Quantification and its accuracy in nontarget screening

Last Tuesday Jon Sobus from EPA presented some of the findings from a recent paper Uncertainty estimation strategies for quantitative non-targeted analysis, where we were also happy to collaborate, on the BP4NTA online meeting. People from BP4NTA meetings have been reaching out with questions on ionization efficiency-based quantification. Some of these are often asked of us and I decided to share the main points with everyone!

  • Is the ionization efficiency tested with flow injection compared with chromatography (e.g. gradient or isocratic)

Yes. We have used both flow injections and chromatographic data both in training data and test data. Chromatography has been exclusively used for the validation of ionization efficiency applicability for concentration predictions.

We have used flow injections to artificially control the mobile phase composition at the time of reaching the ESI source. This is necessary as different chromatographic methods yield slight or significant changes in this composition depending on the changes made (RP vs HILIC, or just a slight tweaking of the column chemistry). The ionization efficiency predictions, however, should not be applicable to one method only. So using chromatography only would not enable accounting for these changes.

The validation – using ionization efficiency for concentration predictions – has been always done for real analysis, though we have incorporated samples and standards usually. Therefore, these have all been recorded with chromatographic resolution.

Interestingly, we just recently saw to our large surprise that the workflow with flow injections really pays off as we were applying the ionization efficiency predictions to SFC data from the Thomas Letzel group and saw results that were comparable, if not better, than what we have seen previously for RP.

  • Do the instruments tested each use a different mass reference calibrant and is the ionization efficiency assessed when each instrument has the mass reference calibrant in use or not in use?

Actually, the performance of the quantification approach is not affected by the mass calibration as long as the peaks can be identified. What matters are the calibration chemicals used to transfer the predicted ionization efficiency values to the instrument-specific response factors. And we have used different calibration chemicals in different instruments. These could be chemicals used for quality control or target chemicals analyzed from the samples in parallel to the nontarget analysis. Our recent experiences showed that the only real requirement for the selection of these chemicals is that they would possess a wide range of response factors. This is necessary for establishing a meaningful regression between the response factors and predicted ionization efficiency values.

  • Does the IE prediction relationship need to be re-developed in each unique matrix? (may have missed it, but assuming this is primarily developed for chemicals in pure solvent)

The model is developed with analytical standards without the matrix. However, the validation for water (study with Been et al. and Kiefer et al., food, and blood samples that we have done does include the matrix as well. The problem is a bit more complex than just retraining. The matrix effect really depends on the very chemical that is coeluting with a specific “analyte”. Meaning that the same chemical can be having suppression of 90% which one chromatographic separation but for another separation, it is suppression of 20%, and in another maybe enhancement of 10%. Also, the matrix effect by blood is not always given the same amount of suppression or enhancement, it depends on the concentration of the coeluting chemical in the sample. As a result, it is almost impossible to model for any matrix as such (not to mention the fact that it is workload-wise almost impossible).

A solution that I support is to model the behavior of the chemical and to incorporate the matrix effect in the uncertainty estimation. One small way to account a bit for it is to run your samples at high concentrations and in dilution. Dilution is the most efficient way to reduce the matrix effect. If you notice that the two predicted concentrations from the neat sample and the dilution do not agree at all, it hints that there is (*) signal saturation and lack of linearity or (**) matrix effect. And this could trigger a further investigation.

  • Following up on this [Kathy Peter’s question] how is ion suppression, micelle formation, and other effects in ESI/mass spec accounted for, and do you imagine these to be significant enough that IE predictions alone are not enough? E.g. also salts in the matrix, multimer formation, and in-source fragmentation, all this will also disperse the signal and alter concentration calculation

Good question. I believe that for the matrix effect a combination of incorporating this in the uncertainty and running a dilution for comparison can do a great job. In our experience, the error arising from imperfect modeling is still rather large compared to the matrix effect. Therefore, the matrix effect will not have a dominant impact on the uncertainty.

Regarding in-source fragmentation – yes the parent ion peak and the in-source fragments need to be added together. This is a bit troublesome but componentization and feature association slowly help us to do this. There are groups of chemicals where failing to account for in-source fragments will cause tremendous problems (phthalates, some acids).

Sodium adduct formation is generally not a problem. We have observed that the relative ionization efficiency is unaffected by the sodium content in the sample. However, if the calibration chemicals and “analytes” are analyzed in two different runs and the sodium concentration significantly changes between the runs, we might encounter problems. The best approach here is to keep in mind this possibility while quantifying the peaks eluting close to the dead time.

Multimer formation is a bit easier if it is saturation triggered, this is the case for many chemicals we have tested and which are incorporated in the model training. As all measurements need to be in the linear range the samples cannot be quantified without dilution anyway, even if the multimers are not observed. However, if multimer formation is a specific binding, it will be very hard to incorporate it purely from the modeling point of view (need for the vast number of chemicals with similar binding properties).