The terms used in model-based bioprocess development, from hybrid modeling and transfer learning to the regulatory vocabulary of ICH and FDA guidance. Each entry opens with a standalone definition, then explains how the concept behaves in a real bioprocess.
No terms match that searchTry a shorter word or an acronym such as CQA or PAT. If a term you need is missing, tell us and we will consider it for the next update.
AI in bioprocessing
Artificial intelligence in bioprocessing covers two things that are frequently conflated: methods that learn how a process behaves from data, and interfaces that let a person direct those methods in natural language.
The distinction determines what a tool can actually do for a development program.
Methods that learn the process. Machine learning models, hybrid models combining learned components with mechanistic structure, and optimization algorithms that select experiments. These determine what can be known about a process from a given set of data: which relationships can be represented, how much data is needed to establish them, how far a prediction can be trusted beyond the tested conditions. This is where the substance sits.
Interfaces and orchestration. Large language models and the agent frameworks built on them, which interpret a request, decide which operations to run and in what order, invoke them, and present the result. These determine how much effort it takes a person to get an answer.
The second category has become the more visible one, and a claim now circulating in this market is that agent-based tooling produces better process development. That claim does not hold, and the reason is worth stating plainly.
An agent decides what to invoke. It does not change what the invoked method is capable of. An agent that orchestrates a multiple linear regression returns multiple linear regression results, obtained with less effort. If the underlying model cannot represent the nonlinearity, the time dependence or the interactions present in the process, automating the decision to use it does not recover them. Any limitation of the method survives its automation intact. Development quality is set by the method; effort is set by the interface.
This is not an argument against agents, which do something genuinely useful. Retrieving data, running an established sequence, assembling a report and handling the repetitive parts of an analysis are real savings, and removing friction is one of the ways expert methods reach people who are not modeling specialists. Implementations also differ in quality: how well a system interprets a request, selects the right operation and handles failure affects whether the analysis a user needed is the analysis that actually ran. What none of that changes is the ceiling set by the method itself.
The practical consequence is architectural. Where analytical methods are reachable through a standard protocol rather than generated ad hoc, the language interface and the method behind it become separable. The interface can be replaced or improved without changing the answer, because the answer comes from the method.
In a regulated industry that separation carries a second consequence, and it is the more important one. An agent that writes its own analysis code produces a result that is difficult to reproduce and harder to audit, because what was done exists only in that run. An agent restricted to invoking defined operations produces a recorded sequence of defined calls, which can be repeated and inspected. The reproducibility comes from those operations being defined, versioned and logged rather than from the protocol itself: a tool-invocation standard such as the Model Context Protocol is the mechanism, not the guarantee.
The question to ask of any tool in this space is therefore not whether it has an agentic layer, but what sits behind it: what class of model, how much data it requires, whether it reports its own uncertainty, and whether what it did can be reconstructed afterwards.
Batch-to-batch variability is the difference in outcome between runs performed under nominally identical conditions.
It arises from raw material lot differences, cell bank passage, small deviations in feeding and control, equipment differences, and assay variability. It is not a defect to be eliminated but a property of a biological process to be quantified, since it sets the margin a process must hold against its specifications.
Classical experimental designs estimate it through replicated center points, which supply the pure-error term against which an effect is judged real.
It also sets a ceiling on model accuracy. A model cannot predict an outcome more precisely than the process reproduces it, so prediction error comparable to natural variability is the practical limit, and error below that level indicates the model has learned noise rather than signal. Reported model performance is only interpretable against this reference.
It is likewise the irreducible part of the uncertainty in any prediction about the process. Uncertainty arising from the model shrinks as more relevant experiments are performed; batch-to-batch variability does not. Separating the two is what determines whether further experiments would help, or whether the remaining spread is simply what the process does.
How a model handles repeated runs that disagree is informative in itself. A model with no representation of uncertainty can only fit a compromise between them. A probabilistic model can recognize the discrepancy, retain the information both runs share, and widen its uncertainty where they diverge, which is a more honest representation of what is actually known.
Also known as: active learning, model-based experimental design
Bayesian optimization is a method for choosing which experiments to run next, by selecting the conditions with the highest joint probability of reaching a stated objective under the current state of knowledge.
The question at any point in development is whether uncertainty can be estimated well enough to support the decision at hand, and then how much each new experiment would contribute to improving that decision. The model is thus not the goal; the uncertainty estimate is. A model is simply the cheapest way to obtain an uncertainty estimate accurate enough to drive a choice, which is why Bayesian optimization requires a model with the properties described under decision-grade model, above all uncertainty that grows outside the tested region.
This differs from classical experimental design in what is being optimized. A factorial design is optimal for inferring the parameters of a model already assumed correct. Bayesian optimization makes no such assumption and instead targets the objective directly, which may be maximizing a performance attribute or maximizing what is learned. Selecting experiments to reduce model uncertainty rather than to reach a performance target is the case usually called active learning; it is the same machinery with information as the objective.
Three properties matter in a bioprocess setting.
Objectives are handled together. Performance and quality attributes are traded against one another rather than optimized in sequence, so treating them jointly is not a refinement but a requirement.
Experiments are proposed as a batch. This reflects a change in what fixes the number of runs. Under a classical design the count comes from the model: thirty-two runs are required because the parameter count requires them, whatever equipment is available. With automated parallel systems the binding constraint is capacity instead — a campaign has a certain number of vessels, and the question becomes what to do with them rather than how many the design demands. A method that selects conditions against an objective adapts to the capacity available; a factorial design does not.
Each experiment carries its own marginal value. Every proposed condition comes with an estimate of how much it adds to the probability of improving the objective. That supports a finer decision than whether to run another round: a user with parallel capacity can judge whether a further experiment is worth including at all, and redirect the remaining capacity to a different objective when it is not. The estimate is least reliable early in screening, when effects are still being identified.
In DataHowLab the method is not exposed directly. A user states objectives and constraints, defines the recipe structure and which process parameters may vary, chooses how many conditions to receive and a risk level, and the algorithm returns candidate recipes. Higher risk settings admit candidates further from established conditions.
Nothing in the underlying algorithms is novel; they are established in the machine learning literature. What is uncommon is the application to pharmaceutical process development, and the combination with hybrid models inside a product built for it.
