assay · intermediate · partly written
Dose-response assay
A dose-response assay runs the same measurement at a series of compound concentrations and fits a curve, so potency and the shape of the effect become numbers.
Also called: dose-response curve, concentration-response curve, IC50 determination, curve fitting
In short
A dose-response curve is a few numbers pulled from a sigmoid: bottom, top, midpoint (IC50 or EC50) and slope. The midpoint is a potency under that experiment's conditions. It is not an affinity. The fit also fails quietly: partial curves, steep slopes, bell shapes and cell division rate all distort the number without any error message. 123
Reported, not yet replicated
His question: why does controllable dosage matter?
He said: "You maybe do dose response. Now you have how potent the compound is, and you're looking for something that you can control the dosage of to get a desired effect, because actually, I'm not sure why."
This is his own words and his own open question, kept as he said it. The next section answers it from the report. 4
Amit's hypothesis
Why controllable dosage matters: what the report says more
Part of the answer is mechanical and part is the report's own reasoning, and they should be kept apart.
The mechanical part: the four fitted parameters are only identified if the data show the floor, the ceiling and the middle, roughly the span from no effect to full effect. A single concentration cannot tell "potent but partial" from "weak but full". A curve with a visible floor, ceiling and middle shows that the effect rises smoothly with dose, and its shape is itself evidence: a steep slope can mean aggregation, a bell shape a three-part complex, a floor above zero partial efficacy.
The therapeutic-window part is the report's own inference (confidence inferred, no source states it): for a drug in a person the same logic holds with one extra edge, a window between nothing and toxic that is wide and smooth. A steep curve is a narrow window; a bell-shaped one means more is not better. The report marks this reading as low confidence.
The report could not open the Assay Guidance Manual text that is said to want points on both sides of the reported IC50, so that statement is not made here. 31567
Our inference
How the curve is fitted more
The model is the four-parameter logistic: Bottom (lower plateau), Top (upper plateau), LogEC50 (the x value halfway between them, called IC50 when the curve falls) and HillSlope, which has no units. A slope of 1.0 is the standard sigmoid; below 1 is shallower, above 1 steeper. Software varies the four numbers until the curve sits as close to the data as possible.
The Hill equation dates from Archibald Hill, 1910, for the sigmoid oxygen binding of haemoglobin, which he read as cooperative binding. The Hill slope is not a reliable count of binding sites except for extreme cooperativity, so treat it as a shape descriptor. 183
Reported, not yet replicated
IC50 is not affinity more
IC50 is the concentration that reduces a process by 50% in vitro. It is a functional measure, not a direct indicator of affinity. Ki is the inhibitor's binding constant. For a competitive inhibitor of an enzyme, Ki = IC50 / (1 + [S]/Km) (the Cheng-Prusoff relation). IC50 depends on substrate and enzyme concentrations, and the relation can over- or underestimate Ki at low agonist concentrations; a 1993 paper (title and abstract only) argues the correction is theoretically invalid for functional antagonist curves and proposes another analysis.
Worked example, the atlas's arithmetic and a hypothetical: if a compound competes with ATP at 1.5 times Km, the formula gives an IC50 about 2.5 times Ki. Double the ATP and the same compound's IC50 rises while Ki does not. The STK33 record's own keywords say "Non-ATP Competitive Inhibitor", so this is not a claim about the STK33 hits. 29103
Reported, not yet replicated
A real programme: the STK33 retest deep
The Broad MLPCN STK33 retest (PubChem AID 2821) took 224 substances: 174 active, 50 inactive. Active meant an IC50 within one log above the highest concentration tested; IC50s were extrapolated up to 1 log over that top concentration; inconclusive fits were flagged when activity was between 30 and 70% as a constant fit, or the fit was invalid (Genedata Screener Condoseo 7.0.3). The record lists 8 test concentrations (0.1, 0.195, 0.38, 0.8, 1.6, 3, 6 and 12 uM), so the top is 12 uM and "within 1 log" allows IC50s up to about 120 uM (the atlas's arithmetic). The record shows 0 inconclusive outcomes. Some of those potencies are therefore extrapolations.
