Phase Separation Methods for Protein Purification: A Meta-Analysis of Purification Performance and Cost-Effectiveness

Abstract

Protein purifications based on phase separations (e.g., precipitation and liquid-liquid extraction) have seen little adoption in commercial protein drug production. To identify barriers, we analyzed the purification performance and economics of 290 phase separation purifications from 168 publications. First, we found that studies using Design of Experiments for optimization achieved significantly greater mean yield and host cell protein log₁₀ removal values than those optimizing one factor at a time (11.5% and 53% increases, respectively). Second, by modeling each reported purification at scales from 10 to 10,000 kg product/yr and comparing its cost-effectiveness vs. chromatography, we found that cost-effectiveness depends strongly on scale: the fraction of phase separations predicted to be cost-effective at the 10, 100, and 1000 kg/yr scales was 8%, 15%, and 43%, respectively. Total cost per unit product depends inversely on input purity, with phase separation being cheaper than chromatography at the 100 kg/yr scale in 100% of cases where input purity was ≤1%, compared to about 25% of cases in the dataset as a whole. Finally, we identified a simple factor that strongly predicts phase separation process costs: the mass ratio of reagents vs. purified product (the “direct materials usage rate”), which explains up to 58% of variation in cost per unit of purified product among all 290 reports, and up to 98% of variation within particular types of phase separation.

GRAPHICAL ABSTRACT

1. Introduction

The total market for recombinant proteins is currently estimated to be more than $270 billion, with a projected value of more than $400 billion by 2026.^[1–3] More than 96% of this is accounted for by protein therapeutics, which are distinguished from other applications by their stringent purity requirements and high manufacturing costs. For such drugs, manufacturing costs typically represent approximately 20% of total sales^[4] and downstream processing typically represents about 75% of total manufacturing costs.^[5] Therefore, protein drug purification processes alone cost more than $40 billion per year. The linchpin of these processes is undoubtedly conventional packed-bed chromatography,^[6,7] which is often also considered their major cost driver and productivity bottleneck.^[8,9]

Despite its limitations, conventional chromatography has remained the preeminent purification operation in biotechnology largely because of three key advantages. First, it is versatile: generally, when developing a purification process for a new protein product, almost all the necessary separations can be performed using some sequence of chromatography steps. Second, it is predictable: most resins in common use have a single dominant selectivity (e.g., electrostatic interactions or hydrophobicity), a relatively small number of easily-controlled variables drives the performance, and optimization strategies are well-established.^[10] Third, it is generally capable of high yield as well as high resolution even among proteins with similar properties. Any alternative separation operations seeking to displace conventional chromatography therefore face the imposing task of overcoming these advantages of versatility, predictability, selectivity and yield. Furthermore, they must do so in a cost-effective manner.

Nevertheless, in certain circumstances, alternative purification methods can dramatically lower process costs without sacrificing product quality, potentially enabling new markets or simpler and more flexible manufacturing processes.^[4,11] One highly active area of development for such alternatives deals with methods that achieve purification by differentially modulating protein solubility among two or more chemical phases. This class of methods can be referred to as phase separations or bulk separations^[12] and includes precipitation and liquid-liquid extraction among the most promising examples. These methods are generally considered attractive because of three potential advantages they offer over chromatography: they can typically achieve rapid separations (e.g., on the order of minutes), they require only simple and relatively inexpensive equipment, and they can be performed with inexpensive chemicals or reagents.

Despite their potential cost advantages, decades of development, and widespread use for non-therapeutic products and specific cases such as plasma proteins,^[13] phase separation methods have seen little uptake into commercial processes for recombinant protein therapeutics. The reasons for this lack of implementation have been partially addressed in other recent reviews, which have surveyed the most common methods for phase separation,^[14–16] discussed the small number of phase separation processes currently in commercial use (e.g., for insulin, erythropoietin, and Apo2L)^[14,17] and the many more under active commercial development,^[14,18,19] and addressed theoretical and engineering aspects of method development.^[14,18,19]

To these valuable prior contributions, we add a meta-analysis of 168 publications reporting 290 distinct purification operations using nearly 40 different types of phase separation method. First, whereas prior reviews have mostly highlighted the most successful examples of phase separations to show the field’s potential, our unbiased approach allows us not only to present a more accurate picture of the range of outcomes that can be expected, but also to quantitatively identify factors associated with increases in purification performance. Second, we address another critical requirement for commercial implementation besides purification performance: cost-effectiveness. No other review has yet provided an understanding of phase separation costs that accounts for various process scales, is quantitatively predictive, or applies to a wide range of phase separation methods. To address this gap, we conduct detailed techno-economic modeling based on our dataset and identify simple and specific conditions under which phase separation methods are likely or unlikely to be cost-effective compared to chromatography. Third, taking into account both purification performance and cost-effectiveness, we assess the current implementation-readiness of various phase separation methods. This provides a more quantitative complement to the assessments of past reviews based on qualitative measures such as technology readiness levels. Finally, we suggest guidelines for future work in the field on the basis of this evidence.