Bioprocess development is the work of establishing how to manufacture a biological product at the required scale, quality and cost, spanning cell line and clone selection, medium and feed development, process parameter optimization, scale-up, and characterization to define operating ranges.
It is expensive in time and material. A mammalian cell culture run takes one to three weeks, each condition consumes materials and analytical capacity, and the number of parameters that could be varied far exceeds the number of runs available. The central constraint is therefore not whether a better process exists but how to find it within the experiments a program can afford, which is what model-based approaches address.
Clone selection is the decision, late in cell line development, of which of a small number of remaining candidate clones will be taken into first-in-human material and full process development.
It is distinct from high-throughput clone screening. Reducing hundreds or thousands of candidates to a shortlist is a statistical exercise carried out at small scale with minimal process context, and it is well served by existing methods. The decision addressed here is the one after it: five or six clones remain, each will be run in a bioreactor system, and one will be committed to. It is close to irreversible, since changing clone after this point means repeating process development and, past certain milestones, regulatory work as well.
The difficulty is that the decision is usually made at conditions nobody has optimized. A typical campaign runs the remaining clones in a 24-vessel system, which fixes the total number of runs, and how many process parameters can be varied within that budget depends on how many model parameters must be estimated. With classical statistical models the arithmetic is unforgiving: whether each clone receives its own model or a single model carries clone-by-parameter interaction terms, every parameter effect has to be estimated separately for each clone. With 24 vessels and six clones either route leaves room for roughly two process parameters, so the campaign compares clones across a narrow, near-fixed set of conditions.
This matters more than it appears. A clone compared at two parameters held near their current values is being assessed at conditions that suit whichever clone happens to perform well there. It is not being assessed on what it could do once its own process is developed, and those are different questions. The risk is not a slower campaign but a wrong selection: a clone rejected because the shared conditions did not suit it, when a process developed around it would have made it the best candidate.
The same problem arises a second way, through what is measured rather than where. A clone campaign is normally judged on endpoint values: titer and quality attributes at harvest. That summary is not sufficient to explain the result, and sometimes it inverts it. Consider a clone that grows faster and produces more through the first week, exhausts its substrate, and declines before harvest. Measured at the end it looks mediocre. Measured across the run it is a strong clone with a feeding problem, and the two are indistinguishable in the number the selection is being made on. One needs a different feed strategy; the other should be dropped. Modeling the full trajectory of every measured variable separates them, and makes inconsistent results diagnosable rather than merely surprising.
Clone identity is a categorical variable with no mechanistic description: nothing in the physics says how one clone differs from another. It is therefore exactly the kind of variable that must be learned from data, and exactly the kind where transfer learning applies. Rather than treating each clone as a separate problem or as a set of interaction terms, the model learns what the clones share and represents the differences as deviations from that common structure. Similarity between clones becomes something the model estimates rather than something the experimenter must specify.
The practical consequence is a change in what a fixed vessel budget can buy. The same 24-vessel campaign can vary four to five process parameters across six clones, with conditions distributed through the parameter space rather than placed at factorial corners, so the levels explored are not restricted to two per parameter. Instead of a ranking at a handful of shared conditions, the campaign supports estimating a favorable operating condition for each clone and comparing the clones on that basis, which is the question the decision was always meant to answer.
Closed-loop control is the automatic adjustment of process inputs based on measured outputs, so that the process corrects itself rather than following a fixed plan.
Bioprocesses already use closed-loop control for physical variables: temperature, pH and dissolved oxygen are held at setpoints by controllers acting on heating, gas flow and base addition. Extending it to biological outcomes such as growth rate, metabolite concentration or product quality is harder, because those quantities are not measured continuously and their responses to available inputs are slow, nonlinear and coupled.
Model-based control addresses this by using a process model to predict the effect of a candidate adjustment before making it, choosing the action expected to keep the process on its intended path. Where the outcome cannot be measured directly, a soft sensor supplies the estimate the controller acts on.
Two requirements distinguish this from control of a physical variable. The model must carry uncertainty, so that an action is not taken on a confident prediction the model has no basis for. And the action must respect the boundaries of the approved design space, since a control move outside it is a deviation regardless of whether it improves the outcome.
Continued process verification is the ongoing assurance during routine production that a process remains in a state of control, forming the third stage of the FDA process validation lifecycle.
It requires collecting and evaluating data from commercial batches to detect drift or unexpected variability. Process models contribute by providing a quantitative expectation against which a new batch can be compared, rather than relying on control charts of individual parameters alone. This is also where process knowledge generated in development pays off, provided it has been retained in a usable form.
Primary source: FDA Process Validation guidance, 2011
Critical process parameter (CPP)
Regulatory term
A critical process parameter is a process parameter whose variability affects a critical quality attribute and which therefore must be monitored or controlled to ensure the process produces the required quality, as defined in ICH Q8.
The regulatory test is impact-based and independent of how tightly the parameter is actually held: a parameter with a strong effect remains critical even under excellent control. An engineering assessment that credits realized control answers a different and equally useful question, which is where operational risk sits, and the two should be reported separately rather than conflated.
A critical quality attribute is a physical, chemical, biological or microbiological property that must be within a defined limit, range or distribution to ensure product quality, as defined in ICH Q8.
For a monoclonal antibody these typically include aggregation, charge variants, glycosylation profile, host cell protein and DNA content, and potency. Criticality is assigned by the attribute’s impact on safety and efficacy, not by how difficult it is to control.
Primary source: ICH Q8(R2), with ICH Q9 for the risk framework
Data contextualization is the recording of each measurement as an instance of a known kind of quantity, so that what it means is fixed when it is captured rather than inferred later by whoever opens the file.
It is not search, and it is not data cleaning. It is the step before both, and it determines whether either is possible.
Three things are fixed by it, in increasing order of consequence.
What the number refers to. Temperature can mean the value the controller was asked to hold or the value actually observed. These are different quantities with different uses, and a system that stores them in one field labeled “temperature” has lost the distinction permanently. The same applies to units, which are a property of the quantity rather than an annotation on it.
What form it is held in. An experiment recorded as a starting temperature, a final temperature and a shift time is not comparable with one recorded as a single constant value, even though both describe temperature over a run. Recorded as profiles, both become comparable. This is the point at which most cross-program reuse is lost, and it is lost silently: each record is individually correct and the set cannot be learned from.
