Properties of Sterilizing Membranes and Prefilters Used in Developing mRNA Lipid Nanoparticle Production

A Micromeritics AutoPore V Model 9620 (Norcross, GA) was used to measure the pore size distribution of the different sterile filters using mercury intrusion porosimetry (MIP) (Gribble et al. 2011). To ensure reproducibility, four clean (unused) 47‐mm or eight 25‐mm filter discs were cut and loaded into the penetrometer, providing at least 0.2–0.3 g of membrane. For dual‐layer filters, the layers were separated and individually characterized. Mercury was incrementally intruded into the sample, with the pressure increased in steps of 2–20 kPa, allowing for equilibration at each step. After testing, mercury was reclaimed under vacuum, and contaminated filters were disposed of as hazardous waste.

MIP was used to evaluate the pore volume distribution for the different sterilizing‐grade membranes and prefilters. In each case, the pore diameters were evaluated from the data at different test pressures based on the Young–Laplace equation (Tanis‐Kanbur et al. 2021): (1)P=4γcos(θ)dpwhere γ is the interfacial surface tension, θ is the contact angle between the fluid and pore, and dp is the effective pore diameter. The pressure data were first converted to pore diameter using Equation 1. The contact angle for mercury was taken as 140° (Calvo et al. 1995) and the surface tension was 0.485 N/m at room temperature. While 140° is a commonly used approximation, the actual contact angle can vary depending on the membrane surface chemistry, and in particular the hydrophobicity; this effect was not considered in the analysis of the MIP data. The volume distributions were then converted to number distributions by dividing the incremental volume of intruded mercury by the volume of a pore, which was modeled as a straight cylinder with the pore length equal to the membrane thickness (as determined by SEM (Taylor et al. 2021)). Note that MIP does not distinguish between "dead‐end pores" (pore volume that does not penetrate entirely through the membrane) and actual through‐pores, and it provides no information on the size of the pore constrictions that will likely determine the retention characteristics of the membrane (i.e., the retention properties of an "hour‐glass" shaped pore will mainly depend on the size of the narrow "neck" in the hour glass). As such, MIP provides a simplified view that may not fully reflect functional performance.

Bubble points measurements were performed following the procedure outlined in (Messerian et al. 2023b) to provide an estimate of the largest pore in each membrane (Hernandez 1996). Briefly, membranes were mounted in 25 mm holders, pre‐flushed with 100 L/m² of deionized water, and pressurized with nitrogen at ~7 kPa/min. The bubble point was recorded as the pressure at which the first bubble appeared in a submerged dip tube. Measurements were performed in triplicate and reported as mean ± standard deviation in Table 3.

Characterization results obtained with the five sterilizing‐grade membranes are summarized in Table 3. Even though all the membranes have a similar nominal pore size rating (0.2 or 0.22 µm), the number‐weighted mean pore diameter measured by MIP ranged from a low of 0.25 µm for the Sterilux membrane to as high as 0.35 µm for the 0.2 µm layer of the Sartopore 2 XLG. The largest pore diameter, based on the measured bubble point, ranged from 0.65 µm for the Stylux to 2.4 µm for the Millipore Express Plus. These values are consistent with results reported by the manufacturers and in the literature (Taylor et al. 2021). The different properties of the membranes are further highlighted by the detailed pore size distributions shown in the Supporting Information (Figure S1). The PVDF membranes (Sterilux and Durapore) exhibit both the smallest mean pore diameter and the lowest water permeability (410–460 LMH/psi), which reflects their tighter pore structure compared to the asymmetric PES membranes, all of which showed significantly greater permeability. Among the PES membranes, the permeability does not appear to directly correlate to mean pore size. Notably, the 0.2 µm Sartopore 2 XLG has the largest mean pore size (0.35 µm), yet its permeability is approximately half that of the Stylux membrane, despite both having similar porosity. This discrepancy may be due to differences in skin layer thickness or internal architecture not captured by MIP. These findings highlight how membrane chemistry and internal structure, beyond nominal pore size, are crucial design features that must be considered when selecting membranes for specific filtration applications.

Figure 1 shows experimental data for the flux during LNP filtration through each of these five membranes at a constant TMP of 20 psi. Although all membranes have similar nominal ratings (0.2 or 0.22 µm) and have high LNP recovery yields ( > 90%), their filtrate flux and capacity vary significantly, with up to three‐fold differences in filtration capacity (defined as the volumetric throughput at which the flux had declined to 10% of its initial value). Despite these differences, the flux decline profiles display a similar pattern of concavity, consistent with the complete pore blockage model reported previously (Messerian et al. 2022). Additionally, filtration capacities did not correlate well with the initial membrane water permeability or the bubble point. For example, despite having a substantially lower permeability than the Stylux membrane, the 0.2 µm layer of the Sartopore 2 XLG has the highest capacity (32 vs 22 L/m² for Stylux).

