Engineering a potent inhibitor of matriptase from the natural hepatocyte growth factor activator inhibitor type-1 (HAI-1) protein

Dysregulated matriptase activity has been established as a key contributor to cancer progression through its activation of growth factors, including the hepatocyte growth factor (HGF). Despite its critical role and prevalence in many human cancers, limitations to developing an effective matriptase inhibitor include weak binding affinity, poor selectivity, and short circulating half-life. We applied rational and combinatorial approaches to engineer a potent inhibitor based on the hepatocyte growth factor activator inhibitor type-1 (HAI-1), a natural matriptase inhibitor. The first Kunitz domain (KD1) of HAI-1 has been well established as a minimal matriptase-binding and inhibition domain, whereas the second Kunitz domain (KD2) is inactive and involved in negative regulation. Here, we replaced the inactive KD2 domain of HAI-1 with an engineered chimeric variant of KD2/KD1 domains and fused the resulting construct to an antibody Fc domain to increase valency and circulating serum half-life. The final protein variant contains four stoichiometric binding sites that we showed were needed to effectively inhibit matriptase with a Ki of 70 ± 5 pm, an increase of 120-fold compared with the natural HAI-1 inhibitor, to our knowledge making it one of the most potent matriptase inhibitors identified to date. Furthermore, the engineered inhibitor demonstrates a protease selectivity profile similar to that of wildtype KD1 but distinct from that of HAI-1. It also inhibits activation of the natural pro-HGF substrate and matriptase expressed on cancer cells with at least an order of magnitude greater efficacy than KD1.

High expression and dysregulated activity of the type II, membrane-anchored serine protease matriptase in the local tumor environment have been shown to correlate with poor patient prognosis in many human cancers, including breast, colorectal, pancreatic, cervical, and prostate cancers (1)(2)(3)(4)(5)(6)(7)(8)(9)(10). This dysregulation is partly driven by the high proteolytic processing and turnover of pro-hepatocyte growth factor (pro-HGF) 4 (9,11) to a form of HGF that activates its cognate receptor, c-Met (12) (Fig. 1A). Matriptase is also known to activate other proteases and growth factors, including urokinase plasminogen activator (11), pro-macrophage-stimulating protein (13), and platelet-derived growth factor-D (PDGF-D) (14), all of which play key roles in cancer growth and metastasis. Furthermore, matriptase has been identified as a critical driver of other diseases, including iron overload disease (15) and osteoarthritis (16), and has been shown to activate the human airway influenza virus (H1N1) (17) and human immunodeficiency virus (HIV) (18). Although the correlation of matriptase overexpression, dysregulation, and disease progression is well established, effective matriptase inhibitors are lacking, highlighting an important clinical need.
Matriptase naturally functions in developmental pathways as well as in tissue regeneration (19 -23). The activity of matriptase is regulated in healthy tissue by the serine protease inhibitor hepatocyte growth factor activator inhibitor type-1 (HAI-1) (Fig. 1A). HAI-1 is primarily expressed on the surface of epithelial cells and naturally blocks the substrate-activating properties of matriptase (24 -26) as well as other structurally related proteases such as the hepatocyte growth factor activator (9,27), hepsin (28,29), and kallakrein-4/5 (9,30). The balance between substrate activation and protease inhibition is critical to the metastatic potential of tumor cells (see Fig. 1B). As such, the ratio of HAI-1 expression to matriptase expression correlates with cancer aggression and patient prognosis and has been established as a key biomarker (2,5,7,10,31,32).