What role it plays. Viable cell density, glucose, titer, measured temperature and pH are all quantities recorded during a run, and they enter a process model in entirely different ways. A metabolite is consumed and appears in a mass balance with an uptake rate tied to cell density. A product is produced. A physical quantity such as temperature or pH influences rates without being conserved. A quality attribute is an outcome to be met rather than a state to be tracked. Classifying a variable is therefore what attaches the right mechanistic prior knowledge to it, and it is the difference between data a model can be built from and data that merely sits in columns.
The classification needed for this is deliberately small: cell-related quantities, metabolites, physical quantities, quality attributes, and the distinction between what was set and what was observed. A comprehensive semantic standard that nobody populates is worth less than a simple scheme actually applied to every dataset.
The reason it has to happen at ingestion is that the information being fixed is not recoverable afterwards. A spreadsheet column headed “temp” cannot be resolved into a setpoint or a measurement by inspection, and a shift recorded as three numbers cannot be reconstructed into a profile without knowing what the numbers meant. Data recorded without this structure is not an asset in waiting; it is data from which the comparability has already been removed.
Storing data, visualizing it, and being able to extract knowledge from it are three different capabilities, and only the third makes data reusable. Contextualization is what makes the third possible.
Data efficiency is the amount of useful process knowledge obtained per experiment.
A data-efficient approach reaches a reliable model from the tens of runs a development program can afford, rather than the hundreds a general-purpose machine learning method would need.
This is the constraint that decides whether model-based development is possible at all. A mammalian cell culture run takes one to three weeks and consumes material and analytical capacity, so a program’s experimental budget is measured in tens of runs across a parameter space with dozens of dimensions. Any method requiring more data than that is not slow; it is unavailable.
Four mechanisms contribute, and they stack.
Less to learn. Relationships known with certainty are imposed rather than inferred. Material is conserved, and rates scale with the amount of biomass present. A model that has to discover conservation from data spends experiments on something nobody doubted. The limit is that only near-certain relationships can be safely imposed.
More from each run. A run is not one observation. Where the model predicts the trajectory of every measured variable rather than a final value, a single experiment contributes the whole time course of cell density, metabolites and product, and the model learns from how the process moved rather than from where it ended. A classical design reads one endpoint per attribute from the same run and discards the rest. This is the largest and least appreciated of the four.
Better-placed runs. Where experiments are chosen against the objective rather than by a fixed pattern, each one is placed where it contributes most. A factorial design distributes runs by a rule set before anything is known; a model-guided selection uses what the previous round established.
A non-empty starting point. A program that inherits a model from related work does not begin at zero, which is why a later program should cost fewer experiments than the first.
What the gain buys is stated more usefully as decisions than as a percentage. A screening campaign can cover four or five process parameters instead of two. A clone selection campaign can compare candidates at conditions suited to each rather than at a single shared setting. A characterization study can fit inside the runs a program already has rather than requiring a budget of its own. In each case what changes is which question the campaign can answer, not only how quickly it answers it.
Two boundaries are worth stating plainly.
Data efficiency does not remove the need for experiments. It changes which experiments are run and how many, concentrating effort on conditions that resolve real uncertainty rather than on filling a design. Confirmation at the conditions that matter is still required.
And there is a floor. No model can predict an outcome more precisely than the process reproduces it, so prediction error comparable to natural batch-to-batch variability is the practical limit, and additional experiments beyond that point buy confidence rather than accuracy. Knowing where that floor sits is what makes it possible to say when a campaign has done its work.
The size of the gain is not a general number. It depends on the process, the stage, the data already available and the campaign, and is properly reported case by case rather than as a ratio.
A decision-grade model is a process model whose output is sufficient to act on: it reports not only a predicted value but how confident it is, what drove the result, how the process behaves over time, and how likely a set of conditions is to satisfy every requirement at once.
Predictive accuracy alone does not meet this bar. A model can be accurate across the conditions already tested and still give no indication of whether it can be trusted at the condition a scientist is actually considering, which is almost always one that has not been tested. The requirements below close that gap. They are properties of the model, not of the interface in front of it.
Uncertainty that behaves correctly. A prediction without an uncertainty estimate cannot be acted on, because nothing distinguishes interpolation inside well-covered territory from extrapolation past the edge of the data. The decisive property is directional: uncertainty must increase with distance from the conditions that have been tested. Two common behaviors violate this. A model built from first principles extrapolates indefinitely, because its equations are assumed valid everywhere. A neural network may saturate far from its training data and return a confident answer in a region where it has no information. Either allows an untested condition to be mistaken for a promising one precisely because the model is confident about it.
Assumptions the model can be confident in. Prior assumptions restricted to what is genuinely well established, with anything less certain learned from data.
The whole time course, not the endpoint. Many process decisions are about timing: when to shift temperature, when to start or change a feed, when to harvest. A model relating setpoints to a final result answers none of them, and two runs reaching the same titer by different routes are indistinguishable to it.
Attribution. Knowing which inputs drove the result and by how much, including inputs acting in combination.
Probability of meeting all criteria. A point prediction of the best condition says nothing about margin.
Evidence of reliability. Goodness of fit is not evidence: a model with enough free parameters will fit almost anything. The meaningful test is prediction error on runs held out of training, measured against the natural batch-to-batch variability of the process, which is the accuracy ceiling no model can exceed in any way the process could detect.
Democratization in bioprocessing is the shift of model-based decision making from specialist data scientists to the process scientists and engineers who own the process, so that the people who will act on a result are the people who obtain it.
The term is used more loosely elsewhere in machine learning, where it usually means lowering the barrier to building models. In bioprocess development the target is different, because a process scientist does not want a model. They want to know which clone to take forward, which five experiments to run next, whether a small-scale system is a valid representation of the large-scale process, and how much margin a set of conditions has against specification. Democratization here means delivering the decision rather than the modeling capability.
The obstacle is not interface design. A model that needs hundreds of runs before it becomes reliable cannot be placed in front of a process scientist however the interface is built, because a development program does not have hundreds of runs. Data efficiency is therefore the precondition rather than a separate feature.
Democratization also fails if knowledge resets at the end of every project, because then each new program needs the specialist again. It holds only when process knowledge persists and is carried forward by transfer learning.
