Digital Twins Cannot Release Your Batch

FDA's January 2025 draft guidance on 21 CFR 211.110 does something most people quoting it only half-remember. At line 136 it says "[i]nnovative technologies allow in-line, at-line, or on-line measurements in place of physical sample removal for laboratory testing". That is the permission slip everyone wanted: you do not have to pull a physical sample to satisfy the in-process testing requirement.

Then, eighty lines later, it closes the second door. "[C]ontrol strategies that rely solely on current process models would be insufficient to satisfy the requirements of § 211.110." Not "would require additional justification". Insufficient. The model is allowed to replace the laboratory, but it is not allowed to replace the measurement.

Everything sold as a digital twin for bioprocessing has to fit through that gap, and most pitches do not. If the value case rests on deleting the sample and the sensor and letting a mechanistic model carry the release decision, there is no version of it FDA has said it would accept. If it rests on a model that consumes a Raman probe, a UV trace and a mass balance and turns them into a decision faster than a QC lab could, that is squarely inside what the draft encourages.

What the January 2025 draft actually permits

The document is short: seven numbered pages, four sections, published in the Federal Register on 6 January 2025 with comments closing 7 April 2025 under docket FDA-2024-D-5374. Every page carries the header "Contains Nonbinding Recommendations. Draft — Not for Implementation." It covers human and animal drug products including biologics, but not active ingredient manufacture.

Three permissions matter. Flexibility on where: § 211.110(c) requires in-process materials to be approved or rejected by the quality control unit "at commencement or completion of significant phases", and FDA declines to define "significant phase" — you define it, FDA assesses your rationale on inspection, and in continuous manufacturing the quality unit can treat two or more unit operations as one significant phase.

Flexibility on how: the preamble to the 1978 CGMP final rule said a sampling plan "can mean both a plan for collection of physical units for testing, or it can mean a schedule by which an examination of some sort is done". FDA leans on that sentence to reach the in-line, at-line and on-line permission at line 136. And models are welcome as a component: "Process models can be a component of the overall control strategy."

What the draft does not contain is worth stating precisely, because it is quoted loosely. Searching the January 2025 draft as posted at fda.gov/media/184825/download on 30 August 2026, the terms "digital twin", "real-time release", "machine learning" and "artificial intelligence" each appear zero times; "release" appears once, in a footnote pointing at § 211.165. This is not an AI guidance and not an RTRT guidance. It is a guidance about whether your in-process control strategy is adequate — the question that decides whether an RTRT proposal is coherent at all.

The sentence that caps the pitch

The reasoning behind the finding is more useful than the finding. FDA sets out two things it has never been shown, and they are testable claims about your model, not about models in general:

"[T]o date, FDA has not been made aware of process models that demonstrate that: (1) the underlying assumptions of the process model will remain valid during routine manufacturing; and (2) the manufacturer can detect if an underlying assumption is no longer valid (e.g., a continuous mixing model that assumes uniform mixing would be unable to detect that uniform mixing is no longer occurring due to material agglomeration on the walls of the mixer)."

Condition two kills most proposals. It is not asking whether the model is accurate. It asks whether the model can tell you it has stopped being accurate. A model fitted on a well-behaved process keeps producing confident, in-specification predictions long after the physical premise underneath it has changed. Agglomeration on a mixer wall is FDA's example. In biologics the equivalent is a column.

FDA then states the constructive version, the sentence to put in a control strategy document: "Process models should incorporate process monitoring or in-process testing to maintain a state of control, facilitate model maintenance, and ensure drug product quality." Three jobs, not one. Monitoring keeps the state of control, and maintains the model, and assures the product. A design that satisfies only the third is incomplete on FDA's own reading.