Inhibition of matriptase-driven cancer progression has been proposed as an attractive strategy for cancer therapy. In one study, induced cell-surface expression of HAI-1 within the tumor environment of an orthotopic adenopancreatic cancer This work was supported in part by NIGMS, National Institutes of Health Training Grant in Biotechnology 5T32GM008412 (to A. C. M.) and NCI, National Institutes of Health Grant R01 CA151706 (to J. R. C.). The authors declare that they have no conflicts of interest with the contents of this article. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health. This article contains Figs. S1-S7 and Table S1. 1  model resulted in reduced tumor size and eliminated metastatic nodule formation (2). In another study, the addition of soluble HAI-1 was shown to significantly lower pro-HGF activation and reduce breast cancer cell invasion in vitro (33), highlighting recombinant HAI-1 as a therapeutic approach. However, the therapeutic utility of HAI-1 is ultimately limited by its nanomolar inhibition constant to matriptase. In contrast, the first Kunitz (KD1) subdomain of HAI-1 (Fig. 1A) (26,34) has been shown to inhibit matriptase activity with significantly greater potency than full-length HAI-1 (35). The small molecular mass of the KD1 domain (ϳ6 versus 58 kDa for HAI-1) confers a short circulating half-life of 20 min, which greatly limits its therapeutic efficacy. Although chemical conjugation of KD1 to polyethylene glycol (PEG) showed significant extension in serum half-life (35), this approach does not further improve the inhibition constant beyond that of wildtype KD1. Alternative approaches to develop matriptase inhibitors include synthetic small molecules (36,37), peptides (38), monoclonal antibodies (39), and constrained peptide scaffolds (40). Although each strategy generated molecules that bound to and inhibited matriptase activity, none address all of the reported therapeutic limitations. An effective therapeutic candidate must bind matriptase with high affinity to effectively outcompete pro-HGF substrate activation as well as possess a long serum half-life to mitigate the need for frequent dosing. To overcome these critical barriers, we used rational and combinatorial approaches to engineer a potent matriptase inhibitor based on a modified variant of the natural HAI-1 protein. In this work, the inactive second Kunitz (KD2) domain of HAI-1 was replaced with a chimeric variant of KD2/KD1 domains. This modified HAI-1 protein was then fused to an antibody crystallizable fragment (Fc) domain, resulting in a final construct with four putative sites that bound additively to matriptase with pM affinity. This engineered protein significantly inhibited pro-HGF activation and matriptase expressed on the surface of lung, breast, and prostate cancer cells.

Engineering HAI-1 as a more potent matriptase inhibitor
We used HAI-1 as a starting scaffold for protein engineering to leverage its intrinsic ability to bind and inhibit matriptase. HAI-1 comprises an N-terminal domain (41), an internal domain (42), KD1, a low-density lipoprotein (LDL)-like domain, KD2, a transmembrane domain, and an intracellular domain (Fig. 1A). KD1 has been well established as the minimal matriptase-binding domain within HAI-1. KD2 has been shown to negatively regulate HAI-1 binding affinity and confer protease specificity (26,34). In this previous work, removal of the KD2 domain resulted in a 10-fold improvement in HAI-1 inhibition of matriptase activity. In addition, matriptase inhibition was proposed to be driven by a 4-amino-acid primary binding interface (Arg-Cys-Arg-Gly) found in KD1 but absent in KD2 ( Fig. 1 and Fig. S1A). To create an improved matriptase inhibitor, we targeted the inactive KD2 domain of HAI-1 to convert it into a matriptase-binding module, effectively doubling the binding sites within the HAI-1 protein.
