High‐throughput amino acid‐level characterization of the interactions of plasminogen activator inhibitor‐1 with variably divergent proteases

Abstract

While members of large paralogous protein families share structural features, their functional niches often diverge significantly. Serine protease inhibitors (SERPINs), whose members typically function as covalent inhibitors of serine proteases, are one such family. Plasminogen activator inhibitor‐1 (PAI‐1) is a prototypic SERPIN, which canonically inhibits tissue‐ and urokinase‐type plasminogen activators (tPA and uPA) to regulate fibrinolysis. PAI‐1 has been shown to also inhibit other serine proteases, including coagulation factor XIIa (FXIIa) and transmembrane serine protease 2 (TMPRSS2). The structural determinants of PAI‐1 inhibitory function toward these non‐canonical protease targets, and the biological significance of these functions, are unknown. We applied deep mutational scanning (DMS) to assess the effects of ~80% of all possible single‐amino acid substitutions in PAI‐1 on its ability to inhibit three putative serine protease targets (uPA, FXIIa, and TMPRSS2). Selection with each target protease generated a unique PAI‐1 mutational landscape, with the determinants of protease specificity distributed throughout PAI‐1's primary sequence. Next, we conducted a comparative analysis of extant orthologous sequences, demonstrating that key residues modulating PAI‐1 inhibition of uPA and FXIIa, but not TMPRSS2, are maintained by purifying selection (also referred to as “negative selection”). PAI‐1's activity toward FXIIa may reflect how protease evolutionary relationships predict SERPIN functional divergence, which we support via a cophylogenetic analysis of secreted SERPINs and their cognate serine proteases. This work provides insight into the functional diversification of SERPINs and lays the framework for extending these studies to other proteases and their regulators.

1. INTRODUCTION

Large paralogous protein families in which members share a common protein structure are ubiquitous across all clades of life. Diversity often results from gene duplication events that occurred during either whole genome duplication or as segmental duplications of one or more genes (Hahn 2009). To acquire and maintain distinct functions despite similarities in sequence and structure, members of these paralogous protein families must acquire unique functional niches (McClune et al. ²⁰¹⁹; McClune and Laub ²⁰²⁰; Nocedal and Laub ²⁰²²). Revealing the shape and complexity of these niches is crucial to understanding the emergence and insulation of protein–protein interactions.

Serine protease inhibitors, or SERPINs, are a protein superfamily with high functional variability and specificity. Inhibitory SERPINs are defined by their ability to efficiently and specifically inhibit one or more members of the highly paralogous serine protease protein family. Both protein families are found across all domains of life, where SERPINs covalently bind serine proteases, irreversibly inhibiting them (Spence et al. 2021). In humans and other vertebrates, SERPINs have been shown to regulate a range of processes, including immune/inflammatory responses, coagulation/fibrinolysis, and extracellular matrix remodeling (Sanrattana et al. 2019). While all SERPINs share a common protein fold, most SERPINs perform their biological function as specific and irreversible inhibitors of serine proteases via a “mousetrap” mechanism (Figure 1) (Law et al. 2006).

FIGURE 1.

SERPINs inhibit serine proteases via a “mousetrap” mechanism to inhibit their primary and non‐canonical protease targets. In its metastable active state, the SERPIN's reactive center loop (RCL, yellow) extends from its central globular core and contains a preferred cleavage site for proteases targeted for inhibition (green). Under conditions favoring the inhibitory pathway (left pathway), after the protease cleaves the scissile bond in the RCL, the acyl intermediate is trapped when the RCL inserts into β‐sheet A forming a fifth β‐strand. Alternatively, the protease cleaves the RCL and the acyl intermediate is hydrolyzed resulting in the insertion of the RCL into β‐sheet A and a non‐functional SERPIN, while the protease remains active (center pathway). Unique among SERPINs, PAI‐1 can also undergo a spontaneous latency transition in which the RCL inserts into β‐sheet in the absence of a proteolytic event, rendering the SERPIN inactive (right pathway).

In their active state, SERPINs exist in a metastable conformation in which the reactive center loop (RCL), which contains an amino acid sequence that mimics that of their target serine protease's preferred cleavage site, extends from the globular core of the protein. Upon SERPIN/protease engagement, irreversible inhibition of the serine protease occurs when the RCL inserts into the SERPIN's central β‐sheet A prior to the resolution of the acyl intermediate; however, if hydrolysis of the acyl intermediate occurs prior to RCL insertion, the SERPIN serves as a substrate rather than an inhibitor of the serine protease—both of these conformations occupy a lower energy space (Figure 1) (Gettins 2002; Huntington ²⁰¹¹; Law et al. ²⁰⁰⁶; Lawrence et al. ¹⁹⁹⁵). Although a SERPIN's protease specificity is derived in part from the RCL amino acid sequence, structural features located distal to the RCL also contribute significantly (Gettins and Olson 2009; Marijanovic et al. ²⁰¹⁹). Some SERPINs appear to be highly specific inhibitors of a single protease, while others exhibit activity toward additional protease targets, often with decreased efficiency compared to their canonical protease target (Gettins 2002). Exosite interactions, as well as a limited number of mutations, are also known to alter SERPIN inhibitory profiles (Gettins 2002; Huntington ²⁰¹¹). The specificity of SERPINs for their target serine protease(s), provides not only a novel system for studying the development of functional niches in the context of paralog evolution, but also one where the determinants of success or failure of the protein–protein interaction are complex.

