A method transfer succeeds when the receiving laboratory, running the written procedure on its own instruments with its own reagents and its own analysts, returns results the sending laboratory would recognise as its own. That is a measurement claim about two laboratories, and it needs what any measurement claim needs: a design fixed before the data exist, a criterion derived from the method rather than from the result, and an explicit statement of what the outcome covers. Most failed transfers were mis-specified rather than mis-run.
Four designs, four different claims
Comparative testing is the default: both units analyse aliquots of common lots to a protocol fixed in advance, and the two data sets are compared against limits set beforehand. The inter-laboratory study is the same design scaled up. Dispas and co-workers ran a supercritical fluid chromatography impurity method across 19 laboratories on four continents and in nine countries, spanning academic, instrument-manufacturer, pharmaceutical and food-industry sites, and estimated method reproducibility by decomposing the variance into replicate, between-day and between-laboratory components, finding repeatability and reproducibility variances similar to or better than those described in the literature for liquid chromatographic impurity determination [1]. The structural lesson carries even though the analytes were small-molecule impurities and the technique was SFC: reproducibility is estimated from a variance decomposition, and the between-day term sits between within-run repeatability and between-laboratory reproducibility. A comparative test run on a single day at each site cannot separate the two, and will under-state the spread the receiving laboratory has to live with.
Co-validation makes the receiving unit a participant in the validation itself, so inter-laboratory precision falls out of the validation data instead of being measured afterwards. Abbatiello and colleagues took this to scale: 11 laboratories using 14 LC-MS systems each developed, determined analytical figures of merit for, and applied highly multiplexed MRM assays targeting 125 peptides derived from 34 proteins in human plasma depleted of 14 abundant proteins, with median intra- and inter-laboratory reproducibility below 20% [2]. Co-validation buys a reproducibility estimate you would otherwise generate twice; it costs the ability to change the method late, since every change re-opens every site.
Revalidation, full or partial, repeats at the receiving site those validation characteristics the transfer could plausibly disturb. Deciding which ones is a judgement about the method's physics, not a checklist exercise: for a gradient reversed-phase peptide method the characteristics at risk are those riding on instrument geometry and on the packing lot, treated below, while identity confirmation by accurate mass generally is not.
A transfer waiver — running the procedure at the receiving site without generating comparative data at all — is defensible only when standing evidence already shows that the receiving unit and its instruments perform this class of separation to the required precision. That evidence is instrument qualification, and for a gradient method it has to be concrete. Legelli and co-workers characterised three supercritical fluid chromatographs from different manufacturers before attempting any transfer — gradient tests for gradient accuracy and precision, each system's dwell volume and pump performance, and injection reproducibility — then transferred a separation of vitamin A acetate configurational isomers onto all three [3]. Because small changes in instrument set-up alter solvent density in SFC and thereby selectivity, they treat SFC transfer as less trivial than transfer in liquid chromatography, so the qualification effort they describe is sized for the harder of the two techniques.
The acceptance criterion comes from the method, and it comes first
A criterion chosen after seeing the comparative data is not a criterion; it is a description. It has to be derived, before the experiment runs, from the precision the method demonstrated during its own validation, and the design has to be sized so that a method genuinely meeting that precision will pass.
Kaminski, Schepers and Wätzig worked through that arithmetic for transfers based on equivalence tests. Their starting point is the fixed transfer design in the ISPE Guide for Technology Transfer, and their finding is that a fixed design can easily bring about high beta-errors, meaning rejection of successful transfers, or high workload from employing many analysts, once the error due to different analysts exceeds 0.6% [4]. Two consequences follow. First, the analyst variance term is not a nuisance to be averaged away; it decides how many analysts and how many series the transfer needs, and the same work supplies decision guidance for choosing that number at a given percentage acceptance limit. Second, the criterion and the design are one object: fixing an acceptance limit without fixing the number of analysts and series determines nothing about whether a good method will pass. The same analysis shows that acceptance criteria previously derived for t-test-based transfers can be scaled by a factor of 1.15, giving broader tolerance without a loss of decision certainty. This is a statistical result about transfer designs for assay methods in general, not one measured on a particular analyte.
Equivalence is not the absence of a significant difference
The commonest statistical error in a transfer report is to run a two-sample t-test on the two data sets, find p above 0.05, and conclude the laboratories agree. The authors above are blunt about it: using the classic two-sample t-test can lead to misjudgments and unsatisfying transfer results because of the test's characteristics, which is the reason ISPE adopted equivalence tests in the first place [4].
The reason is structural. A significance test places "the two means are identical" in the null hypothesis, so failing to reject it is a failure to find evidence, not evidence of sameness. An imprecise comparison, too few replicates or a noisy receiving laboratory all make that failure more likely, and the test rewards them with a pass. Push the other way and it flips: run enough replicates and a difference far smaller than anything analytically meaningful becomes significant, so the more careful the comparison, the likelier a good transfer fails.