Democratization does not mean that an untrained person develops a process. Process expertise remains the scarce and necessary resource. What is removed is the requirement for a second, separate expertise in data science between the scientist and the answer. In current practice some expert supervision is still involved, and natural language interfaces are an emerging mechanism for reducing that dependency further.
Design of experiments is the planning of experimental conditions so that the effects of multiple parameters can be estimated from a limited number of runs rather than varying one factor at a time.
The classical implementations in bioprocessing are factorial and response-surface designs, which place runs at combinations of high and low levels so that main effects and interactions can be separated. Replicated center points supply an estimate of pure error, which is how the design quantifies batch-to-batch variability and decides whether an effect is real. The method is well established, is understood by regulators, and is the reference point against which newer approaches are judged.
What is frequently misunderstood is what such a design is optimal for. A factorial design is optimal for estimating the parameters of a model that is already assumed correct, and specifically of a linear or low-order polynomial model. If the true relationship is not of that form, the design is no longer optimal: the placement of runs is not where information is greatest, it forces experiments at conditions that are often poor with respect to the process outcome, and it requires a substantial number of runs to fit a model that may not describe the process.
Where linearity cannot be assumed, a space-filling design distributes runs more informatively for fitting a flexible model. Where the objective is to reach a target rather than to characterize a surface, Bayesian optimization chooses conditions against that objective directly.
A factorial design also fixes the number of runs in advance, and fixes it from the model rather than from the laboratory. Thirty-two runs are required because the parameter count requires them, independently of whether the experiments are run sequentially or in parallel. That was a reasonable constraint when runs were sequential. With automated parallel systems the binding constraint is capacity: a campaign has a certain number of vessels and the question becomes what to do with them, rather than how many the design demands.
None of this makes DoE obsolete. It remains the right tool when the question is whether a small set of named factors has an effect, when the response is genuinely close to linear over the range studied, and when regulatory familiarity matters. It is less suited to finding optima in nonlinear systems, to handling time-dependent behavior, and to reusing data gathered for another purpose.
The design space is the multidimensional combination of input variables and process parameters that has been demonstrated to provide assurance of quality, as defined in ICH Q8.
Operating within it is not considered a change requiring regulatory approval, which is the practical benefit: it grants operating flexibility without a filing. Establishing one requires demonstrating that quality is maintained throughout the region, not only at the conditions tested. Classical practice defines it deterministically, as the region where predicted attributes sit within specification; a probabilistic reading states the region where the probability of conformance exceeds a chosen level.
Digital bioprocessing is the use of computational models of a biological manufacturing process to guide development and operating decisions, rather than relying on experimentation and operator judgment alone.
The term covers the whole span from a model trained on historical development data to real-time monitoring of a running production batch. What unites those applications is that a validated model stands between the data and the decision, so that questions can be answered by evaluating the model rather than by running the experiment.
The approach exists because biological processes are complex, variable and expensive to test. A mammalian cell culture involves dozens of interacting parameters affecting both yield and product quality, each run takes one to three weeks, and the number of runs a development program can afford is measured in tens. Exhaustive exploration is not available at any budget.
Its practical scope includes deciding which experiments to run, identifying which parameters actually matter, establishing operating ranges and their margins, transferring a process between scales and sites, and monitoring batches while they run.
Digital process assets
Digital process assets are a development organization’s accumulated process knowledge held in a reusable form: the experimental data from every program it has run, the models trained on that data, and the process specifications derived from those models.
The contrast is with how process data is normally treated. A campaign is designed to answer one question, the data is analyzed, a report is written, and the report is what survives. The data behind it remains in a spreadsheet, an instrument export or a batch record, technically retained and practically unavailable for any question other than the one it was gathered for. Treated this way, data is a consumable: bought with experiments, spent once, discarded.
The consequence is that most of the asset already exists and is not being used. An organization that has run programs for a decade holds data from hundreds of runs across dozens of molecules, all of it paid for. Whether it constitutes an asset depends on whether it can still be interrogated, and usually it cannot — not because it was stored carelessly, but because the structure required to compare one program with another was never captured.
Three properties separate an asset from a record.
It stays queryable. Archived data answers the question it was archived for. An asset can be interrogated with questions nobody had in mind when it was generated, which is what makes data from a completed program useful to a new one.
It is derived rather than transcribed. A design space written into a document is true on the day it is written and decays from then on, since every subsequent run changes what is known and nothing updates the document. Held as an asset, the design space is not a stored region but a combination of the model, the specification set and the assurance level it is defined against, so the region is computed whenever it is asked for and reflects everything learned since. Where the regulatory record requires a fixed statement, a version is frozen explicitly for that purpose and the working version continues to develop underneath it, so the submitted record never shifts silently when a model is retrained.
It transfers. An asset that only describes the process it came from has limited value once that process is fixed. Models built on a shared structure carry what they have learned into the next molecule, the next scale and the next stage.
Across a program’s life the derived end of the asset comprises the objectives the process must satisfy, the recipe structure, the ranges explored, the criticality assessment for each parameter together with how it was produced, the design space, the operating region inside it, and the univariate ranges projected from both. These are not separately authored documents; they are answers the model and the specifications produce together, on demand, and they move as the model does.
The consequence for a development organization is that effort compounds. Where assets are retained, a new program starts with a model, a structure and a body of evidence rather than an empty dataset, and needs fewer experiments than the one before it. Where they are not, every program starts from zero regardless of how much work preceded it.
A digital twin is a computational model of a specific process that is kept aligned with that process as it runs, so that it represents the current state of a particular system rather than a generic one.
The term is applied loosely, and a three-way distinction is commonly drawn to separate what is meant, according to how data moves between the process and the model. A digital model has no automatic connection: data is transferred by hand in both directions, and the model is used offline to explore conditions, plan experiments and predict the outcome of changes that have not been made. A digital shadow receives data automatically from the process but sends nothing back, so it reflects the current state of the running process without acting on it. A digital twin exchanges data automatically in both directions, which is what allows it to support real-time monitoring and control rather than observation alone.
The distinction matters because the three place different demands on the model. A digital model is used predictively and must extrapolate credibly to conditions not yet tested, which requires uncertainty that grows outside the explored region. A shadow and a twin must instead reconcile a model prediction with incoming measurements in real time, which is a filtering problem rather than a predictive one. The same underlying model can serve all three, but the discipline applied to it differs.