Plasminogen activator inhibitor‐1 (PAI‐1, encoded by the SERPINE1 gene) is a 379‐amino acid SERPIN canonically considered to be a regulator of fibrinolysis, the enzymatic degradation of fibrin clots, by inhibiting two closely related proteases, urokinase and tissue‐type plasminogen activators (uPA [encoded by the PLAU gene] and tPA [encoded by the PLAT gene], respectively) with high efficiency (second‐order rate constants of ~10⁶–10⁷ M⁻¹ s⁻¹; Sherman et al. ¹⁹⁹²) and a stoichiometry of inhibition (SI) of approximately one (Lawrence et al. 2000). PAI‐1 has also been identified as a possible inhibitor of other hemostatic proteases, including the procoagulant factor (F) XIIa (encoded by the F12 gene), which stands in stark juxtaposition to PAI‐1's primary function as an inhibitor of the clot‐dissolving proteases tPA and uPA (Berrettini et al. 1989; Keijer et al. ¹⁹⁹¹; Puy et al. ²⁰¹⁹; Rezaie ²⁰⁰¹; Sen et al. ²⁰¹¹; Tanaka et al. ²⁰⁰⁹). PAI‐1 also inhibits transmembrane serine protease 2 (TMPRSS2; Figure S1, Supporting Information), through which it mediates the course of viral infections such as influenza and SARS‐CoV‐2 (Dittmann et al. 2015; Shen et al. ²⁰¹⁷), with increased PAI‐1 levels reported in response to influenza (Bouwman et al. 2009; Keller et al. ²⁰⁰⁶) and SARS‐CoV‐2 infections (Al‐Samkari et al. ²⁰²⁰; Zuo et al. ²⁰²¹).

In the present work, we use deep mutational scanning (DMS) to analyze the mutational landscape of PAI‐1 inhibition of several proteases, including uPA and two non‐canonical targets of PAI‐1 (FXIIa and soluble TMPRSS2), characterizing the overall degree and potential regionalization of marginal specificity in these interactions. We follow prior work that defines protein sequence space as the set of all possible amino acids in a protein coupled with the possible evolutionary trajectories to a new sequence (McClune and Laub ²⁰²⁰). Within a sequence space, robust and marginal specificity refer to members of paralogous protein superfamilies for which multiple mutations are required to alter their specificity and cases where a single amino acid substitution can alter a paralog's specificity, respectively (Ghose et al. 2023). We then use comparative evolutionary analyses to contextualize the results of our DMS screen to explore the relationships between protease conservation, paralogous protein relationships, and the sequence space in which these proteins interact.

2. RESULTS AND DISCUSSION

2.1. Construction and characterization of a PAI‐1 variant library on the I91L background

We constructed a novel phage display library as previously described (Haynes et al. 2022; Huttinger et al. ²⁰²¹), though with the addition of an I91L PAI‐1 variant backbone to increase the half‐life of PAI‐1 in its metastable active conformation from 1–2 to ~19 h (Figure 1) (Berkenpas et al. 1995; Haynes et al. ²⁰²²). The I91L background was chosen to isolate the effects of mutants on PAI‐1 serine protease target specificity from those impacting functional stability.

The I91L PAI‐1 variant library exhibited a depth of 6.6 × 10⁶ unique phage clones with an average of three missense or nonsense mutations per clone (Figure S2a; determined as described in section 4) (Huttinger et al. 2021).

Of the 7201 possible single‐amino acid substitutions in PAI‐1 (379 sites × 19 alternative amino acid substitutions), 5688 (79%) are represented in the I91L variant library as determined by high‐throughput sequencing (HTS). The abundance of each variant is highly correlated (slope = 0.95, R ² = 0.85; Figure S2b) with our previous variant library on the wild‐type (WT) PAI‐1 background (Huttinger et al. 2021). Log₂‐fold enrichment scores, as determined by comparing the abundance of each variant in the input versus protease selected library using the DESeq2 statistics package (Haynes et al. 2022; Huttinger et al. ²⁰²¹; Love et al. ²⁰¹⁴; Zhu et al. ²⁰¹⁹), for uPA functional inhibitory variants in the I91L PAI‐1 library are also highly correlated with those previously determined (Huttinger et al. 2021) on the WT PAI‐1 background (Figure S2c; R ² = 0.66, slope = 0.45).

2.2. Single‐amino acid substitutions in PAI‐1 differentially affect its ability to inhibit target proteases

We next screened our I91L PAI‐1 variant library to identify the effects of single‐amino acid substitutions in PAI‐1 on its ability to inhibit FXIIa and TMPRSS2 in addition to uPA (Figure 2a, Data S1). Inhibitory PAI‐1 variant selection was determined by incubating the I91L PAI‐1 variant library with uPA, biotinylated FXIIa (FXIIa‐biotin), or N‐terminal 6xHis tag TMPRSS2 and selecting inhibitory PAI‐1 by immunoprecipitation of the target proteases, as described in section 4. Selection with each protease generated a unique PAI‐1 variant map (Figures 2a, S3, and S4).

FIGURE 2.

PAI‐1 DMS and differential patterns of serine protease inhibition. (a) PAI‐1 specificity fingerprints are shown for uPA, FXIIa, and TMPRSS2. The PAI‐1 amino acid positions are shown along the x‐axis with potential amino acid substitutions shown along the y‐axis. The amino acid residues encompassing the RCL are also highlighted. Tolerated mutations are shown in blue (log₂‐fold enrichment score >0), while loss‐of‐function mutations are shown in red (log₂‐fold enrichment score ≤0). The amino acids of the canonical human PAI‐1 sequence are shown in dark gray. Amino acids that were either not present or not present at sufficient levels to quantify (BaseMean ≥50 as determined by DESeq2; Love et al. ²⁰¹⁴) in the phage library are shown in white. (b) Heatmap showing the common amino acid substitutions across PAI‐1's specificity spaces for uPA, FXIIa, and TMPRSS2 with positive (blue) and negative (red) functional enrichment scores. (c) PCA plot, with individual points representing “n” biological replicates of each condition, comparing the specificity spaces of PAI‐1 for uPA (pink, n = 11), FXIIa (yellow, n = 4), TMPRSS2 (blue, n = 3), and the starting I91L PAI‐1 DMS library (black, n = 11). Groups are annotated with the minimum enclosing ellipse. Venn diagrams depicting the number of significantly (d) enriched and (e) depleted amino acid substitutions shared and unique to each specificity fingerprint.