2. Methods

2.1. Literature Inclusion Criteria

The scope of this analysis was limited to publications in scientific journals from 01/01/2000 to 01/19/2022 (the date of the literature search), in which the following elements could be identified: 1) a process for purifying a protein was reported; 2) a precipitation or aqueous extraction was performed; 3) a product yield was reported; 4) a degree of contaminant removal was reported; 5) original research was reported; and 6) full-text access was available in English and at the authors’ institution.

2.2. Literature Search

The PubMed database (National Library of Medicine (US)) was searched on 01/19/2022 to identify publications potentially matching the inclusion criteria (for the search term, refer to Supplemental Methods). 289 publications matching the inclusion criteria were identified, of which 168 were manually reviewed to achieve a representative sampling.

2.3. Literature Review

Each publication in the dataset was manually reviewed for annotation of key features describing the product, expression system, purification process, methodology, and results. Each publication was further divided into one or more individual records, where each record contained the information relevant to a distinct protein product and/or phase separation unit operation. For example, a publication reporting the use of two sequential precipitation operations to purify one product would comprise two records. In contrast, when a publication reported the evaluation of multiple phase separation unit operations in competition, rather than for sequential use as part of the same process, only the one showing the best results was recorded. From the 168 publications reviewed, the dataset for analysis consisted of 290 unique records of phase separation purification operations.

For each analyzed feature, when possible, data were transformed into standardized units and formats across all records. For a list and definitions of all features considered in this analysis, as well as approaches used to standardize data, refer to Supplementary Methods.

2.4. Classification of alternative separation methods

Information describing the reaction conditions of each purification operation was collected as completely as possible, including, as applicable, the concentrations and identities of each chemical or biological species present and the temperature, time, pH, and volumes involved. However, for the purposes of most analyses, methods were also more broadly grouped: first as using either precipitation or liquid-liquid extraction, and then by the types of agents used to achieve the separation. Phase-forming agents were classified by general chemical properties (e.g., charge, size, polarity) and/or by the features of the product and contaminant proteins that they probe (e.g., affinity recognition or isoelectric point).

2.5. Classification of optimization methods

Records in the dataset were also classified in terms of the methods used to optimize the phase separation unit operation. Optimization methods were categorized in four ways. If nothing in the publication suggested whether or how variables were tested before selecting the final reaction conditions, we classified a study as having no optimization. If each variable was tested independently and final levels of each variable were chosen based on the independent optima, we classified a study as using one-factor-at-a-time optimization (OFAT). If variables were tested simultaneously and final levels of each variable were chosen based on the optimum from the multiple-variables tests, we classified a study as using multiple-factors-at-a-time (MFAT) optimization. Finally, if MFAT experimentation was supplemented by the use of statistical principles to select the tested variable levels, experimental order, or other features of the approach, we classified the study as using Design of Experiments (DoE).

2.6. Data Analysis, Visualization and Statistics

Data analysis, visualization, and statistics were performed in Python using standard methods via the libraries pandas,^[20] seaborn,^[21] Scipy,^[22] and Matplotlib.^[23] Visualization was also performed in GraphPad Prism (GraphPad Software, San Diego, CA).

Linear least-squares regressions were performed after taking the log₁₀ transform of the x and y variables to linearize the data. Statistical differences in product yield were determined using non-parametric tests (the Kruskal-Wallis test and the Mann-Whitney U test for group and pairwise comparisons, respectively) because of the non-normality of the data. Differences in host cell protein removal were determined using one-way ANOVA for group tests and Welch’s t-test for unequal variances for pairwise tests. All multiple comparisons for pairwise tests were corrected by the Bonferroni method. Significance was determined at p < 0.05 or the appropriate corrected value.

2.7. Techno-economic Analysis

Techno-economic analyses were performed in Python using custom unit operation models for a generic phase separation protein purification method and for ion-exchange chromatography. Each operation was modeled to purify four amounts of product through the unit operation per year: 10, 100, 1000, or 10000 kg. Whenever maximum equipment sizes were exceeded, the equipment was divided into as many equally-sized parallel units as necessary. Records from the dataset were included in the techno-economic analysis if they reported a measure of host cell protein removal, a measure of product yield, a product purity level prior to phase separation, and a description of the reaction conditions. For each record, these data were then used to parameterize models of both the phase separation operation as reported and an alternative ion-exchange chromatography operation processing the same input stream. The complete models may be viewed and tested as interactive Python scripts in a web browser at https://mybinder.org/v2/gh/jsd94/Decker-et-al-2022-Techno-economic-analysis/main. Additional details on techno-economic analyses can be found in Supplementary Methods.

3. Results

3.1. Dataset characterization

From 168 reports published between 2000 and January 2022, we identified a dataset of 290 distinct combinations of phase separation methods and product proteins (i.e., 290 data “records”). We began by surveying the existing use-cases of phase separation for protein purification, focusing on four descriptors: the type of method used; the type of product purified; the type of host organism from which the product was purified (or, more generally, the source of the contaminant protein background); and the application for which the research was carried out.