The chosen probe ML281 had an STK33 IC50 of 14 nM, over 700-fold selective against PKA (IC50 above 10 uM) and 550-fold against Aurora B, and at 1 uM hit only FLT3 and KDR beyond STK33 among 83 kinases at 25% or more. In cells it did not meaningfully reduce viability up to 10 uM, with 99.6% (human) plasma protein binding. The authors wrote that synthetic lethality with KRAS cannot be recapitulated with selective STK33 small-molecule inhibitors, while noting that the lack of effect could stem from off-target effects or high plasma protein binding, which they leave unresolved.
The whole-system point: dose-response made the enzyme result trustworthy and made the cell result interpretable, since 10 uM is a bounded statement (the report's reading). 10113
Reported, not yet replicated
qHTS: the dose series inside the primary screen deep
Quantitative high-throughput screening (qHTS) titrates every compound instead of testing one concentration. It was demonstrated on pyruvate kinase, produced concentration-response curves for over 60,000 compounds in one experiment, found activators as well as inhibitors, and the authors say single-point screens miss many actives. A news summary says compounds were tested at seven or more concentrations spanning four orders of magnitude. The abstract says curves were classified by shape and potency; the class definitions were not read and are not stated here. 1213
Reported, not yet replicated
Where the fit fails, and how practitioners check more
Extrapolated midpoint: the curve never reaches a plateau, so the IC50 is a guess beyond the top concentration (the STK33 rule allows up to 1 log). Check: classify partial curves separately.
Steep slope: in 3,048 Lilly biochemical assays, 41 of 61 PAINS alerts showed at least 1.5-fold enrichment for high Hill slope (defined as above 2.0), which the paper links to aggregation.
Hook or bell shape: in degraders, excess compound fills target and ligase separately, so the three-part complex falls and the effect peaks then declines. The standard Hill fit misestimates IC50 and efficacy (vendor statement). In the paper the peak position equals the geometric mean of the two binary dissociation constants, positive cooperativity raises the peak without moving it, and molecular glues (under 500 Da) typically show no measurable hook. Check: a biphasic model and a wide titration.
IC50 mistaken for Ki: substrate concentration moves IC50. Check: Cheng-Prusoff, or Ki from kinetics.
Growth confound: fast-dividing cells look more sensitive. Check: GR metrics.
Irreproducible numbers: compare ranks, not single values.
Non-monotonic biology: U-shaped and inverted-U responses exist, so look at the raw points before trusting any fit. 651415167103
Reported, not yet replicated
How it gets abused deep
Quoting an IC50 without the assay's substrate or ATP level, or comparing IC50s across assays as if they were affinities. Reporting a potency from a curve that never flattens: the STK33 retest rule builds this into the active call. Fitting a sigmoid to a bell. Cherry-picking the readout: ATP light versus cell count gave different potencies for DNA-synthesis inhibitors, a point the report takes from the screening-cascade concept report. 31014
Reported, not yet replicated
Cost and throughput, with basis more
One public rate sheet (an academic core, file dated 2020-05-27, reagents excluded) prices a dose-response confirmation of 60 compounds, 8 points, n=2, on 3 plates at $1,837 internal and $3,490.30 external, about $31 and $58 per compound (the atlas's division). The same sheet prices a single-concentration confirmation of 500 compounds, n=3, on 5 plates at $3,739 internal, about $7.5 per compound. A dose series therefore costs about four times more per compound than a triplicate single-point confirmation on this one sheet. No commercial price was found, and qHTS has no price in the opened sources. 17123
Reported, not yet replicated
Question answered
How potent is this compound in this assay, and does its effect rise smoothly with dose? The fit reduces a whole curve to four numbers: bottom, top, midpoint (IC50 when the curve falls, EC50 when it rises) and Hill slope. The midpoint is a potency under that experiment's conditions, not an affinity. 123
Reported, not yet replicated
Readout physical
Nothing new is measured: it is the same instrument as the primary screen (a plate reader reading light or fluorescence) run on the same compound at a series of concentrations, usually diluted in equal ratios so the x-axis is evenly spaced on a log scale. In the STK33 retest each well reported luminescence from a kinase reaction. The dose is a final concentration in a well, not a dose in a body. 103
Reported, not yet replicated
Not measured
Affinity or mechanism (IC50 shifts with substrate and enzyme concentration); whether the effect is on target (an aggregator or a luciferase inhibitor can give a clean curve); and comparability across labs and days (in a reanalysis of two large cancer cell-line drug screens, only 2 of 15 shared drugs had a Pearson correlation above 0.6 on AUC, and 3 of 15 on IC50, by the rebuttal authors' reading). 2163
Reported, not yet replicated
Setting
Where in cascade
Sample prep