The statistically significant functional log₂‐fold enrichment scores are most similar between uPA and FXIIa, although there is a distinct set of mutations that maintain inhibitory activity toward all three proteases (Figures 2b and S3). Similarly, principal component analyses (PCA) of the differential variant enrichment data for PAI‐1 specificity toward each of the target proteases (Figure 2c) are most similar for uPA and FXIIa, with TMPRSS2 yielding a relatively distinct pattern of enrichment (Figure S4). The large separation of the principal components required for TMPRSS2 inhibition from that of uPA and FXIIa is consistent with the closer phylogenetic relationship between the latter two proteases (Yousef et al. 2003) and indicative of more distinct amino acid substitutions required to render PAI‐1 a specific/efficient TMPRSS2 inhibitor. We further examined the unique set of amino acid substitutions that render PAI‐1 capable (Figure 2d) or incapable (Figure 2e) of inhibiting each of these three target proteases using the statistical thresholds described above. Again, the set of amino acid substitutions in PAI‐1 that promote/maintain or result in a loss of uPA or FXIIa inhibition are more similar to each other than either is to those that impact TMPRSS2 inhibition. More amino acid substitutions pass the statistical threshold (p _adj < 0.1 and BaseMean score ≥ 50) for uPA inhibition (n = 1369) than for either FXIIa (n = 972) or TMPRSS2 (n = 704) (Table 1), suggesting that the more distantly related a protease is to PAI‐1's canonical target uPA, the fewer possible “pathways” to inhibition exist.

TABLE 1.

PAI‐1 variants that lead to maintenance or decrease of inhibitory capacity.

Target protease	Number of variants that maintain/ inhibitory capacity	Number of variants resulting in loss of inhibitory capacity
uPA	541	828
FXIIa	349	623
TMPRSS2	513	191

Note: Variants identified that pass significance thresholds (p _adj < 0.1, base mean score ≥50).

The results of our DMS screens of PAI‐1 inhibition of three distinct serine protease targets suggest that multiple single‐amino acid substitutions may result in PAI‐1 being a TMPRSS2 inhibitor with impaired or reduced inhibition of its canonical target uPA (Figure 2b). In other words, our results suggest that PAI‐1 may exhibit marginal specificity with respect to TMPRSS2 inhibition as multiple single amino acid substitutions appear to alter its protease specificity (Ghose et al. 2023; McClune and Laub ²⁰²⁰). In contrast, the log₂‐fold enrichment scores of the PAI‐1 variant library with respect to uPA and FXIIa inhibition closely mirror each other (Figure 2). We hypothesize that this pattern is the result of PAI‐1 exhibiting robust specificity for the closely related serine proteases uPA and FXIIa in which single‐amino acid substitutions do not dramatically alter its protease inhibition profile. However, it remains unclear if PAI‐1's inhibition of FXIIa results from cross‐reactivity between two closely related proteins or if PAI‐1 is being actively selected to be an inhibitor of both uPA and FXIIa (Conant and Wolfe 2008; Yousef et al. ²⁰⁰³). In the latter scenario, there is an active selection for PAI‐1 to inhibit FXIIa, while in the former, PAI‐1 cross‐reactivity with FXIIa may be the result of an inherent difficulty in SERPINs evolving insular functional niches with respect to closely related serine proteases (McClune et al. ²⁰¹⁹; McClune and Laub ²⁰²⁰).

2.3. Specificity determinants span PAI‐1's primary structure

To interrogate the role of multiple regions of PAI1 in protease specificity, we next performed separate PCAs on each of the twelve 150 base pair (bp) sequencing amplicons (Figure 3 and Table S1) (Haynes et al. 2022; Huttinger et al. ²⁰²¹). Across all 12 amplicons, principal components (PC) 1 and 2 discriminate between the log₂‐fold enrichment scores with respect to inhibition of uPA, FXIIa, and TMPRSS2, demonstrating that SERPIN regions beyond the RCL (residues 331–350 of human PAI‐1 located within amplicon 12) (Gettins and Olson 2009) contribute to target protease specificity. Previous reports of chimeric PAI‐1 variants with RCLs from SERPINs targeting alternative proteases while retaining the ability to inhibit uPA/tPA are consistent with this finding (Lawrence et al. 1990).

FIGURE 3.

Determinants of PAI‐1 specificity are found throughout its sequence space. (a) A cartoon of the PAI‐1 primary amino acid sequence showing which amino acids were included in each amplicon. Amplicon 12 contains PAI‐1's RCL. (b) PCA plots, with individual points representing biological replicates for each condition, comparing the specificity spaces of PAI‐1 for uPA (pink, n = 11), FXIIa (yellow, n = 4), TMPRSS2 (blue, n = 3), and the I91L PAI‐1 variant library (black, n = 11) across PAI‐1's 12 amplicons. Groups are annotated with the minimum enclosing ellipse.

