System Suitability Testing: What a Passing Run Proves

A chromatogram carries no internal record of whether the instrument that produced it was working. Peaks appear at plausible times and integrate to plausible areas whether or not the pump delivered the gradient it was told to, the column bed has voided, or the injector deposited the volume it reported. System suitability is the set of measurements you interpose between acquisition and interpretation to close that gap — a test to verify the adequate working of the equipment used for analytical measurements, run at least at the beginning of a series of routine analyses ^[1]. Its scope is narrower than its reputation: it speaks to the equipment, on the day, on a defined standard.

What each parameter detects when it fails

Resolution

Resolution between the critical pair is the parameter most methods lean on, because it is the parameter that most directly measures whether the separation the method exists to perform is still happening. It is also a composite: the standard expression makes resolution the product of an efficiency term, a selectivity term and a retention term, so a fall in resolution reports that something moved without saying which of the three. That ambiguity is why the number works as a gate and badly as a diagnosis.

Declerck and co-workers defined a system suitability limit for resolution from an intermediate, time-different precision study on four immobilised polysaccharide columns that differed only in manufacturer or particle size, then used it both to decide whether a column had deteriorated and to judge whether a transfer to another column carrying the same chiral selector had succeeded ^[2]. The construction matters more than the setting: in supercritical fluid chromatography on chiral phases, the limit came from the observed variability of the separation itself rather than from a round number.

Tailing and asymmetry

A tailing or asymmetry factor is a single ratio taken across one horizontal slice of a peak, and single ratios hide structure. Wahab, Patel and Armstrong note that existing tests for symmetry indicate skewness in excess — a tail or a front — but not the presence of both, so concurrent fronting and tailing can partly cancel and return a compliant number ^[3]. Their total peak shape analysis extracts peak parameters from the upper region of the peak, above roughly 80% of height, rather than at half height as is done conventionally, and adds a derivative test and a constrained Gaussian fit whose residual pattern shows which region of the peak is misbehaving. They name the situations in which knowing the nature of the asymmetry helps in problem solving: column overloading, slurry packing, buffer mismatch and extra-column band broadening ^[3]. A tailing factor inside its limit therefore licenses one conclusion, that the excess skew is small, and not the broader one that the peak is Gaussian.

Theoretical plates

Plate count is the slow parameter, and it fails in a way threshold tests are poorly shaped to catch. Wu and colleagues tracked system suitability parameters across 19 size-exclusion columns using an in-house monoclonal antibody; a general linear model showed elution time, peak width, asymmetry factor and tailing factor all rising with injection number while peak height, resolution and plate count fell, and after 1000 injections tailing factor and peak width had increased by more than 10% while resolution and plate count had decreased by more than 10% from their starting values ^[4]. That is a drift, not an event: a pass/fail comparison against one limit detects it late by construction, whereas the value recorded at every check contains the trajectory for free.

Repeatability of replicate injections

The most generally applied suitability test considers the precision of the analysis: the repeatability standard deviation across replicate injections must not exceed a predefined value ^[1]. What replicate injections isolate is everything downstream of sample preparation that varies shot to shot — metered volume, valve seal, ionisation stability, detector response — because nothing about the vial changes between them. They say nothing about anything upstream of that vial.

The multisite LC–SID–MRM–MS protocol of Abbatiello and co-workers shows how finely this can be resolved for peptide analytes. Built on a predigested six-protein mixture and evaluated in 11 laboratories on 15 instruments, it monitored peak width, chromatographic resolution, peak capacity, peak area variability and retention time stability, with pass/fail criteria set at peak area coefficient of variation below 0.15, peak width coefficient of variation below 0.15, retention time standard deviation below 0.15 min (9 s) and retention time drift below 0.5 min (30 s) ^[5]. Retention scatter and retention drift carry separate limits there, and that separation is the point: shot-to-shot variation and monotonic movement across a sequence have different origins, and one standard deviation over a whole run blurs them.

Signal-to-noise for a limit test

Signal-to-noise is the one suitability parameter that speaks to detectability rather than separation, and it is the one that matters when the claim you are about to make is that something is absent. Coleman and co-workers set out its awkwardness plainly: because the conventional ratio compares the height of the signal to the height of the noise, it is difficult to relate to the peak area measurements typically used to quantify samples, and peak area at very low concentrations can be more sensitive to all components of noise because of peak shape ^[6]. Their alternative is to take the ratio of area signal to baseline noise for replicate injections of a sample prepared at the limit of quantitation during validation, and to derive the lower limit for that ratio from statistical tolerance intervals, so the criterion inherits a stated coverage probability rather than a convention ^[6].