What separates a digital twin from a model in general is the attachment to a specific system and the currency of that attachment. A model of a generic production process is a model. A model of tank 3 at a named site, updated with the data that tank produces, is a twin.
Feasibility analysis
Feasibility analysis estimates the probability that a given set of process conditions will satisfy every specification simultaneously, across the whole design space rather than at a single point.
The distinction from a point prediction is what makes it a decision rather than a number. “This condition gives the highest predicted titer” and “this condition has an 85% probability of meeting all criteria at once” answer different questions, and only the second accounts for margin. A condition sitting on a steep edge and a condition sitting on a plateau may have identical predicted performance and entirely different survival prospects in routine manufacturing.
The estimate combines three contributions: variability in the process and in the assays, variability in how tightly parameters are actually held, and the model’s own uncertainty about the relationship it has learned. The third is frequently omitted in classical practice, which treats the fitted relationship as known.
Mapped across the design space, feasibility identifies the region within which the probability of conformance exceeds a chosen assurance level. This is a probabilistic reading of the design space, distinct from the classical deterministic overlay in which a region qualifies if the predicted mean sits inside specification.
Feasibility is also the gate before any other development question is meaningful. Asking how much further a process can be improved presupposes that it can meet its requirements at all.
A hybrid model combines a mechanistic description of what is physically certain with machine learning for what is not, rather than choosing between the two.
The mechanistic component carries mass balances, stoichiometry and the structural relationships that hold regardless of the process: material is conserved, rates scale with the amount of biomass present. The learned component carries the kinetics, which depend on cell line, medium, and operating conditions in ways no general theory describes.
The division is chosen by confidence rather than convenience. Assumptions imposed on the model should be ones that are near-certain; anything less certain should be learned. Extensive industrial experience, across more than one hundred study cases, has pushed this boundary toward a smaller mechanistic component than academic hybrid models typically use, because a mechanistic form that is valid under the controlled conditions of a published study is frequently invalid under the conditions of an industrial process.
The practical consequence is data efficiency. A purely data-driven model must learn conservation from data, which wastes experiments on relationships already known with certainty. A purely mechanistic model must have its kinetics assumed, which requires the assumption to be right. A hybrid model reaches useful accuracy on the tens of experiments a development program can realistically afford.
Hybrid models also invert the direction of inference. A mechanistic model asserts a mechanism and tests whether data are consistent with it, and a good fit is weak evidence, since enough adjustable parameters will fit almost anything. A hybrid model does not assert a mechanism. It indicates what functional dependence a quantity such as a growth rate must have in order to explain the observations, with confidence in that shape increasing as data accumulate and concentrated where the process actually operates. This answers the same questions a mechanistic model does — whether growth rate rises with temperature, for instance — while leaving the mechanistic interpretation to the expert rather than presupposing it.
Because hybrid models predict the full trajectory of every measured variable rather than an endpoint, they support decisions about timing, which is what most process decisions are about.
In silico experimentation is the evaluation of process conditions using a computational model instead of running them in the laboratory.
Once a model is available and validated, thousands of conditions can be evaluated in the time a single physical run would take. This is used to narrow a wide parameter space before committing to experiments, to explore combinations that would be impractical to test, and to examine the consequences of a proposed change before making it.
The value depends entirely on the model, and specifically on knowing where the model can be trusted. A prediction inside the region the model has learned is usually reliable; a prediction well outside it may not be, and the model must say which is which. In silico evaluation without an accompanying uncertainty estimate produces confident answers of unknown quality.
It does not replace experiments. It changes which experiments get run, concentrating physical effort on conditions the model identifies as informative or promising, and on confirming the conclusions that matter.
A machine learning model infers relationships directly from data, without a prescribed equation describing the underlying mechanism.
Applied to bioprocesses, such a model can capture nonlinear responses and interactions among many parameters that a polynomial regression would miss, and can work with heterogeneous data including time series, spectra and categorical descriptors such as cell line or medium type.
Two constraints govern its use in development. First, flexibility is paid for in data: a model free to learn any relationship needs enough observations to distinguish the real one from the many that fit, and a development program rarely has them. Second, a purely data-driven model has no representation of what is physically necessary, so it may produce predictions that violate mass conservation or other constraints that are not in question.
A related failure mode matters for decision support specifically. Some architectures become confidently wrong far from their training data, returning a narrow uncertainty estimate in a region where they have no information. Any use of a model to select untested conditions requires the opposite behavior.
These constraints are what motivate hybrid approaches in bioprocessing, rather than any general deficiency of machine learning.
Also known as: first-principles model, ab initio model, white-box model
A mechanistic model describes a process using equations derived from established physical, chemical and biological principles: conservation of mass, reaction kinetics, transport relationships.
Its parameters carry physical meaning, such as a maximum growth rate or a half-saturation constant, which makes the model interpretable and allows knowledge from the literature to be incorporated directly. Because the equations are held to describe the underlying mechanism, the model can in principle be applied outside the conditions where it was fitted.
The limitations are practical rather than conceptual. Biological systems are incompletely understood, so any set of equations is an approximation of what is actually happening. Estimating physically meaningful parameters reliably requires experiments designed for that purpose, and those parameters are frequently correlated, meaning several different combinations fit the data equally well. And the extrapolation property depends entirely on the assumed mechanism being correct: a model whose equations are wrong will extrapolate confidently and incorrectly, because it has no representation of its own ignorance.
In bioprocessing, mechanistic modeling is most successful where the physics is genuinely well characterized, such as chromatography and filtration, and hardest in cell culture, where the governing kinetics are neither known nor stable across cell lines.
Mechanistic prior knowledge is the set of physical and biological assumptions built into a process model before any data is seen: mass balances, reaction stoichiometry, and relationships such as the scaling of consumption and production rates with cell density.
Prior knowledge is what allows a model to be built from few experiments. A model that assumes nothing must learn everything, and learning everything is expensive in runs. The question is therefore not whether to use prior knowledge but which assumptions are safe enough to impose.
The cost of a wrong assumption is not what most people expect. A purely mechanistic model with a wrong assumption loses accuracy, which is visible. A hybrid model does not: the machine-learned component absorbs the correction, so the model still fits. The price is paid in data, because more experiments are needed to correct the imposed structure, and because the learned component is hard to interpret the user may never discover that the assumption was wrong at all.
