12 / Monoclonal antibodies
Titer prediction: making the harvest outlook useful earlier
Editorial date
An upstream planning workflow that turns an evolving culture record into a bounded harvest estimate and a clear review action.
Proposed enterprise workflow · Source-linked analysis · Expert review

THE ENTERPRISE OBJECTIVE
A defined question.
A shared evidence basis.
Give upstream and downstream planners an estimate they can interpret at a defined decision time, including the range of plausible outcomes and the evidence that would change the plan.
Start with the receiving decision
A hypothetical fed-batch program needs an early view of expected harvest concentration to prepare downstream capacity. Historical endpoints exist, but batches differ in process version, duration and sampling coverage. The use case begins with a planning decision at a specified culture age rather than an unrestricted request to predict the final result.
The decision owner
Process development and MSAT own the scientific interpretation, with engineering, analytical and quality colleagues reviewing evidence relevant to their responsibilities.
Why the decision matters
The useful output is a planning choice with a deadline: whether to retain the current downstream preparation, prepare a contingency or request another observation. An estimate delivered after that deadline has limited operational value even if its final error is small. Agree who consumes the forecast, what they can change and the cost of an unnecessary change before selecting the modeling approach.
Is the current harvest outlook reliable enough to revise the downstream plan?
Join cell growth, viability, feeds, metabolites, volume and assay results using the timestamps at which each became available. Record sample collection and result availability separately. Define harvest titer and harvest timing consistently, and keep concentration distinct from recovered product mass.
The evidence boundary
Keep observations, derived quantities and model outputs distinguishable. Missing evidence remains an explicit gap in the review record, with an owner and a proposed way to resolve it.
HOW BIOMEDANA FITS
From source context
to a reviewable decision.
- 01
Define the planning cutoff
Agree the decision time, target endpoint and baseline forecast with the people who use the harvest estimate.
- 02
Prepare the historical view
Bio Hub retains batch context, method versions and result-availability times. Review exclusions and missing measurements before modeling.
- 03
Evaluate the forecasting model
Use a scoped Twin modeling engagement to compare candidates on held-out batches and document where uncertainty prevents a supported estimate.
- 04
Review the planning implication
Share the estimate, interval, data cutoff and change since the previous forecast. A named planner decides whether to adjust the downstream preparation.
Challenge the explanation before using the estimate
Train and evaluate at the same information cutoff used by planners. A measurement returned after that cutoff must not appear as an earlier model input. Compare with simple baselines such as the comparable-cohort median or a last-observation estimate. Keep entire batches together when splitting training and evaluation data.
Evaluation for this decision
Review errors by campaign, process version and forecast horizon. Evaluate interval coverage as well as point error, and inspect poorly performing batches individually. A prediction outside the supported operating region should trigger review or abstention. When the process changes, establish whether the previous evaluation still applies.
Make the model assumptions inspectable
The target definition also determines the scientific question. Concentration at a fixed culture age, concentration at an anticipated harvest time and total product mass are different targets. A change in working volume can alter concentration without an equivalent change in product mass. If harvest time itself is uncertain, report that assumption rather than presenting a forecast for one date as a universal endpoint prediction.
Make the next action explicit
| Review finding | Next action |
|---|---|
| The forecast supports the current planning decision | The named planner reviews the interval and cutoff, records the action and sets the next review time. |
| The plausible range spans the planning boundary | Retain a feasible contingency and identify the next observation that could resolve the decision. |
| The batch is outside the evaluated region | Use the agreed alternative planning basis and investigate applicability before relying on the forecast. |
How to evaluate value
Measure forecast error and interval coverage at each agreed horizon, the frequency of unsupported forecasts and whether the estimate improves the planning decision against its existing baseline.
What your team reviews.
A forecast review record containing the target definition, information cutoff, model version, baseline comparison, uncertainty, applicability assessment and planning decision.
Scope and interpretation
A harvest estimate supports planning. It does not determine release, product quality, achieved recovery or permission to extend a culture. Deployment-specific forecasting capability requires verification.