Yeast surface display is a well established protein engineering technology that has been used for characterizing and screening protein-based inhibitors (40,(43)(44)(45) (Fig. S1B). We found that KD1 and KD2 were well expressed on the surface of yeast as Aga2p mating protein fusions (Fig. S1C). Additionally, we showed that KD1 bound to soluble matriptase with an affinity (K d ) of 13 Ϯ 2 pM, whereas KD2 exhibited no detectable binding (Fig. 2C), a trend in agreement with previous results. KD2 has been suggested to retain the secondary matriptase-binding site conserved from KD1 (26). In an attempt to supplement this binding site, we first introduced the primary matriptase-binding site residues (Arg-Cys-Arg (KD2-graft 1) or Arg-Cys-Arg-Gly (KD2-graft 2)) from KD1 into KD2 (Fig. S1A). Yeast-displayed versions of these constructs revealed high surface expression but a lack of matriptase binding similar to wildtype KD2 (Fig. S1D). These results indicated further engineering was required to effectively convert KD2 into a matriptase-binding domain. Figure 1. A, schematic of the HAI-1 inhibitor, which naturally regulates matriptase activity and levels of pro-HGF activation, thus preventing cancer progression in healthy tissue. B, biological representation of the tumor environment. Dysregulated matriptase cleaves pro-HGF into activated HGF, which is competent to bind to and stimulate its cognate receptor, c-Met. Ligand-receptor dimerization then triggers intracellular signaling pathways that in turn stimulate cellular phenotypic responses, including cell growth, proliferation, and migration.

Engineering a potent matriptase inhibitor
To further explore additional mutation space beyond the grafted primary binding motif, we applied error-prone polymerase chain reaction (PCR) (46) to randomly introduce mutations throughout the KD2-graft 2 gene. The mutated DNA was transformed into Saccharomyces cerevisiae yeast cells, resulting in ϳ5 ϫ 10 7 transformants, which were induced to express a library of yeast surface-displayed KD2 variants, averaging 2 amino acid mutations per gene. The library was screened using fluorescence-activated cell sorting (FACS) to isolate yeast clones that expressed KD2 variants and bound to matriptase ( Fig. 2A). Yeast cells that were collected were grown in culture and induced for KD2 expression for additional screening. Following two rounds of library screening, no detectable matriptase binding was observed. In an effort to increase the mutational load, DNA was recovered from the pooled yeast and subjected to error-prone PCR to introduce additional genetic diversity with an average mutagenic frequency of 2-5 amino acids per gene. A new library of ϳ1 ϫ 10 7 yeast transformants was created and screened by FACS to identify KD2 variants that bound to matriptase. Parallel sorting and analysis for nonspecific binding to secondary antibodies were also performed to reduce occurrence of false-positive binding variants. From these efforts, a yeast population emerged that demonstrated significantly improved binding to matriptase compared with wildtype KD2 (Fig. 2A).

Characterization of sorted library variants reveals a KD2/KD1 chimera
DNA from the final sorted yeast population was isolated and sequenced to identify amino acid mutations that could confer increased matriptase binding compared with wildtype KD1 and KD2 ( Fig. 2B and Table S1). Surprisingly, we identified a chimeric variant that essentially was a fusion of the N terminus of KD2 and C terminus of KD1 (clone 33; named KD2/1). The generation of KD2/1 was likely due to the presence of the wildtype KD1 gene within the library construction and transformation steps, allowing recombination of genetic regions of KD1 and KD2 to generate clone 33.
Select yeast-displayed variants were individually tested for binding to matriptase; however, only the KD2/1 chimera and wildtype KD1 showed any detectable binding signal (Fig. S2). It is likely that additional rounds of screening under more stringent conditions would have resulted in isolation of KD2/KD1 as a clonal yeast population. An equilibrium binding assay showed that yeast-displayed KD2/1 binds to matriptase with an affinity of K d ϭ 220 Ϯ 30 pM (Fig. 2C), which was 20-fold weaker than Figure 2. A, library screening identified a chimera of KD1 and KD2 that binds matriptase. Representative FACS plots are shown from separate yeast library sorting rounds, including sorting gates used to isolate phenotypically improved KD2 variants. B, sequence alignment of clone 33, which is a chimera of KD2/KD1 (termed KD2/1) with KD2 graft 2 in green, wildtype KD1 in orange, and shared sequence space of KD1 and KD2 in white. C, matriptase equilibrium binding of yeast-displayed KD1 wildtype (blue; K d ϭ 13 Ϯ 2 pM), KD2/1 chimera (red; K d ϭ 220 Ϯ 30 pM), and KD2 wildtype (green; no binding), reported as mean and S.D. values. D, yeast-displayed KD domains that bind matriptase also inhibit its activity. Bar graphs indicate quantified matriptase activity for varying number of yeast cells. The color scheme is the same as in C. Significance was quantified with a pairwise t test: *, p Ͻ 0.0001; **, p Ͻ 0.0003; ***, p Ͻ 0.0004; ****, p Ͻ 0.0024. Error bars represent S.D.