The most damaging assumptions are the ones nobody experiences as assumptions. Feeds look trivial to handle, yet in practice feed composition and volume carry real inaccuracies from buffer preparation and delivery, producing a systematic offset rather than random scatter. Basal media are assumed to start at their declared glucose and glutamine concentrations, which is frequently not the case. Without a statistical correction term these produce significant offsets, most visibly between experimental campaigns when a new medium lot is introduced.
Assumptions are also easy to impose and hard to test. Testing a mechanistic hypothesis requires data designed for that purpose, adequate reproducibility, and a mathematical comparison between candidate structures, none of which an industrial development dataset supplies. Even effects as basic as the influence of temperature or pH are typically not known mechanistically for a given process.
Two practical observations from industrial datasets. Separating growth and death into two explicit kinetic terms is textbook-correct and makes models worse when the data cannot deconvolute them, since only the net growth rate is actually informed. And adding saturation kinetics to growth and death rates has not produced measurable improvement, since such forms are simple enough to learn from data and counterproductive when only approximately right.
The deeper point is where the reliable priors sit. Mechanistic knowledge is strongest for the quantities that are already easy to reason about, and weakest exactly where decisions are hard: temperature, pH, clone identity, medium type. The variables a process scientist most needs to understand are the ones for which no dependable prior exists.
Multi-objective optimization is the simultaneous optimization of several process outcomes that cannot all be maximized at once, producing a set of best available compromises rather than a single optimum.
Bioprocesses always have competing objectives. Titer, product quality attributes such as aggregation and glycosylation, medium and feed cost, and process duration respond differently to the same parameters, and conditions that improve one frequently degrade another. Optimizing titer alone and then checking quality afterwards discovers the conflict late, when the remaining design freedom is smallest.
The output is a trade-off surface, known as a Pareto front: the set of conditions for which no objective can be improved without another getting worse. Everything below the front is dominated and should not be chosen; everything on it is a legitimate decision the process team makes according to what the program needs. Reducing that surface to a single weighted score hides the choice rather than making it, because the weights encode a judgment the tool is not positioned to make.
Constraints and objectives are not the same thing. A quality attribute that must stay within specification is a condition on the feasible region rather than a quantity to be traded, and treating it as a further objective produces trade-offs against requirements that were never negotiable.
Multivariate data analysis is a family of statistical methods for analyzing many correlated variables at once, rather than examining each in isolation.
Principal component analysis and partial least squares regression are the established techniques in bioprocessing. Both work by finding a small number of combined directions that capture most of the variation in a large set of correlated measurements, which reduces dimensionality and handles the collinearity that defeats ordinary regression.
They have a long history in this field, are familiar to regulators, and remain effective for batch comparison, fault detection and spectral calibration, where the task is to relate many measurements to an outcome and the relationships are approximately linear.
The constraint is that linearity. These methods describe the process as combinations of linear relationships, so where responses are strongly nonlinear or where the relevant behavior is the shape of a trajectory over time rather than a summary of it, the description becomes limited. They also treat a batch as a set of values rather than as a dynamic system, which makes them well suited to comparing completed runs and less suited to predicting the effect of a change that has not been made.
Pharma 4.0 is the pharmaceutical industry’s adaptation of Industry 4.0 concepts, articulated by ISPE, covering digital integration, data connectivity, automation and model-based decision making across development and manufacturing.
It is a framework and an aspiration rather than a specification, and the term is used broadly. Its practical content in bioprocessing is the combination of connected data infrastructure, process models, and the organizational changes required to act on what they produce.
Primary source: ISPE Pharma 4.0 Operating Model
Process analytical technology (PAT)
Regulatory term
Process analytical technology is a framework, set out in FDA guidance, for designing and analyzing manufacturing through timely measurement of critical quality and performance attributes, with the goal of ensuring final product quality.
In practice it covers online and inline measurement such as spectroscopy and capacitance, the models that turn those signals into meaningful quantities, and the control actions taken in response. The measurement alone is not PAT; the value is in what is done with it.
Primary source: FDA PAT guidance, 2004
Process model validation
Process model validation is the demonstration that a model is fit for the purpose it will be used for: that its predictions are accurate enough, within a stated range of conditions, to support the decisions that will rest on them.
Applied to hybrid and machine learning models, the question is often framed as though these models required a new regulatory framework. They do not. The criteria that establish fitness for purpose are the same ones that apply to any process model, and they do not depend on how the model was constructed.
The precedent is already established. Multivariate statistical models, principally partial least squares, have been used in regulated pharmaceutical manufacturing for decades and are qualified under existing expectations. Those models are purely correlative: they carry no physical structure and their internal components have no mechanistic meaning. A model that additionally enforces mass balances and stoichiometry is not a harder case to justify than one that enforces nothing. The concern attaches to the label rather than to the object.
Qualification rests on predictive performance, not on parameter interpretability. A model is qualified by its error on data withheld from training, evaluated within the conditions it will be used in, and compared against the natural batch-to-batch variability of the process. That criterion applies identically to mechanistic, statistical and hybrid models. It is also the only one that discriminates, since goodness of fit to training data can be manufactured by any model with enough free parameters. Prediction error comparable to the reproducibility of the process is the practical ceiling: a model cannot be more precise than the process it describes, in any way the process could reveal.
Robustness is where the real difference lies. Linearity buys robustness by construction: a model restricted to linear combinations cannot behave erratically, which is part of why such models are accepted, since the structure guarantees the property and nobody has to demonstrate it. A flexible model has no such guarantee and must establish the same property by evidence. That is a genuine additional burden, and the means of meeting it are standard: regularization to penalize unnecessary complexity, ensembling to expose disagreement between equally plausible fits, and validation against data the model has not seen. Each is tested rather than assumed.
A hybrid model is not at the flexible extreme. Its mechanistic component restricts the space of possible behaviors in the same way linearity does, but on physical grounds rather than by functional form, which is a stronger restriction and a more defensible one. What remains free is learned, and constrained by the means above.
This matters because interpretability of parameters is where the comparison is usually made and where it misleads. A mechanistic model is qualified partly through parameters with physical meaning and defensible bounds. A hybrid model does not offer those for its learned component, deliberately, because the point of learning that component is that no reliable prior form exists for it. What it offers instead is a quantitative statement of how confident the model is, which a fitted mechanistic model does not provide, and which is what makes the range of applicability explicit rather than assumed.
