Limit of Detection (LOD)

The limit of detection is the lowest analyte concentration an assay can distinguish from zero with a stated probability, conventionally 95%. It is a property of the assay, the instrument, the reagent lot, and the matrix (serum, plasma, CSF, stool), not a property of the analyte. Below the LOD an assay still produces a number. That number is noise.

How it works

The standard chain, from CLSI EP17, has three rungs.

Limit of blank (LoB): run 60 or more replicates of a blank or analyte-free matrix. LoB is the 95th percentile of those results, or parametrically mean_blank + 1.645 × SD_blank. It is the value a blank exceeds only 5% of the time, the false-positive threshold.

Limit of detection (LOD): run replicates of samples spiked just above LoB. LOD is the concentration at which 95% of measurements exceed LoB, approximated as LoB + 1.645 × SD_low. When blank and low-level SDs are similar this collapses to the familiar mean_blank + 3 × SD_blank. Type I and Type II error are both controlled at 5%.

Limit of quantitation (LoQ): the lowest concentration meeting a predefined precision goal, typically CV ≤ 20% (immunoassays) or ≤ 10% (clinical chemistry), and often a bias goal too. LoQ ≥ LOD, always. If someone reports LoQ below LOD, the definitions were mixed up.

The EPA’s method detection limit (MDL) is the same idea with a different statistic: seven or more replicate spikes at a low level, then MDL = t(n-1, 0.99) × SD. For n = 7, t = 3.143. MDL uses a one-sided 99% bound and no separate blank study, so it controls false positives more strictly and says nothing about false negatives. It is an environmental-methods convention. Do not compare an MDL to a clinical LOD as if they were the same quantity.

For counting assays (PCR, sequencing) LOD is fit as a probit or logistic curve of hit rate against input concentration, and the 95% detection point is read off the fit. FungiQuant’s qPCR LOD was established this way, at single-digit copies per reaction, using a dilution series with many replicates per level.¹ A clinical metagenomic sequencing assay for CSF pathogens was validated the same way, with organism-specific LODs in genome copies per mL, because detection depends on genome size, GC content, and host background.² Cross-platform, the spread is large: rapid SARS-CoV-2 antigen tests detect reliably only at high viral loads, well above RT-PCR, which is why sensitivity looked good in high-load samples and poor in low-load ones.³ Even among molecular assays for the same target, LODs differ several-fold across platforms.⁴

In your own data

Proteomics is where you meet LOD most directly. Olink delivers NPX values plus a per-assay LOD (derived from negative controls, typically background + 3 SD) and a Below LOD flag. Plate LODs shift between runs, so a protein can be flagged on one plate and not the next without changing concentration. Check the flag rate per assay across your timeline before plotting anything: an assay with 60% below-LOD calls is telling you about the assay, not about you.

Clinical chemistry reports censoring directly. hs-CRP <0.2 mg/L, troponin <3 ng/L, TSH <0.01 mIU/L. That string is a left-censored observation. Parsing it to 0 or, worse, 0.2 creates artificial floors in a time series.

In sequencing, LOD is expressed as variant allele fraction at a given depth. At 30x WGS coverage, a 5% mosaic variant is essentially undetectable: the expected alt read count is 1.5. Somatic and mosaic calling needs depth in the hundreds to thousands, and the platform’s own error profile sets the floor. Two-base encoding was introduced specifically to push that error floor down and separate real low-frequency variants from sequencing artifacts.⁵ For copy number, the detection floor is a joint function of probe or read density and segment size, which is why the same deletion is called by one method and missed by another.⁶

RNA-seq has no formal LOD but behaves like one: transcripts under roughly 1 TPM are dominated by counting noise at typical library depths, and zero counts are a mix of true absence and undersampling.

Common mistakes we see, in order of damage done:

1. Substituting LOD/√2 for censored values and then running a paired test. It compresses variance and can manufacture significance.
2. Treating below-LOD as zero, then log-transforming.
3. Comparing values across batches with different LODs without refiltering on the union of flags.
4. Reading longitudinal variation in a near-LOD analyte as biology. The signal is heteroscedastic at the floor.

Our preference: keep a censoring indicator column alongside every value, filter analytes by below-LOD rate (we use ≤ 25% across timepoints) before modeling, and use a censored-regression or Tobit model when the analyte genuinely matters. Never impute silently.

Limitations

LOD is estimated from a specific matrix with a specific interference profile. Your plasma is not the manufacturer’s spiked pool. Assays validated on clean spike-ins routinely lose sensitivity in real samples, and systematic side-by-side evaluation of detection assays shows sensitivity varying by orders of magnitude across methods for the same target.⁷ Reported LODs are also single numbers standing in for a probability curve: detection at 50% of the LOD is not zero, and detection at the LOD is 95%, not 100%.

A below-LOD result is not evidence of absence, and a just-above-LOD result is not a clinical finding. If a measurement near the floor matters to a decision about your health, the next step is a repeat on a more sensitive method ordered and interpreted by a clinician.

Woolf Software 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.