In contrast, PAI‐1 variants with amino acid substitutions in the RCL have been reported to alter PAI‐1's specificity to be either a neutrophil elastase or cathepsin G inhibitor (Stefansson et al. 2004), similar to the naturally occurring Pittsburgh variant of the SERPIN α₁‐antitrypsin (A1AT, SERPINA1), in which a single amino acid substitution at the P1 position in the RCL converts it from an elastase to a thrombin inhibitor (Lewis et al. 1978; Owen et al. ¹⁹⁸³), or engineering efforts on other SERPIN backbones to generate SERPINs with novel inhibitory profiles (Bhakta et al. 2021; Polderdijk et al. ²⁰¹⁷; Sanrattana et al. ²⁰²¹; Scott et al. ²⁰¹⁴; Singh et al. ²⁰²²). Therefore, although single‐amino acid substitutions in a SERPIN's RCL may function in a manner consistent with marginal specificity for its target protease, our data suggest that multiple amino acid substitutions throughout the SERPIN's primary sequence may be required to achieve robust specificity for a SERPIN as an efficient and specific inhibitor of a given protease. Furthermore, these variants may contribute to the coevolution of specific SERPIN: serine protease inhibitory reactions through as of yet not fully understood epistatic mechanisms (Ding et al. 2022; Park et al. ²⁰²²; Spence et al. ²⁰²¹) in which the effect of a given amino acid substitution is dependent upon the identity of one or more other amino acids in the protein's primary structure (Starr and Thornton 2016).

2.4. Evolutionary conservation of PAI‐1 inhibitory functions

Next, to better understand the evolution of PAI‐1's inhibitory niche, we compared the results of our DMS screens to PAI‐1 sequences of extant mammalian species (n = 94). We previously demonstrated that key residues in PAI‐1 for inhibition of uPA are conserved in nature by comparing the results of our DMS screen to natural sequence variation across mammals (Huttinger et al. 2021). We extended these analyses to the present screen to assess whether sites mediating PAI‐1 inhibition of FXIIa and TMPRSS2 exhibit similar selection pressures. An evolutionary conservation score for a given site in the PAI‐1 amino acid sequence was determined (Ashkenazy et al. 2016; Huttinger et al. ²⁰²¹) and correlated with the tolerance of a site to accept mutations that allow PAI‐1 to inhibit a target protease (denoted henceforth as the normalized functional score defined as the mean log₂‐fold enrichment score determined at each position). If there is purifying (negative) selection for PAI‐1 to maintain the inhibition of a given target protease, then we expect that sites in PAI‐1 at which amino acid substitutions do not negatively impact the ability to inhibit the target protease will be less conserved across extant species. Likewise, sites at which amino acid substitutions render PAI‐1 unable to inhibit the target protease will be more conserved, that is, purifying selection should remove harmful genetic variants from the population.

The results of these analyses corroborate our earlier findings that PAI‐1 inhibition of uPA is under purifying selection (Huttinger et al. 2021) with a modest, statistically significant correlation between the normalized functional score and evolutionary conservation score across PAI‐1 specificity space (slope = 0.58, R ² = 0.1427, p = 5.6 × 10⁻¹⁴; Figure 4a). With respect to FXIIa inhibition by PAI‐1, the functional and evolutionary conservation scores associate with a smaller slope and weaker, yet statistically significant, correlation (slope = 0.33, R ² = 0.066, p = 7.5 × 10⁻⁷; Figure 4b). However, no statistically significant association was observed between functional and evolutionary conservation scores with respect to PAI‐1 inhibition of TMPRSS2 (slope = −0.01, R ² = 8.7 × 10⁻⁵, p = 0.87; Figure 4c). First, these results suggest that, as expected, PAI‐1 inhibition of uPA is under purifying selection. Second, either through FXIIa's close evolutionary relationship with uPA or via direct selection for activity against FXIIa, sites impacting FXIIa inhibition are also under purifying selection, despite the poor kinetics of PAI‐1 inhibition of FXIIa compared to uPA. Meanwhile, sites impacting the inhibition of TMPRSS2 are not, on average, under differential selection compared to other sites in the molecule. These results suggest that PAI‐1 inhibition of TMPRSS2 is potentially an example of cross‐functional SERPIN reactivity resulting in inhibition of a non‐canonical target and that inhibition of TMPRSS2 is not a biologically significant function of PAI‐1 that generates detectable signatures of selection across species.

FIGURE 4.

PAI‐1 is under purifying selection to inhibit uPA and FXIIa but not TMPRSS2. Conservation scores (Ashkenazy et al. 2016) for each amino acid position in PAI‐1 (gray dots) are shown as a function of the normalized functional scores determined in our DMS screens of PAI‐1 inhibition of (a) uPA (R ² = 0.14, p = 4.6 × 10⁻¹⁴), (b) FXIIa (R ² = 0.06, p = 7.5 × 10⁻⁷), and (c) TMPRSS2 (R ² = −0.003, p = 0.87). Dashed lines indicated the best fit linear regression with the 95% confidence interval shown in gray shading.

Given that a greater proportion of PAI‐1 sequence variation across extant species is explained by a drive to maintain PAI‐1's ability to inhibit uPA than FXIIa, we speculate that FXIIa inhibition by PAI‐1 may be the result of cross‐reactivity between uPA and FXIIa, given their close phylogenetic relationship, while TMPRSS2 represents a distinct outgroup to the uPA/FXIIa clade (Yousef et al. 2003). Therefore, while neither FXIIa nor TMPRSS2 is inhibited as efficiently by PAI‐1 as uPA (Figure 5c) (Keijer et al. 1991), the close phylogenetic relationship between uPA and FXIIa permits inhibition of both to be retained in the presence of the same amino acid substitutions in the PAI‐1 primary sequence. In contrast, the ability of PAI‐1 to inhibit TMPRSS2 is not retained in the face of amino acid substitutions that are tolerated with respect to both uPA and FXIIa inhibition. The F12 gene appeared in stem tetrapods after their divergence from fish, and therefore interactions between PAI‐1 and uPA predate the existence of FXIIa (Ponczek et al. 2008; Ponczek et al. ²⁰²⁰). Insulation of PAI‐1 from crosstalk to FXIIa may have been prevented by the evolutionary inertia of millions of years of preceding coevolution between PAI‐1 and plasminogen activators.

