What blood test startups sell and how to read the results

Oak
blood-tests biomarkers diagnostics lab-data multi-omics
A small glass-skinned feathered creature on a black plinth with glowing veins, facing an arc of a hundred lit serum ampoules.

Blood test startups like Function Health and Superpower are order-and-interpretation layers on top of the same reference laboratories your physician uses, mostly Quest Diagnostics and Labcorp. For $349-$499 a year they arrange a physician order for a 100-160 analyte panel, route you to a draw site, and present the results in a web app with plain-language annotations and trend lines. They are legitimate in the narrow sense that matters: the assays are run in CLIA-certified, CAP-accredited labs on the same platforms that produce clinical results, and the numbers are real numbers. What you are paying for is the ordering logistics, the bundled price, and the interface. What you are usually not getting is analytical detail, structured data you can compute on, or the statistical framework needed to tell a real change from assay noise.

Who runs the assay, and why that is the only question that matters for accuracy

Nearly every consumer blood panel resolves to a small number of high-volume laboratories. By revenue, Labcorp is the larger of the two national players (roughly $13B in 2024 against Quest’s roughly $10B), and the two of them, plus Sonic Healthcare’s US operations, Mayo Clinic Laboratories, and ARUP, cover most of what a consumer company can order. The startup selects the panel and the venue. The analytical performance belongs to the lab.

This matters because the same analyte name can mean very different measurements. 25-hydroxyvitamin D by chemiluminescent immunoassay and by liquid chromatography-tandem mass spectrometry (LC-MS/MS) disagree, particularly at low concentrations and in the presence of the C3-epimer. Total and free testosterone by immunoassay are unreliable at the low end, which is where women’s and hypogonadal men’s results live, and LC-MS/MS with equilibrium dialysis is the reference approach. Lipoprotein(a) is reported in mg/dL by mass-based assays and nmol/L by particle-number assays, and the two do not convert with a single constant. Apolipoprotein B is measured directly by immunoturbidimetry, whereas LDL cholesterol on a standard lipid panel is often calculated by the Friedewald equation and becomes unreliable above about 400 mg/dL of triglycerides.

So the first thing to ask a vendor is not how many biomarkers are in the panel but which lab, which method, and which LOINC code for each analyte. LOINC (Logical Observation Identifiers Names and Codes) distinguishes methods, specimen types, and units, which is exactly the granularity you need if you intend to compare a result from 2026 against one from 2023.

The panel-size problem

Clinical reference intervals are usually central 95% ranges from a nominally healthy population. Order 100 independent analytes on a healthy person and the probability that at least one falls outside its interval is 1 − 0.95^100, about 99.4%. Biomarkers are correlated, so the true figure is lower, but the conclusion holds: a 150-analyte panel will flag something for nearly everyone, and the app will show you a yellow or red marker. This is the arithmetic of multiple comparisons, not a finding about you.

The second statistical issue is distinguishing a change from noise. For a single analyte, the reference change value is approximately 2.77 × √(CV_a² + CV_i²), where CV_a is analytical imprecision and CV_i is within-subject biological variation. TSH has a within-subject CV around 19-20%, so two TSH values must differ by more than roughly 55% before you should treat the change as real. Total cholesterol has a within-subject CV near 6%, so a much smaller delta is meaningful. Most consumer dashboards draw a line between two points and let the slope speak, which overstates signal for high-variability analytes and understates it for stable ones. The EFLM Biological Variation Database publishes CV_i estimates, and computing your own reference change values per analyte is a twenty-line script that will change how you read your own trend lines.

Pre-analytical conditions dominate more results than most people expect. Fasting state, time of day (cortisol and testosterone have steep diurnal curves), posture during the draw, tourniquet time, hemolysis, and recent exercise all move numbers by more than the assay’s imprecision. If you want comparable longitudinal data, fix the protocol: same draw site, same time of day, same fasting duration, same lab.

Getting your data out in a form you can compute on

Assume you will need to do this yourself. Ask before you pay whether the vendor provides results as structured data or only as a rendered PDF. The formats worth asking for, in descending order of usefulness:

  • FHIR R4 Observation and DiagnosticReport resources, LOINC-coded, with UCUM units and per-analyte reference ranges. This is what you want. Observation.valueQuantity, Observation.referenceRange, and Observation.method give you everything needed for a longitudinal table.
  • HL7 v2.5.1 ORU^R01 messages, which is what the labs emit internally. Parseable with hl7apy in Python; the OBX segments carry the values.
  • CSV export with one row per analyte per date, including units and method.
  • PDF only. Workable but lossy. pdfplumber handles most lab reports better than PyPDF2 because it preserves table geometry, and camelot with flavor="lattice" works when the report has ruled tables. Budget real time for unit normalization and for the fact that reference ranges change between report versions.

Both Quest and Labcorp expose patient-facing portals and developer APIs, and results ordered through a consumer company generally also exist in the lab’s own portal, which is sometimes the cleanest path to the raw report. Store everything in a single long-format table keyed on (LOINC code, collection datetime, method, lab) rather than on the analyte’s display name. Display names drift. LOINC codes do not.

