51
Talk 2: Impact of structural similarity on accuracy of retention time prediction in LC
Next to take the stage was Roman Szucs of Pfizer. Briefly running through their strategy in method development, the main focus was on the ‘Genetic Algorithm’ approach to compute retention ‘in silico’ using molecular descriptors. Decisions at the start of method development are straightforward: reversed-phase LC or GC? From then on, traditionally the experience and training of the separation scientist guide the process, selecting appropriate mobile phase composition: should ‘B’ comprise acetonitrile or methanol? What additives should you use? It was put forward that pH ‘selection’ was the better approach than pH optimisation. Since RPLC is a type of partition chromatography, log D is an important physico-chemical property when predicting retention. It models analytes passing between a layer of more-polar water and less-polar octanol, which simulates analytes retaining on the bonded C18/other stationary phase, and is readily predicted from the range of computational chemistry tools available for free via the internet, such as chemicalize. Using pH values on flatter areas of log D v pH plots create more- robust methods, since here the mobile phase is less sensitive to small pH changes. Their approach to optimising temperature and the gradient involved the use of software such as ACDLabs LC Simulator or Drylab.
It was mentioned that stationary phase selection is much simplified by the use of principal component analysis (PCA), to group together columns according to their observed retention of probe compounds. Using this, they screen on four suitable columns, here listed according to the bonded ligand attached to the silica support. This simple column selection offers alternative selectivity, depending on the separation requirement: C18 for less-polar analytes, phenyl for mildly-
polar analytes with the option of - interaction, or polar embedded for the option of highly aqueous mobile phase.
To model analyte retention, it was presented that the typical options are to treat it as a function of either physico-chemical, ‘Abraham coefficient’ solvation parameters, or molecular descriptors. There was an in-depth comparison between Chromgenius being used to calculate either just physico-chemical properties or using molecular descriptors to take a ‘genetic algorithm’ approach. The ultimate goal was to produce models using multiple linear regression. Genetic algorithms pretend that analyte properties such as log D, log P, number of hydrogen-bond acceptors or donors comprise a chromosome, like a piece of computer code. Each descriptor has a binary value, either 1 for on or 0 for off. The software is inputted with the set of descriptors, for example log D = 1 therefore it is used, log P is 0 therefore it is not used. Then this theoretical technique involves simulating ‘mating’ one combination of descriptors with another, to produce a theoretical offspring. This is repeated, and extra ‘mutations’ in the code are included, for example the random changing of log P from 0 to 1 for a particular ‘chromosome’. The result is a series of combinations of molecular descriptors, each of which is used by the software to predict a compound’s retention time. The software then computes a predicted retention time using this set of varying molecular descriptors.
Correlation coefficient approached R=1 after 20 'generations', and the best results were achieved after 20 'generations'. Using Chromgenius to predict retention, they improved retention predictability from 50% to 80% likelihood. Reducing the number of probe compounds achieved better correlation for a given compound's prediction and observed retention. Unsurprisingly, picking probe compounds more- similar to the target compound gave better predictability. When applied to a set of steroids, the genetic algorithm approach was effective at predicting retention (r2=0.9873; 53 data points).
Talk 3: Modelling of analytical (U)HPLC: An important element in the QbD toolbox
Patrik Petersson from Noro Nordisk presented advances in quality by design (QbD) in (U)HPLC of peptides and proteins. The term QbD was coined by Juran in the 1980's, and this has been applied to
manufacturing processes as part of the six sigma (6σ) frameworks to standardise products and processes. Applying this approach to analytical methods is dictated by specification of the user requirement and internal targets, with only limited guidance from regulators.
The content of their analytical methods is comprised of a choice of 14 different methods, and their target profile is to focus on intermediate precision, product specifications and avoiding out of specification results. Risk-assessment was method-specific, featuring a classical risk matrix. Thus a method is allocated risk according to the severity of a hazard, probability it could happen and the likelihood of detecting it.
Method control employed system suitability tests (SST), defining acceptance limits to determine if data is valid. Rather than defining a method design space, the SST could justify adjustments to a method, such as a change in %B to give more-robust methods.
Again, retention time prediction is a key part of their method understanding process. They start with a generic screening, following up with mobile phase optimisation. To fine-tune a method, they used a semi-theoretical approach, using ACD/LC software to map gradient composition and temperature. They found proteins respond much more strongly to changes in the percentage of acetontrile in mobile phase than small molecules, therefore solvent strength models were employed.
When optimising temperature, they found van’t Hoff plots are somewhat different for proteins compared to small molecules, due to 3-dimensional folding. Plotting the natural logorithm of retention factor (lnk) vs the inverse of temperature (1/T), the plot can be hillock-
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42 |
Page 43 |
Page 44 |
Page 45 |
Page 46 |
Page 47 |
Page 48 |
Page 49 |
Page 50 |
Page 51 |
Page 52 |
Page 53 |
Page 54 |
Page 55 |
Page 56 |
Page 57 |
Page 58 |
Page 59 |
Page 60
