303 Warning Letters in One Year. The Cleaning Limit FDA Stopped Accepting.

On January 21, 2026, the FDA issued Warning Letter 718093 to Signature Formulations, LLC, a contract manufacturer in Phoenix, Arizona. The charge had nothing to do with a mislabeled bottle or a missing signature. It was cross-contamination: the firm made drug products, some containing potent active ingredients, on shared equipment without separation adequate to keep one product out of the next.

The same problem had surfaced six weeks earlier as a recall. On December 8, 2025, Signature Formulations voluntarily pulled ORL Kids bubblegum toothpaste and bubblegum mouthwash off the market over potential contamination with exogenous hormones. Read that again: a children's oral-care product, carrying hormone residue, because the equipment that made it had also made something else. That is what a cleaning-validation failure looks like when it reaches a patient.

The letter puts a blunt question to the firm. Will it stop manufacturing drugs on shared equipment, and if not, what controls will it put in place to prevent cross-contamination? If your facility runs more than one product through the same train, the question is pointed at you too, and "we clean to one-thousandth of a dose" is not the answer the agency wants. It wants a number you can derive, defend, and trace back to a study.

The first two numbers give you the volume and the trajectory. Enforcement did not drift upward; it jumped by nearly six in ten. The third number shows what that volume looks like on the floor: a named firm, a named product, a named contaminant, and a cleaning-validation program that did not hold.

Source: FDA Warning Letter to Signature Formulations, LLC (WL 718093, 01/21/2026); FY2025 vs FY2024 totals via Pharmaceutical Online, "Trends In FDA FY 2025 Warning Letters".

Cross-contamination is the recurring sin

Read enough FY2025 warning letters and a pattern hardens. The most-cited regulation in inspection-based letters last year was 21 CFR 211.22, the rule governing quality-unit responsibilities, cited 62 times. Right behind it sat the written-procedure and process-validation requirements under 211.100(a), cited 51 times. Of the 303 letters, 135 were inspection-based: the kind that follows an investigator walking your floor and watching your changeover.

Shared-facility cross-contamination sits at the intersection of those citations. You cannot prove the cleaning worked, which is a process-validation failure. Someone signed off on a limit that does not defend itself, which is a quality-unit failure. And when residue reaches product, you have a contamination event on top of both. Signature Formulations is the visceral 2026 example, and it is not alone. In February 2026 the FDA cited Bio-Medical Pharmaceutical Manufacturing Corporation in Houston for aseptic-process and site-design failures on sterile products. Different facility, different modality, same theme: when products share space and equipment, the agency wants to see that you have engineered the contamination risk down to a defensible ceiling.

That ceiling has a name. It is the Permitted Daily Exposure.

The limit FDA stopped accepting on its own

For decades, cleaning validation rested on two rules of thumb. The first is the 1/1000th-of-dose rule: set the cleaning limit so that the amount of a previous product carried over into the next stays below one-thousandth of that contaminant's smallest therapeutic dose. Easy to compute, usually conservative, and entirely arbitrary. The factor of 1000 is a convention, not a toxicological finding, and it never asks what the substance actually does at low exposure. The second is the 10 ppm rule: allow no more than 10 parts per million of the previous product in the next. Also easy, also concentration-based, also disconnected from the biology of the molecule.

Each rule attaches a number to a guess. Neither answers the question an inspector now asks first: what daily dose of this specific substance can a patient absorb, every day, for a lifetime, without harm?

A health-based exposure limit answers it, expressed as a Permitted Daily Exposure (PDE), and that limit has been the inspector's baseline expectation for years. The EMA guideline that codifies it, EMA/CHMP/CVMP/SWP/169430/2012, was adopted on November 20, 2014 and came into operation on June 1, 2015. PIC/S built an aide-memoire around inspecting these limits (PI 052-1), and the MHRA has published its own Q&A and inspectorate guidance on the same expectation. The consensus is global and it is settled: derive the PDE, compute the carryover it permits, then compare that against the legacy 1/1000th-dose and 10 ppm figures and adopt the most stringent of the three. The PDE is the scientific anchor, and the old defaults are sanity checks against it rather than substitutes for it. Run rule one or rule two alone, with no derived PDE behind it, and your whole program rests on a number an investigator can dismiss in a single sentence.

What a real PDE actually requires

Here is why most CMC teams do not have one in-house. The EMA formula looks simple:

PDE = (NOAEL × 50 kg) / (F1 × F2 × F3 × F4 × F5)

Source: EMA/CHMP/CVMP/SWP/169430/2012, §4.1

The trouble lives in every term. The NOAEL is the No Observed Adverse Effect Level for the critical effect, which means you first have to decide which toxicological effect drives the limit at all. That decision requires reviewing the candidate effects across multiple studies, species, exposure durations, and endpoints, then justifying why one of them governs. Pick the wrong endpoint and the entire limit is wrong from the first line. If no NOAEL exists, you fall back to a LOAEL and pay for it with an adjustment factor. The 50 kg is the EMA standard adult body weight for the receptor. That term is fixed; nothing else is.

Then come the five adjustment factors, and each one has to be justified on its own.

This is precisely where inspectors find programs wanting. The MHRA has repeatedly documented two failure modes: HBELs generated by staff without toxicological training, and adjustment factors applied as one opaque lumped number rather than a transparent, re-derivable multiplication. A Qualified Person reviewing the file cannot re-derive a limit that arrives as a single divisor with no provenance.