FIGURE 5.

PAI‐1 variants that are most enriched for FXIIa inhibition do not exhibit improved SIs. Volcano plots showing the I91L PAI‐1 variants activity toward (a) FXIIa and (b) uPA with respect to the input variant library. Dark blue circles indicate statistically significant enrichment scores (p < 0.005). The three most enriched PAI‐1 variants (I91L/F98Y, I91L/T251I, and I91L/S331C) are shown in pink. The Log₂‐fold change is plotted on the x‐axis and the statistical significance (−Log₁₀(p _adj)) is plotted on the y‐axis. (c) SIs were determined for each of the variants for both uPA (green) and FXIIa (pink) inhibition. For each variant, the SI for uPA was set to one to account for inactive PAI‐1 present in the protein preparation. T‐tests were used to determine statistical significance (*p < 0.05; **p < 0.01; ***p < 0.001). (d) FXIIa normalized functional mutation scores are compared to the uPA normalized functional mutation scores for each variant. The linear regression of the data is shown with the dashed black line (slope = 0.78, R ² = 0.43, p = 2.2 × 10⁻¹⁶), while the conservation score is indicated by the color of each point as indicated in the key. Residues 98, 251, and 331 are labeled and outlined in black.

2.5. Comparing the effects of single amino acid substitutions in PAI‐1 on the inhibition of uPA and FXIIa

Although PAI‐1 inhibits both uPA and FXIIa, PAI‐1 inhibition of uPA is anti‐fibrinolytic (stabilizing the thrombus), while inhibition of FXIIa is anti‐thrombotic. One potential explanation for these opposite effects is that due to the recent divergence of uPA and FXIIa, PAI‐1's ability to inhibit uPA has been “carried over” into its ability to inhibit FXIIa, as evidenced by the similarity in PAI‐1's specificity fingerprints for both uPA and FXIIa (Figures 2 and 5). To further address the similar effects of amino acid substitution in PAI‐1 on its ability to inhibit uPA and FXIIa, we next tested the effects of the most enriched amino acid substitutions in PAI‐1 with respect to the inhibition of both of these proteases. We identified the most enriched single‐amino acid substitutions in our DMS screens for PAI‐1 inhibition of both uPA and FXIIa on the I91L background as being F98Y, T251I, and S331C (Figure 5a,b). The WT human amino acids at these positions are highly conserved in extant mammalian species, possibly due to the requirement of the I91L substitution or a general extension of PAI‐1's functional half‐life to open up the functional accessibility of these substitutions.

These variants were expressed as recombinant proteins, and their SIs with respect to uPA and FXIIa were determined (Figure 5c). For a SERPIN to function as a specific and efficient inhibitor of a given serine protease, it must have a sufficiently rapid second‐order rate constant and an SI close to “1” such that virtually all the reactions proceed along the inhibitory pathway with little to no substrate pathway observed (Olson and Gettins 2011). The SI for FXIIa inhibition was consistently 2–3‐fold higher than that of uPA inhibition—indicating that when interacting with FXIIa, in contrast to uPA, PAI‐1 is more likely to utilize the substrate versus the inhibitory pathway (Figure 1) and is consistent with previous findings that the rate constant of PAI‐1 inhibition of FXIIa is several orders of magnitude lower than that for uPA inhibition (Keijer et al. 1991). Of note, these variants were not the most enriched variants in our original DMS screen on the WT PAI‐1 background (Huttinger et al. 2021), with the F98Y and S331C variants actually classified as non‐functional against WT PAI‐1 (Figure S2c). This finding suggests that these mutations are epistatic with the I91L substitution with respect to PAI‐1 function (Berkenpas et al. 1995) or that these variants decrease the functional stability of PAI‐1 such that their half‐lives are too short to be detected on the WT background in our assay, despite being otherwise inhibitory variants. One potential implication of epistasis between the I91L and T98Y and/or S331C substitutions, with respect to PAI‐1 inhibitory function, is that although the T98Y and S331C substitutions are functional inhibitors on the I91L background, this functionality of these substitutions is not available to PAI‐1 in the absence of the I91L substitution.

As the three most highly enriched PAI‐1 variants identified in our DMS screens for both uPA and FXIIa inhibition were the same (Figure 5a,b), we next compared the evolutionary conservation scores for PAI‐1 with respect to both uPA and FXIIa inhibition (Figure 5d). Not only are the normalized functional scores for uPA and FXIIa highly correlated with each other (slope = 0.78, R ² = 0.43, p = 2.2 × 10⁻¹⁶) as expected (Figure 2b); but amino acid sites that are conserved across extant mammalian species are less likely to accept mutations in both DMS screens (quadrant I, QI), while sites that are more evolutionarily labile are more likely to accept mutations (quadrant III, QIII). The mean evolutionary conservation scores (±standard deviation) between QI (0.45 ± 1.21) and QIII (−0.36 ± 0.70) are highly statistically significantly different (p _adj < 0.0001, Tukey's HSD test), suggesting that amino acid substitutions in PAI‐1 that would impact its protease specificity for either uPA or FXIIa are likely to affect inhibition of both proteases. However, given that PAI‐1 is a more efficient inhibitor of uPA than of FXIIa (Keijer et al. 1991), these results suggest that, as a whole, the amino acid sequence of PAI‐1 is tuned to preferably inhibit uPA compared to FXIIa. In contrast, the normalized functional scores for uPA and TMRPSS2 (Figure S5) are not correlated with each other (slope = 0.14, R ² = 0.01, p = 0.052). Furthermore, the mean evolutionary conservation score between those sites at which amino acid substitutions maintain/promote inhibition of uPA and TMPRSS2 (QI; 0.45 ± 1.13) is more evolutionarily labile than those that promote TMPRSS2 but not uPA inhibition (Q2; −0.24 ± 0.77, p = 3.6 × 10⁻⁶)—emphasizing that these function‐altering amino acid substitutions do not promote the canonical inhibition of uPA while promoting TMPRSS2 inhibition (Figure 2b).

