{"schema_version": "1.0", "id": "z-prime-detector-linearity", "canonical_url": "https://discoveryinpractice.com/articles/z-prime-detector-linearity/", "title": "A better Z′ can hide a worse measurement", "author": {"name": "Andrew Stewart", "url": "https://discoveryinpractice.com/about/#andrew-stewart"}, "language": "en", "summary": "A higher Z′ can reflect nonlinear detector compression rather than a better assay. This article explains how bright-well variation can shrink while effect sizes become distorted, and how to check proportionality without confusing detector behavior with assay chemistry.", "takeaways": ["Z′ measures control separation and variability; it does not establish detector linearity or accuracy between the controls.", "Shorter integration can reduce accumulated counts without reducing the photon arrival rate that causes counting losses.", "Check low, intermediate, and bright signals using manufacturer-supported attenuation or a calibrated verification method; preserve the original well data."], "limitations_summary": "Compression does not always increase Z′. The numerical example is a constructed model, not a measurement of any particular reader.", "topics": ["Assay quality", "Detector linearity", "Luminescence", "Z-prime"], "publication_status": "published", "date_published": "2026-09-25", "date_modified": "2026-09-25", "license": "CC-BY-4.0", "license_url": "https://creativecommons.org/licenses/by/4.0/", "license_status": "Published under CC BY 4.0", "content_version": "1.0", "body_html": "

The bright controls have become unusually consistent. Their coefficient of variation has dropped, the Z-prime factor has improved, and the plate passes QC. Before crediting the assay optimization, check whether the reader is still measuring those wells proportionally.

In a luminescence assay, detector compression can make different bright signals look more alike. Under some conditions, that improves Z′ while making the measurements less accurate. The controls look tidier because part of their variation has been lost in detection.

What Z-prime measures

The Z-prime factor, usually written Z′, compares the separation of two control populations with their variability. Using H and L for high- and low-signal controls:

Z′ = 1 − 3(σH + σL)|μH − μL|

Here, μ is the mean and σ is the standard deviation. Larger separation raises Z′; larger standard deviations lower it. The original formulation by Zhang, Chung and Oldenburg remains a useful way to assess screening controls. [1]

The formula contains no check that light input and reported signal are proportional. It uses the numbers the reader supplies. It also says nothing about the response between the two controls, where many compounds will fall.

Simply reducing every value by 20% cannot improve Z′. Both the standard deviations and the difference between the means fall by 20%, leaving their ratio unchanged. The problem considered here is nonlinear compression: the brighter signals lose a greater fraction of their counts.

How photon counting loses proportionality

Photon counting works by resolving and counting individual detector pulses. The detector and its electronics need a finite interval to distinguish successive events. At sufficiently high arrival rates, pulses overlap and some events are missed. Hamamatsu describes this count-rate dependence and uses a paralyzable detector model in its photon-counting simulator. [2]

Compression can begin while the output is still increasing. A numerical result below an obvious ceiling therefore does not establish linearity. Warning behavior depends on the instrument: Hamamatsu describes both photon-counting heads with over-light detection and heads without it. [3] Check what a reader's warning actually monitors and how that relates to its specified linear range.

This also explains why shortening the integration time may fail to help. For a stable glow, counting for one-tenth as long reduces the accumulated counts, but it does not reduce the rate at which pulses arrive. A remedy for a total-count limit is not necessarily a remedy for pulse overlap. Whether a particular setting changes either limit depends on the acquisition system.

A worked example of misleadingly good controls

Take five constructed values in each control group. The high control has a mean of 100 and a sample standard deviation of 10. The low control has a mean of 10 and a standard deviation of 1. Their Z′ is:

Z′ = 1 − 3(10 + 1)100 − 10 = 0.633

Now pass those same values through an illustrative compression curve, f(x) = x × exp(−0.00223144x). This is the shape of the paralyzable mean-response model, expressed in arbitrary units and chosen so that an input of 100 produces an output of 80. It represents no particular reader. [2]

Constructed example: linear and compressed control measurements
QuantityLinear responseCompressed response
High-control mean100.0079.87
High-control standard deviation10.006.22
Low-control mean10.009.78
Low-control standard deviation1.000.96
Z′0.6330.693

The inputs have not become more reproducible. Compression has reduced the spread of the high controls enough to outweigh the loss of control separation. Z′ rises from about 0.63 to 0.69, and the high-control CV falls from 10% to about 7.8%.

The same curve distorts intermediate values. An input of 55 lies halfway between the original control means and represents 50% inhibition under high-to-low normalization. After compression, it reads about 48.65. Normalize that value against the compressed control means in the table and it reports about 44.5% inhibition.

The curve is still increasing across this example, so it preserves the order of noiseless inputs. Distorted effect sizes do not require shuffled rankings. A cutoff applied to normalized inhibition can move a compound between hit and non-hit categories even if every well retains its place in the ranking.

These are deterministic calculations on invented well values. They illustrate distortion of between-well differences, without adding photon-counting noise or modeling an instrument's correction electronics. Nor does compression always raise Z′: it changes both the spread and the separation. If low-control variability dominates, the loss of separation can make Z′ worse.

The general statistical concern is already recognized in the Assay Guidance Manual, which notes that saturation artifacts can reduce control variability in imaging assays. [4] Imaging saturation and photon-counting losses have different physical causes, but both can make control variability a misleading guide to measurement quality.

Test detector linearity without changing the chemistry

A falling slope in a concentration series is a reason to investigate, but it does not identify the detector as the cause. The assay itself may be nonlinear. In CellTiter-Glo, for example, Promega notes that ATP content per cell can change with cell density, altering the relationship between cell number and luminescence. [5]

Optical attenuation helps separate those possibilities. It changes the light reaching the detector while leaving the sample chemistry in place. Hamamatsu describes a check using a neutral-density filter with known transmission: a 10% transmitting filter should reduce the measured count rate to one-tenth within the linear range. [3]

For a microplate reader, use a manufacturer-supported attenuation setting or a suitable calibrated verification device. A piece of dark plastic over the plate is not a calibrated filter. Transmission must be known over the relevant emission wavelengths, and the arrangement must preserve the measurement geometry.

Compare several brightness levels, including the brightest expected samples and intermediate controls. Account for dark/background contributions. Confirm whether the software reports raw attenuated output or automatically compensates for transmission; the expected numerical ratio differs. Hold temperature and assay age steady, and balance measurement order so that glow decay does not masquerade as nonlinearity.

If you use dilution instead, keep the final volume and matrix composition matched as far as the assay permits. Verify that the dilution changes light production proportionally. Diluting an active luciferase mixture can change substrate conditions, inhibitors, and reaction kinetics; diluting cells can change biology. A dilution series tests the combined assay-and-reader response unless those contributions are controlled.

Changing the settings requires another check

Lowering PMT gain is not interchangeable with reducing the light arriving at a photon counter. It changes pulse amplitudes and can affect which pulses cross the counting threshold. [3] Use the instrument's validated settings and assess linearity again; a smaller displayed number alone is insufficient.

Readers can extend useful range through faster counting electronics, appropriate count-loss correction, optical attenuation, or an alternative acquisition mode. Hamamatsu documents both count-rate correction and the hardware dependence of counting linearity. [3] The relevant specification is the linear range of the measurement mode you will use, with an acceptable error over the signals your assay produces.

An advertised range spanning many orders of magnitude is incomplete information without the test conditions. Ask whether it applies within one acquisition setting, how measurements are combined when settings change, and what happens at the transition. If two modes overlap, compare them there using stable samples. An alternative mode needs its own linearity and background checks.

Choose settings that preserve weak signals while accommodating the brightest plausible wells. Verify intermediate responses before comparing Z′. Otherwise, optimization can favor a setting that suppresses bright-well variation at the expense of accurate effect sizes.

What to include in assay validation

Inspect the raw control distributions. Save means, standard deviations, CVs, acquisition settings, and the individual well values alongside Z′. Unexpected tightening of the bright controls deserves investigation, but does not by itself prove compression.

Include intermediate controls or a reference concentration series. The two endpoints can look acceptable while the response between them is distorted. The Assay Guidance Manual recommends examining raw trends and reference compounds at multiple concentrations during pilot screening. [6]

Check more than one brightness range. Cover ordinary controls, the brightest expected samples, and enough higher signal to test the margin you intend to rely on.

Repeat the relevant checks after changes. A brighter reagent formulation, a different plate, or new acquisition settings can move the assay outside the range you previously established.

Investigate suspect runs before correcting them. Preserve the original data. Establish the response experimentally and assess whether stored samples can be remeasured under suitable conditions; dividing everything by a guessed recovery factor will not undo nonlinear compression.

References

\n

1. Zhang JH, Chung TDY, Oldenburg KR. A simple statistical parameter for use in evaluation and validation of high throughput screening assays. Journal of Biomolecular Screening. 1999;4(2):67–73. DOI: 10.1177/108705719900400206. Formula also presented in reference 4, section 2.

\n

2. Hamamatsu Photonics. Photon Counting SNR Simulator. Count Rate Linearity section and paralyzable response equation. Accessed September 24, 2026.

\n

3. Hamamatsu Photonics. Photomultiplier Tubes: Basics and Applications, fourth edition. Printed pp. 145–147 on pulse-height discrimination, pp. 149–150 on count-rate linearity/correction, and p. 181 on over-light detection and attenuation checks; PDF pp. 158–160, 162–163 and 194.

\n

4. Bray MA, Carpenter A. Advanced Assay Development Guidelines for Image-Based High Content Screening and Analysis. Assay Guidance Manual. July 8, 2017. Section 2, Z′ and V-factor discussions.

\n

5. Promega. CellTiter-Glo 2.0 Assay technical manual TM403. Revised January 2023. Section 4.B, Cellular ATP Content; PDF p. 11 / printed p. 10.

\n

6. Assay Development for Protein Kinase Enzymes. Assay Guidance Manual. Pilot Screens section. Accessed September 24, 2026.