The compound is diluted into a concentration series. In the STK33 retest it was pinned in nanolitres from a 10 mM stock: 25 nL by pintool into a 384-well plate, or 2.5 nL into a 1536-well plate. 10
Reported, not yet replicated
Conditions
What changes the number: the substrate or ATP level (for an enzyme assay, IC50 rises with substrate for a competitive inhibitor), the top concentration tested (a curve that never flattens gives an extrapolated midpoint), the readout chosen (the report notes ATP light versus cell count gave different potencies for DNA-synthesis inhibitors), and in cells the number of divisions during the assay (fast-dividing cells look more sensitive). In the STK33 assay the ATP level was about 1.5 times its Km, according to the record's overview. 210153
Reported, not yet replicated
Throughput
On the one public rate sheet, a dose-response confirmation is priced for 60 compounds at 8 points, in duplicate, on 3 plates; a single-concentration confirmation on the same sheet covers 500 compounds in triplicate on 5 plates. The qHTS variant titrates every compound inside the primary screen: over 60,000 compounds gave concentration-response curves in one experiment. The class "medium" is the atlas's own sorting for the confirmation use, inferred from these figures; qHTS would sort higher. 1712
Our inference
Cost
Per sample
31 to 58 USD per compound, 8-point dose-response in duplicate (MSU Assay Development and Drug Repurposing Core recharge sheet: 60 compounds, 8 points, n=2, 3 plates costs $1,837 internal and $3,490.30 external; per-compound figures are the atlas's own division. Academic recharge, reagents excluded. A single-concentration triplicate confirmation on the same sheet is about $7.5 per compound internal, so the dose series costs about four times as much per compound. This is an academic core-facility recharge rate, not a commercial or contract-lab price; no commercial price was found.; as of 2020 (rate sheet file dated 2020-05-27; likely higher now))173
Per campaign
1,837 to 3,490.3 USD per 60-compound, 8-point, n=2, 3-plate confirmation (Same MSU core rate sheet: internal rate (low) and external rate (high). Reagents excluded. No commercial CRO price was found. This is an academic core-facility recharge rate, not a commercial or contract-lab price; no commercial price was found.; as of 2020 (rate sheet file dated 2020-05-27; likely higher now))17
Engineerable variants
Dose spacing and range
Equal-ratio dilutions over a wider span. The qHTS design puts the whole concentration series inside the primary screen instead of after it; the news summary says compounds were tested at seven or more concentrations spanning four orders of magnitude. 1213
Reported, not yet replicated
A different y-axis: GR50
GR50 normalises growth by the untreated cells' growth rate, because IC50 and Emax depend heavily on how many divisions occur during the assay and create artefactual genotype-sensitivity correlations. The metrics are GR50, GRmax and an area over the curve, and calculators are public. Only the abstract and a LINCS summary were opened, so the formula is not given here. 1518
Reported, not yet replicated
Use cases
STK33 retest to the probe ML281 (Broad MLPCN)
The retest of 224 substances in an 8-concentration series called 174 active and 50 inactive; active meant an IC50 within one log above the top concentration tested. The series decided which hits went on to chemistry. The chosen probe ML281 had an STK33 IC50 of 14 nM. In cells it did not meaningfully reduce viability up to 10 uM, so the cell result is a bounded statement; a single-point cell test could not distinguish no effect from no exposure (the report's reading). 10113
Reported, not yet replicated
qHTS of pyruvate kinase
Titrating every compound gave concentration-response curves for over 60,000 compounds in one experiment, found activators as well as inhibitors, and the authors say single-point screens miss many actives. Abstract only. 12
Reported, not yet replicated
ML role
Does
No machine-learning method specific to dose-response fitting or to dose-response data was found in the opened reports. The curve is fitted by software that varies the four numbers until the curve sits as close to the data as possible (nonlinear least squares; the GraphPad page itself does not describe the method). 31
Reported, not yet replicated
Cannot
Defaults applied at scale give wrong numbers silently when the curve is not a sigmoid: for degrader data a vendor states the standard Hill fit misestimates IC50 and efficacy. Any model that learns from IC50 labels also learns their noise and their conditions: a fitted midpoint can be an extrapolation, and potencies differ between labs. This second point is the atlas's own inference, not a sourced finding. 141610
Our inference
Data availability
PubChem BioAssay AID 2821 (and AID 2661, the primary screen) public
The STK33 retest record lists the 8 test concentrations in its result columns, the active call rules and the curve-fitting software, so each potency can be read together with the top concentration that produced it. AID 2661 is the single-concentration primary screen it followed. 1019
Caveats. A deposited potency may be an extrapolation up to one log above the top concentration; the record's own call rule says so. One programme, one target.
