Ultrafiltration Process AnalysisPreliminary Studies to Simplify the Process and Reduce the Number of pCPP

For starting the UF step was analyzed (Fig. 1b), and some tests were planned to verify the possibility to further simplify the process and reduce the number of pCPP.

The UF step consists of two diafiltration steps differing in number of diavolumes and diafiltration buffers:

Furthermore, processed material is concentrated before and after the first diafiltration step (Fig. 1b).

We evaluated the possibility of simplifying the process, understanding the impact that concentration and diafiltration steps could have on the purification and on the quality of the final product. In particular, we evaluated the possibility of avoiding the use of two different diafiltration buffers (Supporting Information (SI), Run 1 vs Run 2) and eliminating an intermediate concentration step (SI Run 2 vs Run 3).

GMMA coming from these three runs showed similar characteristics, as reported in Table S3 and Figure S1; therefore, the process was simplified as in Run 3 (Fig. 1c): growth supernatant was adjusted to pH 7.2, concentrated directly 12 × , and diafiltered against saline only (16 DV). Moreover, samples collected as intermediates during each purification were analyzed for soluble proteins and DNA impurities to improve our knowledge on the overall process (Table S4).

Selection of Critical Quality Attributes (CQA) to Monitor

Table I lists the potential GMMA CQAs: some of them refer to GMMA as particle; other relate to key antigens presented on GMMA surface, e.g., OAg of LPS molecules. Such list was established on the basis of prior knowledge, current structure–function understanding, strategic nonclinical studies performed with S. Typhimurium GMMA.

Table I Classification of Impact of TFF Step on S. Typhimurium GMMA Potential CQAs and Potential Secondary Interactions

We have evaluated if the UF step can have an impact on each the potential CQAs.

Depending on the possible impact, the potential CQAs are classified as impacted (I)/not impacted (NI)/potentially impacted (PI). The classification has been carried out based on product and process understanding know-how.

In Table I, the secondary interactions between potential CQAs, assigned based on logical reasoning and the chemical-physical knowledge of the drug substance, are reported. As an illustrative example, with a change in molecular size distribution of the OAg, a change also in OAg quantification and OAg to protein ratio can be expected.

Definitions of the Process

Based on the assessment reported above, only CQAs directly or indirectly impacted by the UF step were considered for establishing a relationship between CQAs and pCPP (Table IIA). This first assessment was based on current know-how and historical data collected. The impact of each pCPP on each CQA has been classified following the criteria reported in Table IIB, and the type of relationship (if known) has been indicated:

ID: if the increase of the pCPP value causes a decrease of CQA value and vice versa (decrease-increase)

II: if the increase of the pCPP value causes an increase of CQA value and vice versa (decrease-decrease)

Data analysis: if the type of the relationship is not clear and additional data are needed.

Table II Scoring Impact of Ultrafiltration pCPP on CQAs. (A) Relationship/Scoring Between CQAs and pCPP; (B) Classification Criteria of the Impact of Each pCPP on Each CQA

GMMA concentration in the material to be processed (coming from the previous purification step) could also have an impact, but it has not been considered at this stage.

Other parameters such as UF membrane cutoff and material, diafiltration buffer composition and raw materials that clearly can have an impact on the CQAs but are considered fixed at this stage of the study as they are already defined in the process and cannot be easily modified.

From the total scores calculated in Table IIA, the most critical process parameters that impact CQAs resulted to be transmembrane pressure (TMP), number of DV, and starting pH.

Measurement System Analysis (MSA)

For the purpose of this study, it was important to identify appropriate analytical methods to characterize the intermediate samples coming from the UF step.

HPLC-SEC was used to determine soluble proteins, DNA impurities, and GMMA yield. Total OAg was estimated by HPAEC-PAD, while total protein content was determined by micro-BCA, and their ratio was calculated. OAg size distribution (molecular weight and OAg to core ratio) and lipid A content were determined by HPLC-SEC/semicarbazide on extracted OAg. Protein pattern was qualitatively investigated by SDS-PAGE as well as appearance. O-acetyl content was quantified by 1H NMR. Two different methods, HPLC-SEC/MALLS and DLS, were used to determine particle size.

It was deemed appropriate to characterize DLS, HPLC-SEC methods, micro-BCA, and HPAEC-PAD for their reproducibility to verify if appropriate for building statistical models through the DoE approach, considering the minimum signal change detectable/noise ratio that has been predicted from the power calculation in DoE planning.