It has been speculated that the high SI for PAI‐1 inhibition of FXIIa effectively results in FXIIa inactivation of PAI‐1, further reducing its potential as an antifibrinolytic (Tanaka et al. 2009). As the inhibition of both uPA and FXIIa occupies similar PAI‐1 sequence spaces, the amino acid substitutions that were identified as the best FXIIa inhibitors in our DMS screen were still more efficient uPA inhibitors (Figure 5)—consistent with previous studies (Berrettini et al. 1989; Sherman et al. ¹⁹⁹²). Overall, our data indicate that PAI‐1 maintains its function as an inhibitor of these two proteases (although there is a “preference” to inhibit uPA over FXIIa), with few single‐amino acid substitutions (n = 16) conferring a preference for inhibition of FXIIa over uPA (Figures 2b and 5a,b). This is in contrast to PAI‐1 inhibition of TMPRSS2, in which multiple single‐amino acid substitutions (n = 184) appear to improve specificity toward TMPRSS2 over FXIIa/uPA (Figure 2b). We, therefore, hypothesized that PAI‐1 has maintained inhibitory activity toward FXIIa in addition to uPA as a function of the close evolutionary relationship between these serine proteases (McClune et al. ²⁰¹⁹; Yousef et al. ²⁰⁰³). In essence, our results suggest closely related serine proteases form a crowded sequence space with regard to SERPIN inhibition. An expectation arising from these findings is that, beyond PAI‐1, a given SERPIN will inhibit groups of closely related serine proteases with patterns of SERPIN and protease diversification mirroring one another.

2.6. Coevolution of SERPINs and their cognate serine proteases

To test the role of phylogenetic relatedness and codiversification in driving interactions between SERPINs and their cognate serine proteases, we next performed a co‐phylogenetic analysis of human secreted SERPINs and their cognate serine proteases. Cophylogenetic analyses are typically used to study coevolution in host–parasite interactions, where the null hypothesis is that the shapes of the two phylogenies of each putatively coevolving group are not mutually informative of one another (Legendre et al. 2002). We reasoned that a similar statistic of mutual information in phylogenetic topologies would quantify SERPIN: serine protease codiversification. We therefore curated a list of human secreted SERPINs (n = 13) with well‐described cognate serine proteases (n = 16) (Godinez et al. 2022; Heiker et al. ²⁰¹³; Janciauskiene et al. ²⁰²⁴; Olson and Gettins ²⁰¹¹).

Our cophylogenetic analysis (Figure 6a) supports the hypothesis that the secreted SERPINs and serine proteases have co‐diversified, as the maximum likelihood gene phylogenies for each family were significantly mutually informative of one another (p < 0.039). This result is robust to phylogenetic uncertainty, as the majority of 1000 random pairs of SERPIN and serine protease bootstrapped gene trees are also mutually informative (Figure 6b). The bulk of this mutual signal appears to reside deep in the evolution of these groups, reflected in the SERPINA clade largely inhibiting the KLK/elastase clade, whereas the other SERPINs, such as the clade formed by group I/E/G SERPINs, largely inhibit the related complement system and coagulation proteases. Notably for the current study, the closely related SERPINE1, SERPINE2, and SERPINI1 all inhibit uPA (encoded by the PLAU gene) and tPA (encoded by the PLAT gene) and most likely coevolved from an ancestral SERPIN, with the emergence of the two plasminogen activators from a common serine protease ancestor that also gave rise to FXIIa (F12) (Jendroszek et al. 2019). Similarly, the closely related serine proteases thrombin (encoded by the F2 gene) and FXa (encoded by the F10 gene) are both inhibited by a common SERPIN, antithrombin (encoded by the SERPINC1 gene), with which they also coevolved (Figure 6a) (Yousef et al. 2003). Therefore, we anticipate that a DMS screen of antithrombin's inhibition of thrombin and FXa would also reveal overlapping effects of single amino acid substitutions.

FIGURE 6.

Cophylogenetic analysis reveals that human secreted SERPINs coevolved with secreted serine proteases. (a) A representative phylogenic tree of human secreted SERPINs (left) is mutually informative with that of human secreted serine proteases (right) as determined with the parafit R package (p = 0.039) (Legendre et al. 2002). SERPINs and proteases are denoted by their human gene names. For SERPINs, colors denote closely related proteins (SERPIN C clade, red; SERPIN E and I clades, blue; SERPIN A clade, orange; SERPIN G and F clades, green; SERPIN D clade, purple), while for serine proteases, colors denote the group of SERPINs most commonly inhibiting a given protease. Proteases equally likely to be inhibited by SERPINs from more than one group are shown in black. (b) Histogram showing the distribution of p‐values testing for significant mutual information between 1000 random pairs of secreted SERPIN and serine protease phylogenies. The mean p‐value is indicated by the vertical dashed line and is less than 0.05, indicating that on average, the bootstrapped phylogenies are mutually informative of each other.