As shown in Figure 2, the capacity of the different sterilizing‐grade membranes is highly correlated with the mean pore diameter evaluated using MIP, with a simple linear regression fit giving R² = 0.86. The greatest capacity was obtained with the filter having the largest mean pore diameter (0.2 µm layer of the Sartopore 2 XLG) while the lowest capacity was obtained with the filter having the smallest mean pore diameter (Sterilux). This relationship is consistent with the expectation that smaller pores would be more easily blocked by the LNPs. Membranes with very small mean pore sizes, (i.e., more overlap with the size distribution of the LNPs) would likely show a disproportionately lower capacity, leading to the very low capacity observed with the Sterilux membrane. Additionally, the Durapore and Stylux membranes show similar capacity despite having different pore symmetry and membrane chemistry, indicating that these factors are not as impactful to LNP filtration capacity as is the mean pore size.

The filtration capacities of the single‐layer sterile filters (data in Figure 1) were all significantly smaller than the capacity of the dual‐layer Sartopore 2 XLG, which in previous LNP filtration studies using similar feed streams was shown to exhibit approximately two‐fold higher capacity (Griffiths et al. 2016). This observation suggests that the 0.8 µm prefilter provides significant capacity improvement by removing foulants responsible for limiting the capacity of the 0.2 µm layer. To investigate this phenomenon further, constant TMP filtration experiments were performed for the LNP with the individual layers of the Sartopore 2 XLG and compared to the dual‐layer configuration. As shown in Figure 3, the initial flux through the dual‐layer filter and the 0.2 µm layer are very similar, and much smaller than that for the 0.8 µm layer alone, indicating that the initial flux through the dual‐layer filter is dominated by the properties of the 0.2 µm layer. The flux decline through the 0.8 µm layer is negligible for the volume tested, indicating minimal fouling over the course of the filtration due to the large pore size relative to the foulants in this LNP formulation. The mean pore diameter of the 0.8 µm layer determined by MIP is 1.1 µm, which is three‐fold greater than that of the 0.2 µm layer. The capacity of the dual‐layer filter was approximately twice that of the 0.2 µm membrane alone, similar to previous findings with a similar LNP formulation (Messerian et al. 2022), demonstrating the effectiveness of the 0.8 µm prefilter.

To evaluate differences between various modes of prefiltration, experiments were performed using the 0.8 and 0.2 µm layers of the Sartopore 2 XLG in three configurations: (1) batch prefiltration, in which the feed is first filtered through the prefilter, with the permeate collected and then separately filtered through the sterilizing‐grade filter; (2) inline prefiltration, where the prefilter and sterilizing‐grade filter are placed in series in the same flow path; and (3) the integral (dual‐layer) configuration, in which both membranes are part of a single device. Results from these experiments are summarized in Figure 4. Although the addition of a prefilter significantly enhanced the filtration performance in all three modes, by far the best performance was obtained with the integral configuration. The inline configuration performed marginally better than the batch configuration, which provided only a 17% increase in capacity. Note that the difference in results for the integral configuration shown in Figures 3 and 4 is due to the different aliquot of LNP used for these experiments (which were subjected to different freeze‐thaw cycles).

These differences in fouling and capacity between configurations suggest that LNP foulants are not permanently removed by the prefilter and that LNP fouling may be influenced by time‐ and/or shear‐dependent mechanisms. The "lag time" between filtration through the prefilter and the sterilizing‐grade filter likely plays a critical role in determining the overall fouling behavior. This lag time varies from tens of minutes for the batch configuration, to several seconds for the inline configuration, to only milliseconds for the integral configuration. The lack of significant improvement in performance for the inline prefiltration compared to the batch prefiltration process suggests that LNP foulants may reaggregate fairly quickly (i.e., on a timescale shorter than several seconds). This behavior is consistent with previous findings (Griffiths et al. 2016), where sequential filtration of LNPs through two sterilizing‐grade filters using a batch setup failed to significantly improve filter capacity, suggesting that foulant removal during prefiltration is not permanent and complete. A similar discrepancy in performance between prefiltration modes was reported during virus filtration of monoclonal antibodies (mAbs), where the addition of a 5 µm Durapore prefilter significantly enhanced the flux and increased the capacity but only when used in the integral configuration. In that case, the improvement in performance was thought to be due to the time‐dependent self‐association of mAbs, with the mAbs re‐associating after passage through the prefilter when used in either the batch or inline configurations (Billups et al. 2022).

Alternatively, the integral configuration may help distribute flow more evenly across the sterilizing‐grade layer, which could reduce localized fouling and thereby improve overall filtration efficiency. The integral configuration could also have a preconditioning effect on the LNPs, with the shear forces involved in passage through the prefilter influencing the LNPs (or LNP aggregates) so that it is easier for them to pass through the sterilizing‐grade membrane. A preconditioning phenomenon has been previously reported during ultrafiltration of plasmid DNA (pDNA), with the pDNA effectively "pre‐stretched" by passing it through a larger pore size membrane (Li et al. 2015).