A qualified model is qualified somewhere. The range of applicability is part of the qualification, not a caveat on it. A model that reports uncertainty growing with distance from the conditions it has seen makes that boundary visible and testable, rather than leaving it to be inferred from where the experiments happened to be. A model that extrapolates without widening its uncertainty offers no such boundary, which is a weakness of the older approach rather than a strength.
Updating is the genuinely new question. A model that is retrained as more data arrives is more useful and requires an explicit control, since a regulatory commitment cannot move silently underneath a submission. The answer is separation: the working model continues to develop as knowledge accumulates, while anything submitted is frozen as a specific version, with the model, the dataset and the derived regions recorded exactly as filed. The dossier reflects what was submitted; the working model reflects what is known. Changes to the frozen version are then handled as changes, through the change control the quality system already defines.
This is a statement about methodology, not about precedent. That a modeling approach meets established criteria for fitness for purpose is a question of method and evidence, and is separate from what any regulatory authority has accepted in any particular submission.
Quality by Design is a development approach, defined in ICH Q8, in which product quality is built into a process by design rather than established by testing the finished product.
It requires identifying the quality attributes that matter to the patient, understanding how process parameters affect them, and defining an operating region and control strategy on that basis. The approach is well established in regulatory expectation; the practical obstacle has been generating enough process understanding to support it within a realistic experimental budget.
Real-time predictive monitoring is the use of a model during a running batch to estimate its current state and project how it will end.
It differs from conventional monitoring in the projection. Conventional monitoring compares current measurements against expected ranges and raises an alarm when a limit is crossed, which is a statement about now. Predictive monitoring uses the process model to propagate the current state forward, producing an estimate of the final outcome while there is still time to influence it.
This changes what an operator can act on. A deviation detected at the point of harvest is information for the investigation report. The same deviation identified on day five, with a projection showing the batch trending toward specification failure, is a decision about whether and how to intervene.
The projection must carry uncertainty, and that uncertainty naturally narrows as the batch progresses and more of it is observed. A projection presented without a confidence range invites either over-reaction to normal variability or misplaced confidence in a forecast made early with little information.
Scale-up and scale-down modeling
Scale-up and scale-down modeling is the use of process models to relate behavior at one operating scale to behavior at another.
Scale-up predicts how a process developed at small scale will behave in production. Scale-down establishes whether a small-scale system is a valid representation of the large-scale process, which is required if development and characterization work performed at small scale is to support decisions about manufacturing.
The difficulty is that scale changes the physical environment rather than the biology. Mixing time, oxygen transfer, shear and gradients in dissolved gases and pH all change with vessel geometry and volume, and cells respond to the conditions they experience rather than to the setpoints. A model in terms of setpoints alone therefore transfers poorly.
What models offer here is not free prediction across scales but rapid correction from data. A model carrying the structure common to both scales can be adjusted to large-scale behavior using a small number of large-scale runs, which is a different and more realistic claim than predicting large-scale performance without any large-scale data. Where models carry uncertainty, the same machinery quantifies how confident the transfer is.
Sensitivity analysis quantifies how much a specific process outcome changes when a specific process parameter is varied, evaluated at a stated operating point or across a stated region.
It is always defined for a named outcome and a named parameter. A parameter may strongly affect aggregation while barely affecting titer, so a single ranking of parameters without reference to an attribute is not meaningful.
Two forms are used. A local analysis measures the rate of change at one operating point, which answers how the process responds to a small deviation from where it currently sits. A global analysis measures the contribution of each parameter to the variation of the outcome across the whole region of interest, which captures nonlinearity and interactions that a local measure at a single point cannot see. The two are the same quantity computed over different window widths rather than competing methods.
Sobol indices are the standard model-agnostic way of computing the global form. They decompose the variance of the outcome into the contribution of each parameter and of each combination of parameters, without assuming anything about the model that produced it. This is related to but distinct from Shapley value analysis: Sobol indices decompose the variance of an outcome across a region, while Shapley values attribute a single prediction to its inputs.
Sensitivity is silent about whether the movement matters. A parameter can be highly sensitive and entirely unproblematic if it is held tightly and the outcome has ample margin, or weakly sensitive and critical if the margin is small. Turning sensitivity into a decision requires a specification to compare against.
A related distinction is frequently collapsed. How much a model relies on a variable in order to make its predictions reflects how that variable happened to vary in the available data. How much the outcome would respond if the variable were moved reflects the process. A parameter held constant in all historical runs will register as unimportant and may still be highly sensitive.
Shapley value analysis attributes a model’s prediction to its individual inputs, assigning each one a share of the difference between the prediction and a baseline.
The method originates in cooperative game theory, where it distributes a shared payoff among participants according to their average marginal contribution across all possible groupings. Applied to a model, the participants are the input variables and the payoff is the prediction.
Its usefulness in bioprocessing is that it works on any model, including ones with no closed-form structure to inspect, and that it handles interaction rather than assuming effects add up. A parameter that only matters in combination with another is credited accordingly rather than appearing irrelevant.
Two things it does not do. It attributes the prediction rather than establishing causation: an input the model relies on is not necessarily a lever on the process. And it is a statement about the model, so its trustworthiness is bounded by the model’s.
A soft sensor is a model that estimates a process variable that is not directly measured, using variables that are.
In bioprocessing the typical targets are viable cell density, metabolite concentrations and product titer, estimated from online signals such as capacitance, off-gas composition, spectroscopic measurements and control system data. The motivation is that the quantities most worth knowing continuously are usually the ones requiring offline sampling and laboratory analysis, which produces a handful of values per run and a delay before each.
A soft sensor closes that gap by producing a continuous estimate. This supports observing a trend as it develops rather than after the fact, and makes feedback control possible for variables that could not otherwise be controlled in real time.
Two properties determine whether a soft sensor is usable in practice. It must report the confidence of its estimate, since an estimate acted upon without knowing its reliability is a hazard rather than an aid. And it must degrade predictably when the process moves outside the conditions it was built from, rather than continuing to report confidently.
A space-filling design distributes experimental conditions as evenly as possible throughout the parameter space, rather than concentrating them at the corners and center of the region as a factorial design does.