\n
", "body_text": "Introduction\nThe bright controls have become unusually consistent. Their coefficient of variation has dropped, the Z-prime factor has improved, and the plate passes QC. Before crediting the assay optimization, check whether the reader is still measuring those wells proportionally.\n\nIn a luminescence assay, detector compression can make different bright signals look more alike. Under some conditions, that improves Z′ while making the measurements less accurate. The controls look tidier because part of their variation has been lost in detection.\n\nWhat Z-prime measures\nThe Z-prime factor, usually written Z′, compares the separation of two control populations with their variability. Using H and L for high- and low-signal controls:\n\nZ prime equals one minus three times the sum of the high and low control standard deviations divided by the absolute difference between their means.\n\nHere, μ is the mean and σ is the standard deviation. Larger separation raises Z′; larger standard deviations lower it. The original formulation by Zhang, Chung and Oldenburg remains a useful way to assess screening controls. [1]\n\nThe formula contains no check that light input and reported signal are proportional. It uses the numbers the reader supplies. It also says nothing about the response between the two controls, where many compounds will fall.\n\nSimply reducing every value by 20% cannot improve Z′. Both the standard deviations and the difference between the means fall by 20%, leaving their ratio unchanged. The problem considered here is nonlinear compression: the brighter signals lose a greater fraction of their counts.\n\nHow photon counting loses proportionality\nPhoton counting works by resolving and counting individual detector pulses. The detector and its electronics need a finite interval to distinguish successive events. At sufficiently high arrival rates, pulses overlap and some events are missed. Hamamatsu describes this count-rate dependence and uses a paralyzable detector model in its photon-counting simulator. [2]\n\nCompression can begin while the output is still increasing. A numerical result below an obvious ceiling therefore does not establish linearity. Warning behavior depends on the instrument: Hamamatsu describes both photon-counting heads with over-light detection and heads without it. [3] Check what a reader's warning actually monitors and how that relates to its specified linear range.\n\nThis also explains why shortening the integration time may fail to help. For a stable glow, counting for one-tenth as long reduces the accumulated counts, but it does not reduce the rate at which pulses arrive. A remedy for a total-count limit is not necessarily a remedy for pulse overlap. Whether a particular setting changes either limit depends on the acquisition system.\n\nA worked example of misleadingly good controls\nTake five constructed values in each control group. The high control has a mean of 100 and a sample standard deviation of 10. The low control has a mean of 10 and a standard deviation of 1. Their Z′ is:\n\nZ prime equals one minus three times the quantity ten plus one divided by the quantity one hundred minus ten, which equals 0.633.\n\nNow pass those same values through an illustrative compression curve, f(x) = x × exp(−0.00223144x). This is the shape of the paralyzable mean-response model, expressed in arbitrary units and chosen so that an input of 100 produces an output of 80. It represents no particular reader. [2]\n\nConstructed example: linear and compressed control measurementsQuantity | Linear response | Compressed response | High-control mean | 100.00 | 79.87 | High-control standard deviation | 10.00 | 6.22 | Low-control mean | 10.00 | 9.78 | Low-control standard deviation | 1.00 | 0.96 | Z′ | 0.633 | 0.693 |\n\nThe inputs have not become more reproducible. Compression has reduced the spread of the high controls enough to outweigh the loss of control separation. Z′ rises from about 0.63 to 0.69, and the high-control CV falls from 10% to about 7.8%.\n\nThe same curve distorts intermediate values. An input of 55 lies halfway between the original control means and represents 50% inhibition under high-to-low normalization. After compression, it reads about 48.65. Normalize that value against the compressed control means in the table and it reports about 44.5% inhibition.\n\nThe curve is still increasing across this example, so it preserves the order of noiseless inputs. Distorted effect sizes do not require shuffled rankings. A cutoff applied to normalized inhibition can move a compound between hit and non-hit categories even if every well retains its place in the ranking.\n\nThese are deterministic calculations on invented well values. They illustrate distortion of between-well differences, without adding photon-counting noise or modeling an instrument's correction electronics. Nor does compression always raise Z′: it changes both the spread and the separation. If low-control variability dominates, the loss of separation can make Z′ worse.\n\nThe general statistical concern is already recognized in the Assay Guidance Manual, which notes that saturation artifacts can reduce control variability in imaging assays. [4] Imaging saturation and photon-counting losses have different physical causes, but both can make control variability a misleading guide to measurement quality.\n\nTest detector linearity without changing the chemistry\nA falling slope in a concentration series is a reason to investigate, but it does not identify the detector as the cause. The assay itself may be nonlinear. In CellTiter-Glo, for example, Promega notes that ATP content per cell can change with cell density, altering the relationship between cell number and luminescence. [5]\n\nOptical attenuation helps separate those possibilities. It changes the light reaching the detector while leaving the sample chemistry in place. Hamamatsu describes a check using a neutral-density filter with known transmission: a 10% transmitting filter should reduce the measured count rate to one-tenth within the linear range. [3]\n\nFor a microplate reader, use a manufacturer-supported attenuation setting or a suitable calibrated verification device. A piece of dark plastic over the plate is not a calibrated filter. Transmission must be known over the relevant emission wavelengths, and the arrangement must preserve the measurement geometry.\n\nCompare several brightness levels, including the brightest expected samples and intermediate controls. Account for dark/background contributions. Confirm whether the software reports raw attenuated output or automatically compensates for transmission; the expected numerical ratio differs. Hold temperature and assay age steady, and balance measurement order so that glow decay does not masquerade as nonlinearity.\n\nIf you use dilution instead, keep the final volume and matrix composition matched as far as the assay permits. Verify that the dilution changes light production proportionally. Diluting an active luciferase mixture can change substrate conditions, inhibitors, and reaction kinetics; diluting cells can change biology. A dilution series tests the combined assay-and-reader response unless those contributions are controlled.\n\nChanging the settings requires another check\nLowering PMT gain is not interchangeable with reducing the light arriving at a photon counter. It changes pulse amplitudes and can affect which pulses cross the counting threshold. [3] Use the instrument's validated settings and assess linearity again; a smaller displayed number alone is insufficient.\n\nReaders can extend useful range through faster counting electronics, appropriate count-loss correction, optical attenuation, or an alternative acquisition mode. Hamamatsu documents both count-rate correction and the hardware dependence of counting linearity. [3] The relevant specification is the linear range of the measurement mode you will use, with an acceptable error over the signals your assay produces.\n\nAn advertised range spanning many orders of magnitude is incomplete information without the test conditions. Ask whether it applies within one acquisition setting, how measurements are combined when settings change, and what happens at the transition. If two modes overlap, compare them there using stable samples. An alternative mode needs its own linearity and background checks.\n\nChoose settings that preserve weak signals while accommodating the brightest plausible wells. Verify intermediate responses before comparing Z′. Otherwise, optimization can favor a setting that suppresses bright-well variation at the expense of accurate effect sizes.\n\nWhat to include in assay validation\nInspect the raw control distributions. Save means, standard deviations, CVs, acquisition settings, and the individual well values alongside Z′. Unexpected tightening of the bright controls deserves investigation, but does not by itself prove compression.\n\nInclude intermediate controls or a reference concentration series. The two endpoints can look acceptable while the response between them is distorted. The Assay Guidance Manual recommends examining raw trends and reference compounds at multiple concentrations during pilot screening. [6]\n\nCheck more than one brightness range. Cover ordinary controls, the brightest expected samples, and enough higher signal to test the margin you intend to rely on.\n\nRepeat the relevant checks after changes. A brighter reagent formulation, a different plate, or new acquisition settings can move the assay outside the range you previously established.\n\nInvestigate suspect runs before correcting them. Preserve the original data. Establish the response experimentally and assess whether stored samples can be remeasured under suitable conditions; dividing everything by a guessed recovery factor will not undo nonlinear compression.\n\nReferences\n1. Zhang JH, Chung TDY, Oldenburg KR. A simple statistical parameter for use in evaluation and validation of high throughput screening assays. Journal of Biomolecular Screening. 1999;4(2):67–73. DOI: 10.1177/108705719900400206. Formula also presented in reference 4, section 2. 2. Hamamatsu Photonics. Photon Counting SNR Simulator. Count Rate Linearity section and paralyzable response equation. Accessed September 24, 2026. 3. Hamamatsu Photonics. Photomultiplier Tubes: Basics and Applications, fourth edition. Printed pp. 145–147 on pulse-height discrimination, pp. 149–150 on count-rate linearity/correction, and p. 181 on over-light detection and attenuation checks; PDF pp. 158–160, 162–163 and 194. 4. Bray MA, Carpenter A. Advanced Assay Development Guidelines for Image-Based High Content Screening and Analysis. Assay Guidance Manual. July 8, 2017. Section 2, Z′ and V-factor discussions. 5. Promega. CellTiter-Glo 2.0 Assay technical manual TM403. Revised January 2023. Section 4.B, Cellular ATP Content; PDF p. 11 / printed p. 10. 6. Assay Development for Protein Kinase Enzymes. Assay Guidance Manual. Pilot Screens section. Accessed September 24, 2026.", "sections": [{"id": "introduction", "heading": "Introduction", "blocks": [{"type": "p", "text": "The bright controls have become unusually consistent. Their coefficient of variation has dropped, the Z-prime factor has improved, and the plate passes QC. Before crediting the assay optimization, check whether the reader is still measuring those wells proportionally.", "html": "

The bright controls have become unusually consistent. Their coefficient of variation has dropped, the Z-prime factor has improved, and the plate passes QC. Before crediting the assay optimization, check whether the reader is still measuring those wells proportionally.

"}, {"type": "p", "text": "In a luminescence assay, detector compression can make different bright signals look more alike. Under some conditions, that improves Z′ while making the measurements less accurate. The controls look tidier because part of their variation has been lost in detection.", "html": "

In a luminescence assay, detector compression can make different bright signals look more alike. Under some conditions, that improves Z′ while making the measurements less accurate. The controls look tidier because part of their variation has been lost in detection.

"}]}, {"id": "what-z-prime-measures", "heading": "What Z-prime measures", "blocks": [{"type": "p", "text": "The Z-prime factor, usually written Z′, compares the separation of two control populations with their variability. Using H and L for high- and low-signal controls:", "html": "

The Z-prime factor, usually written Z′, compares the separation of two control populations with their variability. Using H and L for high- and low-signal controls:

"}, {"type": "div", "text": "Z prime equals one minus three times the sum of the high and low control standard deviations divided by the absolute difference between their means.", "html": "
Z′ = 1 − 3(σH + σL)|μH − μL|
"}, {"type": "p", "text": "Here, μ is the mean and σ is the standard deviation. Larger separation raises Z′; larger standard deviations lower it. The original formulation by Zhang, Chung and Oldenburg remains a useful way to assess screening controls. [1]", "html": "

Here, μ is the mean and σ is the standard deviation. Larger separation raises Z′; larger standard deviations lower it. The original formulation by Zhang, Chung and Oldenburg remains a useful way to assess screening controls. [1]

"}, {"type": "p", "text": "The formula contains no check that light input and reported signal are proportional. It uses the numbers the reader supplies. It also says nothing about the response between the two controls, where many compounds will fall.", "html": "

The formula contains no check that light input and reported signal are proportional. It uses the numbers the reader supplies. It also says nothing about the response between the two controls, where many compounds will fall.

"}, {"type": "p", "text": "Simply reducing every value by 20% cannot improve Z′. Both the standard deviations and the difference between the means fall by 20%, leaving their ratio unchanged. The problem considered here is nonlinear compression: the brighter signals lose a greater fraction of their counts.", "html": "

Simply reducing every value by 20% cannot improve Z′. Both the standard deviations and the difference between the means fall by 20%, leaving their ratio unchanged. The problem considered here is nonlinear compression: the brighter signals lose a greater fraction of their counts.