Engineering a potent matriptase inhibitor
that of KD1. Notably, inhibitors that demonstrate protease binding do not always demonstrate protease inhibition of the target active site (47,48). Thus, proteins were tested for functional matriptase inhibition by incubating an increasing number of yeast cells displaying KD2/1 or wildtype KD1 or KD2 with soluble matriptase and substrate. Fluorescent matriptase substrate activation was quantified over time and revealed that both yeast-displayed KD1 and KD2/1 domains significantly inhibited matriptase activity in a cell number-dependent manner, whereas the yeast-displayed KD2 domain did not inhibit matriptase, even with up to 10 5 yeast cells (Fig. 2D).

Development of a soluble matriptase inhibitor
After validating that yeast-tethered KD2/1 could bind to and inhibit matriptase, we next created a soluble, recombinant matriptase inhibitor. We first replaced the wildtype KD2 domain of full-length HAI-1 with the sequence of the engineered KD2/1 chimera. The construct was fused to the Fc domain of an antibody, which is an established protein engineering strategy that confers therapeutic properties, including circulating half-life extension, immune system recruitment, and elevated binding affinity through avidity (49,50). This protein was termed KD1-KD2/1-Fc (Fig. 3). In addition, given the favorable binding properties of yeast-displayed KD1 (Fig. 2C), an alternative HAI-1 design was created where the wildtype KD2 domain was replaced with a second wildtype KD1 domain, termed KD1x2-Fc. Wildtype KD1 monomer, full-length HAI-1 monomer, and an HAI-1-Fc fusion (HAI-Fc) were also produced as controls. The function of the KD1 domain can be diminished by introducing an R260A point mutation, which is known to ablate matriptase binding (34). Thus, we used this point mutation to create constructs containing a non-inhibitory KD1 domain to allow us to parse the importance of each Kunitz domain in matriptase inhibition ( Fig. 3; HAI-R260A-Fc and KD1-R260A-KD2/1-Fc). Each protein construct was expressed in a transient mammalian cell expression system and underwent two-step purification. The resulting size exclusion chromatograms demonstrate expression and purification of all proteins (Fig. S3, A and B) with the exception of KD1x2-Fc (Fig.  S3, C and D). These findings suggest that the 6 N-terminal amino acids from wildtype KD2, found in the KD2/1 chimera (Table S1), were critical for effective protein folding/expression in the context of full-length HAI-1.

KD1-KD2/1-Fc is a potent and selective inhibitor of matriptase
We next evaluated the purified proteins for their ability to inhibit matriptase activity. We first tested each protein construct using an in vitro kinetic inhibition assay. Dose-response plots were generated for each inhibitor (Fig. S4A), and inhibition constant (K i ) values were determined using Equations 1 and 2 as reported previously (36,38,40). Table 1 lists the resulting K i value for each inhibitor construct and the number of functional Kunitz domains present. As expected, HAI-R260A-Fc has no detectable inhibition of matriptase due to the ablating R260A mutation that disrupts wildtype KD1 function, confirming that the KD2 domain or Fc domain does not participate in matriptase inhibition. In contrast, the KD1-R260A-KD2/1-Fc protein exhibited a K i of 550 Ϯ 50 pM, indicating that the KD2/1 domain is a functional inhibitor of matriptase when incorporated into the HAI-1-Fc fusion protein. The HAI-1 monomer has a moderate K i of 9.1 Ϯ 1 nM, which improves slightly by 2-fold as a bivalent HAI-1-Fc fusion (K i ϭ 4.2 Ϯ 0.5 nM). Additionally, the wildtype KD1 monomer has a K i of 310 Ϯ 20 pM, which is in agreement with previous K i measurements for matriptase (34,42), and is a 30-fold more potent inhibitor relative to the full-length HAI-1 monomer. This improvement further demonstrates the negative regulation of KD2 in the context of the full-length HAI-1 inhibitor.