We found that nearly 40 distinct types of phase separation method were in use (Figure S1). However, the level of evidence available for each method varies widely: of the 39 methods observed, only 15 were reported in five or more records, while 20 methods were reported in only one or two records. Therefore, most of the observed methods lack the level of replication or generalization to diverse products that would be desirable for making conclusions about their readiness for commercial implementation or comparing them to other methods.

We also broadly classified products purified in the analyzed studies as monoclonal antibodies or antibody fragments, crude immunoglobulin fractions (Ig), enzymes, or none of the former (Figure S2A). While it would be ideal to further classify products based on structural features, the great variety of products in the dataset and the sparsity of sequence or structural information made this classification infeasible within the scope of this study. We additionally classified hosts as bacteria, mammalian cell culture, fungi, plants, or animal-derived samples (e.g., serum), with over half of the data approximately equally divided between the first two categories (Figure S2B). In terms of research application, by far the most common was medical products, representing more than half of the data (Figure S2C). Additional characterizations of the dataset by these categories are available in Figures S3-S7.

In the remainder of this analysis, we will interpret the data with a particular emphasis on using phase separation methods to purify medical products, for three reasons. The first is that this scenario represents such a large fraction of the dataset. Second, medical products make up by far the largest fraction of the overall recombinant protein market. Third, this scenario is associated with the most stringent and most uniform purification requirements as well as the highest manufacturing costs, making the comparison to established alternative methods such as packed-bed chromatography both easiest and most meaningful.

To begin the analysis, we sought to extract information pertinent to the cost and the purification method description and performance (Table S1). However, with respect to key outcome variables, we found that most were reported in 50% of cases or less (Figure S8). Furthermore, this sparsity of important data was not the result of a subset of highly incomplete records. Rather, almost all records in the dataset were substantially incomplete (Figure S9). This represents an important barrier to the progress of the field as well as to the uptake of these methods into commercial processes. For the field to progress, there needs to be more consistent reporting of the key metrics required for a thorough evaluation. Nevertheless, sufficient information could be found to permit quantitative analyses of both purification performance and economics.

3.2. Purification performance of phase separation vs chromatography

Having generally described the contents of the dataset, we next turned to a quantitative evaluation of protein purification performance. The goal of this evaluation was to guide considerations of which phase separation-based methods should be considered for commercial adoption, and which require further study. To this end, we compared phase separation methods both to each other and to two of the most common chromatography-based methods in biopharmaceutical production, ion-exchange (IEX) and Protein A chromatography (ProA). For the purposes of comparison, we obtained data on IEX and ProA performance from the literature (see Tables S2-S4).The first aspect of purification performance we considered was removal of host cell proteins (HCPs), because no other measure of contaminant removal was widely reported in a manner that allowed direct comparison among different studies. HCP removal values (here shown as the log₁₀ reduction value in absolute HCP mass, or LRV) could be identified for 21 different methods (Figure 1A). Of these, 14 were only represented by three or fewer datapoints. For the other 7 methods, the variation within methods was generally greater than the variation among methods. Thus, it was not clear that any one phase separation method or set of methods could be reliably predicted to perform better than another for an arbitrary new product.

Figure 1: Performance of phase separation-based protein purification methods.

(A) Host cell protein removal by method type for phase separation-based methods and two common types of chromatography. Methods are precipitations unless ending with ATPS (aqueous two-phase separation). (B) Product yields of phase separation-based protein purification methods showing that many have demonstrated high yields. LLE: liquid-liquid extraction. ATPS: aqueous two-phase separation. ELP: elastin-like polypeptide. TPP: three-phase partitioning. For both panels, box and whisker plots show the minimum, 25^th percentile, 75^th percentile and maximum values, respectively. Individual datapoints are shown as black circles. The median of each group is highlighted in orange.

When compared to chromatography-based alternatives, 11 and 14 phase separation methods had at least one report of HCP removal greater than the median value for ProA or IEX, respectively. However, median performance compared much less favorably. Among methods with two or more datapoints, only 2 and 5 methods had median HCP LRVs greater than the median for ProA and IEX, respectively. Moreover, 3 of these 5 methods were represented by only two or three datapoints each. Thus, while it is clear that phase separation-based methods can sometimes outperform chromatography in HCP removal, more work is required to validate the robustness and predictability of this performance under product or process variations.

Since monoclonal antibody (mAb) products are the most economically significant area of the reported research in the field, we conducted this evaluation separating mAb products (Figure S10A) from others (i.e., all non-mAb protein products including those produced recombinantly as well as those harvested from animal, plant, or fungal sources) (Figure S10B). For mAb products specifically, while several individual datapoints remained above the median values for ProA and IEX, only three methods had median values greater than the median for IEX. Only one method, also based on affinity interactions, had a median HCP LRV greater than that of ProA. Therefore, in terms of protein contaminant removal, phase separation-based purifications reported to date generally compare more favorably with standard chromatographic methods for non-mAb products than for mAbs.