The 0.2 µm layer of the Sartopore 2 XLG has a highly asymmetric pore structure, with the pore size decreasing from the upstream side to the downstream side ("skin‐side"). Asymmetric pore structures are intentionally designed to maximize permeability by reducing the thickness of the size‐selective skin (Tan and Rodrigue 2019). In addition, the more open pore structure in the support can increase filtration capacity by capturing foulants of different sizes, acting much like a depth filter (Li et al. 2009).

To demonstrate the effectiveness of this design, additional prefiltration experiments were performed in which the orientation of the 0.2 µm layer was reversed so that the side with the smaller pore size was facing towards the feed flow; the 0.8 µm layer was used in its normal orientation. Filtration experiments were again conducted at a constant TMP of 20 psi, with results summarized in Figure 5. The capacities achieved in this reverse orientation are all smaller than those in the standard orientation. For the 0.2 µm layer alone, the initial flux in the reverse orientation is more than two‐fold smaller than in the standard orientation (3700 vs 8100 LMH) and the capacity was only 9 L/m². In addition, the batch and integral prefiltration have minimal effect on the capacity of the 0.2 µm layer when used in the reverse orientation, demonstrating the critical role of the more open (larger pore size) upstream section of the asymmetric sterilizing‐grade membrane in reducing the extent of fouling. Note that the difference in results between Figures 4 and 5 is due to the use of a different aliquot of LNP for the experiments shown in Figure 4; Figures 3 and 5 were generated using the same LNP aliquots.

The filtration experiments performed with the different sterilizing‐grade membranes in Table 1 suggest that sterilizing‐grade membrane chemistry may not have a large effect on fouling and capacity for this LNP formulation. To further investigate the potential contributions from membrane chemistry, experiments were performed in which either a hydrophilic (HVLP02500) or a hydrophobic (HVHP02500) version of the 0.45 µm PVDF Durapore membrane was used as an integral prefilter for the 0.2 µm layer of the Sartopore 2 XLG (in its standard orientation). Although hydrophobic membranes are most commonly used for gas separations (Claramunt et al. 2021), the 0.45 µm pore size is sufficiently large that it is possible to wet the hydrophobic Durapore membrane with the aqueous LNP feed solution at relatively low TMP ( < 2 psi). The Durapore prefilters were selected because they are symmetric and have essentially identical pore size distributions (see Supporting Information, Figure S2). As shown in Figure 6, the initial flux through the dual‐layer membrane formed using the hydrophobic prefilter is approximately 2.5‐fold lower than that formed using the hydrophilic prefilter (2300 vs. 5500 LMH, respectively). Interestingly, the flux through the hydrophobic prefilter slightly increases during the initial stages of filtration before declining at higher throughput. Similar effects have been observed by Girones et al. (2005) when studying hydrophobic microsieve membranes, with this increase in flux caused by changes in the wetting of the hydrophobic pores. In this case, the change in wetting during the initial phase of the filtration may be due to interactions with the lipid components in the LNP feed. Overall, the dual‐layer membrane formed using the hydrophilic prefilter has a slightly higher capacity than that formed using the hydrophobic prefilter. However, the difference in performance between these membrane chemistries is relatively small (capacities within 20%), suggesting that the chemistry of the prefilter does not play a major role in the effectiveness of the prefiltration, at least for the LNP feed studied in this work.

To investigate the impact of the size of the foulants in this LNP feed, several custom combination dual‐layer membranes were created by stacking hydrophilic PVDF Durapore prefilters of varying pore sizes directly on top of the 0.2 µm layer of the Sartopore 2 XLG in an integral configuration. The Durapore membranes were chosen for their homogeneous pore morphology and availability across a range of pore sizes. Figure 7 illustrates the performance of these dual‐layer membranes with prefilter nominal pore sizes ranging from 0.45 to 5 µm. The initial fluxes through all of the combination filters are very similar to that for the 0.2 µm layer of the Sartopore 2 XLG since the sterilizing grade layer is the primary contributor to the overall permeability. The filtration capacity increases with increasing prefilter pore size up to 0.8 µm, from 31 L/m² with no prefilter to 54 L/m² with the addition of a 0.45 µm prefilter, to 62 L/m² when using a 0.65 µm prefilter, and to 83 L/m² with the addition of a 0.8 µm prefilter (the upper layer in the commercial Sartopore 2 XLG). However, the 5 µm Durapore prefilter has almost no effect on the filtrate flux or capacity. These data strongly suggest that the foulants involved in LNP filtration are in the sub‐micron size range, being effectively removed by the 0.65 µm Durapore prefilters but showing minimal removal by the 5 µm Durapore prefilter. This behavior is in stark contrast to data obtained during sterile filtration of a glycoconjugate vaccine drug substance, where the addition of the same 5 µm Durapore prefilter led to more than a 100‐fold increase in capacity (Du et al. 2021).