"}]}, {"id": "how-photon-counting-loses-proportionality", "heading": "How photon counting loses proportionality", "blocks": [{"type": "p", "text": "Photon counting works by resolving and counting individual detector pulses. The detector and its electronics need a finite interval to distinguish successive events. At sufficiently high arrival rates, pulses overlap and some events are missed. Hamamatsu describes this count-rate dependence and uses a paralyzable detector model in its photon-counting simulator. [2]", "html": "

Photon counting works by resolving and counting individual detector pulses. The detector and its electronics need a finite interval to distinguish successive events. At sufficiently high arrival rates, pulses overlap and some events are missed. Hamamatsu describes this count-rate dependence and uses a paralyzable detector model in its photon-counting simulator. [2]

"}, {"type": "p", "text": "Compression can begin while the output is still increasing. A numerical result below an obvious ceiling therefore does not establish linearity. Warning behavior depends on the instrument: Hamamatsu describes both photon-counting heads with over-light detection and heads without it. [3] Check what a reader's warning actually monitors and how that relates to its specified linear range.", "html": "

Compression can begin while the output is still increasing. A numerical result below an obvious ceiling therefore does not establish linearity. Warning behavior depends on the instrument: Hamamatsu describes both photon-counting heads with over-light detection and heads without it. [3] Check what a reader's warning actually monitors and how that relates to its specified linear range.

"}, {"type": "p", "text": "This also explains why shortening the integration time may fail to help. For a stable glow, counting for one-tenth as long reduces the accumulated counts, but it does not reduce the rate at which pulses arrive. A remedy for a total-count limit is not necessarily a remedy for pulse overlap. Whether a particular setting changes either limit depends on the acquisition system.", "html": "

This also explains why shortening the integration time may fail to help. For a stable glow, counting for one-tenth as long reduces the accumulated counts, but it does not reduce the rate at which pulses arrive. A remedy for a total-count limit is not necessarily a remedy for pulse overlap. Whether a particular setting changes either limit depends on the acquisition system.

"}]}, {"id": "a-worked-example-of-misleadingly-good-controls", "heading": "A worked example of misleadingly good controls", "blocks": [{"type": "p", "text": "Take five constructed values in each control group. The high control has a mean of 100 and a sample standard deviation of 10. The low control has a mean of 10 and a standard deviation of 1. Their Z′ is:", "html": "

Take five constructed values in each control group. The high control has a mean of 100 and a sample standard deviation of 10. The low control has a mean of 10 and a standard deviation of 1. Their Z′ is:

"}, {"type": "div", "text": "Z prime equals one minus three times the quantity ten plus one divided by the quantity one hundred minus ten, which equals 0.633.", "html": "
Z′ = 1 − 3(10 + 1)100 − 10 = 0.633
"}, {"type": "p", "text": "Now pass those same values through an illustrative compression curve, f(x) = x × exp(−0.00223144x). This is the shape of the paralyzable mean-response model, expressed in arbitrary units and chosen so that an input of 100 produces an output of 80. It represents no particular reader. [2]", "html": "

Now pass those same values through an illustrative compression curve, f(x) = x × exp(−0.00223144x). This is the shape of the paralyzable mean-response model, expressed in arbitrary units and chosen so that an input of 100 produces an output of 80. It represents no particular reader. [2]

"}, {"type": "div", "text": "Constructed example: linear and compressed control measurementsQuantity | Linear response | Compressed response | High-control mean | 100.00 | 79.87 | High-control standard deviation | 10.00 | 6.22 | Low-control mean | 10.00 | 9.78 | Low-control standard deviation | 1.00 | 0.96 | Z′ | 0.633 | 0.693 |", "html": "
Constructed example: linear and compressed control measurements
QuantityLinear responseCompressed response
High-control mean100.0079.87
High-control standard deviation10.006.22
Low-control mean10.009.78
Low-control standard deviation1.000.96
Z′0.6330.693
"}, {"type": "p", "text": "The inputs have not become more reproducible. Compression has reduced the spread of the high controls enough to outweigh the loss of control separation. Z′ rises from about 0.63 to 0.69, and the high-control CV falls from 10% to about 7.8%.", "html": "

The inputs have not become more reproducible. Compression has reduced the spread of the high controls enough to outweigh the loss of control separation. Z′ rises from about 0.63 to 0.69, and the high-control CV falls from 10% to about 7.8%.

"}, {"type": "p", "text": "The same curve distorts intermediate values. An input of 55 lies halfway between the original control means and represents 50% inhibition under high-to-low normalization. After compression, it reads about 48.65. Normalize that value against the compressed control means in the table and it reports about 44.5% inhibition.", "html": "

The same curve distorts intermediate values. An input of 55 lies halfway between the original control means and represents 50% inhibition under high-to-low normalization. After compression, it reads about 48.65. Normalize that value against the compressed control means in the table and it reports about 44.5% inhibition.

"}, {"type": "p", "text": "The curve is still increasing across this example, so it preserves the order of noiseless inputs. Distorted effect sizes do not require shuffled rankings. A cutoff applied to normalized inhibition can move a compound between hit and non-hit categories even if every well retains its place in the ranking.", "html": "

The curve is still increasing across this example, so it preserves the order of noiseless inputs. Distorted effect sizes do not require shuffled rankings. A cutoff applied to normalized inhibition can move a compound between hit and non-hit categories even if every well retains its place in the ranking.

"}, {"type": "p", "text": "These are deterministic calculations on invented well values. They illustrate distortion of between-well differences, without adding photon-counting noise or modeling an instrument's correction electronics. Nor does compression always raise Z′: it changes both the spread and the separation. If low-control variability dominates, the loss of separation can make Z′ worse.", "html": "

These are deterministic calculations on invented well values. They illustrate distortion of between-well differences, without adding photon-counting noise or modeling an instrument's correction electronics. Nor does compression always raise Z′: it changes both the spread and the separation. If low-control variability dominates, the loss of separation can make Z′ worse.

"}, {"type": "p", "text": "The general statistical concern is already recognized in the Assay Guidance Manual, which notes that saturation artifacts can reduce control variability in imaging assays. [4] Imaging saturation and photon-counting losses have different physical causes, but both can make control variability a misleading guide to measurement quality.", "html": "

The general statistical concern is already recognized in the Assay Guidance Manual, which notes that saturation artifacts can reduce control variability in imaging assays. [4] Imaging saturation and photon-counting losses have different physical causes, but both can make control variability a misleading guide to measurement quality.

"}]}, {"id": "test-detector-linearity-without-changing-the-chemistry", "heading": "Test detector linearity without changing the chemistry", "blocks": [{"type": "p", "text": "A falling slope in a concentration series is a reason to investigate, but it does not identify the detector as the cause. The assay itself may be nonlinear. In CellTiter-Glo, for example, Promega notes that ATP content per cell can change with cell density, altering the relationship between cell number and luminescence. [5]", "html": "

A falling slope in a concentration series is a reason to investigate, but it does not identify the detector as the cause. The assay itself may be nonlinear. In CellTiter-Glo, for example, Promega notes that ATP content per cell can change with cell density, altering the relationship between cell number and luminescence. [5]

"}, {"type": "p", "text": "Optical attenuation helps separate those possibilities. It changes the light reaching the detector while leaving the sample chemistry in place. Hamamatsu describes a check using a neutral-density filter with known transmission: a 10% transmitting filter should reduce the measured count rate to one-tenth within the linear range. [3]", "html": "

Optical attenuation helps separate those possibilities. It changes the light reaching the detector while leaving the sample chemistry in place. Hamamatsu describes a check using a neutral-density filter with known transmission: a 10% transmitting filter should reduce the measured count rate to one-tenth within the linear range. [3]

"}, {"type": "p", "text": "For a microplate reader, use a manufacturer-supported attenuation setting or a suitable calibrated verification device. A piece of dark plastic over the plate is not a calibrated filter. Transmission must be known over the relevant emission wavelengths, and the arrangement must preserve the measurement geometry.", "html": "

For a microplate reader, use a manufacturer-supported attenuation setting or a suitable calibrated verification device. A piece of dark plastic over the plate is not a calibrated filter. Transmission must be known over the relevant emission wavelengths, and the arrangement must preserve the measurement geometry.

"}, {"type": "p", "text": "Compare several brightness levels, including the brightest expected samples and intermediate controls. Account for dark/background contributions. Confirm whether the software reports raw attenuated output or automatically compensates for transmission; the expected numerical ratio differs. Hold temperature and assay age steady, and balance measurement order so that glow decay does not masquerade as nonlinearity.", "html": "

Compare several brightness levels, including the brightest expected samples and intermediate controls. Account for dark/background contributions. Confirm whether the software reports raw attenuated output or automatically compensates for transmission; the expected numerical ratio differs. Hold temperature and assay age steady, and balance measurement order so that glow decay does not masquerade as nonlinearity.

"}, {"type": "p", "text": "If you use dilution instead, keep the final volume and matrix composition matched as far as the assay permits. Verify that the dilution changes light production proportionally. Diluting an active luciferase mixture can change substrate conditions, inhibitors, and reaction kinetics; diluting cells can change biology. A dilution series tests the combined assay-and-reader response unless those contributions are controlled.", "html": "

If you use dilution instead, keep the final volume and matrix composition matched as far as the assay permits. Verify that the dilution changes light production proportionally. Diluting an active luciferase mixture can change substrate conditions, inhibitors, and reaction kinetics; diluting cells can change biology. A dilution series tests the combined assay-and-reader response unless those contributions are controlled.

"}]}, {"id": "changing-the-settings-requires-another-check", "heading": "Changing the settings requires another check", "blocks": [{"type": "p", "text": "Lowering PMT gain is not interchangeable with reducing the light arriving at a photon counter. It changes pulse amplitudes and can affect which pulses cross the counting threshold. [3] Use the instrument's validated settings and assess linearity again; a smaller displayed number alone is insufficient.", "html": "

Lowering PMT gain is not interchangeable with reducing the light arriving at a photon counter. It changes pulse amplitudes and can affect which pulses cross the counting threshold. [3] Use the instrument's validated settings and assess linearity again; a smaller displayed number alone is insufficient.

"}, {"type": "p", "text": "Readers can extend useful range through faster counting electronics, appropriate count-loss correction, optical attenuation, or an alternative acquisition mode. Hamamatsu documents both count-rate correction and the hardware dependence of counting linearity. [3] The relevant specification is the linear range of the measurement mode you will use, with an acceptable error over the signals your assay produces.", "html": "

Readers can extend useful range through faster counting electronics, appropriate count-loss correction, optical attenuation, or an alternative acquisition mode. Hamamatsu documents both count-rate correction and the hardware dependence of counting linearity. [3] The relevant specification is the linear range of the measurement mode you will use, with an acceptable error over the signals your assay produces.