3. CONCLUSION

Inhibitory SERPINs have evolved to fulfill unique functional niches in which they inhibit one or a few proteases with residual inhibitory activity toward additional proteases (Gettins 2002; Olson and Gettins ²⁰¹¹). Here, we have harnessed the power of DMS and combined it with comparative evolutionary approaches to assess PAI‐1's inhibitory function toward its canonical target, uPA, as well as non‐canonical targets that are both closely (FXIIa) and distantly (TMPRSS2) related to uPA. Our data demonstrate that the determinants of PAI‐1 specificity are scattered throughout its amino acid sequence space, with no one region of the primary sequence completely dictating the specificity of a target protease, suggesting that residues within the PAI‐1 structure mediate long‐range effects that help dictate its specificity as a protease inhibitor. Furthermore, we find that residues modulating PAI‐1 inhibition of uPA and FXIIa are under purifying selection, with no such signature for inhibition of TMPRSS2. PAI‐1 instead exhibits marginal specificity toward TMPRSS2, but robust specificity that prevents it from becoming a FXIIa specialist. As uPA and FXIIa are closely related serine proteases, the ability of PAI‐1 to be a specific inhibitor of uPA plasminogen activators, but not FXIIa, is restricted by a locally crowded sequence space. Coevolution of SERPINs with their cognate target serine protease may explain many overlapping and non‐specific SERPIN: serine protease interactions and could help guide future efforts to engineer SERPIN specificity. In conclusion, we have demonstrated the power of DMS to interpret the evolution and functional specificity of members of the SERPIN protein superfamily and anticipate that similar analyses have the potential to improve understanding of the mechanisms driving the diversification and specialization of other large paralogous protein families.

4. MATERIALS AND METHODS

4.1. Phage displayed PAI‐1 variant library preparation

A phage display PAI‐1 variant library was constructed on the I91L PAI‐1 background using error‐prone PCR, as previously described (Haynes et al. 2022; Huttinger et al. ²⁰²¹) with the GeneMorph II Random Mutagenesis Kit (Agilent Technologies, Santa Clara, CA) using primers that preserved the AscI and NotI restriction sites, and cloned into a modified pAY‐FE plasmid (Genebank #MW464120) in which an amber stop codon (TAG) immediately preceding the SERPINE1‐gIII fusion construct was mutated to glutamine (CAG, Gln) to increase expression of the PAI‐1‐pIII coat fusion protein (pLMH1; Table S1) (Haynes et al. 2022; Huttinger et al. ²⁰²¹). Following cloning of the I91L PAI‐1 variant library into the pLMH1 plasmid, the library was transformed into electrocompetent XL‐1 Blue MRF′ E. coli (Agilent Technologies). Library depth was determined by counting the number of ampicillin‐resistant colonies. Mutation frequency was estimated by Sanger sequencing of the PAI‐1 (SERPINE1) inserts from randomly selected individual phage clones (n = 24).

The I19L PAI‐1 phage displayed library was produced as described previously (Haynes et al. 2022; Huttinger et al. ²⁰²¹). Briefly, E. coli harboring the I91L PAI‐1 library was grown in LB media supplemented with 2% glucose and ampicillin (0.1 mg/mL) to mid‐log phase at 37°C, infected with M13KO7 helper phage (Cytiva) for 1 h, and transferred to 2xYT media supplemented with ampicillin (0.1 mg/mL), kanamycin (0.03 mg/mL), and IPTG (0.4 mM) to induce expression of the PAI‐1 phage displayed library and grown for 2 h at 37°C. Escherichia coli were removed by centrifugation (4200g and 4500g at 4°C for 10 min) and phage were precipitated from the supernatant with PEG‐8000 (2.5% w/v) and NaCl (0.5M) overnight at 4°C. Phage were pelleted by centrifugation (20,000g for 20 min at 4°C) and resuspended in 50 mM Tris containing 150 mM NaCl at pH 7.4 (TBS).

4.2. PAI‐1 variant selection

uPA selection assays were performed as previously described (Haynes et al. 2022; Huttinger et al. ²⁰²¹). Human coagulation FXIIa (Innovative Research, Novi, MI) was biotinylated, and the degree of labeling was assessed using the EZ‐Link Sulfo‐NHS‐LC‐Biotin labelling kit (ThermoFisher Scientific) with 1–2 molecules of biotin per protein molecule. Soluble recombinant TMPRSS2 (aa 106–492) with an N‐terminal 6xHis tag expressed in yeast was purchased from Creative BioMart (Shirley, NY) and was shown to be inhibited by WT PAI‐1 (Figure S1). FXIIa‐biotin (100 nM) or TMPRSS2 (1 μM) (Shrimp et al. 2020) were incubated with a 1:10 dilution of the input I91L PAI‐1 phage display library for 30 min at 37°C in TBS (50 mM Tris base, 150 mM NaCl, pH 7.4) containing 5% BSA (TBS‐BSA). All reactions were quenched with the addition of cOmplete EDTA‐free protease inhibitor cocktail (MilliporeSigma) for 10 min at 37°C. The biotinylated and 6xHis‐tagged screens were immunoprecipitated overnight with streptavidin or Ni‐NTA magnetic beads (New England Biolabs, Ipswich, MA), respectively. phPAI‐1:enzyme complexes were further washed, eluted, and quantified as described previously (Haynes et al. 2022; Huttinger et al. ²⁰²¹). Replicas represent the selection of independent cultures of the input phage display PAI1 variant library.

4.3. High‐throughput sequencing

Phage displayed PAI‐1 cDNA were sequenced in twelve 150 bp amplicons using primers listed in Table S1. Amplicons were prepared for HTS as previously described (Haynes et al. 2022; Huttinger et al. ²⁰²¹) and sequenced with 8 × 10⁵–2 × 10⁶ 150 bp paired‐end reads per amplicon. HTS data were analyzed using the DESeq2 software package (Love et al. 2014; Zhu et al. ²⁰¹⁹). Significance thresholds for DESeq2 were set to p _adj < 0.1 (default significance in DESeq2) and BaseMean score ≥50 to only select those PAI‐1 variants with significant non‐zero counts in the bimodal distribution of Base. MA plots showing the distribution of amino acid substitutions in PAI‐1 with respect to these thresholds are shown in Figures S6–S8. Notably, when amber stop codons (TAG) are read through as Gln, those variants are enriched in the selected libraries relative to other nonsense mutations (TAA and TGA) introduced by error‐prone PCR, indicating that these variants are functional and further demonstrating that the screens discriminate between loss‐of‐function and gain‐/maintenance‐of‐function mutations.