Reported, not yet replicated
Academic assay core (MSU Assay Development and Drug Repurposing Core) in-house
A published recharge sheet prices a 60-compound, 8-point, duplicate dose-response confirmation at $1,837 internal and $3,490.30 external, and a 500-compound single-concentration confirmation at $3,739 internal. 17
Caveats. Rate sheet file dated 2020-05-27, not current; reagents excluded; academic recharge, not an industry price; internal rates are for the home institution. No commercial CRO price was found.
Reported, not yet replicated
Teach this
In one sentence
A dose-response curve fits four numbers to a sigmoid, and its midpoint is a potency under that assay's conditions, not a binding strength. 12
Reported, not yet replicated
The five-minute version
Plate the compound across a log series of concentrations, read the same signal as the primary screen, and fit four numbers: bottom, top, midpoint and slope. Keep the compounds whose curves look like titrations, and test the survivors with assays that cannot be fooled by the same artefact. The curve's shape, not only its midpoint, is evidence: a steep slope can mean aggregation, a bell can mean a three-part complex, a floor above zero can mean partial efficacy. In the STK33 retest the active call allowed an IC50 up to one log above the top concentration of 12 uM, so some potencies were extrapolations. 31065
Reported, not yet replicated
What a listener will ask
Why not just test one concentration?
Is the IC50 how tightly the compound binds?
No. IC50 is a functional measure that depends on substrate and enzyme concentrations; for a competitive enzyme inhibitor Ki = IC50 / (1 + [S]/Km), and the relation fails for other modes. 2
Reported, not yet replicated
Why does the report say controllable dosage matters?
Mechanically, a curve with a visible floor, ceiling and middle shows the effect rises smoothly with dose. The therapeutic-window reading, a wide smooth range between nothing and toxic, is the report's own inference, not a sourced finding. 3
Our inference
The common wrong idea
A lower IC50 means the compound binds the target more tightly.
An IC50 is a potency under the assay's conditions; change the substrate level and the same compound's IC50 moves while its Ki does not.
The one citation to hand over: 1
Still open here
- Time per run (sample to result) was not found in any opened source, so time_per_run is unwritten.
- No commercial CRO price for a dose-response series was found; the only price basis is one academic core rate sheet from 2020 (reagents excluded). A current price range with a second basis is needed.
- Capital: no source on funding, company formation or platform deals for dose-response or qHTS was opened, so the capital field is unwritten.
- before_after is unwritten. The only supported shift is single-point screening to qHTS titration, and the report gives the authors' claim and the scale but not what forced it or what changed afterwards; the four steps need a source that tells that history.
- The Assay Guidance Manual chapter was blocked in this run. Unconfirmed, therefore not stated as claims: that it wants points on both sides of the reported IC50; that fixing Top, Bottom and slope is acceptable and under what conditions; that repeating with detergent and reporting the slope with its confidence interval are the standard checks for a steep curve.
- The qHTS curve classes (how curves were classified by shape and potency in Inglese 2006) are not stated: only the abstract was opened.
- The Hill slope of ML281 and the final-curve concentrations in the probe paper body were not checked; the full STK33 hit-to-probe curve data are not covered here.
- No public figure was found on how often a fitted IC50 is an extrapolation beyond the top concentration, so how common the STK33 retest rule is in other programmes is unknown.
- No machine-learning method specific to curve fitting or to dose-response data was found in the opened reports.