What is binding today, and what is not

The most common error here is treating a draft as a rule. The instrument map as it stands on 30 August 2026:

Two entries deserve emphasis. Annex 17, in operation since 26 December 2018, is not permissive in the way people assume. Clause 3.10: "In the event that the results from RTRT fail or are trending toward failure, a RTRT approach may not be substituted by end-product testing." The EMA guideline it derives from (EMA/CHMP/QWP/811210/2009-Rev1, adopted by CHMP 15 March 2012, effective 1 October 2012) goes further: returning to end-product testing after RTRT approval requires a variation to the marketing authorisation. The comfortable mental model — put the twin in, keep the lab as a safety net, fall back when the numbers look odd — is not available in the EU. Annex 17 clause 3.3 requires a contingency procedure for sensor or equipment failure, which is not the same as a discretionary fallback when you dislike a result.

Draft Annex 22 is the other. Published for consultation on 7 July 2025 alongside a draft revision of Annex 11, consultation closed 7 October 2025, it has not been adopted. The computerised-systems text you are inspected against today is still the 2011 Annex 11. Anyone quoting an Annex 22 clause number at you as a requirement is quoting a consultation document.

A Protein A model that decays with the resin

Take the concrete case. A capture step model predicts antibody breakthrough during Protein A load, so the column can be loaded closer to its true capacity and the load phase ended on prediction rather than on a conservative fixed volume or a UV threshold. This is the archetypal bioprocess soft sensor, and ICH Q13's glossary gives it the only definition ICH offers: "A model that is used in lieu of physical measurement to estimate a variable or attribute … based on measured data".

The model's load capacity term is calibrated against dynamic binding capacity, and DBC is not a constant. It falls with cycling — ligand degradation under alkaline clean-in-place conditions, ligand leaching, pore occlusion, fouling. Feidl and colleagues (Journal of Chromatography A, 2020) needed two separate ageing parameters in a hybrid lumped kinetic model to describe it: one for capacity deterioration, one for the decline in mass transfer caused by fouling. A model with a single fixed capacity term is already the wrong shape.

Now the numbers. Ramakrishna and Rathore (Journal of Chromatography B, 2024) monitored Protein A cycling by principal component analysis of the UV data and saw a deviation at the 120th cycle and an out-of-specification around the 149th, corroborated by yield decline. Their headline finding should worry any control strategy author: significant resin ageing was detectable 20 to 25 cycles before yield decline was observable. For twenty-odd cycles the routine indicator said nothing while the physical premise of the model moved.

It gets worse for anyone relying on DBC as the guard. Pathak and colleagues (Biotechnology Progress, 2018) cycled an agarose-based Protein A resin over 50 cycles against three feeds of differing histone, protease, DNA and host cell protein content. DBC did not vary between the conditions — yet particle porosity fell in all of them, and resins cycled in HCP- and histone-rich feeds showed greater capacity loss by other measures. The single number most lifetime protocols track was the number that missed it.

Smaller effects hide even more easily. Beattie and colleagues (Applied Spectroscopy, 2023) measured a 4% to 5% decrease in DBC between cycle 6 and cycle 29 on MabSelect PrismA. A five percent drift in a capacity term, inside a load model carrying a safety margin, produces no alarm and no deviation. It produces slightly optimistic breakthrough predictions, batch after batch, in the direction that matters. And the degradation is chemically real, not statistical drift: Zhang and colleagues (Antibodies, 2026) used the multi-attribute method on four Cytiva Protein A resins to quantify ligand deamidation, isomerisation and fragmentation induced by repeated CIP cycles, despite the engineered alkaline stability of the newer chemistries.

This is FDA's condition two, in a column. The assumption that the resin binds as it did when the model was fitted degrades continuously on a timescale of tens of cycles, and the routine in-process signal lags it.

Why no change control ever fires

Here is the governance hole. Resin lifetime is validated up front: ICH Q5A(R2), adopted 1 November 2023, requires at section 6.2.6 that "[c]hromatography media/resin lifetime use should be indicated, and parameters with potential impact on viral clearance should be defined". A maximum cycle number exists, it sits in the dossier, and a counter enforces it.

The model, though, was validated as a separate object — usually inside a PAT or process validation package — against data collected at some point in resin life. Nothing connects the two. Cycle 87 is not a change. Nobody raises a change control for a column's ordinary passage through its approved life, because nothing has changed in the sense a change-control SOP recognises: no procedure, no equipment, no material, no parameter. The resin got older, which is what it was qualified to do.