"}, {"type": "p", "text": "An advertised range spanning many orders of magnitude is incomplete information without the test conditions. Ask whether it applies within one acquisition setting, how measurements are combined when settings change, and what happens at the transition. If two modes overlap, compare them there using stable samples. An alternative mode needs its own linearity and background checks.", "html": "

An advertised range spanning many orders of magnitude is incomplete information without the test conditions. Ask whether it applies within one acquisition setting, how measurements are combined when settings change, and what happens at the transition. If two modes overlap, compare them there using stable samples. An alternative mode needs its own linearity and background checks.

"}, {"type": "p", "text": "Choose settings that preserve weak signals while accommodating the brightest plausible wells. Verify intermediate responses before comparing Z′. Otherwise, optimization can favor a setting that suppresses bright-well variation at the expense of accurate effect sizes.", "html": "

Choose settings that preserve weak signals while accommodating the brightest plausible wells. Verify intermediate responses before comparing Z′. Otherwise, optimization can favor a setting that suppresses bright-well variation at the expense of accurate effect sizes.

"}]}, {"id": "what-to-include-in-assay-validation", "heading": "What to include in assay validation", "blocks": [{"type": "p", "text": "Inspect the raw control distributions. Save means, standard deviations, CVs, acquisition settings, and the individual well values alongside Z′. Unexpected tightening of the bright controls deserves investigation, but does not by itself prove compression.", "html": "

Inspect the raw control distributions. Save means, standard deviations, CVs, acquisition settings, and the individual well values alongside Z′. Unexpected tightening of the bright controls deserves investigation, but does not by itself prove compression.

"}, {"type": "p", "text": "Include intermediate controls or a reference concentration series. The two endpoints can look acceptable while the response between them is distorted. The Assay Guidance Manual recommends examining raw trends and reference compounds at multiple concentrations during pilot screening. [6]", "html": "

Include intermediate controls or a reference concentration series. The two endpoints can look acceptable while the response between them is distorted. The Assay Guidance Manual recommends examining raw trends and reference compounds at multiple concentrations during pilot screening. [6]

"}, {"type": "p", "text": "Check more than one brightness range. Cover ordinary controls, the brightest expected samples, and enough higher signal to test the margin you intend to rely on.", "html": "

Check more than one brightness range. Cover ordinary controls, the brightest expected samples, and enough higher signal to test the margin you intend to rely on.

"}, {"type": "p", "text": "Repeat the relevant checks after changes. A brighter reagent formulation, a different plate, or new acquisition settings can move the assay outside the range you previously established.", "html": "

Repeat the relevant checks after changes. A brighter reagent formulation, a different plate, or new acquisition settings can move the assay outside the range you previously established.

"}, {"type": "p", "text": "Investigate suspect runs before correcting them. Preserve the original data. Establish the response experimentally and assess whether stored samples can be remeasured under suitable conditions; dividing everything by a guessed recovery factor will not undo nonlinear compression.", "html": "

Investigate suspect runs before correcting them. Preserve the original data. Establish the response experimentally and assess whether stored samples can be remeasured under suitable conditions; dividing everything by a guessed recovery factor will not undo nonlinear compression.

"}]}, {"id": "references", "heading": "References", "blocks": [{"type": "references", "text": "1. Zhang JH, Chung TDY, Oldenburg KR. A simple statistical parameter for use in evaluation and validation of high throughput screening assays. Journal of Biomolecular Screening. 1999;4(2):67–73. DOI: 10.1177/108705719900400206. Formula also presented in reference 4, section 2. 2. Hamamatsu Photonics. Photon Counting SNR Simulator. Count Rate Linearity section and paralyzable response equation. Accessed September 24, 2026. 3. Hamamatsu Photonics. Photomultiplier Tubes: Basics and Applications, fourth edition. Printed pp. 145–147 on pulse-height discrimination, pp. 149–150 on count-rate linearity/correction, and p. 181 on over-light detection and attenuation checks; PDF pp. 158–160, 162–163 and 194. 4. Bray MA, Carpenter A. Advanced Assay Development Guidelines for Image-Based High Content Screening and Analysis. Assay Guidance Manual. July 8, 2017. Section 2, Z′ and V-factor discussions. 5. Promega. CellTiter-Glo 2.0 Assay technical manual TM403. Revised January 2023. Section 4.B, Cellular ATP Content; PDF p. 11 / printed p. 10. 6. Assay Development for Protein Kinase Enzymes. Assay Guidance Manual. Pilot Screens section. Accessed September 24, 2026.", "html": "
\n

1. Zhang JH, Chung TDY, Oldenburg KR. A simple statistical parameter for use in evaluation and validation of high throughput screening assays. Journal of Biomolecular Screening. 1999;4(2):67–73. DOI: 10.1177/108705719900400206. Formula also presented in reference 4, section 2.

\n

2. Hamamatsu Photonics. Photon Counting SNR Simulator. Count Rate Linearity section and paralyzable response equation. Accessed September 24, 2026.

\n

3. Hamamatsu Photonics. Photomultiplier Tubes: Basics and Applications, fourth edition. Printed pp. 145–147 on pulse-height discrimination, pp. 149–150 on count-rate linearity/correction, and p. 181 on over-light detection and attenuation checks; PDF pp. 158–160, 162–163 and 194.

\n

4. Bray MA, Carpenter A. Advanced Assay Development Guidelines for Image-Based High Content Screening and Analysis. Assay Guidance Manual. July 8, 2017. Section 2, Z′ and V-factor discussions.

\n

5. Promega. CellTiter-Glo 2.0 Assay technical manual TM403. Revised January 2023. Section 4.B, Cellular ATP Content; PDF p. 11 / printed p. 10.

\n

6. Assay Development for Protein Kinase Enzymes. Assay Guidance Manual. Pilot Screens section. Accessed September 24, 2026.

\n
"}]}], "references": [{"id": "ref-1", "citation": "1. Zhang JH, Chung TDY, Oldenburg KR. A simple statistical parameter for use in evaluation and validation of high throughput screening assays. Journal of Biomolecular Screening. 1999;4(2):67–73. DOI: 10.1177/108705719900400206. Formula also presented in reference 4, section 2.", "urls": ["https://journals.sagepub.com/doi/10.1177/108705719900400206"]}, {"id": "ref-2", "citation": "2. Hamamatsu Photonics. Photon Counting SNR Simulator. Count Rate Linearity section and paralyzable response equation. Accessed September 24, 2026.", "urls": ["https://www.hamamatsu.com/eu/en/resources/interactive-tools/photon-counting-snr-simulator.html"]}, {"id": "ref-3", "citation": "3. Hamamatsu Photonics. Photomultiplier Tubes: Basics and Applications, fourth edition. Printed pp. 145–147 on pulse-height discrimination, pp. 149–150 on count-rate linearity/correction, and p. 181 on over-light detection and attenuation checks; PDF pp. 158–160, 162–163 and 194.", "urls": ["https://www.hamamatsu.com/content/dam/hamamatsu-photonics/sites/documents/99_SALES_LIBRARY/etd/PMT_handbook_v4E.pdf"]}, {"id": "ref-4", "citation": "4. Bray MA, Carpenter A. Advanced Assay Development Guidelines for Image-Based High Content Screening and Analysis. Assay Guidance Manual. July 8, 2017. Section 2, Z′ and V-factor discussions.", "urls": ["https://www.ncbi.nlm.nih.gov/sites/books/NBK126174/?report=reader"]}, {"id": "ref-5", "citation": "5. Promega. CellTiter-Glo 2.0 Assay technical manual TM403. Revised January 2023. Section 4.B, Cellular ATP Content; PDF p. 11 / printed p. 10.", "urls": ["https://worldwide.promega.com/-/media/files/resources/protocols/technical-manuals/101/celltiterglo-2-0-assay-protocol.pdf?la=en"]}, {"id": "ref-6", "citation": "6. Assay Development for Protein Kinase Enzymes. Assay Guidance Manual. Pilot Screens section. Accessed September 24, 2026.", "urls": ["https://www.ncbi.nlm.nih.gov/sites/books/NBK91991/?report=reader"]}], "equations": [{"description": "Z prime equals one minus three times the sum of the high and low control standard deviations divided by the absolute difference between their means.", "mathml": "Z′ = 1 − 3(σH + σL)|μH − μL|"}, {"description": "Z prime equals one minus three times the quantity ten plus one divided by the quantity one hundred minus ten, which equals 0.633.", "mathml": "Z′ = 1 − 3(10 + 1)100 − 10 = 0.633"}], "tables": [{"caption": "Constructed example: linear and compressed control measurements", "rows": [["Quantity", "Linear response", "Compressed response"], ["High-control mean", "100.00", "79.87"], ["High-control standard deviation", "10.00", "6.22"], ["Low-control mean", "10.00", "9.78"], ["Low-control standard deviation", "1.00", "0.96"], ["Z′", "0.633", "0.693"]], "data_kind": "constructed example"}], "provenance": {"kind": "Author-provided article converted to structured text", "examples": "Constructed calculations, not a raw experimental dataset", "editorial_additions": "Summary, takeaways, topic tags, and limitations summary"}, "display_additions": {"figures": [{"kind": "illustrative", "description": "Schematic control distributions; not measured wells"}, {"kind": "constructed model", "description": "f(x) = x × exp(−0.00223144x), no specific reader"}], "bench_check": true, "html": "
Fig. 1 — Same plate, two gain settings
Gain in linear rangeProportional
Gain near saturationCompressed
Low controlHigh controlRelative signal →
Illustrative distributions, not measured wells. They show how compression can narrow a bright control. Whether changing gain produces this behavior depends on the reader; no universal gain or saturation threshold is implied.
Fig. 2 — Detector response vs. input
Illustrative detector compressionThe model response curves below the ideal linear response as input increases. Input 100 gives output 80 in arbitrary units.Ideal linear responseInput 100 → output 80INPUT · ARBITRARY UNITS →REPORTED SIGNAL →0120
Constructed model from the article: f(x) = x × exp(−0.00223144x). The response loses proportionality while still increasing across this range. It represents no particular reader.

Bench check

Before trusting the statistic
01

Inspect the raw control distributions

Retain individual well values, means, standard deviations, CVs, and acquisition settings alongside Z′.

02

Check proportionality over the signal range

Use manufacturer-supported attenuation or a calibrated verification device. A dilution series also changes the assay unless its chemistry is controlled.

03

Recheck after changing the setup

Include low, intermediate, and bright signals when changing reagents, plates, or acquisition settings. A smaller displayed number alone does not establish linearity.

