Concentration Prediction estimates the concentration of an unknown sample by inverting a previously measured standard curve. A titration of known concentrations is fit with a four-parameter logistic (4PL), a three-parameter logistic (3PL) whose lower asymptote is fixed at zero, or a through-origin scaled cubic model (y ∝ x³). Each sensor reading from the unknown sample is then mapped back through the curve to a concentration. Unlike the per-phase evaluations above, this is launched as a project-level workflow (action card on the project home) so a calibration measurement can be reused across many unknown samples acquired under matched conditions.

For the physics of dose-response measurements and how to design a useful titration, see BIA → Equilibrium. For 4PL and 3PL, the underlying logistic fit is the same one documented in Equilibrium Evaluation.

When to choose it

• You have a stable, well-characterised analyte/ligand pair and want to quantify how much of that analyte is present in an unknown sample (potency assay, batch QC, supernatant titer).
• You can run a titration of known concentrations once, then probe many unknowns against the same standard curve.
• The unknown sample is dilutable: running the same sample at several dilutions and combining the back-calculated values is the most reliable way to land within the curve's linear range and confirm internal consistency.

Data requirements

• Calibration source. A measurement, processed data, or saved Concentration Determination evaluation with a Calibration Curve / Titration Series phase containing at least three ASSOCIATION injections at known concentrations. Aim for ≥6 concentrations bracketing the expected target. Logistic fits should include ≥2 concentrations below and ≥2 above the inflection; cubic fits should cover the full interval in which the cubic response is expected to hold. Sources without that phase and processed data explicitly scoped to another phase cannot be selected. For processed data covering the full measurement, saved processing is applied first and the Calibration Curve phase is then used for the prediction.
• Unknown source(s). One or more measurements, processed data, or saved Concentration Determination evaluations with an Unknown Concentration phase. Each association injection should carry a Dilution Factor set in the injection editor; if a sample was diluted before being injected, the dilution factor is what back-converts the curve-derived value to the original sample concentration. Processed data explicitly scoped to another phase cannot be selected. For full-measurement processed data, saved processing is applied first and the Unknown Concentration phase is then used for the prediction.
• Matched conditions. Buffer, flow rate, ligand layout, and analyte chemistry must match between calibration and unknown for the curve to apply. The system does not detect or correct condition mismatches.
• Normalization injections. By default the injection before the first IMMOBILIZATION injection is mapped to 0, and that IMMOBILIZATION injection is mapped to 1. Calibration and unknown traces should both contain comparable reference injections. If they do not, choose Custom reference injections and set the 0 and 1 reference injections manually, or choose No normalization.

How to run it

1. Open the project home and click the Concentration Determination goal card.
2. In Details, enter a study name and add optional tags or a description.
3. In Data, select one calibration source and one or more unknown sources. Use the Calibration and Unknown samples tabs to switch the full-width selection table between the two roles. If the calibration measurement also carries an Unknown Concentration phase, the wizard automatically selects it as an unknown sample too. Selecting an evaluation as an unknown reuses all of its unknown inputs.
4. Sources without the phase required for their role and processed data explicitly scoped to another phase are disabled. Next remains unavailable until one valid calibration source and at least one valid unknown source are selected.
5. In Cleanup, switch between Calibration and Unknown samples to inspect the full trace preview and cleanup actions for each input. Each input keeps its own cleanup actions, and the edited inputs are used for both the prediction preview and the saved evaluation. Cleanup actions modify individual traces; final concentration normalization is configured once in Parameters.
6. In Parameters, review concentration normalization, sensor aggregation, and automatic outlier removal, or keep the recommended defaults. The reference value percentiles are entered as percentages. The info buttons beside concentration normalization and Aggregate sensors open the guidance below. Optional known concentration (nM) is in the collapsed Advanced Parameters section.
7. In Fitting, choose the calibration model and inspect the calibration curve, concentration boxplot, trace plots, and findings. The default 4PL keeps both asymptotes flexible; use 3PL only when baseline subtraction represents a true zero response. Use Cubic only when the assay is expected to follow a non-saturating x³ response over the working interval.
8. In the Calibration Curve tab's Median view, select a calibration injection to exclude the whole injection; switch to Sensors to select individual sensor observations. Manual exclusions remain visible as gray × markers and can be restored individually or reset together from Manual Exclusions. They are identified by injection UUID and optional sensor name, compose with automatic exclusions, and are not cleared when automatic outlier removal is toggled.
9. In the Unknownstab, use the checkbox at the right of a group's adjusted median and confidence interval to exclude or restore the whole unknown injection, or select an individual sensor row to change only that observation. Rows outside the fitted calibration boundaries remain visible with an unavailable concentration and zero weight. At least one contributing sensor result must remain included; the control that would remove the final contributor is disabled. Use Distribution to inspect the concentration boxplot.
10. Continue to Review. The prediction runs automatically and shows the study setup plus the calibration, concentration, and trace plots. Cleanup actions configured for calibration and unknown inputs are listed separately. Each list shows up to three action names followed by an ellipsis when more actions were applied; select the list to show or hide all applied actions. Review also reports the overall concentration, confidence interval, contributing unknown count, and warnings. You can inspect this preview without saving it.
11. Click Save to persist the named run as a study under the project (writes concentration_statistics.csv plus calibration, boxplot, and trace plots), then return to the Evaluations table.

