layer that is dense in time rather than dense in features.

There are more layers than these (epigenomics, lipidomics, single-cell, extracellular vesicle omics), and databases exist for each. SVAtlas, for instance, assembles single extracellular vesicle profiles as a dedicated resource ². Disease-focused aggregations do the same thing for one condition, pulling genomic, transcriptomic, proteomic, and methylation biomarkers into one queryable store ³.

Which databases to query

For a personal profile, this is the short list we use and why.

Variant annotation. Run Ensembl VEP with a cache, not the web interface, once you exceed a few hundred variants:

vep -i sample.vcf.gz -o annotated.vcf --vcf --cache --offline \
  --assembly GRCh38 --everything --fork 8 \
  --plugin CADD,whole_genome_SNVs.tsv.gz \
  --custom clinvar.vcf.gz,ClinVar,vcf,exact,0,CLNSIG,CLNREVSTAT

--everything turns on SIFT, PolyPhen, gnomAD frequencies, and regulatory consequences. The two fields that do most of the filtering work are gnomAD popmax allele frequency (drop anything above 1% if you are hunting rare coding variation) and ClinVar review status. Treat CLNREVSTAT as mandatory: a “pathogenic” assertion with one submitter and no assertion criteria is a different object from a two-star reviewed one. Alternative: Annovar, faster to set up, weaker on transcript consequence handling for multi-transcript genes.

Population frequency. gnomAD v4 directly, via its GraphQL API or the per-chromosome Hail tables if you are doing this at scale. The important caveat is ancestry: if your genetic ancestry is poorly represented in gnomAD, “rare” in the aggregate frequency can mean “common in your population,” and that changes interpretation completely.

Pharmacogenomics. PharmGKB and CPIC for star-allele definitions. Call diplotypes with PharmCAT rather than hand-matching rsIDs. The reason is that CYP2D6 involves copy number and hybrid alleles that a naive rsID lookup will get wrong. Nothing in a PGx report should change a medication without a clinician and, in most cases, a clinical-grade confirmatory test. Research-grade sequencing is not a prescribing instrument.

Expression context. GTEx v8 median TPM by tissue, to answer “is this gene even expressed where I think it is.” Human Protein Atlas for protein-level tissue distribution. For your own blood RNA, the useful comparison is longitudinal, your sample against your prior samples. Do not compare against GTEx, because cross-study batch effects and library prep differences swamp the biological signal you care about.

Pathways and networks. Reactome and KEGG for pathway membership, STRING for interactions. If you want to see multiple omics layers on one graph, Omics Visualizer in Cytoscape handles the case where several measurements map to the same node. Examples are multiple phosphosites on one protein, or protein plus transcript for one gene. Plain Cytoscape does not do this well ⁴.

Proteomics repositories. PRIDE for raw mass-spec, and UniProt for entry-level annotation. If you have Olink data, the assay-to-UniProt mapping is in the platform’s own annotation file and you should keep it in version control, because panel composition changes between releases.

How to structure your own data so queries work

The thing that makes multi-omics usable is a single identifier spine. Pick one and convert everything to it at ingest.

• Genes: Ensembl gene IDs with version stripped (ENSG00000141510). Do not use HGNC symbols, because symbols get renamed and Excel corrupts them.
• Proteins: UniProt accessions, not gene symbols, because one gene maps to multiple isoforms and one Olink assay to one epitope.
• Variants: chromosome, position, ref and alt on GRCh38, left-normalized with bcftools norm -f GRCh38.fa -m -any. The rsIDs are not stable keys.
• Samples: your own sample ID plus collection timestamp in UTC, on every row of every table.

Store the wide matrices as Parquet, one file per omics layer, with features as rows and samples as columns, plus a long-format observations table for anything sparse. DuckDB over the Parquet files gives you SQL joins across layers without standing up a server:

SELECT r.gene_id, r.tpm, p.npx, v.consequence
FROM 'rna/*.parquet' r
LEFT JOIN 'prot/*.parquet' p USING (gene_id, sample_id)
LEFT JOIN 'variants/annotated.parquet' v USING (gene_id)
WHERE r.tpm > 1;

Keep provenance columns: pipeline version, reference genome build, annotation database release date. An annotation is a snapshot of consensus at a moment. ClinVar reclassifies variants, and a result you recorded in 2024 may not be the current assertion. Re-running annotation against a fresh cache every six months and diffing the output is the single highest-value maintenance job in a personal profile.

