Which Biomarkers Are Worth Measuring, and Why

A biomarker is worth measuring if it has low within-person variability relative to its between-person spread, if it changes for reasons you can identify, and if a repeat measurement means something. By that standard the short list is smaller than most panels suggest. It covers apolipoprotein B and lipoprotein(a) once in a lifetime. It covers HbA1c with fasting insulin and glucose. It covers ALT and GGT. It covers ferritin with transferrin saturation. It covers creatinine with cystatin C. It covers the urine albumin-to-creatinine ratio and TSH with free T4. It ends with a complete blood count and hs-CRP read as a trend rather than a point. Everything else on a 100-marker menu is either a research assay, a marker with variability so large that one draw tells you little, or a number that only becomes informative in a clinical context where a physician is interpreting it alongside you.

This page covers how we rank biomarkers, the numbers that decide whether a single measurement is interpretable, and what the genome, transcriptome, proteome, and continuous glucose layers add that a blood panel cannot.

The criteria we use to rank a biomarker

The first criterion is the index of individuality: the ratio of within-person biological variation to between-person variation. When that ratio is low, your own value sits in a narrow band and a population reference interval is nearly useless for you. HbA1c is the clean example. Within-person variation is on the order of two percent, so a move from 5.1 to 5.6 percent is a real change even though both values fall inside the reference range.

The second criterion is analytical variation. Immunoassays for ferritin and hs-CRP carry method-dependent bias, so results from two labs on the same specimen can differ by more than the change you are trying to detect. Fix the lab and the assay platform before you start reading trends. Record the method on every result.

The third criterion is whether the marker responds to anything identifiable. Lipoprotein(a) is largely set by the LPA gene, which is why one measurement is usually enough and why the result belongs in a conversation with a clinician rather than in a dashboard. Fasting insulin responds to sleep, training load, and diet within weeks. That makes it a useful repeated measure.

The blood panel we would order

Start with the lipid layer. ApoB in mg/dL counts atherogenic particles directly and has a within-person variation of roughly seven percent, which is tighter than calculated LDL cholesterol. Order Lp(a) in nmol/L once, on a molar assay rather than a mass assay, because mass units depend on isoform size.

For glucose and insulin, take HbA1c, fasting glucose, and fasting insulin from the same draw and compute HOMA-IR yourself. HbA1c is biased by red cell lifespan, so read it next to the CBC. A high reticulocyte count or a low haptoglobin will pull it down for reasons unrelated to glucose.

For the liver and iron, take ALT, GGT, ferritin, and transferrin saturation together. Ferritin rises with inflammation, so a ferritin above the reference interval with a normal transferrin saturation and a raised hs-CRP is usually an inflammatory signal rather than iron loading. Transferrin saturation above about 45 percent on two fasting draws is the pattern that warrants a clinician and a look at HFE genotype.

For the kidney, order creatinine and cystatin C in the same tube. The two equations disagree in people with unusual muscle mass, and the disagreement itself is informative. Add a urine albumin-to-creatinine ratio, which detects early glomerular change that serum markers miss.

Treat hs-CRP as a trend. Within-person variation is large, plausibly forty percent or more, so a single value of 3 mg/L means little. Two draws two weeks apart, both above 2 mg/L and with no recent infection, is a pattern to bring to a physician.

Continuous glucose data is the highest-resolution layer you can buy

A 14-day continuous glucose monitor produces roughly 1,300 to 4,000 readings, depending on sampling interval. That density supports statistics no blood draw can reach. We compute mean glucose, standard deviation, and coefficient of variation. We also compute time between 70 and 140 mg/dL and the shape of the postprandial curve, meaning peak height above baseline and minutes to return within 15 mg/dL of baseline.

The last of those is the most useful and the least reported. Two people can share a mean of 98 mg/dL while one returns to baseline in 90 minutes and the other takes three hours. Export the raw stream rather than the app summary. Dexcom Clarity gives CSV with a timestamp and a value per row, and Libre exports a similar file with a record-type column that distinguishes scan values from historic values. Filter on record type before averaging or you will double-count.