Where acceptance criteria come from

There are two defensible sources for a limit and one indefensible one. The first is a robustness study. Hund and colleagues examined the strategy of deducing suitability limits from the responses observed when method factors are deliberately varied, and tested how far it carries for complex samples of microbial origin ^[1]. The logic is direct: if you know how far resolution moves when composition or column temperature moves by the amount your laboratory can hold, you know what a passing resolution has bought you.

The second source is your own precision data. Ermer and Ploss argue that establishing an acceptable analytical variability is of central importance precisely because many other acceptance criteria can be derived from it, and that acceptable precision ranges can be assembled from validation, transfer exercises, control charts and routine applications such as batch release or stability studies ^[7]. A reliable target variability is also what lets you recognise atypical behaviour when it appears, and maintaining validated status through long-term routine use means monitoring performance parameters — system suitability tests, control charts — rather than treating validation as a finished event ^[7].

The indefensible source is a number lifted from a monograph for an unrelated analyte because it looked standard. A limit derived from your own distribution has a false-failure rate you can compute. An inherited limit has one you cannot.

Why compendial criteria may not fit a peptide method

McCalley's comparison of a purely polymeric column with a Type B silica C18 for basic peptides makes the mismatch concrete. Good peak shapes were obtained for small masses of bradykinin and some analogues, 0.1 µg or less, on both columns with acidic mobile phases, and both showed a similar and serious deterioration in peak shape as sample mass rose ^[8]. The magnitude depends on the additive: peak capacity for a 2.5 µg load of one bradykinin on the polymeric column fell to 0.38 times its 0.1 µg value with 0.02 M formic acid, against 0.77 times with the same concentration of trifluoroacetic acid, which McCalley attributes to the ion-pair effect of TFA and its higher ionic strength reducing mutual repulsion between charged peptides held on the hydrophobic surface ^[8].

Two consequences follow for suitability design. A tailing factor or plate count measured on a dilute standard does not predict the peak shape of a concentrated sample injection, so the suitability injection has to sit at a load resembling the one the method will actually run. And the reflex diagnosis of peptide tailing as silanol activity does not survive this data set: the general similarity in behaviour between the polymeric and the silica column led McCalley to suggest that silanol groups have little involvement in the retention and overload behaviour of these peptides on highly inert Type B silica phases ^[8]. A monograph limit written around a small molecule whose tailing mechanism is silanol interaction has no authority over a peptide method where the mechanism is something else.

The neighbouring error is trusting a metric because it is customary. Zhou, Gucinski and Boyne spiked a bovine serum albumin digest standard with synthetic peptides spanning 0.1% to 100% of the BSA digest peptide concentration, and found that BSA sequence coverage — a commonly used indicator of instrument performance — did not effectively identify settings that led to limited dynamic range or poorer absolute mass accuracy on two separate LC–MS systems; metrics focused on detection limit and on sensitivity for peptide identification were the ones needed instead ^[9]. A metric earns its place by being shown responsive to the failure you care about.

The mixture you inject is part of the test

Mutton and co-workers built a seven-component small-molecule mix for generic reversed-phase LC–MS because many test mixtures are formulated to characterise column quality at neutral pH, whereas theirs was designed to permit an overall suitability assessment of the whole system — column performance plus aspects of the individual instrumental components — and to alert the user when any part, instrument or column, degrades overall performance ^[10].

How far that can be pushed is visible in the 46 system performance metrics Rudnick and co-workers defined for LC–MS/MS, covering chromatographic performance, electrospray source stability, MS1 and MS2 signals, dynamic sampling of ions for MS/MS, and peptide identification. Applied to replicate analyses, the metrics responded consistently to controlled perturbations and typically varied by less than 10%, which is what allows them to reveal even subtle differences in the performance of system components; across interlaboratory studies run under a common standard operating procedure they identified outlier data and gave clues to specific causes ^[11]. Bereman's survey of the tools that implement such monitoring separates two families: peptide identification metrics — numbers of proteins, peptides or peptide spectral matches from a complex mixture — and identification-free metrics such as retention time reproducibility, full width half maximum, ion injection times and integrated peptide intensities, with implementation reducing to three choices: which metrics, what standard, and how often ^[12].

The design rule this implies: a probe should occupy the same retention, loading and ionisation regime as the analyte, because a parameter measured outside that regime characterises the instrument in a condition your sample never experiences.