"}, "content_sha256": "39831090c495e52f8564ced712390c9e2b63c03d39f5af4c42d5ef46be9206cc"} {"schema_version": "1.0", "id": "luminescence-integration-time", "canonical_url": "https://discoveryinpractice.com/articles/luminescence-integration-time/", "title": "When reading longer stops helping", "author": {"name": "Andrew Stewart", "url": "https://discoveryinpractice.com/about/#andrew-stewart"}, "language": "en", "summary": "Longer luminescence reads help when photon collection limits precision, but they cannot remove persistent differences between wells. This article shows how to compare repeated reads with independently prepared replicates, account for background noise, and choose a useful stopping point.", "takeaways": ["Under a stable, linear, shot-noise-limited model, halving photon-counting CV requires four times as many detected photons.", "Compare repeated reads of the same wells with independent preparations to investigate measurement noise and persistent well differences.", "Evaluate uncertainty in the reported endpoint, including background subtraction, and balance read order against assay age and temperature."], "limitations_summary": "A precision plateau does not identify its cause by itself. The examples assume specific statistical models and use constructed values, not experimental measurements.", "topics": ["Measurement & precision", "Luminescence", "Integration time", "Background noise"], "publication_status": "published", "date_published": "2026-09-25", "date_modified": "2026-09-25", "license": "CC-BY-4.0", "license_url": "https://creativecommons.org/licenses/by/4.0/", "license_status": "Published under CC BY 4.0", "content_version": "1.0", "body_html": "

The plate looks noisy, so you increase the integration time. This is often a sensible first move in a luminescence assay. It is also an excellent way to spend sixteen times longer measuring a problem that was already in the wells.

Microplate assay variability has several origins. Some fluctuations arise while the reader collects light. Others come from differences in cell number, reagent delivery, temperature or biology. Longer integration can improve photon statistics. It cannot put the missing cells back into well H17.

The useful question is how much of the variability your next second of reading can actually remove.

What photon shot noise costs

Even a perfectly steady light source does not deliver precisely the same number of detected photons in every measurement. For independent photon arrivals described by a Poisson distribution, the standard deviation is the square root of the mean count. With negligible background and other noise, the relative uncertainty is:

CVphoton = 1N

Here, N is the mean number of detected photons accumulated during the measurement. CV is expressed as a fraction; multiply by 100 for a percentage. At 100 photons, the photon-counting CV is 10%. At 1,000 it is about 3.2%; at 10,000, 1%. This is the shot-noise limit under those assumptions. Real measurements can be worse. [1, 2]

The square root is expensive. Halving the CV requires four times as many detected photons. If the source is stable and detection remains linear, that means four times the integration time, or four times the useful photon collection rate, or some combination.

Do not insert a displayed value of 10,000 RLU into this equation. Relative light units are instrument-dependent, and counts per second are a rate, not the accumulated count. A bigger displayed number after changing gain does not establish that you collected more photons.

Efficient light collection can save substantial time when photon statistics dominate. Merely multiplying an existing signal multiplies its fluctuations as well.

A small improvement can consume a large afternoon

Suppose a one-second read contributes 6% measurement CV, while persistent differences between nominally identical wells contribute 8%. Assume these contributions are independent, the signal is stable and the measurement component decreases with the square root of integration time.

The variances add. The percentages do not:

CVtotal2 ≈ CVmeasurement2 + CVbetween wells2

This constructed example gives the following result:

Constructed example: diminishing precision gains from longer integration
Integration per wellMeasurement CVPersistent well CVTotal CV
1 second6.0%8.0%10.0%
4 seconds3.0%8.0%8.5%
16 seconds1.5%8.0%8.1%

The detector measurement improves fourfold between the first and last rows. Overall precision barely improves after four seconds because most of the remaining variance belongs to the wells.

For a reader measuring wells one at a time, sixteen seconds across 1,536 wells is nearly seven hours of integration alone, before motion and other overheads. That is a large operational penalty for moving from 8.5% to 8.1%. The signal might also change appreciably while you wait.

An 8% floor is an assumption in this example, not a diagnosis. A plateau in an actual experiment could include dispensing variation, persistent optical effects, imperfect blank correction or sample instability. The shape of the curve tells you where to investigate; it does not identify the culprit by itself.

Cells have counting statistics too

Imagine dispensing a well-mixed suspension in which cells arrive independently, with an average of 100 cells per well. Under a Poisson loading model, the cell count has a standard deviation of 10 cells: a 10% CV before the assay chemistry has done anything.

At 1,000 cells per well, the corresponding count CV is 3.2%; at 10,000, 1%. The arithmetic resembles photon counting because both examples use the same probability model. The interventions are different. More detected photons improve the estimate of the light coming from those particular cells. They do not redraw the cell population.

Poisson seeding is more than a classroom convenience. Chang and colleagues used it to describe initial occupancy in microwell arrays, then followed substantial differences in subsequent clonal growth. Their system supports the loading principle; it does not establish a universal CV for a conventional cell-based screen. [3]

Clumping, settling during dispensing, unequal delivered volumes and differential growth can all change the distribution. Deliberately dispensing a known number of cells changes the model too. For cultured assays, the relevant population is the one present when the signal is generated, which may differ considerably from the starting population.

Cells also contribute unequal amounts of signal. If cell number is Poisson with mean n, and each cell contributes an independent signal with single-cell CV c, the model becomes:

CVcell signal2 = 1 + c2n

With 100 cells and a single-cell CV of 100%, the predicted well-signal CV is about 14.1%, rather than the 10% obtained by assuming identical cells. This calculation excludes detector noise and assumes cell brightness is independent of cell number and of other cells. Density-dependent biology can break those assumptions.

That distinction matters for ATP assays. Promega's CellTiter-Glo 2.0 manual discusses changes in ATP per cell with cell density and physiological state. Luminescence can be proportional to cell number over a validated range without being a literal cell counter under every treatment. [4]

Adding cells may reduce sampling variation, but it can also change the biology you intended to measure. Check the response to compounds and the useful assay window before accepting the prettier CV.

Repeated reads and replicate wells answer different questions

Ten reads of one stable well repeatedly measure the same preparation. Ten independently prepared wells include preparation differences. Calling both exercises “reproducibility” can conceal the most useful information in the experiment.

A practical integration-time study should contain both. Use representative low, middle and high signals, along with blanks. Include independently prepared replicates and repeated measurements where the assay tolerates them. Compare the variability within each well with the variability between wells at each integration time.

If repeated reads improve while between-well variation remains nearly unchanged, investigate the preparation and persistent spatial effects. If both improve, photon collection may still be limiting. If both deteriorate with time, investigate drift or damage before calculating a noise floor.

Balance read order and elapsed time across settings. Reading every well briefly first and every well slowly last confounds integration time with assay age. Fluorescence excitation may itself perturb the sample; use matched fresh preparations when repeated exposure changes the signal.

For a stable assay, plotting CV squared against inverse integration time can be informative. A simple model gives a straight line whose intercept represents variation that does not decrease with longer reading. Use it as a diagnostic approximation. Correlated noise, drift and nonlinearity can defeat the interpretation.

Subtracting background leaves its noise behind

Blank subtraction removes an estimate of the average background. It does not remove the random background photons collected in the sample well.

Consider an ideal counting measurement with 1,000 signal photons and 1,000 background photons. Even if you knew the mean background perfectly, the sample measurement fluctuates according to all 2,000 detected photons. After subtraction, the net signal is 1,000, but its standard deviation is about 44.7: a CV of 4.5%, rather than the 3.2% expected without background. An experimentally estimated blank adds uncertainty of its own. [1, 2]

This is why reducing optical background and leakage from neighboring bright wells can be worth more than extending the read. Neither a clean-looking baseline nor a large raw signal guarantees a precise small difference after subtraction. Check performance at the low signals and intermediate responses that determine compound ranking, as well as at the bright control.

Spend the next second where it helps

Longer reading also extends the period during which temperature, evaporation or reaction progress can change the plate. Promega specifically identifies temperature as affecting CellTiter-Glo 2.0 light intensity and decay, and cautions about temperature gradients within plates. [4] A reproducible read schedule and stable sample temperature therefore belong in the precision experiment.

Choose optics, plate geometry and focus settings that collect useful light efficiently while controlling background and interwell leakage. Verify detector linearity over the actual signal range. Compression can make bright wells appear less variable, which is a separate problem from genuinely better photon statistics.

During optimization:

Increase integration time when repeated-read data show that measurement noise remains important.

Fix dispensing, mixing or cell-loading problems when the variation persists between preparations.

Compare uncertainty in the reported endpoint, including any subtraction or ratio, rather than judging only the brightest raw channel.

Recheck the chosen timing under a realistic plate sequence, including temperature equilibration and queue delays.

Stop extending the read when the improvement no longer changes the decisions the assay must support. At that point, the next useful experiment is likely to happen at the dispenser, in the incubator or in the assay design.

References

\n

1. Hamamatsu Photonics. Photon Counting SNR Simulator. Photon statistics and signal/background contributions; accessed September 24, 2026.

\n

2. Owicki JC. Fluorescence Polarization and Anisotropy in High Throughput Screening: Perspectives and Primer. Journal of Biomolecular Screening. 2000;5:297–306.

\n

3. Chang TC and colleagues. Microwell arrays reveal cellular heterogeneity during the clonal expansion of transformed human cells. Technology. 2015;3:163–171. Initial seeding statistics and subsequent clonal behavior in their microwell system.

\n

4. Promega. CellTiter-Glo 2.0 Cell Viability Assay Technical Manual TM403. Revision 1/23, section 4; temperature and ATP-per-cell considerations. Worked examples in this article are constructed calculations, not measurements from this manual.