The instruments that would catch this are precise, and their status differs sharply.

Binding today. Annex 11 clause 11, in operation since 30 June 2011, requires computerised systems to be periodically evaluated to confirm they remain in a valid state, taking in performance, reliability and validation status. That is the hook an EU inspector already has. It is generic, and at most sites it is annual — a cadence that cannot see a twenty-cycle blind spot.

Endorsed and expected, not binding. The ICH Quality IWG Points to Consider, dated 6 December 2011, defines a high-impact model as one where "prediction from the model is a significant indicator of quality of the product", and says lifecycle verification should include "a risk-based frequency of comparing the model's prediction with that of the reference method, triggers for model updates (e.g., due to changes in raw materials or equipment), procedures for handling model-predicted Out of Specification (OOS) results, periodic evaluations, and approaches to model recalibration". Fifteen years old, still the sharpest statement of the requirement, and a breakthrough model used to end the load phase is high impact by that definition.

Draft, not law. Draft Annex 22 clause 10.3 would require model performance to be monitored regularly to detect changes, and 10.4 would require monitoring of whether input data remain within the model's sample space, with defined drift metrics. Those two clauses name the failure directly, and neither is in force: consultation closed 7 October 2025 and no adopted text exists as of 30 August 2026.

The honest position for a QA director in 2026: the requirement to catch resin-driven model decay is real, it is inferable from ICH Q13's model maintenance obligation and Annex 11 clause 11, and it is written down nowhere binding in the form you need. You have to write it yourself.

Where a model lives in the dossier, and can it ever replace the sample?

ICH Q13 says where model information goes: material ensuring correct application of process models in the control strategy, "including contingency plan when the model is not available", in 3.2.S.2.4 or 3.2.P.3.4; models tied to release-testing analytical procedures in 3.2.S.4 or 3.2.P.5 with their contingency testing plans; and 3.2.R for "[v]alidation data for high impact process models, if used". A filed model therefore has an approved form, and material changes to it are regulatory changes, not IT changes. Q13 also states that RTRT "is not a regulatory requirement for CM implementation" — the twin does not have to carry release for continuous manufacturing to be approvable.

Can the model ever replace the sample outright? FDA's position is provisional rather than closed. The draft ends by saying FDA "anticipates that these scientific advancements can be leveraged to pursue in-process control strategies that increasingly rely on process models", and directs anyone interested to CDER's Emerging Technology Team or CBER's Advanced Technologies Team, "as early in the development process as possible". That is the mechanism by which the two conditions get demonstrated for a specific process. If you believe your model can prove its assumptions hold and can detect when they do not, the guidance names the door to knock on.

What this means in practice

Three things.

Re-scope the business case away from sample elimination. The saving in an in-line or at-line strategy comes from cycle time and QC lab load, not from removing measurement. McKinsey's smart QC work puts the digitally-enabled cost reduction at 25% to 45% for a chemical QC lab and 15% to 35% for microbiology — lab-productivity numbers, not release-decision numbers. A programme justified on deleting probes is justified on something FDA has called insufficient.

Write the decay trigger into the control strategy and tie it to the cycle counter. The artefact: a model verification frequency expressed in column cycles rather than months, an acceptance criterion on the residual between prediction and reference measurement, a rule for what happens when that residual trends, and a named owner. Ramakrishna and Rathore's 20-to-25-cycle detection lead is the design input — verification spaced further apart than the lead time of the failure is decorative. Under Annex 17 clause 3.4 any change that could affect the validated status of the process is assessed for risk to product quality, and a model retrain is such a change.

Decide who signs, before the vendor asks. Section 211.110(c) requires the quality control unit to approve or reject in-process materials during the production process. A model output that ends a load phase is an input to that decision, not a substitute for the person making it. If nobody has decided whether the quality unit or process engineering owns the model's acceptance criteria, the pilot will surface it at the worst moment.

The framing that survives contact with an inspector is narrow. The model is not the control strategy. It is a component of one, sitting on top of measurement, with a documented way of noticing when it has drifted from the process it claims to represent. Everything else is a pitch.