An equivalence test inverts the hypotheses. The null is that the difference between sites is at least as large as a pre-set limit, and evidence is required to reject it, so the burden falls on the party claiming sameness and precision is rewarded rather than punished. In the protocol that means stating the equivalence limit in the units of the method's own precision, reporting the confidence interval for the between-site difference rather than a p-value, and making the pass condition that the interval lies wholly inside the limit.
What moves a peptide gradient when the instrument changes
Dwell volume — the volume between the point where mobile phase is proportioned and the column inlet — is the geometric difference most likely to break a gradient transfer. Guillarme and co-workers, deriving transfer rules for gradient separations onto short columns packed with small particles, state that special attention was paid to the dwell, or gradient delay, volume, which causes the most detrimental effect for transferring a gradient method, and that they therefore decreased the dwell volume simultaneously with the column dead volume, applying this to a separation of a pharmaceutical drug and eight related impurities that under optimal conditions ran ten times faster with an acceptable loss in resolution [5]. Milenković and colleagues reach the same diagnosis from the other direction, identifying dwell volume difference between HPLC systems as one of the main reasons gradient transfers need troubleshooting with an uncertain outcome, and building the dwell volumes of the instruments intended for the method's lifecycle into the optimisation itself; screening with a Plackett-Burman design and optimising with a D-optimal design and Monte Carlo simulation, they obtained conditions for dabigatran etexilate mesylate and its nine impurities that held across three (U)HPLC instruments chosen for their large dwell volume differences [6]. Designing against the instruments you will actually use is cheaper than transferring and then troubleshooting.
Correcting for dwell volume alone is not sufficient, and the residual has been measured. Bos and co-workers modelled the gradient each instrument actually delivers as a response function rather than as a delayed copy of the programmed profile, and report that the average relative error in predicted retention times on two other instruments fell from 9.8% and 12.2% under a dwell-volume-only correction to 2.1% and 6.5% when the instrument-specific, geometry-induced deformation was modelled [7]. Two systems with the same nominal dwell volume can still deliver different gradient shapes, and the authors demonstrated the significance and impact of that deformation for fast gradients, noting that accurate retention modelling becomes challenging in particular when very fast separations are pursued.
Extra-column volume acts on a different quantity. Zhou and co-workers, comparing configurations of a 1500 bar UHPLC system with 2.1 mm i.d. by 100 mm columns packed with 1.5 µm core-shell particles, found that changing the internal diameter of connection tubing and flow cells affects elution time, dispersion characteristics and the kinetic performance limits of the system; decreasing tubing i.d. from 100 to 75 µm gained separation efficiency, and connection tubing of smaller internal diameter reduced residence time through its reduced fluidic volume, while retention factors rose through a pressure-induced retention effect that was significant for late-eluting analytes [8]. Peak position and peak width move for different reasons, so a transfer check that looks only at retention will miss the resolution loss extra-column dispersion inflicts on early, narrow peaks.
Where the receiving instrument genuinely differs, the new conditions can be calculated rather than guessed. Kormány and colleagues transferred an optimised gradient method onto columns of different length, internal diameter and particle size running on instruments with larger dwell volumes, using retention modelling software to compute the new gradient programme, and obtained average retention time errors between predicted and observed chromatograms of −0.11% and 0.45% for the transfers to 100 mm by 3 mm, 2.6 µm and 150 mm by 4.6 mm, 5 µm columns respectively [9]. The test system was loratadine and seven related pharmacopeial impurities, so those figures belong to a small-molecule separation; what generalises is that a modelled transfer yields a predicted chromatogram you can falsify, and a rule-of-thumb transfer does not.
For peptide work there is a complementary route: stop comparing absolute retention. Vialas and co-workers evaluated dimensionless relative retention times, calculated against retention time standard peptides, in a multicentric setting spanning 28 laboratories, and found them a useful tool for transferring and sharing peptide retention times across different chromatographic set-ups both within and between laboratories, with very low variability over long time periods [10]. A normalised retention scale turns an instrument-geometry problem into a calibration problem. It does not make two gradients identical, and it will not rescue a method whose resolution of a critical pair depends on gradient shape.
When a transfer fails: column lot, reagent, technique
Column-to-column variation within a packing batch and batch-to-batch variation between batches are different quantities, separated experimentally long ago. Kele and Guiochon measured both on Vydac 218TP C18, a polymeric wide-pore silica-based phase of the type used for peptide separations, comparing five columns packed from the same batch against six columns packed from six different batches, under five sets of chromatographic conditions, tracking retention time, retention and separation factors, hydrophobic and steric selectivity, efficiency and tailing factor [11]. The probe set was 30 neutral, acidic and basic test compounds, so the magnitudes are not peptide magnitudes; the design point is what transfers. If the receiving laboratory opens a column from a different batch, batch is a variable in the transfer, and the only way to know whether it matters is to have run the outgoing method on both.