We also compared phase separation-based protein purification methods on the basis of product yield (Figure 1B). Despite significant variability as for HCP removal, almost all methods had median yields greater than 80%, with most greater than 90%. This suggests that the yield of these methods should not be a barrier to their adoption in commercial protein drug processes, where step yields of 80–90% are generally considered acceptable.

Finally, we evaluated the methods in the dataset based on their ability to remove the other major classes of contaminants frequently encountered: high- and low-molecular weight variants of the product (HMW and LMW, respectively), endotoxin or lipopolysaccharide (LPS), and DNA. These data are summarized in Figure S11. For each of these classes of contaminants, it was found that phase separation-based methods could in some cases achieve reductions of multiple orders of magnitude, making them generally competitive with chromatography. Once again, however, more complete reporting of results is needed to facilitate evaluation of the field against current standard methods.

3.3. Effect of experimental design on method performance

Having observed large variations in purification performance within most of the phase separation methods, we began to investigate whether other factors could explain performance differences. First, we assessed the effect of different approaches to experimental design and optimization. Although the within-groups variation remained large, we were nevertheless able to identify statistically significant differences in the means of both HCP removal and product yield based on the experimental design method used. For HCP removal (Figure 2A), we found that the use of DoE approaches was significantly better than one-factor-at-a-time (OFAT) optimization, with a 53% increase in the mean log₁₀ reduction value (one-way ANOVA followed by Welch’s t-test with Bonferroni correction). For yield (Figure 2B), DoE was significantly better than either no optimization or OFAT optimization, with 16.1% and 11.5% increases in the mean, respectively (one-way Kruskal-Wallis test followed by Mann-Whitney U test with Bonferroni correction). Surprisingly, for both yield and HCP removal, OFAT or multiple-factors-at-a-time (MFAT) optimization did not lead to statistically significant better results than conducting no optimization.

Figure 2: Use of DoE results in improved method performance.

A) Effect on host cell protein removal showing that use of DoE significantly improves performance compared to one-factor-at-a-time optimization. B) Effect on product yield showing that use of DoE significantly improves performance compared to no optimization or one-factor-at-a-time optimization. For both panels, box and whisker plots show the minimum, 25^th percentile, 75^th percentile and maximum values, respectively. Individual datapoints are shown as black circles. Comparisons not shown were not significant (corrected p > 0.05).

3.4. Diversity of use of phase separation agents

One potential advantage of phase separation over chromatography is being able to use and independently control several selective agents with different selectivities in the same operation. Thus, to further examine the diversity in approaches to phase separation-based purification and the use of different selective modes in combination, we separated each method into its constituent selective agents and counted the number of times each agent was used either alone or in each of its possible pairwise combinations. The results of this analysis are shown in Figure 3A. We found 16 distinct agents used to achieve phase separations, which were employed in 45 different combinations of either one or two agents. Thus, the pairwise combination space of agents used to purify proteins by phase separation is only approximately 1/3^rd explored. Furthermore, for the majority of agents, by far the most common case was that the agent was used alone (i.e., lying on the diagonal of Figure 3A). Finally, over 75% of methods using a combination of agents had 5 or fewer records. Therefore, although phase separation-based protein purification methods have a potential advantage over chromatography in their ability to independently and simultaneously select on multiple axes of protein properties, this potential remains largely unexplored or underutilized.

Figure 3: Use of selective agents in phase separation-based protein purification.

(A) The pairwise combination space of selective agents used in phase separation-based protein purification is largely unexplored or underutilized. Each square shows the number of reports that utilize that agent combination. Because the matrix is symmetric, only the upper half is shown for clarity. Tag: fusions to the product protein conferring a non-native selectivity. ELP: elastin-like polypeptides either fused to the product protein or to a binding partner. Responsive polymer: polymers exhibiting sharp phase change behavior as a function of something other than polymer concentration (e.g., pH or temperature). Annotations outside the matrix show both the type of phase-forming agent used and approximate groupings based on predominant mechanisms indicated by the literature. (B) Improvement of HCP removal using an increasing number of separation agents. Box and whisker plots show the minimum, 25^th percentile, 75^th percentile and maximum values, respectively. Individual datapoints are shown as black circles. Following a significant one-way ANOVA, groups were compared by Welch’s t-test with Bonferroni correction for multiple comparisons. Comparisons not shown were not significant (corrected p > 0.05).

Given this, we sought to determine if the use of an increasing number of separation agents in combination could, in general, lead to an increased performance in terms of host cell protein removal. We grouped phase separation methods for which HCP data was available by the number of separation agents they utilized (1 to 3) (Figure 3B). Additionally, a method employing one or two agents was only included if at least one of those agents was also employed in other methods utilizing a greater number of separation agents (two or three agents, respectively). Although we identified only three records employing a combination of three separation agents in the dataset, all three methods demonstrated superior performance compared to those utilizing fewer agents. This underscores the significance of delving into the combinatorial space of phase separation agents for further exploration.