\n
", "body_text": "Introduction\nThe plate looks noisy, so you increase the integration time. This is often a sensible first move in a luminescence assay. It is also an excellent way to spend sixteen times longer measuring a problem that was already in the wells.\n\nMicroplate assay variability has several origins. Some fluctuations arise while the reader collects light. Others come from differences in cell number, reagent delivery, temperature or biology. Longer integration can improve photon statistics. It cannot put the missing cells back into well H17.\n\nThe useful question is how much of the variability your next second of reading can actually remove.\n\nWhat photon shot noise costs\nEven a perfectly steady light source does not deliver precisely the same number of detected photons in every measurement. For independent photon arrivals described by a Poisson distribution, the standard deviation is the square root of the mean count. With negligible background and other noise, the relative uncertainty is:\n\nPhoton coefficient of variation equals one divided by the square root of the mean detected photon count N.\n\nHere, N is the mean number of detected photons accumulated during the measurement. CV is expressed as a fraction; multiply by 100 for a percentage. At 100 photons, the photon-counting CV is 10%. At 1,000 it is about 3.2%; at 10,000, 1%. This is the shot-noise limit under those assumptions. Real measurements can be worse. [1, 2]\n\nThe square root is expensive. Halving the CV requires four times as many detected photons. If the source is stable and detection remains linear, that means four times the integration time, or four times the useful photon collection rate, or some combination.\n\nDo not insert a displayed value of 10,000 RLU into this equation. Relative light units are instrument-dependent, and counts per second are a rate, not the accumulated count. A bigger displayed number after changing gain does not establish that you collected more photons.\n\nEfficient light collection can save substantial time when photon statistics dominate. Merely multiplying an existing signal multiplies its fluctuations as well.\n\nA small improvement can consume a large afternoon\nSuppose a one-second read contributes 6% measurement CV, while persistent differences between nominally identical wells contribute 8%. Assume these contributions are independent, the signal is stable and the measurement component decreases with the square root of integration time.\n\nThe variances add. The percentages do not:\n\nTotal coefficient of variation squared is approximately measurement coefficient of variation squared plus between-well coefficient of variation squared.\n\nThis constructed example gives the following result:\n\nConstructed example: diminishing precision gains from longer integrationIntegration per well | Measurement CV | Persistent well CV | Total CV | 1 second | 6.0% | 8.0% | 10.0% | 4 seconds | 3.0% | 8.0% | 8.5% | 16 seconds | 1.5% | 8.0% | 8.1% |\n\nThe detector measurement improves fourfold between the first and last rows. Overall precision barely improves after four seconds because most of the remaining variance belongs to the wells.\n\nFor a reader measuring wells one at a time, sixteen seconds across 1,536 wells is nearly seven hours of integration alone, before motion and other overheads. That is a large operational penalty for moving from 8.5% to 8.1%. The signal might also change appreciably while you wait.\n\nAn 8% floor is an assumption in this example, not a diagnosis. A plateau in an actual experiment could include dispensing variation, persistent optical effects, imperfect blank correction or sample instability. The shape of the curve tells you where to investigate; it does not identify the culprit by itself.\n\nCells have counting statistics too\nImagine dispensing a well-mixed suspension in which cells arrive independently, with an average of 100 cells per well. Under a Poisson loading model, the cell count has a standard deviation of 10 cells: a 10% CV before the assay chemistry has done anything.\n\nAt 1,000 cells per well, the corresponding count CV is 3.2%; at 10,000, 1%. The arithmetic resembles photon counting because both examples use the same probability model. The interventions are different. More detected photons improve the estimate of the light coming from those particular cells. They do not redraw the cell population.\n\nPoisson seeding is more than a classroom convenience. Chang and colleagues used it to describe initial occupancy in microwell arrays, then followed substantial differences in subsequent clonal growth. Their system supports the loading principle; it does not establish a universal CV for a conventional cell-based screen. [3]\n\nClumping, settling during dispensing, unequal delivered volumes and differential growth can all change the distribution. Deliberately dispensing a known number of cells changes the model too. For cultured assays, the relevant population is the one present when the signal is generated, which may differ considerably from the starting population.\n\nCells also contribute unequal amounts of signal. If cell number is Poisson with mean n, and each cell contributes an independent signal with single-cell CV c, the model becomes:\n\nCell-signal coefficient of variation squared equals the quantity one plus c squared divided by n.\n\nWith 100 cells and a single-cell CV of 100%, the predicted well-signal CV is about 14.1%, rather than the 10% obtained by assuming identical cells. This calculation excludes detector noise and assumes cell brightness is independent of cell number and of other cells. Density-dependent biology can break those assumptions.\n\nThat distinction matters for ATP assays. Promega's CellTiter-Glo 2.0 manual discusses changes in ATP per cell with cell density and physiological state. Luminescence can be proportional to cell number over a validated range without being a literal cell counter under every treatment. [4]\n\nAdding cells may reduce sampling variation, but it can also change the biology you intended to measure. Check the response to compounds and the useful assay window before accepting the prettier CV.\n\nRepeated reads and replicate wells answer different questions\nTen reads of one stable well repeatedly measure the same preparation. Ten independently prepared wells include preparation differences. Calling both exercises “reproducibility” can conceal the most useful information in the experiment.\n\nA practical integration-time study should contain both. Use representative low, middle and high signals, along with blanks. Include independently prepared replicates and repeated measurements where the assay tolerates them. Compare the variability within each well with the variability between wells at each integration time.\n\nIf repeated reads improve while between-well variation remains nearly unchanged, investigate the preparation and persistent spatial effects. If both improve, photon collection may still be limiting. If both deteriorate with time, investigate drift or damage before calculating a noise floor.\n\nBalance read order and elapsed time across settings. Reading every well briefly first and every well slowly last confounds integration time with assay age. Fluorescence excitation may itself perturb the sample; use matched fresh preparations when repeated exposure changes the signal.\n\nFor a stable assay, plotting CV squared against inverse integration time can be informative. A simple model gives a straight line whose intercept represents variation that does not decrease with longer reading. Use it as a diagnostic approximation. Correlated noise, drift and nonlinearity can defeat the interpretation.\n\nSubtracting background leaves its noise behind\nBlank subtraction removes an estimate of the average background. It does not remove the random background photons collected in the sample well.\n\nConsider an ideal counting measurement with 1,000 signal photons and 1,000 background photons. Even if you knew the mean background perfectly, the sample measurement fluctuates according to all 2,000 detected photons. After subtraction, the net signal is 1,000, but its standard deviation is about 44.7: a CV of 4.5%, rather than the 3.2% expected without background. An experimentally estimated blank adds uncertainty of its own. [1, 2]\n\nThis is why reducing optical background and leakage from neighboring bright wells can be worth more than extending the read. Neither a clean-looking baseline nor a large raw signal guarantees a precise small difference after subtraction. Check performance at the low signals and intermediate responses that determine compound ranking, as well as at the bright control.\n\nSpend the next second where it helps\nLonger reading also extends the period during which temperature, evaporation or reaction progress can change the plate. Promega specifically identifies temperature as affecting CellTiter-Glo 2.0 light intensity and decay, and cautions about temperature gradients within plates. [4] A reproducible read schedule and stable sample temperature therefore belong in the precision experiment.\n\nChoose optics, plate geometry and focus settings that collect useful light efficiently while controlling background and interwell leakage. Verify detector linearity over the actual signal range. Compression can make bright wells appear less variable, which is a separate problem from genuinely better photon statistics.\n\nDuring optimization:\n\nIncrease integration time when repeated-read data show that measurement noise remains important.\n\nFix dispensing, mixing or cell-loading problems when the variation persists between preparations.\n\nCompare uncertainty in the reported endpoint, including any subtraction or ratio, rather than judging only the brightest raw channel.\n\nRecheck the chosen timing under a realistic plate sequence, including temperature equilibration and queue delays.\n\nStop extending the read when the improvement no longer changes the decisions the assay must support. At that point, the next useful experiment is likely to happen at the dispenser, in the incubator or in the assay design.\n\nReferences\n1. Hamamatsu Photonics. Photon Counting SNR Simulator. Photon statistics and signal/background contributions; accessed September 24, 2026. 2. Owicki JC. Fluorescence Polarization and Anisotropy in High Throughput Screening: Perspectives and Primer. Journal of Biomolecular Screening. 2000;5:297–306. 3. Chang TC and colleagues. Microwell arrays reveal cellular heterogeneity during the clonal expansion of transformed human cells. Technology. 2015;3:163–171. Initial seeding statistics and subsequent clonal behavior in their microwell system. 4. Promega. CellTiter-Glo 2.0 Cell Viability Assay Technical Manual TM403. Revision 1/23, section 4; temperature and ATP-per-cell considerations. Worked examples in this article are constructed calculations, not measurements from this manual.", "sections": [{"id": "introduction", "heading": "Introduction", "blocks": [{"type": "p", "text": "The plate looks noisy, so you increase the integration time. This is often a sensible first move in a luminescence assay. It is also an excellent way to spend sixteen times longer measuring a problem that was already in the wells.", "html": "

The plate looks noisy, so you increase the integration time. This is often a sensible first move in a luminescence assay. It is also an excellent way to spend sixteen times longer measuring a problem that was already in the wells.

"}, {"type": "p", "text": "Microplate assay variability has several origins. Some fluctuations arise while the reader collects light. Others come from differences in cell number, reagent delivery, temperature or biology. Longer integration can improve photon statistics. It cannot put the missing cells back into well H17.", "html": "

Microplate assay variability has several origins. Some fluctuations arise while the reader collects light. Others come from differences in cell number, reagent delivery, temperature or biology. Longer integration can improve photon statistics. It cannot put the missing cells back into well H17.

"}, {"type": "p", "text": "The useful question is how much of the variability your next second of reading can actually remove.", "html": "

The useful question is how much of the variability your next second of reading can actually remove.

"}]}, {"id": "what-photon-shot-noise-costs", "heading": "What photon shot noise costs", "blocks": [{"type": "p", "text": "Even a perfectly steady light source does not deliver precisely the same number of detected photons in every measurement. For independent photon arrivals described by a Poisson distribution, the standard deviation is the square root of the mean count. With negligible background and other noise, the relative uncertainty is:", "html": "

Even a perfectly steady light source does not deliver precisely the same number of detected photons in every measurement. For independent photon arrivals described by a Poisson distribution, the standard deviation is the square root of the mean count. With negligible background and other noise, the relative uncertainty is:

"}, {"type": "div", "text": "Photon coefficient of variation equals one divided by the square root of the mean detected photon count N.", "html": "
CVphoton = 1N
"}, {"type": "p", "text": "Here, N is the mean number of detected photons accumulated during the measurement. CV is expressed as a fraction; multiply by 100 for a percentage. At 100 photons, the photon-counting CV is 10%. At 1,000 it is about 3.2%; at 10,000, 1%. This is the shot-noise limit under those assumptions. Real measurements can be worse. [1, 2]", "html": "

Here, N is the mean number of detected photons accumulated during the measurement. CV is expressed as a fraction; multiply by 100 for a percentage. At 100 photons, the photon-counting CV is 10%. At 1,000 it is about 3.2%; at 10,000, 1%. This is the shot-noise limit under those assumptions. Real measurements can be worse. [1, 2]

"}, {"type": "p", "text": "The square root is expensive. Halving the CV requires four times as many detected photons. If the source is stable and detection remains linear, that means four times the integration time, or four times the useful photon collection rate, or some combination.", "html": "

The square root is expensive. Halving the CV requires four times as many detected photons. If the source is stable and detection remains linear, that means four times the integration time, or four times the useful photon collection rate, or some combination.