Bracketing, and why a control sample is a different instrument

A suitability standard is a defined mixture injected from a vial that never went through sample preparation; a control sample is carried through that preparation alongside the samples. They measure different things and neither substitutes for the other, which is easiest seen from a case where the two disagreed.

Briscoe, Stiles and Hage describe a generic suitability test integrated throughout a bioanalytical LC/MS/MS run — tests for signal stability, carryover and instrument response — applied across more than 25,000 sample batches. In two case studies, quality control samples and standards met every acceptance criterion and the batch nonetheless failed the suitability test and was rejected; on reanalysis, the calculated concentrations of more than 35% of the samples in those batches changed by more than 15% ^[13]. Their conclusion: poor performance of the system can adversely affect the calculated concentrations of unknown samples even when the quality control results appear acceptable.

Tsantilas and co-workers describe the complementary architecture. System suitability samples are measured repeatedly and longitudinally with targeted methods — on three instrument platforms, used to identify severe system failures and to track function over months to years — while internal controls incorporated at the protein and peptide levels assess sample preparation and, critically, differentiate system failures from sample-specific issues, and external controls prepared alongside the experimental samples verify consistency and quantitative potential during batch correction and normalisation ^[14]. Three layers answer three different questions; collapsing them leaves you able to say something is wrong and unable to say where.

Bracketing follows from the same logic. An opening-only check cannot see anything that develops during the sequence, which is precisely the class of failure the separate retention-drift criterion above was written to catch, and the shape of the injection-number dependence measured on the ageing columns. Suitability injections at the start, at intervals and at the end convert a single verdict into an interval over which it holds; without the closing bracket, the last sample in a long sequence is interpreted on a measurement made hours earlier.

What a passing test does not establish

It does not establish that the method is selective. Suitability shows that the equipment can deliver the separation the method describes; whether that separation resolves the impurities actually present in your material is a property of the method, demonstrated with orthogonal evidence. Demond and co-workers built two HPLC methods for pramlintide, a 37-amino-acid peptide — one reversed-phase, one strong cation exchange — which differ in separation mode and therefore give orthogonal information, with a limit of quantitation for degradation products of approximately 0.1% relative to intact peptide. Both separated the peptide from D-isomers at Ala5, Ala8, Ala5-Ala8 and Leu12; neither resolved it from diastereomers with D-isomers near the C- and N-termini ^[15]. A resolution criterion on the peaks a method does separate says nothing about the pair it does not.

It does not establish that the sample behaved. Suitability and sample injections differ in matrix and usually in load, and the peptide overload results above bound how far peak shape can move between the two with nothing about the instrument having changed.

It does not establish performance across the working range. A suitability check normally runs at one or two concentrations, so detectability at the low end is a separate question — which is why the tolerance-interval construction above sits at the limit of quantitation.

It does not by itself replace instrument qualification, though the relationship runs in a less obvious direction than usually given. Kaminski and colleagues argue that periodic performance qualification carried out under standard reference test conditions provides only a snapshot of system performance, and that system suitability data together with performance data from daily use can supply a continuous performance qualification instead; they compiled a list of twelve critical parameters, judged some — injector accuracy, thermostatting accuracy — redundant, and incorporated others into the proposed approach ^[16]. Suitability data can carry qualification weight, in other words, but only where someone has done that mapping deliberately.

Finally, it does not validate the integration. Resolution, plate count and tailing factor are all computed from the baseline the software drew, so a misplaced baseline propagates into all three and can return a mutually consistent, criterion-compliant set of numbers describing a peak that was never there. Suitability parameters constrain the integration far less than the integration constrains them.

Reading a failure

Because the parameters overlap in what they sense, the pattern of what moved is more informative than any single breach. The joint movement of those parameters recorded in the column-ageing study above is what a bed wearing with use looks like across a suitability record. Area repeatability degrading while retention holds implicates what replicate injections isolate by construction, downstream of the separation. Retention moving while efficiency and peak area hold implicates whatever sets retention rather than the bed that generates the plates. None of these readings survives storing only a verdict; they need the numeric value at every check — the practical argument for the monitoring and control-charting habit that keeps a validated procedure validated ^[7].

One discipline is worth stating, because it is where suitability testing is most often quietly defeated. A failed check invalidates the interpretation of the run; it does not invalidate the sample, and it does not license repeating the injection until a passing result appears. A test re-run without an assignable cause until it passes has stopped being a measurement.