4.4. Principal component analysis

PCA was performed using the DESeq2 software package. For the analysis of variants across the entire PAI‐1 sequence (Figure 2), the 12 sequencing amplicons were concatenated as described previously, prior to the implementation of DESeq2 and subsequent generation of PCAs. For variants within a single amplicon (Figure 3), these analyses were implemented without concatenation of the amplicons.

4.5. Evolutionary variability of PAI‐1

As previously described, evolutionary conservation scores at each amino acid position in PAI‐1 were calculated using ConSurf (https://consurf.tau.ac.il/), where higher ConSurf scores indicate more evolutionarily variable positions (Ashkenazy et al. 2016; Huttinger et al. ²⁰²¹). To relate the results of our DMS screen to evolutionary conservation scores, we define the normalized functional score as the sum of the mean log₂‐fold determined at each position. Conservation scores were then analyzed as a function of the normalized functional score at each position using the R software package (version 4.3.3).

4.6. Expression and characterization of recombinant PAI‐1 variants

PAI‐1 variant cDNA (I91L, I91L/F98Y, I91L/T251I, and I91L/S331C PAI‐1) in the pET‐24(+) expression plasmid with a C‐terminal Gly‐Ser‐Gly hinge and 6XHis‐tag was purchased from Twist Bioscience (South San Francisco, CA) and transformed into NiCo21(DE3) chemically competent E. coli (New England Biolabs, Ipswitch, MA). PAI‐1 variants were expressed as previously described (Haynes et al. 2022), and the concentration was determined by absorption at 280 nm (ε_1%,280nm = 7.0). SI for each variant was determined by incubating uPA (2.5 nM) or factor XIIa (100 nM) with PAI‐1 variant concentrations ranging from 0 to 4 nM and 0 to 10 nM, respectively. The SIs for each variant were normalized to an SI of uPA inhibition defined as 1.

4.7. PAI‐1 inhibition of TMPRSS2

WT PAI‐1 was expressed and purified as described above. TMPRSS2 (1 μM total protein, yet only fractionally active; Shrimp et al. ²⁰²⁰) was incubated with WT PAI‐1 (1 nM effective active concentration) or vehicle control for 30 min at room temperature (~25°C). TMPRSS2 activity was determined by monitoring its ability to cleave the peptidyl substrate boc‐QAR‐AMC (λ _ex = 370 nm, λ _em = 440 nm; VWR International) for 10 min (Shrimp et al. 2020). Following this monitoring, samples were electrophoresed on a 4%–20% Tris‐glycine gel (Invitrogen) under non‐reducing conditions, transferred to nitrocellulose, and blotted using a rabbit‐anti‐PAI‐1 primary antibody (Abcam).

4.8. Cophylogenetic analyses

A list of human secreted SERPINs (n = 13) and serine proteases (n = 16) was curated from the literature (Godinez et al. 2022; Heiker et al. ²⁰¹³; Janciauskiene et al. ²⁰²⁴; Olson and Gettins ²⁰¹¹). We defined secreted SERPINs and serine proteases as those containing a signal peptide. We used the COBALT tool to produce a domain‐centric alignment of both the SERPINs and serine proteases in the lists, and then pruned the alignment manually to include only sites with fewer than 90% gaps. We used TreeBeST (Vilella et al. 2009) to produce phylogenetic trees for both SERPINs and serine proteases, using 1000 bootstrap replicates, and recovered the maximum likelihood tree for our focal analyses. A cophylogenetic analysis was then performed using the Parafit package in R with the null hypothesis that the evolution of serine proteases and SERPINs was independent of each other (Legendre et al. 2002). The p‐value was derived from 1000 random permutations of the tip states of the two trees, and the robustness of the result was then assessed by randomly pairing the bootstrap replicate trees to generate 1000 pairs of SERPIN and serine protease trees to determine the frequency of significant associations.

AUTHOR CONTRIBUTIONS

Laura M. Haynes: Conceptualization; investigation; writing – original draft; methodology; formal analysis; funding acquisition. Matthew L. Holding: Investigation; writing – review and editing; formal analysis; methodology; visualization. Hannah L. DiGiovanni: Investigation. David Siemieniak: Formal analysis; investigation; data curation; software. David Ginsburg: Resources; supervision; conceptualization; writing – review and editing; funding acquisition.

CONFLICT OF INTEREST STATEMENT

D.G. is a member of MDI Therapeutics' Clinical and Scientific Advisory Board, which is developing therapeutic PAI‐1 inhibitors.

Supporting information

Data S1. Supporting Information.

Data S2. Supporting Information.

Haynes LM, Holding ML, DiGiovanni HL, Siemieniak D, Ginsburg D. High‐throughput amino acid‐level characterization of the interactions of plasminogen activator inhibitor‐1 with variably divergent proteases. Protein Science. 2025;34(4):e70088. 10.1002/pro.70088

DATA AVAILABILITY STATEMENT

Bioinformatics code and associated information required to execute the code (amino acid substitution counts from high‐throughput sequencing and ConSurf scores) are available at https://github.com/hayneslm/PAI-1_and_divergent_proteases.git. Raw sequencing files and associated code can be accessed at https://doi.org/10.7302/r2wk-3n35.