Deriving a PDE properly is toxicology work, and a spreadsheet does not do it for you. It demands a critical-effect review across the primary literature and regulatory dossiers, a defensible point of departure, and an explicit justification for all five factors that a reviewer can follow. Most formulation and CMC teams are not staffed for that, and it is exactly the work the agency now expects to see.

No PDE, no defensible MACO

The PDE is not where the chain ends. It feeds the number that actually gates your equipment. The Maximum Allowable Carryover (MACO) is how much of a previous product may legitimately remain on shared equipment before the next product runs. In its standard form:

MACO = (PDE of previous product × minimum batch size of next product) / (maximum daily dose of next product)

Source: standard MACO derivation in cleaning validation practice

MACO then converts into a per-area residue limit in milligrams per square centimeter of shared product-contact surface, and into a ppm figure in the next batch. Those two numbers set your swab and rinse acceptance criteria, and they dictate how sensitive your analytical method has to be to prove the surface is clean.

Follow the dependency backward. Your analytical method specification depends on your residue limit, your residue limit depends on your MACO, and your MACO depends on your PDE. If the PDE is arbitrary, everything downstream of it is arbitrary too: the swab spec, the rinse limit, the method sensitivity, the pass/fail line your operators read at every changeover. The whole validation rests on the toxicological anchor at the top, and if that anchor is a rule of thumb, one observation pulls the entire structure down.

The MHRA flags a quieter failure mode that follows directly from a weak limit. Some firms stop analytical testing at changeover and rely on visual inspection alone, even when the HBEL-safe residue limit sits below the threshold the human eye can detect. A clean-looking surface proves nothing when the safe limit is invisible. Without a derived PDE and the MACO it produces, you have no way to know whether visual inspection is anywhere near adequate, and an investigator who does that arithmetic will know before you do.

The PDE Agent: the derivation, done to the framework

DeepC's PDE Agent is built to close this gap. It is one of six regulatory safety-assessment specialists, dispatched by the Safety Assessment Agent, and its single job is to derive Permitted Daily Exposure values for cross-contamination control in shared facilities, per EMA/CHMP/CVMP/SWP/169430/2012, and to carry that limit through to MACO. It maps onto the enforcement gap point by point.

It performs the core derivation to the guideline. PDE = (NOAEL × 50 kg) / (F1 × F2 × F3 × F4 × F5), exactly as EMA §4.1 specifies, with each of the five factors shown as an explicit, individually justified multiplication: species for F1, inter-individual variability for F2, duration for F3, severity for F4, NOAEL-to-LOAEL adjustment for F5. Nothing arrives as a lumped divisor, and a Qualified Person can re-derive every step.

It also selects and defends the critical effect, which is the single hardest judgment in the whole derivation. The agent tabulates the candidate effects across studies, species, durations, and endpoints, then states which one drives the limit and why. The decision that determines whether the limit is right at all shows up as reasoned work rather than an assertion. Route mismatches are handled rather than papered over: when the available NOAEL comes from a different route of exposure than the shared-facility scenario, the agent produces bioavailability-corrected PDE values instead of silently importing a number from the wrong route.

From there it carries the limit through to MACO, producing worked examples for representative next-product scenarios, expressed in milligrams per equipment area and ppm in the next-product batch. Those are the exact figures that set your swab limit and your method sensitivity, so the PDE does not sit in isolation; it lands as the cleaning numbers your validation program actually uses.

It also knows when the standard derivation does not apply, which is where teams quietly sink an inspection by running the wrong method. Highly sensitising APIs get a dedicated-facility recommendation per GMP §3.6 rather than a carryover limit. Therapeutic macromolecules and biologics are often exempt under EMA §5.3 with denaturation evidence, so the agent assesses eligibility instead of forcing a PDE. Genotoxicants without a threshold fall back to the TTC per EMEA/CHMP/QWP/251344/2006. Cytotoxic anticancer APIs get their own handling. The veterinary parallel swaps in VICH GL18(R) Appendix 3 with a 1 kg body weight. The agent routes each case to the right method.

The output is a file built to survive review: a regulatory PDF with a stamped cover page, a Klimisch-scored reliability annex (Ks 1 to Ks 4 on every primary study citation), numbered references, and an unsigned Qualified Person sign-off block left for your QP. Every primary toxicology citation carries the test species and strain, the dose levels actually tested, GLP status, vehicle, route, exposure duration, and a source-verbatim conclusion. A bare "NOAEL = X mg/kg" line does not clear QP review, and the agent does not produce one.

Underneath all of this, the NOAEL and the critical effect trace back to real, cited studies rather than model recall. The agent harvests toxicology from ECHA's IUCLID REACH study records, roughly 1.5 million experimental records filterable by Klimisch score, alongside EPA ToxValDB, ToxRefDB and IRIS, EFSA OpenFoodTox, JECFA, and primary literature from PubMed, with physicochemistry from PubChem. The point of departure that drives your limit traces back to a study an investigator can pull and read.

The requests it fields read like the work itself

• "Derive a PDE for lacosamide for a shared oral-solid-dose facility."
• "We need a MACO for switching from this API to a 100 mg/day OTC product on the same equipment."
• "Run a PDE for this monoclonal antibody: is it eligible or do we need a dedicated facility?"