Sensor accuracy sets the floor on interpretation. Mean absolute relative difference of around nine percent means a reading of 140 mg/dL is compatible with a venous value near 127. Compare curves within a sensor, not across sensors, and discard the first 24 hours of each wear.

What the omics layers add

Whole-genome sequencing at 30x mean depth is a one-time measurement. Ask for the CRAM or BAM aligned to GRCh38 plus a gVCF. A filtered variant list is not enough. DeepVariant or GATK HaplotypeCaller will reproduce the small-variant calls, and you will need the alignment itself for copy number and for the pharmacogenes. CYP2D6 in particular requires star-allele calling with a tool like PyPGx or PharmCAT, because hybrid alleles and whole-gene duplications are invisible to a plain VCF. Pathogenic findings in the ACMG secondary-findings genes require a genetic counselor or physician to interpret and, in most cases, confirmation on an orthogonal assay.

Blood RNA sequencing measures state rather than sequence. Draw into PAXgene tubes and insist on globin depletion, because HBB and HBA transcripts otherwise consume a large share of reads. Quantify with salmon in selective-alignment mode against a decoy-aware index, then read TPM values with the knowledge that most of the signal reflects cell composition. A shift in neutrophil fraction will move hundreds of genes at once. Deconvolution with a reference signature matrix separates composition from per-cell expression, and without that step a differential expression list is mostly a blood count in disguise.

Circulating nucleic acids are a distinct class. Plasma microRNAs are stable and measurable, and they have been studied extensively as cancer markers, though assay standardization and normalization remain the limiting problems in moving them from research to routine use.¹ Treat any miRNA panel as exploratory data rather than a screening result.

Affinity proteomics gives you 3,000 to 7,000 protein measurements per sample. Olink reports NPX on a log2 scale and SomaScan reports relative fluorescence units, and neither is an absolute concentration. Values are comparable within a batch and require bridging samples across batches. The use of a proteomic panel today is as a longitudinal self-comparison: run the same platform on samples six to twelve months apart and look at what moved. Broader multi-omic profiling is advancing fast as a research approach to disease characterization.²

The failure modes that ruin longitudinal panels

Pre-analytical handling dominates. Potassium rises with hemolysis, glucose falls in a tube left at room temperature without fluoride, and platelet counts drift with delayed processing. Draw at the same time of day, after a similar fast, and off intense training by at least 48 hours.

Lab switching is the second failure. A different immunoassay for ferritin or TSH can introduce a step change that looks like a biological trend. Keep a column for lab and method in your own table and annotate every switch.

The third is multiple comparisons. A 100-marker panel will produce about five out-of-range values by chance in a healthy person. Decide in advance which markers you will act on, and treat the rest as context that a clinician reads with you.

Questions people also ask

How often should I retest? Match the interval to the marker’s biology. HbA1c reflects roughly the prior three months, so quarterly is the fastest useful cadence. ApoB and liver enzymes respond within four to six weeks of a change. Lp(a) needs one measurement, ever.

Is a 100-biomarker panel better than a 20-marker panel? Only if you know what to do with the extra 80. More markers raise the chance of a spurious out-of-range result without adding decision-relevant information. We would rather have twenty markers measured four times than a hundred measured once.

What file formats should I insist on? For the genome, CRAM or BAM plus gVCF. For RNA, FASTQ plus a transcript-level quantification table. For proteomics, raw NPX or RFU with batch identifiers. For glucose, CSV with timestamps. A PDF report is not data.

Can an AI agent interpret these files for me? It can compute on them, plot them, and flag patterns across layers. That is genuinely useful. It cannot replace a clinician for a pathogenic variant, an abnormal result, or any decision about treatment.

Footnotes

1. Vanashri Nawale, Avinash Darekar, Ravindranath Saudagar. Micro-RNA: biomarker for the treatment of cancer. Journal of Drug Delivery and Therapeutics, 2019. https://doi.org/10.22270/jddt.v9i2-s.2464 ↩

2. Prerna Mehta. Revolutionizing therapeutics: Exploring novel biotechnological methods for disease management and treatment. Biophilia Insights, 2025. https://doi.org/10.52679/bi.e202531004 ↩