"}, {"type": "p", "text": "Do not insert a displayed value of 10,000 RLU into this equation. Relative light units are instrument-dependent, and counts per second are a rate, not the accumulated count. A bigger displayed number after changing gain does not establish that you collected more photons.", "html": "

Do not insert a displayed value of 10,000 RLU into this equation. Relative light units are instrument-dependent, and counts per second are a rate, not the accumulated count. A bigger displayed number after changing gain does not establish that you collected more photons.

"}, {"type": "p", "text": "Efficient light collection can save substantial time when photon statistics dominate. Merely multiplying an existing signal multiplies its fluctuations as well.", "html": "

Efficient light collection can save substantial time when photon statistics dominate. Merely multiplying an existing signal multiplies its fluctuations as well.

"}]}, {"id": "a-small-improvement-can-consume-a-large-afternoon", "heading": "A small improvement can consume a large afternoon", "blocks": [{"type": "p", "text": "Suppose a one-second read contributes 6% measurement CV, while persistent differences between nominally identical wells contribute 8%. Assume these contributions are independent, the signal is stable and the measurement component decreases with the square root of integration time.", "html": "

Suppose a one-second read contributes 6% measurement CV, while persistent differences between nominally identical wells contribute 8%. Assume these contributions are independent, the signal is stable and the measurement component decreases with the square root of integration time.

"}, {"type": "p", "text": "The variances add. The percentages do not:", "html": "

The variances add. The percentages do not:

"}, {"type": "div", "text": "Total coefficient of variation squared is approximately measurement coefficient of variation squared plus between-well coefficient of variation squared.", "html": "
CVtotal2 ≈ CVmeasurement2 + CVbetween wells2
"}, {"type": "p", "text": "This constructed example gives the following result:", "html": "

This constructed example gives the following result:

"}, {"type": "div", "text": "Constructed example: diminishing precision gains from longer integrationIntegration per well | Measurement CV | Persistent well CV | Total CV | 1 second | 6.0% | 8.0% | 10.0% | 4 seconds | 3.0% | 8.0% | 8.5% | 16 seconds | 1.5% | 8.0% | 8.1% |", "html": "
Constructed example: diminishing precision gains from longer integration
Integration per wellMeasurement CVPersistent well CVTotal CV
1 second6.0%8.0%10.0%
4 seconds3.0%8.0%8.5%
16 seconds1.5%8.0%8.1%
"}, {"type": "p", "text": "The detector measurement improves fourfold between the first and last rows. Overall precision barely improves after four seconds because most of the remaining variance belongs to the wells.", "html": "

The detector measurement improves fourfold between the first and last rows. Overall precision barely improves after four seconds because most of the remaining variance belongs to the wells.

"}, {"type": "p", "text": "For a reader measuring wells one at a time, sixteen seconds across 1,536 wells is nearly seven hours of integration alone, before motion and other overheads. That is a large operational penalty for moving from 8.5% to 8.1%. The signal might also change appreciably while you wait.", "html": "

For a reader measuring wells one at a time, sixteen seconds across 1,536 wells is nearly seven hours of integration alone, before motion and other overheads. That is a large operational penalty for moving from 8.5% to 8.1%. The signal might also change appreciably while you wait.

"}, {"type": "p", "text": "An 8% floor is an assumption in this example, not a diagnosis. A plateau in an actual experiment could include dispensing variation, persistent optical effects, imperfect blank correction or sample instability. The shape of the curve tells you where to investigate; it does not identify the culprit by itself.", "html": "

An 8% floor is an assumption in this example, not a diagnosis. A plateau in an actual experiment could include dispensing variation, persistent optical effects, imperfect blank correction or sample instability. The shape of the curve tells you where to investigate; it does not identify the culprit by itself.

"}]}, {"id": "cells-have-counting-statistics-too", "heading": "Cells have counting statistics too", "blocks": [{"type": "p", "text": "Imagine dispensing a well-mixed suspension in which cells arrive independently, with an average of 100 cells per well. Under a Poisson loading model, the cell count has a standard deviation of 10 cells: a 10% CV before the assay chemistry has done anything.", "html": "

Imagine dispensing a well-mixed suspension in which cells arrive independently, with an average of 100 cells per well. Under a Poisson loading model, the cell count has a standard deviation of 10 cells: a 10% CV before the assay chemistry has done anything.

"}, {"type": "p", "text": "At 1,000 cells per well, the corresponding count CV is 3.2%; at 10,000, 1%. The arithmetic resembles photon counting because both examples use the same probability model. The interventions are different. More detected photons improve the estimate of the light coming from those particular cells. They do not redraw the cell population.", "html": "

At 1,000 cells per well, the corresponding count CV is 3.2%; at 10,000, 1%. The arithmetic resembles photon counting because both examples use the same probability model. The interventions are different. More detected photons improve the estimate of the light coming from those particular cells. They do not redraw the cell population.

"}, {"type": "p", "text": "Poisson seeding is more than a classroom convenience. Chang and colleagues used it to describe initial occupancy in microwell arrays, then followed substantial differences in subsequent clonal growth. Their system supports the loading principle; it does not establish a universal CV for a conventional cell-based screen. [3]", "html": "

Poisson seeding is more than a classroom convenience. Chang and colleagues used it to describe initial occupancy in microwell arrays, then followed substantial differences in subsequent clonal growth. Their system supports the loading principle; it does not establish a universal CV for a conventional cell-based screen. [3]

"}, {"type": "p", "text": "Clumping, settling during dispensing, unequal delivered volumes and differential growth can all change the distribution. Deliberately dispensing a known number of cells changes the model too. For cultured assays, the relevant population is the one present when the signal is generated, which may differ considerably from the starting population.", "html": "

Clumping, settling during dispensing, unequal delivered volumes and differential growth can all change the distribution. Deliberately dispensing a known number of cells changes the model too. For cultured assays, the relevant population is the one present when the signal is generated, which may differ considerably from the starting population.

"}, {"type": "p", "text": "Cells also contribute unequal amounts of signal. If cell number is Poisson with mean n, and each cell contributes an independent signal with single-cell CV c, the model becomes:", "html": "

Cells also contribute unequal amounts of signal. If cell number is Poisson with mean n, and each cell contributes an independent signal with single-cell CV c, the model becomes:

"}, {"type": "div", "text": "Cell-signal coefficient of variation squared equals the quantity one plus c squared divided by n.", "html": "
CVcell signal2 = 1 + c2n
"}, {"type": "p", "text": "With 100 cells and a single-cell CV of 100%, the predicted well-signal CV is about 14.1%, rather than the 10% obtained by assuming identical cells. This calculation excludes detector noise and assumes cell brightness is independent of cell number and of other cells. Density-dependent biology can break those assumptions.", "html": "

With 100 cells and a single-cell CV of 100%, the predicted well-signal CV is about 14.1%, rather than the 10% obtained by assuming identical cells. This calculation excludes detector noise and assumes cell brightness is independent of cell number and of other cells. Density-dependent biology can break those assumptions.

"}, {"type": "p", "text": "That distinction matters for ATP assays. Promega's CellTiter-Glo 2.0 manual discusses changes in ATP per cell with cell density and physiological state. Luminescence can be proportional to cell number over a validated range without being a literal cell counter under every treatment. [4]", "html": "

That distinction matters for ATP assays. Promega's CellTiter-Glo 2.0 manual discusses changes in ATP per cell with cell density and physiological state. Luminescence can be proportional to cell number over a validated range without being a literal cell counter under every treatment. [4]

"}, {"type": "p", "text": "Adding cells may reduce sampling variation, but it can also change the biology you intended to measure. Check the response to compounds and the useful assay window before accepting the prettier CV.", "html": "

Adding cells may reduce sampling variation, but it can also change the biology you intended to measure. Check the response to compounds and the useful assay window before accepting the prettier CV.

"}]}, {"id": "repeated-reads-and-replicate-wells-answer-different-questions", "heading": "Repeated reads and replicate wells answer different questions", "blocks": [{"type": "p", "text": "Ten reads of one stable well repeatedly measure the same preparation. Ten independently prepared wells include preparation differences. Calling both exercises “reproducibility” can conceal the most useful information in the experiment.", "html": "

Ten reads of one stable well repeatedly measure the same preparation. Ten independently prepared wells include preparation differences. Calling both exercises “reproducibility” can conceal the most useful information in the experiment.

"}, {"type": "p", "text": "A practical integration-time study should contain both. Use representative low, middle and high signals, along with blanks. Include independently prepared replicates and repeated measurements where the assay tolerates them. Compare the variability within each well with the variability between wells at each integration time.", "html": "

A practical integration-time study should contain both. Use representative low, middle and high signals, along with blanks. Include independently prepared replicates and repeated measurements where the assay tolerates them. Compare the variability within each well with the variability between wells at each integration time.

"}, {"type": "p", "text": "If repeated reads improve while between-well variation remains nearly unchanged, investigate the preparation and persistent spatial effects. If both improve, photon collection may still be limiting. If both deteriorate with time, investigate drift or damage before calculating a noise floor.", "html": "

If repeated reads improve while between-well variation remains nearly unchanged, investigate the preparation and persistent spatial effects. If both improve, photon collection may still be limiting. If both deteriorate with time, investigate drift or damage before calculating a noise floor.

"}, {"type": "p", "text": "Balance read order and elapsed time across settings. Reading every well briefly first and every well slowly last confounds integration time with assay age. Fluorescence excitation may itself perturb the sample; use matched fresh preparations when repeated exposure changes the signal.", "html": "

Balance read order and elapsed time across settings. Reading every well briefly first and every well slowly last confounds integration time with assay age. Fluorescence excitation may itself perturb the sample; use matched fresh preparations when repeated exposure changes the signal.

"}, {"type": "p", "text": "For a stable assay, plotting CV squared against inverse integration time can be informative. A simple model gives a straight line whose intercept represents variation that does not decrease with longer reading. Use it as a diagnostic approximation. Correlated noise, drift and nonlinearity can defeat the interpretation.", "html": "

For a stable assay, plotting CV squared against inverse integration time can be informative. A simple model gives a straight line whose intercept represents variation that does not decrease with longer reading. Use it as a diagnostic approximation. Correlated noise, drift and nonlinearity can defeat the interpretation.

"}]}, {"id": "subtracting-background-leaves-its-noise-behind", "heading": "Subtracting background leaves its noise behind", "blocks": [{"type": "p", "text": "Blank subtraction removes an estimate of the average background. It does not remove the random background photons collected in the sample well.", "html": "

Blank subtraction removes an estimate of the average background. It does not remove the random background photons collected in the sample well.