Finally, KD1-KD2/1-Fc had the lowest K i of 70 Ϯ 5 pM, making it the most effective inhibitor in our panel. More specifically, KD1-KD2/1-Fc demonstrates a relative improvement in K i of 4-fold compared with the wildtype KD1 monomer, a  Engineering a potent matriptase inhibitor 60-fold improvement compared with HAI-Fc, and a 120-fold improvement compared with the HAI-1 monomer. The increased potency is due to the replacement of the KD2 domain with the matriptase-binding KD2/1 chimera as well as the homodimeric nature of the Fc fusion construct. These combined engineering efforts expand the number of matriptasebinding sites from one domain in wildtype HAI-1 to four domains within the final KD1-KD2/1-Fc construct. Further experiments also determined that KD1-KD2/1-Fc follows a competitive inhibition modality for matriptase ( Fig. S5) similar to that of wildtype KD1 monomer (26). We next tested the selectivity of KD1-KD2/1-Fc against a panel of naturally soluble or cell-anchored serine proteases, including trypsin 3 (51), urokinase (11), kallikrein 4 (30), and hepsin (28,29); each of these proteases has a range of native affinities to wildtype HAI-1. We found that none of the proteins tested could inhibit trypsin 3 or urokinase activity (Fig. S4, B-D, and Table 2). In contrast, KD1-KD2/1-Fc and wildtype KD1 monomer inhibit kallikrein 4 similarly at 8.0 Ϯ 2 and 9.3 Ϯ 2 nM, respectively, whereas the wildtype HAI-1 monomer more weakly inhibits kallikrein 4 with K i values above 100 nM. Additionally, KD1-KD2/1-Fc and wildtype KD1 monomer inhibit hepsin with K i values of 1.5 Ϯ 0.4 and 5.4 Ϯ 2 nM, respectively, whereas the wildtype HAI-1 monomer more weakly inhibits hepsin with a K i value of 72 Ϯ 30 nM. Overall, the relative selectivity of KD1-KD2/1-Fc for matriptase remains at Ͼ1,000-fold over trypsin 3 and urokinase, 110-fold over kallikrein 4, and 20-fold over hepsin.

KD1-KD2/1-Fc inhibits matriptase expressed on cancer cells
After demonstrating potent and selective inhibition of the in vitro form of matriptase, we next tested the ability of KD1-KD2/ 1-Fc to inhibit matriptase expressed on human cancer cell lines. We first confirmed expression and functional activity of matriptase on the surface of three human cancer cell lines, MDA-MB-231 (breast), A549 (lung), and PC3 (prostate), using a matriptase-specific antibody and a commercial matriptase substrate. Positive matriptase expression levels correlating with matriptase functional activity were identified for each cell line tested (Fig. S7, A and B). KD1-KD2/1-Fc was then tested and compared with wildtype KD1 and HAI-1 monomer proteins for inhibition of fluorescent matriptase substrate activation. Doseresponse curves of matriptase inhibition were then generated and fit to quantify IC 50 values for each inhibitor tested (Fig. 4A). The results demonstrate that KD1-KD2/1-Fc inhibits matriptase up to 10-and 80-fold better compared with wildtype KD1 and HAI-1 inhibitors, respectively. KD1-KD2/1-Fc was also confirmed to bind to the surface of cancer cell lines, further confirming specific interactions with cell-associated matriptase (Fig. S7C).