3.5. Techno-Economic Analysis

In addition to purification performance, process economics are a key factor affecting the readiness of phase separation methods for adoption into commercial processes. This is especially true for protein drug products, for which manufacturing costs are especially high and current adoption of phase separation methods is low. Therefore, for each record in the dataset with sufficient information, we conducted a techno-economic analysis by generating a pair of in silico unit operation cost models: one for the prescribed phase separation operation and another for an alternative IEX chromatography operation designed to process the same input stream. We did not model affinity-based chromatography methods, which are not available for most non-mAb products and therefore would not have allowed a relevant comparison to phase separation methods for most of our dataset. The dataset for techno-economic analysis consisted of 72 records representing 14 phase separation methods.

3.5.1. Effect of process scale on cost-effectiveness

After constructing the unit operation models for each record in the dataset, we simulated each at four different scales: with phase separation or chromatography unit operations sized to yield 10, 100, 1000, or 10000 kg of purified product per year. The first three of these scales were chosen to represent reasonable current values for protein drug production because they are approximately the 25th percentile, median, and maximum of current mAb drug production scales (Figure S12). The largest scale is perhaps reasonable for a very small number of products such as insulin,^[24] but was also chosen as a speculative test case that may become relevant for more products in the future.^[25] After obtaining an estimated operating cost per kg of purified product for each model, we then obtained a measure of cost-effectiveness by further normalizing costs per kg of product by the degree of HCP removal achieved, using either the reported HCP LRV for the phase separation method or a median HCP LRV of 1.03 for IEX (see Table S2).

The first noteworthy finding of our techno-economic analysis was that, across all methods, phase separation-based protein purifications were highly unlikely to be cost-effective compared to IEX chromatography at scales of 10 kg/yr or less (Figure 4A). However, each increase in process scale showed a major increase (50% – 182%) in the fraction of phase separation models that were cost-effective compared to their chromatographic counterparts. By far the largest such increase in proportional terms occurred between the 100 and 1000 kg/yr scales (Figure 4B and C, respectively), suggesting that this range of scales may represent a rule-of-thumb “tipping point” for phase separation cost-effectiveness. However, it is also noteworthy that more than 30% of records in the dataset reported phase separation methods that were not predicted to be cost-effective even at extremely large scales of 10000 kg/yr (Figure 4D).

Figure 4: Cost effectiveness of phase separation vs chromatography.

Phase separation methods become more cost-effective than ion-exchange chromatography-based alternatives as production scale increases. Panels show results at four different levels of annual product throughput. Each point represents the cost per kg of product purified normalized by the log₁₀-reduction value (LRV) of host cell protein contaminants, as calculated for a given phase separation operation (x-axis) or the alternative chromatographic operation (y-axis). The red line shows equality of cost-effectiveness between chromatography and phase separation; points above the line are cost-effective compared to chromatography while those below the line are not. Annotations show the number of records in each group.

3.5.2. Effect of input stream purity on cost effectiveness

Another major trend we observed was that phase separation models were much more likely to be less expensive per kg of product yielded than their paired chromatographic models when the product purity in the input stream was low rather than when it was high. For example, of the 16 records in the dataset with an input purity of less than 1%, 100% of the phase separation models were less expensive per unit product yielded than their chromatographic alternatives at all scales above 100 kg/yr (Figure 5). By contrast, among the entire dataset, the fraction of phase separation models which were less expensive than chromatography per unit product was less than 80% even at the largest process scale (Figure 5B). This effect also showed a clear interaction with process scale, with initial purity showing the largest effect on cost at scales on the order of 100 kg/yr (Figure 5B, compare slopes of the curves across scales).

Figure 5: Effect of initial purity on cost-effectiveness of phase separation.

A) Cumulative distribution of records in the dataset by initial purity (i.e., purity of the product in the input stream just prior to the phase separation operation). This information is provided to assist with interpretation of the fractions in panel B. B) Cumulative distribution of records in the dataset for which total cost per kg of product purified is less for the reported phase separation method than for an alternative ion-exchange chromatography operation, as a function of initial purity and annual product throughput. As initial purity of the input stream decreases, phase separation methods become more likely to be less expensive per kg of product purified than an alternative ion-exchange chromatography operation. The fraction at each purity level considers only records with an initial purity less than that value. E.g., at 1% initial purity, the denominator is 16 records (see panel A). Note that the x-axis is a logarithmic scale.

3.5.3. Cost of raw materials and consumables drives process cost

Having observed two trends in the total unit cost and cost-effectiveness of phase separation purification methods vs. chromatographic alternatives, we next sought to explain the mechanisms behind these trends. We found that the best explanation for both trends across the entire dataset was a variation in the direct materials cost (from either chromatographic resin and buffers or phase-forming materials). First, we observed that as process scale increases, the cost of these direct materials makes up an increasing fraction of the total unit operation costs for both chromatography and phase separation models (Figure S13). For all chromatography models and almost all phase separation models, direct materials represent the single largest cost at the 10000 kg product/yr scale. Even at smaller scales, many of the phase separation models are dominated by these costs.