Where the blood panel genuinely stops

A standard chemistry-plus-immunoassay panel measures a few dozen to a couple hundred analytes chosen because they have decades of clinical validation. The research frontier is measuring thousands of molecular features and looking for patterns, which is a different activity with different statistics. Affinity proteomics platforms now quantify thousands of plasma proteins per sample, and the case for early cancer detection rests on multi-protein signatures rather than single markers 1. Circulating cell-free DNA methylation profiles are being evaluated as noninvasive markers for conditions that currently require imaging or invasive sampling, including multiple sclerosis 2. Menstrual effluent proteomics has been proposed as a route to diagnosing endometriosis, which today often takes years and a laparoscopy 3. Single-ratio electronic assays are being developed for neurodegenerative disease, such as the GPNMB/GRN ratio for early Parkinson’s disease 4. Even non-molecular signals are under study as biomarkers, with voice acoustics being one active area 5.

None of these are in a $400 consumer panel, and most are not yet validated for individual decision-making. Their relevance to you is that they define what a molecular baseline will look like in a few years, and that a wide baseline taken now is the comparison point for any of them. Interpretation of any panel result that falls outside its reference interval, or any trend you believe is real, belongs with a clinician who can see your history and examine you. Data ownership and clinical judgment are separate goods, and buying the first does not substitute for the second.

What we would do

Buy the cheapest legitimate route to the assays you want, insist on structured output, and control the pre-analytical variables yourself. Concretely: pick your analytes deliberately rather than accepting a 150-item bundle, prefer LC-MS/MS methods where they exist (vitamin D, testosterone, thyroid hormones), request direct apoB and a particle-number Lp(a) rather than calculated LDL-C alone, draw at the same time of day after the same fasting interval, and use the same lab every time. Then compute reference change values before you interpret any delta. The membership model is fine value for the ordering convenience. The interpretation layer is the part you should be willing to replace with your own code.

Questions people also ask

Is Superpower testing legit, and is it worth it? The company is real and the assays are run at established reference labs, mostly through Quest draw sites, so the results are as valid as the underlying methods. Whether it is worth roughly $349-$499 a year depends on whether you would otherwise pay cash for a comparable panel (frequently more) and on whether you can get structured data out. Confirm the export format before subscribing.

Who is bigger, Quest or Labcorp? Labcorp is larger by revenue, on the order of $13B versus roughly $10B in recent years, though the two are comparable in US clinical testing volume. For your purposes they are interchangeable in quality and not interchangeable in results: switch labs and you introduce a method-level discontinuity in your own time series.

Would Theranos ever be possible? Some of it, eventually, and for a narrower set of analytes than was claimed. The barrier was never miniaturization alone but the sample: a fingerstick drop gives you a few dozen microliters of variably hemolyzed capillary blood, and separating plasma at that scale is a genuine engineering problem that researchers are still solving with on-strip plasma separation units in paper microfluidic devices 6. High-sensitivity single-analyte platforms on small volumes are advancing 4. A full chemistry panel from one drop, with clinical-grade precision across four orders of magnitude of analyte concentration, remains out of reach.

Does Theranos still exist? No. Its Newark, California lab was sanctioned and closed after 2016, the company dissolved in 2018, and Elizabeth Holmes and Ramesh Balwani were convicted of fraud and imprisoned.

What are the major blood-testing companies? For clinical volume in the United States: Labcorp, Quest Diagnostics, Sonic Healthcare USA, Mayo Clinic Laboratories, and ARUP Laboratories, with BioReference and academic hospital labs also significant. Consumer brands sit on top of these and rarely operate their own wet lab.

Oak builds longitudinal molecular profiles of individuals: whole-genome sequencing, RNA sequencing, proteomics, blood biomarkers, and continuous glucose data, integrated into one model of you. Build your profile.

Footnotes

  1. Lubna Therachiyil, Anju Surendranath, Anjana Anand, et al. Proteomics-Driven Cancer Biomarkers for Early Detection and Targeted Therapy: Insights from the Middle East. Journal of Proteome Research, 2025. https://doi.org/10.1021/acs.jproteome.5c00949 ↩

  2. Hailu Fu, Jocelyn Liang, Kevin Huang, et al. Circulating cell-free DNA methylation profiles as candidate noninvasive multiple sclerosis biomarkers. Communications Medicine, 2026. https://doi.org/10.1038/s43856-026-01825-x ↩

  3. Carrie Arnold. Menstrualome aids difficult endometriosis diagnosis. Nature Biotechnology, 2026. https://doi.org/10.1038/s41587-026-03265-3 ↩

  4. Xiaohong Jiang, Jintao Zheng, Jiacheng Yu, et al. Rapid Hematological Profiling of the GPNMB/GRN Ratio via Bioelectronic Platform for Early Diagnosis of Parkinson’s Disease. Advanced Materials, 2026. https://doi.org/10.1002/adma.74326 ↩ ↩2

  5. Tanja Knaus, Susanne Bauer. From genome to voiceome: the quest for voice-based biomarker technologies in health research. BioSocieties, 2026. https://doi.org/10.1057/s41292-026-00394-5 ↩

  6. Xuefei Gao, Jennifer Boryczka, Sujan Kasani, et al. Enabling Direct Protein Detection in a Drop of Whole Blood with an “On-Strip” Plasma Separation Unit in a Paper-Based Lateral Flow Strip. Analytical Chemistry, 2020. https://doi.org/10.1021/acs.analchem.0c02555 ↩