"}, {"type": "p", "text": "Consider an ideal counting measurement with 1,000 signal photons and 1,000 background photons. Even if you knew the mean background perfectly, the sample measurement fluctuates according to all 2,000 detected photons. After subtraction, the net signal is 1,000, but its standard deviation is about 44.7: a CV of 4.5%, rather than the 3.2% expected without background. An experimentally estimated blank adds uncertainty of its own. [1, 2]", "html": "

Consider an ideal counting measurement with 1,000 signal photons and 1,000 background photons. Even if you knew the mean background perfectly, the sample measurement fluctuates according to all 2,000 detected photons. After subtraction, the net signal is 1,000, but its standard deviation is about 44.7: a CV of 4.5%, rather than the 3.2% expected without background. An experimentally estimated blank adds uncertainty of its own. [1, 2]

"}, {"type": "p", "text": "This is why reducing optical background and leakage from neighboring bright wells can be worth more than extending the read. Neither a clean-looking baseline nor a large raw signal guarantees a precise small difference after subtraction. Check performance at the low signals and intermediate responses that determine compound ranking, as well as at the bright control.", "html": "

This is why reducing optical background and leakage from neighboring bright wells can be worth more than extending the read. Neither a clean-looking baseline nor a large raw signal guarantees a precise small difference after subtraction. Check performance at the low signals and intermediate responses that determine compound ranking, as well as at the bright control.

"}]}, {"id": "spend-the-next-second-where-it-helps", "heading": "Spend the next second where it helps", "blocks": [{"type": "p", "text": "Longer reading also extends the period during which temperature, evaporation or reaction progress can change the plate. Promega specifically identifies temperature as affecting CellTiter-Glo 2.0 light intensity and decay, and cautions about temperature gradients within plates. [4] A reproducible read schedule and stable sample temperature therefore belong in the precision experiment.", "html": "

Longer reading also extends the period during which temperature, evaporation or reaction progress can change the plate. Promega specifically identifies temperature as affecting CellTiter-Glo 2.0 light intensity and decay, and cautions about temperature gradients within plates. [4] A reproducible read schedule and stable sample temperature therefore belong in the precision experiment.

"}, {"type": "p", "text": "Choose optics, plate geometry and focus settings that collect useful light efficiently while controlling background and interwell leakage. Verify detector linearity over the actual signal range. Compression can make bright wells appear less variable, which is a separate problem from genuinely better photon statistics.", "html": "

Choose optics, plate geometry and focus settings that collect useful light efficiently while controlling background and interwell leakage. Verify detector linearity over the actual signal range. Compression can make bright wells appear less variable, which is a separate problem from genuinely better photon statistics.

"}, {"type": "p", "text": "During optimization:", "html": "

During optimization:

"}, {"type": "p", "text": "Increase integration time when repeated-read data show that measurement noise remains important.", "html": "

Increase integration time when repeated-read data show that measurement noise remains important.

"}, {"type": "p", "text": "Fix dispensing, mixing or cell-loading problems when the variation persists between preparations.", "html": "

Fix dispensing, mixing or cell-loading problems when the variation persists between preparations.

"}, {"type": "p", "text": "Compare uncertainty in the reported endpoint, including any subtraction or ratio, rather than judging only the brightest raw channel.", "html": "

Compare uncertainty in the reported endpoint, including any subtraction or ratio, rather than judging only the brightest raw channel.

"}, {"type": "p", "text": "Recheck the chosen timing under a realistic plate sequence, including temperature equilibration and queue delays.", "html": "

Recheck the chosen timing under a realistic plate sequence, including temperature equilibration and queue delays.

"}, {"type": "p", "text": "Stop extending the read when the improvement no longer changes the decisions the assay must support. At that point, the next useful experiment is likely to happen at the dispenser, in the incubator or in the assay design.", "html": "

Stop extending the read when the improvement no longer changes the decisions the assay must support. At that point, the next useful experiment is likely to happen at the dispenser, in the incubator or in the assay design.

"}]}, {"id": "references", "heading": "References", "blocks": [{"type": "references", "text": "1. Hamamatsu Photonics. Photon Counting SNR Simulator. Photon statistics and signal/background contributions; accessed September 24, 2026. 2. Owicki JC. Fluorescence Polarization and Anisotropy in High Throughput Screening: Perspectives and Primer. Journal of Biomolecular Screening. 2000;5:297–306. 3. Chang TC and colleagues. Microwell arrays reveal cellular heterogeneity during the clonal expansion of transformed human cells. Technology. 2015;3:163–171. Initial seeding statistics and subsequent clonal behavior in their microwell system. 4. Promega. CellTiter-Glo 2.0 Cell Viability Assay Technical Manual TM403. Revision 1/23, section 4; temperature and ATP-per-cell considerations. Worked examples in this article are constructed calculations, not measurements from this manual.", "html": "
\n

1. Hamamatsu Photonics. Photon Counting SNR Simulator. Photon statistics and signal/background contributions; accessed September 24, 2026.

\n

2. Owicki JC. Fluorescence Polarization and Anisotropy in High Throughput Screening: Perspectives and Primer. Journal of Biomolecular Screening. 2000;5:297–306.

\n

3. Chang TC and colleagues. Microwell arrays reveal cellular heterogeneity during the clonal expansion of transformed human cells. Technology. 2015;3:163–171. Initial seeding statistics and subsequent clonal behavior in their microwell system.

\n

4. Promega. CellTiter-Glo 2.0 Cell Viability Assay Technical Manual TM403. Revision 1/23, section 4; temperature and ATP-per-cell considerations. Worked examples in this article are constructed calculations, not measurements from this manual.

\n
"}]}], "references": [{"id": "ref-1", "citation": "1. Hamamatsu Photonics. Photon Counting SNR Simulator. Photon statistics and signal/background contributions; accessed September 24, 2026.", "urls": ["https://www.hamamatsu.com/eu/en/resources/interactive-tools/photon-counting-snr-simulator.html"]}, {"id": "ref-2", "citation": "2. Owicki JC. Fluorescence Polarization and Anisotropy in High Throughput Screening: Perspectives and Primer. Journal of Biomolecular Screening. 2000;5:297–306.", "urls": ["https://doi.org/10.1177/108705710000500501"]}, {"id": "ref-3", "citation": "3. Chang TC and colleagues. Microwell arrays reveal cellular heterogeneity during the clonal expansion of transformed human cells. Technology. 2015;3:163–171. Initial seeding statistics and subsequent clonal behavior in their microwell system.", "urls": ["https://pmc.ncbi.nlm.nih.gov/articles/PMC4854201/"]}, {"id": "ref-4", "citation": "4. Promega. CellTiter-Glo 2.0 Cell Viability Assay Technical Manual TM403. Revision 1/23, section 4; temperature and ATP-per-cell considerations. Worked examples in this article are constructed calculations, not measurements from this manual.", "urls": ["https://worldwide.promega.com/-/media/files/resources/protocols/technical-manuals/101/celltiterglo-2-0-assay-protocol.pdf?la=en"]}], "equations": [{"description": "Photon coefficient of variation equals one divided by the square root of the mean detected photon count N.", "mathml": "CVphoton = 1N"}, {"description": "Total coefficient of variation squared is approximately measurement coefficient of variation squared plus between-well coefficient of variation squared.", "mathml": "CVtotal2 ≈ CVmeasurement2 + CVbetween wells2"}, {"description": "Cell-signal coefficient of variation squared equals the quantity one plus c squared divided by n.", "mathml": "CVcell signal2 = 1 + c2n"}], "tables": [{"caption": "Constructed example: diminishing precision gains from longer integration", "rows": [["Integration per well", "Measurement CV", "Persistent well CV", "Total CV"], ["1 second", "6.0%", "8.0%", "10.0%"], ["4 seconds", "3.0%", "8.0%", "8.5%"], ["16 seconds", "1.5%", "8.0%", "8.1%"]], "data_kind": "constructed example"}], "provenance": {"kind": "Author-provided article converted to structured text", "examples": "Constructed calculations, not a raw experimental dataset", "editorial_additions": "Summary, takeaways, topic tags, and limitations summary"}, "display_additions": {"figures": [], "bench_check": false, "html": ""}, "content_sha256": "014655d21b70af281a7698e0b7f32df1fd2a8867f123f84d3b3bf5fcc97a05ee"} {"schema_version": "1.0", "id": "plate-map", "canonical_url": "https://discoveryinpractice.com/bench-tips/plate-map/", "title": "Read the plate map before you read the statistic", "author": {"name": "Discovery in Practice", "url": "https://discoveryinpractice.com/about/", "@type": "Organization"}, "language": "en", "summary": "Inspect spatial patterns alongside control statistics, then use those patterns to choose a follow-up check.", "takeaways": ["Preserve raw wells and their positions.", "Compare patterns with the dispensing layout and read order.", "A spatial pattern alone does not establish its cause."], "limitations_summary": "Editorial adaptation from the related article. The plate map is illustrative, not experimental data.", "topics": ["Bench tips", "Plate maps"], "publication_status": "published", "date_published": "2026-09-25", "date_modified": "2026-09-25", "license": "CC-BY-4.0", "license_url": "https://creativecommons.org/licenses/by/4.0/", "content_version": "1.0", "body_html": "
BT·01Bench tip

Read the plate map before you read the statistic.

One summary number can hide where a measurement changes across the plate.

384 wells · illustrativeCols 1–2 high · 23–24 low

This map is an illustration, not experimental data.

Start with the raw wells

Inspect the spatial pattern alongside the control distributions. Look for edge patterns, row or column shifts, and unusual individual wells. Compare the map with the dispensing layout, control positions, and reading order.

Use the pattern to choose a check

A pattern is a reason to investigate, not a diagnosis. Keep preparation and reader settings with the original data so you can compare runs without losing the context.

For the full discussion of misleadingly consistent controls, read A better Z′ can hide a worse measurement.

", "body_text": "BT·01Bench tipRead the plate map before you read the statistic.One summary number can hide where a measurement changes across the plate.384 wells · illustrativeCols 1–2 high · 23–24 lowThis map is an illustration, not experimental data.Start with the raw wellsInspect the spatial pattern alongside the control distributions. Look for edge patterns, row or column shifts, and unusual individual wells. Compare the map with the dispensing layout, control positions, and reading order.Use the pattern to choose a checkA pattern is a reason to investigate, not a diagnosis. Keep preparation and reader settings with the original data so you can compare runs without losing the context.For the full discussion of misleadingly consistent controls, read A better Z′ can hide a worse measurement.", "sections": [], "references": [{"id": "related-article", "citation": "Andrew Stewart. A better Z′ can hide a worse measurement.", "urls": ["https://discoveryinpractice.com/articles/z-prime-detector-linearity/"]}], "equations": [], "tables": [], "provenance": {"kind": "Editorial adaptation", "examples": "Illustrative map, not experimental measurements", "editorial_status": "Approved for publication"}, "license_status": "Published under CC BY 4.0", "content_sha256": "763d8831adbbabb97e874b5035d5aee2c06a8cd6aa1125a1361f0578b6a3c1cc"}