This scenario describes a proposed enterprise workflow. Deployment-specific integration, review and model behavior require verification. Business and process impact are evaluated against an agreed baseline and measured outcomes.
Explore the modality ↗The evidence needed to make this comparison useful
Build the evidence set around the receiving question. The following review map identifies what each source contributes and the uncertainty it is intended to resolve.
| Evidence layer | What to retain | Question it helps answer |
|---|---|---|
| Information available at cutoff | Collection time, result-availability time, source refresh and missingness. | Would a planner genuinely have had this input at the stated decision time? |
| Endpoint definition | Harvest timing, concentration basis, assay version and volume context. | Are all training labels measuring the same outcome? |
| Batch applicability | Cell line, process version, feeds, scale and campaign context. | Is the new batch represented by the evaluated cohort? |
| Planning dependency | Preparation deadline, flexible capacity and the action owner. | What decision can change because of the forecast? |
A wider forecast can be more useful than a precise-looking number
Imagine a planning review before the final culture samples are available. A point estimate suggests that the expected harvest sits close to the downstream team’s planning boundary. A dashboard that shows only that value invites a binary decision based on apparent precision. The more useful record shows the estimated range, the information cutoff and which unresolved observations dominate the uncertainty.
The scientist checks whether the newest titer result was actually available when the forecast was generated. If it arrived later, it belongs in the next forecast cycle. The planner checks whether the assumed harvest date and working volume match the active plan. A mismatch in either can explain disagreement without implying that the culture itself has changed.
Suppose the plausible range spans both sides of the planning boundary. The right action may be to retain a flexible preparation option and review again when a specified assay returns. If the range is narrow and the batch remains within the evaluated process region, the team may have a stronger basis to update the plan. If the batch is outside that region, the record should say that the forecast is unsupported and identify an alternative planning basis.
Later, evaluate whether the chosen action was sensible using the evidence available at the time. An outcome near the point estimate does not prove the decision process was sound, and a single surprising batch does not characterize overall model reliability. Keep the forecast history so reviewers can examine how estimates and uncertainty evolved without replacing earlier versions.
A staged path from the first review to repeatable use
Four delivery stages
- 01
Historical replay
Reconstruct past planning cutoffs using only inputs available at each cutoff. Compare with the planning baseline and record where source timing cannot be recovered.
- 02
Prospective shadow use
Issue forecasts alongside the existing planning process without making them the sole decision basis. Record latency, missing data, uncertainty and reviewer interpretation.
- 03
Bounded planning adoption
Agree the process versions and decisions for which the evidence supports use. Define an escalation path when the estimate is late, unstable or outside scope.
- 04
Ongoing evaluation
Review forecast behavior by horizon and campaign. Revisit applicability after material process or assay changes rather than assuming the original evaluation still holds.
What travels with the decision
Each forecast record should contain the generation time, source cutoff, target definition, model version, estimate, interval and applicability status. Add the planner’s action and the evidence requested for the next update. Scientific interpretation and production planning may belong to different people; make the receiving responsibility clear enough that the estimate cannot be mistaken for an approved harvest instruction.
Questions to settle before starting
How early should the forecast be produced?
At a time that can influence an actual preparation decision. Evaluate several agreed horizons and keep their error and uncertainty separate; later forecasts should not be used to advertise the reliability of earlier ones.
Does more historical data always improve the estimate?
Additional batches help only when their target definitions and process context are usable. Pooling incompatible versions can conceal systematic errors. Evaluate whether including a cohort improves independent performance.
Can predicted titer determine product release?
No. The proposed forecast informs planning. Release, quality assessment and harvest execution require their own evidence and responsible review. A concentration estimate cannot substitute for those decisions.
Scientific background and further reading
These references provide background for the scientific methods and review considerations discussed here. The scenario, delivery plan and decision examples are proposed BioMedAna workflows; the publications do not establish BioMedAna deployment or performance.
START WITH YOUR PROCESS QUESTION
Turn the evidence
into a reviewable next step.
Bring your program context, available evidence and the decision your team needs to make.