We next performed a cancer cell invasion assay to further test the ability of KD1-KD2/1-Fc to inhibit cell-expressed matriptase activation of huPro-HGF. Invasion assays are often used to measure the phenotypic behavior of cancer cells in response to growth factor stimulation and protease inhibition involving the matriptase-HAI-1/pro-HGF-Met pathway (Fig. 1B) (33,53). To stimulate cancer invasion, we first transfected human embryonic kidney (HEK) cells to overexpress soluble huPro-HGF (Fig.  S6, A and B). The HEK-huPro-HGF cell line was used to construct a coculture assay to model cancer cell invasion in response to paracrine growth factor stimulation. A significant increase in invasion of both breast (MDA-MB-231) and lung (A549) cancer cells was observed upon incubation with HEK-huPro-HGF cells (Fig. 4B) compared with controls of cancer cells alone or with untransfected HEK cells. Furthermore, when KD1-KD2/1-Fc was added to the reaction containing HEK-hu-Pro-HGF cells, the invaded cell number decreased significantly compared with conditions without inhibitor. Collectively, these results indicate that KD1-KD2/1-Fc can potently bind to and

-Fc protease selectivity profile
Shown is a summary of the K i values quantified from dose-response plots (Fig. S4B) for each soluble protease and inhibitor tested. KD1 WT, wildtype KD1 monomer; HAI-1 WT, wildtype HAI-1 monomer. Values were fit and calculated using Equations 1 and 2 and are reported as the mean and S.D. of triplicate measurements. Approximate -fold selectivity values for wildtype KD1 monomer and HAI-1 monomer are reported relative to KD1-KD2/1-Fc. Ͼ100,000 pM K i values indicate no inhibition measured, and dashes (-) indicate the K i ratio could not be calculated.

Engineering a potent matriptase inhibitor
inhibit soluble and cell-expressed matriptase and block huPro-HGF activation to impede cell migration and invasion.

Discussion
Matriptase represents a critical protease that when unregulated participates in the aggressive advancement of many human maladies. In the context of cancer, loss of the natural matriptase inhibitor HAI-1 has been found to result in more aggressive cancer progression and poor patient prognosis. Previous attempts to develop potent and selective matriptase inhibitors centering around wildtype HAI-1 or KD1 have limitations due to weak matriptase binding affinity or short circulating half-life (34,35). Previous reports (26,34) and our current study have also supported that the KD2 domain imposes negative regulation on the inhibitory capacity of the KD1 domain (up to 30-fold) within the context of wildtype HAI-1. To overcome these limitations, we applied rational and combinatorial protein engineering strategies in an attempt to convert the KD2 domain of HAI-1 into a matriptase-binding domain.
Despite the sequence similarities between KD1 and KD2, we did not identify KD2 variants that could bind to matriptase. Homology modeling has previously been used to study key differences between KD1 and KD2 domains (26). With the exception of Leu-284, all secondary binding regions were proposed to be conserved between the two domains. Thus, the main difference between KD1 and KD2 appears to be the primary matriptase-binding site (Arg-Cys-Arg-Gly) found in KD1 but absent in KD2. We found that grafting this dominant matriptase-binding sequence from KD1 into KD2 was not sufficient to convert it into a binding domain. These results indicate that structural features beyond the primary binding site additionally influence matriptase binding to KD1 and are absent in KD2. Moreover, error-prone PCR was not sufficient to introduce favorable mutations into this modified KD2 domain to confer matriptase binding. One possible explanation is that the relatively small sequence space explored in our screens compared with theoretical library size limited our outcome, especially if multiple cooperative mutations are needed. Alternatively, a recent study indicated the importance of concerted interdomain interactions within HAI-1 for overall folding and function (42). In particular, the internal domain of HAI-1 was shown to interact with KD1 but also improves its availability for reaction with protease (i.e. stimulates its inhibitory activity), a finding also observed in earlier work (34). Furthermore, these findings may explain why KD2 mutants screened in isolation did not reveal matriptase-binding attributes and why the 6 KD2 residues that precede the KD1 sequence were critical in the context of the KD2/KD1 chimera for folding and functional activity.