What integration across layers does and does not give you

The appeal of multi-omics is causal direction: a variant, an expression change, a protein change, and a biomarker shift lining up in one pathway is more convincing than any one of them. The practical difficulty is that layers have different noise structures, different missingness, and different batch behavior, and naive concatenation lets the highest-variance layer dominate everything ⁵. Metabolomics in particular carries annotation ambiguity and technical variance that need explicit handling before it joins the other layers ¹.

For one person, the strongest analyses are within-layer and longitudinal: your CGM response to a repeated meal, your hs-CRP trajectory across twelve draws, your blood transcriptome before and after a change you made. Cross-layer work at n=1 is hypothesis generation. The cohort databases exist because the statistics need many people, and the same tooling built for cross-sample screening generalizes across domains, from plastic-degrading enzymes in metagenomes ⁶ to algal bloom prediction from environmental samples ⁷. Your own data feeds into that logic as a single point, and that point is most useful when you sample it repeatedly.

Questions people also ask

What are omics data types? Genomics (DNA sequence), transcriptomics (RNA abundance), proteomics (protein abundance), metabolomics (small-molecule concentrations), and epigenomics (DNA methylation, chromatin state). Each has its own file formats, its own reference databases, and its own failure modes. Clinical chemistry and continuous glucose are not usually called omics but belong in the same store because they anchor the molecular layers to physiology.

What is meant by omics data? Measurement of an entire molecular class in one assay rather than one analyte at a time. A cholesterol test measures one thing. An untargeted metabolomics run measures thousands of features, most of which are unannotated. The tradeoff is breadth against per-measurement reliability.

What is a multi-omics database? A resource holding two or more molecular layers on the same samples, so you can ask whether a variant, a transcript change, and a protein change coincide. Disease-specific examples aggregate published biomarkers across layers for one condition ³. Discovery indexes like OmicsDI do something different: they federate metadata across repositories so you can find datasets, without hosting the measurements themselves.

Which omics database should I query first for my own variants? Ensembl VEP with a local cache, gnomAD for frequency, ClinVar with review status attached. Those three cover most of what you can responsibly conclude from a personal genome. Anything that looks clinically consequential goes to a genetic counselor or physician for confirmatory testing. A spreadsheet is not where it gets settled.

Footnotes

1. Su H. Chu, Mengna Huang, Rachel S. Kelly, et al. Integration of Metabolomic and Other Omics Data in Population-Based Study Designs: An Epidemiological Perspective. Metabolites, 2019. https://doi.org/10.3390/metabo9060117 ↩ ↩²

2. Zhongkai Wei, Na Zhou, Ming Jing, et al. SVAtlas: a comprehensive single extracellular vesicle omics resource. Nucleic Acids Research, 2025. https://doi.org/10.1093/nar/gkaf1189 ↩

3. Jianhua Li, Xiaotian Yang, Qinjie Chu, et al. Multi-omics molecular biomarkers and database of osteoarthritis. Database, 2022. https://doi.org/10.1093/database/baac052 ↩ ↩²

4. Marc Legeay, Nadezhda T. Doncheva, John H. Morris, et al. Visualize omics data on networks with Omics Visualizer, a Cytoscape App. F1000Research, 2020. https://doi.org/10.12688/f1000research.22280.1 ↩

5. Tonia S. Schwartz. The Promises and the Challenges of Integrating Multi-Omics and Systems Biology in Comparative Stress Biology. Integrative and Comparative Biology, 2020. https://doi.org/10.1093/icb/icaa026 ↩

6. Anna Krzynowek, Jasper Snoeks, Karoline Faust. PlasticEnz: An integrated database and screening tool combining homology and machine learning to identify plastic-degrading enzymes in meta-omics datasets. PLoS Computational Biology, 2026. https://doi.org/10.1371/journal.pcbi.1013892 ↩

7. Gwenn M. M. Hennon, Sonya T. Dyhrman. Progress and promise of omics for predicting the impacts of climate change on harmful algal blooms. Harmful Algae, 2019. https://doi.org/10.1016/j.hal.2019.03.005 ↩