Mobile-phase additive is the second usual suspect, and for peptides it is not a small offset. Shibue, Mant and Hodges examined phosphoric acid, trifluoroacetic acid, pentafluoropropionic acid and heptafluorobutyric acid from 1 to 60 mM against two mixtures of 18-residue synthetic peptides, one of uniform +4 charge and one spanning +1 to +4. Peptides of the same charge were resolved very similarly whatever reagent was used, with overall retention increasing with the hydrophobicity of the anion in the order phosphate, TFA−, PFPA−, HFBA−; but peptides of differing charge moved at differing rates relative to each other depending on reagent concentration, and above 10 mM HFBA the elution order of the four-peptide mixture reversed relative to the order the other three reagents produced across the whole range [12]. Additive identity and concentration therefore set peptide selectivity, not merely retention. A method that says "0.1% acid" is under-specified when two laboratories prepare stock by different routes or dispense by different means — and 0.1% (v/v) TFA sits at roughly 13 mM, inside the range those authors scanned.
Analyst technique is the third, and it is more concrete than the phrase suggests. Bongaerts and co-workers, quantifying the neuropeptides neuromedin U-8 and neuromedin S by LC-MS/MS, found non-specific binding dominating the analysis and traced it to the individual sample preparation steps, namely the solvents applied and the pipetting protocol, with addition of 0.05% plasma as an adsorption competitor described as primordial to avoid peptide loss. In the same work, column temperature was optimal at 35 °C for NmU-8 and 45 °C for NmS, with higher temperatures substantially decreasing sensitivity, and starting the gradient at 20% rather than 5% organic modifier significantly improved peak shape for both [13]. Every one of those is a detail a written method can omit and two analysts can perform differently. The authors also highlight the difficulties faced when quantifying larger neuropeptides of 23 to 36 residues compared with smaller ones under 15 residues — a warning against extrapolating a transfer across peptide size.
What the receiving laboratory needs in writing
The transfer package is the set of things that would otherwise be reconstructed by guesswork. Beyond the procedure: the system suitability test with its acceptance limits, and the instruction to treat it as a live check rather than a start-of-day formality. In the 11-laboratory peptide study above, a system suitability protocol built into the experimental design enabled real-time monitoring of LC-MRM-MS performance during assay development and implementation, allowing early detection and correction of chromatographic and instrumental problems [2]. Alongside it: the exact column part number and the batch used for the sending laboratory's data; additive identity, concentration and preparation route; solvent and labware specification wherever adsorption is in play; the receiving instrument's measured dwell and extra-column volumes rather than nominal figures; the re-equilibration volume between injections; and the pre-specified equivalence limit with the design that supports it.
Common material is what makes a comparison a comparison. For peptides that can be a value-assigned reference material rather than a shared aliquot of a working stock. Deng and co-workers characterised natural and isotope-labelled angiotensin I certified reference materials, assigning content by mass balance and quantitative NMR — with moisture by Karl Fischer titration, anions by ion chromatography and inorganic elements by ICP-MS feeding the balance — verifying the assignment by isotope dilution mass spectrometry, and naming support for method validation, quality control and inter-laboratory comparison as an explicit aim [14]. A shared, value-assigned material converts "our numbers differ" into "one of us is biased, and by this much".
What a passed transfer licenses you to say
A successful transfer licenses a narrow claim: this procedure, on these analytes, in this matrix, over this concentration range, gives results at the receiving site equivalent within the pre-set limit to results at the sending site. It does not extend to a different matrix, a longer peptide, a lower concentration or a different column chemistry, and says nothing about a third laboratory.
The multi-site literature is careful about that scope. Addona and colleagues showed, using common materials and standardised protocols, that multiplexed MRM-based assays can be highly reproducible within and across laboratories and instrument platforms and are sensitive to low microgram-per-millilitre protein concentrations in unfractionated plasma, and presented the output as data and benchmarks against which individual laboratories can compare their own performance [15]. Collins and co-workers, evaluating SWATH mass spectrometry across 11 sites worldwide, consistently detected and reproducibly quantified more than 4000 proteins from HEK293 cell lysate, and used synthetic peptide dilution series to show that the sensitivity, dynamic range and reproducibility established by the method were uniformly achieved [16]. In each case the demonstrated object is a specified workflow executed to a common protocol on common material, not a property of the technique in the abstract.
Which points at the corollary worth designing for. When a transfer fails, the informative output is not the pass or fail flag but the variance decomposition behind it — the argument for building enough analysts, days and column batches into the design that a failure can be attributed rather than merely recorded.