Our protein engineering efforts serendipitously identified a chimeric domain comprising KD2 and KD1 regions that demonstrated subnanomolar matriptase binding and potent matriptase inhibition compared with the wildtype KD2 domain. Further creation of a bivalent KD1-KD2/1-Fc fusion results in a protein construct that contains four matriptase-binding domains: two wildtype KD1 domains and two engineered KD2/KD1 (KD2/1) chimeric domains, accentuating its potency 120-fold relative to wildtype HAI-1 (Table 1). This dramatic improvement can be attributed to replacement of the sterically regulating wildtype KD2 domain with the engineered KD2/1 matriptase-binding chimera.

Engineering a potent matriptase inhibitor
The KD2/1 chimeric domain was critical to this work as a protein variant created by replacing the KD2 domain with another wildtype KD1 domain (KD1x2-Fc) was unable to be recombinantly expressed in mammalian cell culture. Interestingly, this loss of recombinant expression was also observed in a previous study in which disruption of the internal domain and KD1 domain interface of HAI-1 resulted in a 10-fold loss in expression level (42).
Notably, the K i for KD1-KD2/1-Fc improves 4-fold relative to wildtype KD1 monomer, suggesting stoichiometric binding of one matriptase molecule to one functionally inhibiting domain of KD1-KD2/1-Fc. This relationship is also observed in the 2-fold relative difference in K i between wildtype HAI-1 monomer and HAI-Fc. Stoichiometric 1:1 wildtype HAI-1 and matriptase complexes have been previously observed to occur naturally (54). This model of inhibition assumes that all four KD1-KD2/1-Fc functional domains are equally accessible for simultaneous matriptase inhibition. Results obtained with the KD1-R260A-KD2/1-Fc inhibitor support this possibility in which the KD2/1 domains functionally inhibit matriptase within the context of the Fc fusion construct with a K i of 550 Ϯ 50 pM. It is important to note that, although non-inhibitordepleting conditions were used to quantify the K i value for all inhibitors tested, K i values may be an overestimate as the concentration of matriptase used (50 pM) was at the lowest limit for assay detection and may also not be 100% active under the assay conditions. Further testing using solid-phase or solution-based assay formats could help confirm the K i for this very tight matriptase inhibitor, assuming improved detection limits. However, the K i of wildtype KD1 monomer we report (310 Ϯ 20 pM) closely aligns with prior values for rat (328 Ϯ 181 pM) and human (380 Ϯ 70 pM) matriptase in the literature from studies using similar assay methods (34,42), thus strongly validating our results.
KD1-KD2/1-Fc additionally retains the highest selectivity to matriptase among a panel of serine proteases tested (Table 2). Interestingly, although KD1-KD2/1-Fc and wildtype KD1 inhibit kallikrein 4 with low nanomolar K i values, wildtype HAI-1 monomer only weakly inhibits kallikrein 4 with a K i Ͼ100 nM. This result further demonstrates the regulatory role that KD2 plays on the KD1 domain within the context of fulllength HAI-1. This regulation is also observed for hepsin where KD1-KD2/1-Fc and wildtype KD1 monomer inhibit hepsin more effectively compared with the wildtype HAI-1 monomer. Like matriptase, the relative difference in K i values for hepsin between KD1-KD2/1-Fc and wildtype KD1 monomer is also 4-fold, which further supports the stoichiometric inhibition hypothesis stated above. Hepsin is reported to share a role similar to that of matriptase in cancer progression, and thus dual targeting of hepsin and matriptase by KD1-KD2/1-Fc may serve as an attractive therapeutic feature (28,29,35).