Further insights into the design of dual‐layer membranes for the sterile filtration of LNP were gained by testing various prefilters (Table 2) with different chemistries and pore size distributions (see Supporting Information, Figures S3 and S4) in combination with the 0.2 µm layer of the Sartopore 2 XLG. Similarly to the combinations evaluated in Figure 7, these custom, integral dual‐layer membrane combinations are not available commercially. As shown in the left panel of Figure 8, several of the evaluated custom dual‐layer configurations achieve a superior capacity compared to that of the dual‐layer Sartopore 2 XLG membrane (0.8 µm + 0.2 µm). The best performing prefilters included the 0.45 µm cellulose acetate (CA) prefilter sold as part of the Sartobran P (52307H‐47‐P‐B), the 0.45 µm polyethersulfone (PES) layer of the Sartopore Platinum (54907H‐47‐B), a 0.8 µm mixed cellulose ester membrane from MilliporeSigma (AAWG04700), and a 0.45 µm PES membrane from Sterlitech (PES4525100). These high‐performing prefilters span a range of membrane chemistries, again suggesting that it is the pore size and/or morphology, not the chemistry itself, that is critical in determining the performance of these novel dual‐layer filters. There were also many dual‐layer configurations evaluated that had worse performance than that of the dual‐layer Sartopore 2 XLG membrane, with several of these shown in the right panel of Figure 8. In particular, the relatively poor performance of the 0.35 µm prefilter (Sartopore 2 XLI) may well be due to its small pore size compared to the rest of the prefilters.

The detailed pore size distributions of the best‐performing prefilters were evaluated by MIP. In each case, the number distribution of pores was evaluated by considering three size bins: pores smaller than 400 nm, pores between 400 and 800 nm in size, and pores larger than 800 nm. The left panel of Figure 9 presents the percentage of pores within each size range for the different prefilters. All the prefilters have broad pore size distributions, with significant numbers of pores below 400 nm as well as a significant number of pores above 800 nm. For example, the 0.8 µm layer of the Sartopore 2 XLG has approximately 20% of its pores (based on the number distribution) smaller than 400 nm and over 60% of its pores larger than 800 nm. Conversely, the 0.45 µm cellulose acetate layer from the Sartobran P has a distribution skewed towards smaller pore sizes, with nearly 50% of its pores falling below 400 nm and only 20% exceeding 800 nm.

The prefilters that show superior performance are consistently found to have a higher percentage of pores within the 400–800 nm range, which is the approximate size range previously identified as being critical for effective prefiltration of this LNP formulation (Figure 7). This behavior is examined more closely in the right panel of Figure 9, which shows the improvement in capacity (defined as the relative increase in the capacity for the dual‐layer filter compared to that for the dual‐layer Sartopore 2 XLG) plotted against the percentage of pores in the 400–800 nm size range. The data show a high degree of correlation suggesting that the pore size distribution of the prefilter is a key factor determining the prefilter performance. Other metrics including porosity, median pore size, and the percentage of pores in broader or adjacent size ranges were also evaluated but did not show meaningful correlation with capacity. Note that this correlation does not hold for all prefilters; the Sartopore 2 XLI shows relatively poor performance even though 47% of its pores are in the 400–800 nm range (Supporting Information, Figure S4), likely due to the relatively high percentage of very small pores in this prefilter (38% of the pores are less than 400 nm in size).

Interestingly, despite differences in membrane chemistry, supplier, and configuration, the fouling behavior of the sterilizing‐grade membranes is largely consistent. In all cases, except when the Sartobran P was used as a prefilter, the data fit well to the complete pore blockage model (Messerian et al. 2022), suggesting that fouling for the mRNA‐LNP formulations examined in this work is due primarily to the deposition of larger LNP foulants onto membrane pores. Despite sharing the same fouling mechanism, substantial differences in capacity are observed across the tested membranes, driven by structural factors such as pore size distribution, layer orientation, and configuration. These results imply that membranes originally optimized for protein‐based biologics may not be well suited for other feed streams, including large particle size therapeutics like LNPs. In particular, the greater increase in capacity provided by prefilters enriched in 400–800 nm pores suggests that achieving a closer match between the size of the foulant and the size of the pores in the prefilter is key to reducing the rate of fouling. This highlights a new design space for membrane manufacturers, focusing on fine‐tuning sub‐micron pore size distributions to accommodate the unique filtration behaviors of large particle size therapeutics. Future studies will be required to demonstrate whether this behavior also holds for more rigid particles like viral vectors.