The characterization design was planned with two different operators for a total of six independent analytical sessions (three for each operator). In the analysis of variance, the sessions were nested into the operator as it was not possible for the operators to perform the analysis in parallel.

For each analytical method, samples were prepared in triplicate at multiple concentrations starting from a GMMA standard lot to investigate the performance of the analytical method at different concentration levels (i.e., close to the lowest, the middle and the highest level of the analytical method calibration curve).

For GMMA protein content and purity (soluble protein and nucleic acid) by HPLC-SEC, two samples differing in the content of impurities, to represent also the worst case, were analyzed in triplicate: bacteria culture supernatant (unprocessed material, thus containing a high percentage of soluble proteins and high 260/280 nm Abs ratio) and TFF retentate after 13 DV (not completely purified GMMA, with a certain percentage of soluble proteins), analyzed respectively neat and with a twofold dilution.

For each analysis session, the preparation and the analysis order of the samples were randomized.

Variance component analyses were performed with mixed effects model, and results are reported in Table III (per each analytical method/sample type, statistical analysis is reported in SI Tables S5-S18). The precision of each method is expressed as coefficient of variation (CV), considering the average of all measurements for each concentration tested.

Table III Analytical Precision CharacterizationOptimization Phase (DoE)Design, Factors, and Responses

At this point, the DoE approach was used to investigate the impact of selected pCPP on GMMA CQAs. The following three key factors were used: starting pH, TMP, and number of DV. As mainly quadratic models were expected for the responses under evaluation, RSM design was selected from the beginning without performing a preliminary screening design and then a characterization design.

This choice was made because only three input factors were selected based on the assessment reported in Table IIA.

Parameter ranges were selected on the following basis:

Starting pH: suitable range to avoid O-acetyl groups hydrolysis and protein precipitation based on prior knowledge (26) (setting: low 5.3, middle 6.25, high 7.2);

TMP: instrumental technical feasibility (setting: low 0.5, middle 0.875, high 1.25 bar);

DV: from preliminary study which results are reported in Table S4 (low 5; middle 13, high 21).

A face centered design, instead of a rotatable one, was chosen as the parameters cannot be set far outside the design space due to GMMA product stability (starting pH), instrumental boundaries (TMP), and based on technical knowledge (DV).

With the design described, a quadratic model would be able to detect a signal/noise ratio of 2 standard deviations at 5% alpha level with a power of 97.8% for order 1 term, 96.3% for two-way interactions, and 87.9% for quadratic terms. The related Fraction of Design Space (FDS) graph is reported in Figure S2.

Thirteen output responses among the GMMA CQAs have been evaluated: soluble proteins (1), DNA content (2), GMMA size by HPLC-SEC MALLS (3), GMMA size by DLS (4), GMMA size polydispersity index by DLS (5), pH of product (6), osmolality (7), GMMA yield (8), OAg/total protein ratio (9), O-acetyl content (10), OAg size (11), OAg to core ratio (12), and lipid A to GMMA protein ratio (13). GMMA content (8) was also evaluated to get insight on process yield.

DoE Elaboration

After having obtained models for the different attributes (only for lipid A and DLS Z-average a model was not obtained, Table IV) and identified the dependence between CPP and each of CQA, we concentrated our attention on the quality of the drug substance and not on the yield, to identify the design space to purify GMMA with CQAs inside set boundaries. GMMA yield was investigated, but it was outside the scope of this work that aimed to ensure good quality of the final product and not to improve yields. This is related to the fact that manufacturability of GMMA is simple and at low cost and yields obtained are already satisfactory.

Table IV DoE Response Surface Results

Among all the responses considered, only the following ones were used: soluble proteins, DNA content, pH, OAg/total protein ratio, and O-acetyl content for process optimization. Particle size distribution by MALLS, osmolality, OAg size, and OAg/core ratio gave small variation respect to the boundaries fixed in the entire design space.

Increasing the number of DV, impurities (nucleic acids and soluble proteins) are reduced and consequently the OAg/total protein ratio increased.

Starting pH, as expected, has the major impact in determining the purified GMMA pH (in case of higher starting pH, a higher number of DV contributes to achieve a lower purified GMMA final pH). Starting pH values close to 6.5 resulted the optimal to preserve O-acetylation level that is also negatively impacted by higher number of DV.