Second, we observed that there is a strong inverse relationship between initial product purity and direct materials costs per kg of product purified for chromatography, but not for phase separations (Figure 6). Moreover, direct materials costs per unit product purified decrease sharply with increasing scale for phase separation methods but much less significantly for chromatography (Figure 6, compare among panels). There is therefore an interaction between initial purity and process scale that affects the difference in direct materials unit costs between chromatography and phase separations. As scale rises and initial purity decreases, more phase separation models fall below their counterpart chromatographic models on the direct materials unit cost axis.

Figure 6: Effect of initial purity and scale on direct materials cost.

Costs of direct materials per kg of product purified are inversely related to initial product purity for chromatography methods, but not for phase separation methods. Direct materials: for phase separation methods, phase-forming materials and buffers; for chromatography, resin and buffers. Panels show four different levels of annual product throughput. Phase separations and chromatography are shown in blue and gray, respectively.

3.5.4. Direct materials usage rate as a predictor of total costs

We have established the importance of direct materials costs in explaining the trends of increasing economic favorability of phase separations over chromatography with increasing process scale and decreasing initial product purity. Next, we wanted to assess the degree to which differences in direct materials usage could also explain differences in cost among phase separation methods. To do so, we turned to the direct materials usage rate defined as the ratio of the mass of phase-forming agents and buffer components to the mass of product yielded. We found that, for all 72 records and at all four modeled process scales, total costs per kg of product purified were significantly correlated with the direct materials usage rate. Least-squares linear regressions of the log-transformed data are shown in Figure 7. Depending on scale, the direct materials usage rate alone explained between 51.4% and 58.4% of the variation in total unit operation costs per mass of product purified (R² of 0.514 – 0.584). In other words, the direct materials usage rate provides a strong and simple predictor of total costs per kg of product purified across all phase separation models. It is also worth noting that the slope of the correlation increased from 0.144 to 0.32 with increasing process scale, such that at the 10000 kg/yr scale, each 10-fold increase in direct materials usage rate would produce on average a doubling (10^0.32 ≈ 2.1) in the total operating costs. Because the direct materials usage rates observed in the dataset spanned more than four orders of magnitude, this is an important effect.

Figure 7: Cost of product as a function of direct materials use.

Cost per kg of purified product for all phase separation methods is directly related to the ratio of purification reagent mass to purified product mass. Black lines and inset boxes show the line of best fit and parameters for a linear regression of the log-scale data. Panels show results at four different levels of annual product throughput.

Having identified a good global predictor for the total cost per unit of product purified for phase separations as a whole, we next explored the variability in cost and cost-effectiveness when comparing different methods for achieving phase separation. With respect to total unit operation costs (Figure S14), we found three trends. First, the relative ordering of various phase separation methods by cost varied notably with scale. In general, as scale increased, although absolute costs for all methods declined, methods using no or few phase-forming materials (e.g., those based on factors such as pH and temperature) became relatively less expensive while those relying on phase-forming chemicals (e.g. inorganic salt-based precipitations) became relatively more expensive. Second, the variability within a given method generally increased with increasing scale. Finally, the range of costs across method types was either nearly or completely overlapping at all scales. This suggests again, as for HCP removal, that there is no single best or worst phase separation method in general, but that the within-method variation is most important. With respect to cost-effectiveness ($ per kg of product per log₁₀ reduction in HCPs), we note that the same themes of within-method variability applied (Figure S15), while the rank order of methods was substantially different than for cost alone (Figure S14-15).

Next, we sought to test whether direct materials usage rate could also explain the cost variations among different examples of the same phase separation method. We selected four methods that showed the highest cost variability based on Figure S14—polymer/salt ATPSs as well as precipitations based on polymers, polyelectrolytes, and inorganic salts—and used linear least-squares regression to fit the cost predictions for all examples of each method to each example’s direct materials usage rate. The results of this analysis at both the largest and smallest modeled process scales are shown in Figure S16. Of these 8 regressions, all were highly significant (p < 0.005) except that for polymer/salt ATPSs at the largest process scale, which also approached significance (p ≈ 0.15). More importantly, of the seven significant regressions, all explained more than 80% of the within-method variability in process costs, while the best explained more than 98% (R² = 0.81 – 0.984). Therefore, the direct materials usage rate almost entirely explained within-method cost differences.

3.5. Implementation-Readiness of Phase Separation Methods

Finally, having compared the phase separation methods in the dataset on the basis of purification performance and economics, we sought to combine these measures to form a single assessment of the degree to which phase separation methods were currently ready to be implemented into commercial processes for protein drugs. Therefore, we defined three categories of implementation readiness. The first includes methods that are not cost-effective compared to conventional IEX chromatography, as measured by cost per kg of purified product per log₁₀ reduction in HCPs. These methods are therefore not generally expected to be suitable at present for adoption in protein drug processes, because adding additional separation capacity by chromatography would be more cost-effective. The second includes methods that are cost-effective compared to IEX chromatography but provide less overall HCP removal, and therefore may be considered as supplements or partial replacements to chromatography. Finally, the third category contains methods that have the potential to entirely replace at least one standard chromatographic unit operation in a downstream process, because they are both cost-effective compared to IEX and have HCP LRVs greater than the median for IEX.