Latin hypercube sampling is the common starting point: each parameter’s range is divided into equal intervals and the runs are arranged so that every interval is sampled once. Sampling alone is not sufficient, since a randomly generated Latin hypercube can still leave substantial correlation between parameters. Practical designs are optimized for two properties at once — coverage, so that no region of the space is left empty, and near-orthogonality between the parameters.
Orthogonality is the property a space-filling design shares with a factorial design, and it matters as much here as it does there. Where two parameters happen to vary together across the runs, their effects cannot be separated and the model will attribute to one what belongs to the other. A design that covers the space well but confounds two parameters produces spurious correlations that no amount of modeling recovers from.
The rationale for filling the space is otherwise different from that of a factorial design. A factorial design places runs where they are most informative for a linear model, which is at the extremes. A space-filling design makes no assumption about the form of the relationship and samples the interior as well, which is where a flexible model needs information in order to learn curvature and interaction. Where linearity has not been established, this is frequently the better choice.
Two practical consequences. Every run gathers information about a distinct region rather than replicating a corner, which matters when the total number of runs is small. And the resulting dataset supports later reuse for questions that were not asked when it was designed, because it covers the space rather than a specific hypothesis.
Spectroscopic model calibration is the process of relating a spectroscopic measurement to a concentration or attribute of interest, so that the instrument can be used to quantify rather than merely to observe.
Raman, near-infrared and infrared probes produce a spectrum containing contributions from every component present. Calibration establishes the relationship between the spectrum and the quantity sought, using a set of samples where that quantity has been measured by a reference method.
The practical difficulties are well known. Calibration requires reference samples spanning the range of conditions the model will encounter, and generating them is expensive. The relationship is affected by the cell line, the medium and the equipment, so a calibration built on one process may not transfer to another. And spectral responses drift as instruments age or are serviced.
Transfer learning is directly relevant here, since it addresses the transfer problem specifically: a calibration built for one process can be adapted to a new one with far fewer reference samples than building from scratch would require.
Transfer learning is the reuse of what a model has learned from one process, product or scale to accelerate modeling of another, so that a new program does not start from zero.
In bioprocessing the transfer usually runs along one of three axes: between molecules produced by related cell lines, between scales from small-scale systems to production, and between stages from early development to manufacturing. The same machinery supports all three.
Transfer requires a shared basis. Knowledge can only be carried across datasets if what is being corrected is the same underlying structure in each case, which means the mechanistic component of the model must be common to all of them and must be safe for all of them. This is a constraint on model design as much as an application of it: the larger the assumed mechanistic component, the greater the risk that it is valid for one product and not another, and the less transferable the model becomes. It is one of the strongest arguments for keeping the assumed part small.
Because the models are probabilistic, the differences between datasets can themselves be represented and inspected. Learned descriptors, such as embedding vectors, formalize what makes two products similar or different and allow them to be compared and grouped rather than merely observed to disagree.
This is the property a purely mechanistic model lacks. Where the same nominal process behaves differently at another scale or on another site, a mechanistic model can register the discrepancy but cannot carry knowledge across unless the cause is identified and expressed mechanistically. A probabilistic model can acknowledge the difference, use the similarity that remains, and adjust its predictions accordingly.
The practical effect is cumulative: a later program should require fewer experiments than the first.
Uncertainty quantification is the estimation of how much confidence a model’s prediction deserves, reported alongside the prediction itself.
Two sources are distinguished, and the distinction is practical rather than academic. Aleatoric uncertainty is the inherent variability of the process and its measurements: two runs at identical setpoints give different results, and no amount of additional data removes this. Epistemic uncertainty is the model’s own ignorance, arising from finite data, and it shrinks as more relevant experiments are performed. Only the second is reducible, so telling them apart determines whether running more experiments would help.
For decision support the behavior of the estimate matters more than its presence. Uncertainty must widen as conditions move away from those already tested. Without that property, a model may report high confidence about a region it knows nothing about, and an optimization routine will be drawn toward exactly those conditions because they appear both promising and certain.
Reported as a band around a prediction, uncertainty converts a single number into a range a process team can reason about. Combined with specification limits, it becomes a probability of conformance rather than a pass or fail verdict.
A hybrid model combines mechanistic models, based on first principles, with data-driven machine learning. It learns the full bioprocess, from the initial design through its evolution over time to the final critical quality attributes (CQAs). This gives deeper insight into how process design and product quality are connected. DataHow pioneered hybrid modeling for bioprocessing, and it is the core technology of its software, DataHowLab.
Compared with a standard Design of Experiments (DoE), hybrid models can cut experimental timelines by 40 to 70%. They also lead to processes with higher yield and better product quality. A case study with Bristol Myers Squibb (BMS) showed that hybrid models needed less data than DoE to understand CQA behavior.
DataHowLab does not need large amounts of historical data. Teams can start with a small number of experiments and let the model guide them step by step toward the best process conditions. The exact number of runs depends on the goal of the project.
Yes. DataHowLab uses transfer learning to apply insights from comparable products and from other scales. This can speed up development and reduce the number of costly large-scale runs. Use cases have shown that up to 80% fewer experiments could be needed when DataHowLab supports process development from early to late stage.
JMP and SIMCA are general-purpose tools that serve many industries and often need a data scientist to use them well. DataHowLab is built specifically for bioprocessing and uses hybrid models, which work with limited data and capture complex, nonlinear process behavior. One solution covers the full process life cycle, from clone selection and process optimization to scale-up, tech transfer, and manufacturing.
Hybrid models predict both the final outcome of a process and how the process will evolve from any point in time. This makes them useful in development and in manufacturing, for applications such as monitoring, predictive alarming, control, and root cause analysis. This is why hybrid models are a key building block of digital twins in the pharmaceutical industry.
DataHowLab is designed for bioprocess specialists, including process scientists and technicians, not only data scientists. No modeling or coding expertise is needed. As a cloud solution, it lets the whole team work on the same data and make decisions together.
Yes. DataHowLab handles both online and offline data, and combining the two is a common use case. The measurement frequency of online data can be adjusted to fit the process.
DataHow works with 14 of the 20 largest pharmaceutical companies and has worked with more than 40 companies in total. Customer names are not shared, out of respect for confidentiality. Public use cases and success stories are available on the DataHow website.
Put these methods to work on your process
Talk to a bioprocess expert about where hybrid modeling, transfer learning and model-guided experiments fit in your development program.