Configuration

How concentrations are calculated

The pipeline runs in three stages:

1. Build the calibration curve. The titration phase is normalized (see above), the baseline (zero-concentration) response is subtracted from each higher-concentration response to give a delta response, and the selected 4PL, 3PL, or scaled through-origin cubic model is fit to delta-response vs. concentration — either per sensor (default) or aggregated. Manually excluded sensor observations and whole injections, together with any automatically rejected calibration outlier, do not contribute to the fit.
2. Invert per sensor. For each association injection of the unknown trace, the delta response of each sensor is inverted through its fitted curve. Predictions between the lowest and highest retained standards used in the final fit use interpolation, including the two boundaries. A response outside the fitted response at those boundaries has no predicted concentration and zero effective weight; it is reported as below or above range rather than extrapolated or clamped to an endpoint. The retained value is the per-sensor diluted concentration x_d, which is multiplied by the injection's dilution factor d to recover the original sample concentration X = d · x_d.
3. Combine across dilutions. Each sensor also carries a sensitivity weight |x · dy/dx| at its reading point on the curve — high where the curve is steep and low where it is flat. Per dilution group, the median of Xacross in-range sensor predictions with a positive effective weight is the group estimate; positive weights do not otherwise change each sensor's rank in that median. Manually excluded observations and out-of-range responses have zero effective weight and do not enter the group median, CI, or count. Group estimates are then combined into a single Overallvalue as a weighted mean, with each group's weight equal to the mean of its per-sensor sensitivities — including zeros — divided by dilution². That mean uses the sensors retained by the shared cross-dilution quality mask as its denominator: a sensor unsupported only in one dilution contributes zero to that group's mean, while a sensor rejected globally for invalid data or outlier behavior is absent from every group's denominator.