First, we compared implementation readiness as a function of the year of publication of each study in the techno-economic analysis dataset (N = 72) as well as of process scale (Figure S17). As an encouraging sign for the field, we observed that the rate of “implementation-ready” phase separations has been increasing over time. For example, at the 10 kg/yr scale, only one such example was published in the 16 years between January 2000 and December 2015, while 5 were published in the 6 years between January 2016 and January 2022.

Next, we assessed whether differences in phase separation method were associated with differences in implementation readiness (Figure 8). At the 10 kg/yr and 100 kg/yr production scales, respectively, only 3 and 5 of the 14 method types had any implementation-ready examples. While most methods had at least one implementation-ready example at the 1000 kg/yr scale, we note that the great majority of current protein drug processes operate below this scale (Figure S12). Notably, the methods most likely to be cost-effective or potential chromatography replacements were all precipitations, and included those based on polyelectrolytes, inorganic salts, polymers, and the combination of heat, inorganic salts, and extreme pH. For more details on which methods are implementation-ready or not, refer to Table S6.

Figure 8: Implementation readiness as a function of method type and annual product throughput.

Methods are precipitations unless ending in ATPS (aqueous two-phase separation). For further information on the definitions of technology readiness categories, refer to the text.

Then, because the approach to experimental design and optimization was found to significantly affect product yield and HCP removal, we tested whether the same factors were associated with increased likelihood of implementation-readiness. In general, it did not appear that more advanced optimization methods (e.g., MFAT and DoE approaches) were associated with greater rates of technology readiness than simple OFAT optimization when optimizing either for HCP removal (Figure S18) or for product yield (Figure S19). Furthermore, optimization of these variables by any method was not clearly better than not optimizing them at all, except at large scales (e.g., 1000 kg/yr, Figure S18C and S19C). We note that there were no examples in the entire analysis dataset (N = 290) where the authors optimized directly for cost-effectiveness or for lowering direct materials usage.

4. Discussion

Methods of protein purification based on modulating solubility among different chemical phases (i.e., phase separations), vs. standard alternatives such as chromatography, have the potential to simplify protein manufacturing and reduce its costs. This is especially important for protein drug manufacturing, where process costs and complexity are highest. However, while phase separations are in common use in other areas such as industrial enzyme production, they have not been adopted in protein drug manufacturing outside of a small number of cases.^[14,17] Our purpose in this study was to identify the reasons for this lack of implementation, to assess the current implementation-readiness of technologies in the field, and to offer evidence-based recommendations for future work to help increase implementation-readiness.

Our key findings can be summarized as a set of three guidelines.

4.1. Phase separations often do not cost less than chromatography; implementation should focus on applications with large scale and low input purity

Previous reviews have primarily offered two reasons for the lack of uptake of phase separation methods in protein drug manufacturing: lack of mechanistic and process engineering knowledge;^{[9,14,16,18,19]} and insufficient or inconsistent selectivity.^[19] Meanwhile, the economic aspects of using phase separation vs. standard alternatives such as chromatography in the protein drug context are usually treated only briefly. In fact, it is often seemingly taken for granted that phase separation methods will be less expensive than chromatography. Although some recent studies do treat this question in depth, they only consider processing scales much larger than those commonly used for protein drugs^[26] and/or only one case study of a particular phase separation method and product protein.^[4,26]

In contrast to previous work in the field, we present the first in-depth techno-economic analysis of the full range of demonstrated use-cases for phase separations in protein purification. To our surprise, and in contrast with the conventional understanding, we found that phase separation methods are in fact often economically unfavorable vs. chromatography under a wide range of process conditions applicable to most currently marketed protein drugs. Furthermore, although our definition of cost-effectiveness incorporates the degree of HCP removal achieved, this unfavorable economic comparison is not due solely to a lack of selectivity in phase separation methods, because it persists when accounting only for total costs rather than cost-effectiveness (Figure 5B).

Although the total costs and cost-effectiveness of phase separation methods varied widely, the evidence provides good rules of thumb to predict situations that favor phase separation vs. chromatography. We found that the most favorable process conditions for phase separation cost-effectiveness were large scale (Figure 4), especially greater than 1000 kg/yr, and low initial purity of the product in the input stream (Figure 5B). Moreover, these two factors have an interaction, where initial purity is most important at moderately large scales (e.g., 100 and 1000 kg/yr) (see Figure 5B).

Therefore, near-term efforts to implement phase separation methods in protein drug processes should focus primarily on cases with large process throughputs (e.g., 100s of kg product/yr or more) where a phase separation can replace or supplement a chromatography step that receives relatively low-purity input material. Furthermore, and especially outside of this regime of large scale and low input purity, researchers should not take the cost advantages of phase separation methods for granted. Rather, some kind of economic analysis should be routinely considered during phase separation method development alongside analysis of purification performance.