By increasing TMP probably, there is an impact on GMMA integrity resulting in an increase of soluble proteins especially with lower DV number.

Variability of the Process Factors

For the process optimization, also the POE was considered in order to find factor settings that minimize variation transmitted to the response from each factor; by this way, the process will be more robust to variations in input factors.

To perform the model optimization with POE, we considered the following variability of the factors:

The equipment measures the DV by weighing the TFF permeate and the standard deviation reported for the instrumental weighing system is of 0.1 g. Considering that 1 DV corresponds to 25 g, the standard deviation considered for the DV factor corresponds to 0.004 DV (25 g:1dv = 0.1:x).

The TMP is measured with a pressure transducer and the standard deviation reported for the device used is of 0.02 bar.

For the starting material pH measurement, the pH meter variability reported in the instrument specification is not deemed appropriate as the sample is very dirty, so it has been measured on real sample. The characterization design was planned with one operator for three independent analytical sessions (one per day); in each analytical session, three different measurements were performed on the same real sample. In Table S32, the variance component analysis, obtained with general linear model, is reported. The standard deviation calculated from the total variance is 0.11 pH unit.

Optimization with POE

Optimal conditions for performing the UF step resulted to be starting pH 6.7, TMP 0.56 bar, and DV 17.3 (the optimization without POE minimization would lead instead to the following conditions: starting pH 7.2, TMP 0.50 bar, DV 18.6). In Table S33, the 95% CI for mean and tolerance interval for the predicted values of CQAs in the optimal conditions identified are reported.

In Figure S36, desirability ramps for each factor and each response, as well as the combined desirability, are reported. In Fig. 2, the combined desirability function calculated in the design space is reported showing the process parameter (PP) ranges to obtain purified GMMA with CQA mean responses in the specifications set. In Figure S37, S38, and S39, the graphs of the CQA models close to the optimized conditions are reported. In Figure S40, S41, and S42, the graphs of the POE for CQA close to the optimized conditions are reported.

Fig. 2

Combined desirability function plots. Each plot reports desirability function outcomes in the range of number of DV and starting pH values, respectively, in y and x axes at three different levels of TMP (low, middle, and high)

Model Confirmation

To confirm the model, an additional UF run with the same starting material was performed using the optimized conditions identified. The TFF retentate was fully characterized (Table S34), and the responses obtained were evaluated respect to the predictions (Table S35): all attributes fell in the predicted ranges.

The study was conducted by ignoring possible variability coming from the starting material, and the runs were performed at small scale. Of course, it will be important to verify that expected results are obtained on material coming from different fermentation runs and by performing the TFF at bigger scale. Working at small scale allowed to execute multiple runs in relatively short period of time and to understand impact that process parameters can have on GMMA quality. By scaling up the process, changes could be expected, but the knowledge acquired will allow to rapidly identify optimal working conditions if needed. It will be interesting to apply conditions identified in this study for S. Typhimurium to GMMA from other pathogens and confirm platform potential of the process. We could expect need for certain adjustments based on specific antigen characteristics, but again, the info acquired will allow to accelerate the work on GMMA from different organisms.

Capability Analysis

We evaluated, in the optimized CPPs settings identified by DoE, the effect of CPPs variabilities on CQAs variabilities.

Monte Carlo simulations were run on CPPs, and, by using the transfer functions found with the DoE, the resulting CQAs distributions were identified (Fig. 3 and Table V) and studied with the capability analysis to understand the probability of an out-of-specification. Results of the capability analysis are reported in Table VI.

Fig. 3

CQA distribution obtained from transfer function found by DoE on Monte Carlo simulations of CPPs. a Soluble proteins %; b DNA impurities (reported as 260/280 ratio by HPLC-SEC); c pH; d OAg/total protein w/w ratio; e O-acetyl content %

Table V Description of CQA DistributionsTable VI Summary Results from Capability Analyses for Each CQA

For all parameters, except pH and OAg/protein ratio, Ppk is higher than 1.33, meaning that the process is appropriate for meeting the boundaries set. For pH and OAg/protein ratio, Ppk values are close to 1, and in addition, the process is not centered on the target. Therefore, the tolerances for the key factors that influence the process (starting pH, TMP, DV) as well as the MSA values for the CQAs need to be kept under control with time.

The results obtained are considered acceptable for a biological process as the highest probability of out of spec is 0.16%.