Results

• Fitting workspace. Calibration Curve contains the standard curve and its Median/Sensors control. The Result card above the workspace contains the overall result on every tab.Unknowns contains full-width per-injection sensor tables.Distribution contains the concentration boxplot. Accuracy appears only when a known concentration was supplied and an accuracy assessment is available. Active range and dilution-response findings appear above the calibration plot in a compact warning strip; Review presents the same decision-relevant warnings with summary text only.
• Overall concentration in nM, with 95% confidence interval and the number of unknown groups that contribute positive effective weight.
• Per-group statistics.One percentage-labelled tab per reported unknown / dilution group: median of the positive-effective-weight back-calculated sample concentrations with its confidence interval and a per-sensor table in the Unknowns workspace (measured response, prediction status, raw concentration, adjusted-for-dilution concentration, and the sensor's normalized share of that group's sensitivity). Point to a tab to see its full injection name. Its color marker matches the group in the plots, and the table expands without an internal scrollbar. Manually excluded rows remain visible with a 0% share so they can be restored. A Contribution chip on each tab shows the share of the Overall result that dilution group accounts for — computed as mean(weights) / dilution²normalized across all groups, which is the same formula the combiner uses. If one group's contribution is overwhelmingly high relative to the others, the Overall is effectively driven by that group alone.
• Calibration plot.The Calibration Curve tab shows each sensor fit as solid over that sensor's retained-standard interval; no fitted extension is drawn beyond the calibration boundaries. The plot also contains calibration standards and one colored median marker for each unknown / dilution group with a contributing prediction. Groups with at least two positive-effective-weight predictions also get an x-range box. Each group keeps the same color on the calibration plot and the boxplot so the same sample can be followed across both views. Select a whole calibration injection in Median view or an individual calibration observation in Sensors view. Individual sensor fits and finite, in-range sensor prediction dots are available through the Median/Sensors option above the chart but hidden by default to keep the QC view readable. The logarithmic concentration axis labels the actual positive standard concentrations in nM. The response axis shows normalized calibration data in a.u.; without normalization, coherent mass density is displayed in pg/mm² using the same single display conversion as the measured traces.
• Boxplot. The Distribution tab shows positive-effective-weight back-calculated concentrations per dilution group, side-by-side, with a dashed line marking the Overall predicted result and a shaded band marking its 95% CI. Hover the line to see the numeric prediction and CI range.
• Accuracy. When an independently known concentration was supplied and the prediction returns an accuracy assessment, the Accuracy tab reports agreement with that reference value. Runs without an accuracy assessment omit the tab.
• Saved trace plots. The completed evaluation includes a calibration trace and one plot per unknown trace as artifacts, useful for spotting injection artefacts before trusting the numbers.
• CSV export (when saved as a study): concentration_statistics.csv contains the Overall result, per-group statistics (with a Contribution (%) column matching the chip in the UI), and a per-sensor section listing raw value, adjusted value, and Weight Share (%) normalized within each reported group. Group statistics use contributing sensor results; explicit inclusion-status and exclusion-reason columns distinguish manually excluded injection and sensor rows from naturally zero-weight rows.
• Saved evaluation statistics. Opening the completed evaluation from the Evaluations table shows findings first, followed by the Hologram Identifier and Filter by tags controls. The saved Overall result, dilution-group summaries, contributions, per-sensor values, plot artifacts, and downloadable CSV appear below those controls. Tag selections filter both the saved statistics view and the artifacts.
• Deterministic findings. Every completed evaluation reports how many unknown groups remain usable. For logistic models, it also reports whether the calibration curves cover at least 90% of their fitted high-concentration plateau; this plateau assessment does not apply to the non-saturating cubic model. When any response is outside the fitted boundary or unavailable, a Calibration rangefinding reports aggregate counts with bounded per-group detail plus the affected sensor's measured response and status. Incomplete or unassessable logistic curves trigger a warning that includes the affected curve count and, when at least one curve is assessable, its median coverage, threshold, and highest retained standard used in the final fit. Automatic quality evidence records the exact injection and sensor trigger, its response or adjusted concentration when finite, and whether it was rejected for a non-finite value or IQR outlier behavior.
• Dilution response compression. This advisory check uses the three highest sample fractions when they span at least 4× in dilution factor. For each sensor, their measured responses are normalized against the fitted response interval between the lowest and highest retained calibration standards. A sensor is compressed when all three responses stay at or above the 75% calibrated response position while spanning no more than 15% of that interval. The warning appears when at least four sensors are assessable and at least half are compressed. It indicates possible response saturation or a dilution-series inconsistency; it does not exclude observations or alter prediction weights.
• Re-evaluation. Use the Re-evaluate icon for a completed Concentration Determination in the Evaluations table to open the workflow on its Fitting step. Each data role is preselected on the raw data that role ran on, carrying the clean-up actions stored with it, so the fit reproduces the one the evaluation reports. Its model, normalization, aggregation, automatic outlier setting, and manual calibration exclusions keyed by injection UUID and optional sensor name are restored, and the name carries the next version number. Saving creates a new evaluation and leaves the original unchanged.

Good data example

A healthy standard-curve prediction should show dilution groups that agree after back-calculation, with the Overall result inside the shaded 95% CI band on the boxplot. The AP_Ligelizumab-1 regression fixture is a representative good run: the true sample concentration is 50 nM and the weighted combine reports 46.60 nM with a 95% CI of 45.56–48.48 nM across 255 sensor predictions.

The group medians are not identical, but they stay in the same range and bracket the combined result. That is the pattern to look for before trusting the Overall prediction: agreement across dilutions, no single group dominating the contribution chips, and unknown points landing on the informative middle of the calibration curve.

Interpretation and troubleshooting