Although the trends we have highlighted hold true for our dataset, it is also important to note certain limitations of our economic analysis. First, the dataset is simply not as large as would be desirable: only 72 of 290 records we reviewed reported the information necessary to analyze cost-effectiveness. Second, because of practical limitations and the unavailability of relevant data, our analysis accounts only for the operating costs directly incurred by a phase separation method or by its hypothetical chromatographic alternative and only for performance in removing HCPs, not other contaminants. Therefore, phase separations may be more favorable than indicated by our analysis when they can provide additional advantages beyond direct operating costs. For example, they may reduce capital expenses, enable more flexible manufacturing, or simultaneously remove multiple classes of contaminants that are not easily separated by a single chromatography step.

4.2. The direct materials usage rate can be used alongside detailed techno-economic analysis as a simple predictor of phase separation process costs

Although it is important to consider phase separation methods from an economic perspective, because even those with good purification performance may not be cost-effective, it is not feasible that every study should include a detailed economic analysis. Therefore, it is highly desirable to have simple and effective predictors of phase separation operating costs which can facilitate approximate estimation and optimization. Our findings revealed one such predictor in the direct materials usage rate, i.e., the mass ratio of phase-forming materials and buffer components to purified product. This metric is easy to calculate, requires no information besides the mass balance of the phase separation operation, and quantitatively explains cost variations both across phase separation method types and within a given method type. Of course, it does not necessarily apply to methods that use strictly or primarily non-chemical means such as temperature or pressure to induce phase changes.

While we believe that the importance of the direct materials usage rate for phase separation cost-effectiveness has not been clearly stated before, we find that it does emerge from a combination of well-known phenomena in the field. First, as process scale increases, direct materials costs come to dominate total costs for both phase separations and chromatography (Figure S13). This is because these materials scale more closely with the amount of product processed than do other costs such as equipment or labor.^[8] Second, as process scale increases, the costs of direct materials per unit of product purified falls sharply for phase separations but much less strongly for chromatography (Figure 6). This is due to the fact that chromatographic resins do not enjoy significant economies of scale in these scale regimes,^[25] whereas raw materials used in phase separations more commonly do.^[8,27] Third, there is a strong dependence of the direct materials costs on the purity of the input stream for chromatography, but not for phase separations (Figure 6). This is because chromatography involves a direct interaction between the resin and both the product as well as at least some contaminants, so that the resin mass required is directly related to the mass of protein in the input stream and inversely related to its purity. In contrast, many materials used to induce phase separations are based on changing the properties of the solvent rather than on interacting with proteins directly.^[28] Therefore, the required mass of these materials scales directly with the process volume and inversely with initial product concentration, but not necessarily with the protein mass or purity of the input stream. While this difference in scaling behaviors has been noted before,^[14,29] neither its implications for initial product purity nor its detailed impact on total process costs have previously been reported. The net effect of these three factors is to make the direct materials usage rate a key driver of cost differences between phase separation and chromatography, especially as scale increases and input purity falls.

4.3. Design of Experiments, multi-modal techniques, and greater mechanistic understanding present key opportunities for improvement of phase separation methods

It has been frequently noted that phase separation methods generally involve complex, poorly understood mechanisms and are not amenable to predictive design.^[16,19] These facts intersect with our results in at least three ways.

First, in line with previous reports suggesting the value of DoE for phase separation method development, we found that DoE-based approaches to method optimization were associated with significantly greater HCP removal and yield than other approaches. This makes sense given that DoE is most valuable precisely when an outcome depends on many factors and rational design is not feasible. Interestingly, we did not find that the approach used to optimize phase separation methods predicted implementation-readiness. This may be because researchers were optimizing only for purification performance and not for cost-effectiveness, because important optimizations had already been done prior to a particular publication and so were not reported, or because the sample size is insufficient.

Second, we observed high variability in the purification performance and predicted costs of phase separations, even among examples of the same general method. In fact, the variation within each method was often even greater than the variation between methods (see, e.g., Figure 1, Figure S14). This observation fits with the notion that mechanistic understanding is lacking: without predictive models to match known properties of a protein with known selectivity mechanisms of a phase separation method, sub-optimal methods will often be chosen for a given protein, and optimal conditions may not be reached even when the best general method is chosen; this would result in a general spreading of the range of outcomes for all methods, which is what is observed.

Third, we found that phase separation methods using three or more different selective agents (“multi-modal” phase separations) had much greater HCP removal than those using only one or two agents. Although this finding is based on limited available data, it presents both an interesting challenge and opportunity. The challenge is that adding more selective modes only adds more complexity to a situation already suffering from a lack of mechanistic understanding. On the other hand, since there are so many ways to achieve phase separation (see Figure 3A) and testing them in combination is often as simple as adding more components to a solution, the opportunities for combining selective modes seem to be vast and likely greater than for chromatography. Therefore, as mechanistic understanding and design rules for phase separations continue to advance, a greater and greater potential advantage of multi-modal phase separations over chromatography should be unlocked.

Highlights.

• Phase separation purification methods are not always cheaper than chromatography

• The use of a combination of phase separating agents remains largely underexplored/underutilized

• Lower initial purity and increasing process scale favor phase separation over chromatography

• The direct material usage rate is an important predictor of phase separation cost-effectiveness

• Future work in the field should report, and ideally optimize, the direct material usage rate