- Which optimal replicate number and plate design to use was not covered by any opened source.
- The report removed its own figures on Hill-slope confidence intervals and on aggregation-detergent checks because their sources were not re-opened.
Not yet written: Time per run, Before after, Capital.
Linked from
- Enzyme activity assay
- High-content imaging and morphological profiling
- Luminescent cell viability assay (CellTiter-Glo)
- Reporter gene assay
- Target engagement in cells (CETSA, NanoBRET)
- Western blot
- Dose response
- High-throughput screening
- Mechanism-confirmation assays
- Medicinal chemistry: potency, selectivity, ADMET
- The assay-error ceiling
- Therapeutic window and therapeutic index
- A hit series that binds and then fails in cells
- Potent on the enzyme, weak in cells: the compound or the cell?
Sources
- GraphPad Prism guide: classic dose-response curve with variable slope. GraphPad. web, not peer reviewed; opened 2026-11-17 all records citing this
- IC50 (Wikipedia). Wikipedia. web, not peer reviewed; opened 2026-11-17 all records citing this
- HowStuffWorks: dose-response and curve fitting (our own earlier write-up, unpublished). Kept in the atlas repository; not published.
- Amit's dictation of 2026-11-17: what the atlas must contain. Kept in the atlas repository; not published.
- Han. Han, J Biol Chem 2020: hook effect and ternary complex in degraders. Journal of Biological Chemistry (2020). paper; opened 2026-11-17 all records citing this
- Lilly 2018 analysis of PAINS in biochemical assays (PMC full text). PubMed Central (2018). paper; opened 2026-11-17 all records citing this
- Dose-response relationship (Wikipedia). Wikipedia. web, not peer reviewed; opened 2026-11-17 all records citing this
- Hill equation (biochemistry) (Wikipedia). Wikipedia. web, not peer reviewed; opened 2026-11-17 all records citing this
- Lazareno, Birdsall. Lazareno and Birdsall 1993 on the Cheng-Prusoff correction (Europe PMC listing). Europe PMC (1993). paper; opened 2026-11-17 all records citing this
- PubChem BioAssay AID 2821: STK33 inhibitor retest, 8-point dose response. PubChem (2010). database, not peer reviewed; opened 2026-11-17 all records citing this
- Weiwer et al.. Weiwer et al. 2012: the STK33 probe ML281 (PMC full text). PubMed Central (2012). paper; opened 2026-11-17 all records citing this
- Inglese et al.. Quantitative high-throughput screening (qHTS), Inglese et al., PNAS 2006 (abstract via Europe PMC). PNAS (2006). paper; opened 2026-11-17 all records citing this
- ScienceDaily summary of the 2006 qHTS paper, 2006-07-25. ScienceDaily (2006). news, not peer reviewed; opened 2026-11-17 all records citing this
- Genedata: automated analysis of complex biphasic PROTAC dose-response data. Genedata. vendor-page, not peer reviewed; opened 2026-11-17 all records citing this
- Hafner et al.. Hafner et al., Nature Methods 2016: growth rate inhibition metrics (abstract via Europe PMC). Nature Methods (2016). paper; opened 2026-11-17 all records citing this
- Safikhani et al.. Safikhani et al., F1000Research 2016: reanalysis of two large cancer cell-line drug screens. F1000Research (2016). paper; opened 2026-11-17 all records citing this
- MSU Assay Development and Drug Repurposing Core recharge rates (file dated 2020-05-27). Michigan State University (2020). report, not peer reviewed; opened 2026-11-17 all records citing this
- LINCS summary of Hafner et al. Nature Methods 2016. LINCS, Harvard Medical School (2016). web, not peer reviewed; opened 2026-11-17 all records citing this
- PubChem BioAssay AID 2661: STK33 primary screen (ADP-Glo). PubChem. database, not peer reviewed; opened 2026-11-17 all records citing this
Where it sits on the map
History of this record (7)
This record has been changed 7 times, on the dates below. The note kept against each change is written for the people who maintain the atlas, in their own shorthand, so it is not shown here; it is in the record itself.
- Throughput class
- medium
- Emerging
- no
- Status
- partly written
- Last updated
- 2026-11-19
- Record
- content/assays/dose-response-assay.json