Cell-based activity assays further demonstrate the superior potency of KD1-KD2/1-Fc in inhibiting matriptase activity of both soluble and cell-associated forms. The huPro-HGF activation assay qualitatively confirmed that KD1-KD2/1-Fc inhibits matriptase activation of the prodomain form of HGF with reduced MDCK cell scattering at lower inhibitor concentrations than of wildtype KD1 monomer (Fig. S6). KD1-KD2/1-Fc efficacy is further observed by a 10-fold greater inhibition of matriptase activity on cancer cells relative to wildtype KD1 monomer (Fig. 4A). The greater magnitude of improvement in cancer cell matriptase inhibition (10-fold) compared with soluble matriptase (4-fold) is possibly due to avidity effects of the KD1-KD2/1-Fc construct on cell-anchored matriptase. KD1-KD2/1-Fc also significantly reduces lung and breast cancer cell invasion in vitro (Fig. 4B). The extent of reduced invasion aligns with previous inhibition results using this standard invasion model (33,53). Notably, the invasion assay also included medium containing 2% fetal bovine serum to maintain HEK cell viability, identified to contain significant levels of active proteases capable of cleaving commercial matriptase substrate. This high background of protease activity, combined with the heterogeneity of other serum proteins and constitutively overexpressed huPro-HGF, likely contributes to underestimating the reduction in cancer cell invasion observed with KD1-KD2/ 1-Fc. Further testing using optimized media conditions might result in a greater extent of matriptase inhibition in cancer cells upon treatment with KD1-KD2/1-Fc.
The KD1-KD2/1-Fc construct has a K i for matriptase of 70 Ϯ 5 pM, which is among the tightest K i values measured for protein-based matriptase inhibitors. In addition to wildtype KD1 (K i ϭ 310 pM Ϯ 20 pM) (42,53), previous efforts have generated matriptase inhibitors based on constrained peptides (K i ϭ 830 Ϯ 140 pM (40) and K i ϭ 290 Ϯ 54 pM (38)) and antibodies (K i ϭ 720 pM (39)). Peptide-and KD1-based inhibitors as well as synthetic small-molecule inhibitors (36, 37) have short circulating half-lives and restricted molecular surface area for binding matriptase with high affinity. The engineered KD1-KD2/1-Fc protein fulfills several attractive design criteria. First, use of the native HAI-1 as a starting point for therapeutic development leverages the affinity and specificity of the natural inhibitor. Second, replacing the KD2 domain with an active matriptasebinding domain is expected to be minimally perturbing to native HAI-1. Third, fusion of the engineered construct to an Fc domain creates a bivalent protein, in this case resulting in four matriptase-binding sites that improve protease inhibition. Fourth, fusion to an Fc domain is expected to increase serum half-life through increased molecular weight and FcRn-mediated recycling (50), requiring less frequent therapeutic dosing and allowing manufacturing processes that are similar to those used to produce antibodies. Matriptase imaging experiments have also suggested that cell-anchored HAI-1 can serve as a natural reservoir for secreted proteases, effectively increasing their local concentration and activity at the leading edge of cancer invasion (55,56). High-affinity inhibitor binding is therefore critical to effectively outcompete the interaction of proteases with native cell-surface HAI-1. The engineered matriptase-binding protein described here thus has the potential to function as an HAI-1 "decoy" in cancer and other disorders where matriptase underlies disease pathophysiology.

Yeast cell-surface binding assays and library screening conditions
Induced EBY100 yeast cells were counted (A 600 of 1 ϭ 10 7 cells/ml), and 1 ϫ 10 5 cells/sample were washed with 1 ml of matriptase assay buffer and then mixed with soluble matriptase (final concentration, 0 -1 nM, serial dilution) and matriptase assay buffer (final concentration, 200 l to 100 ml). Liganddepleting conditions were addressed using methods described previously and estimated by assuming each yeast displayed 50,000 copies of inhibitor per cell (57,58). Reactions were incubated for 48 h at room temperature to reach equilibrium. Samples were then incubated with a 1:250 dilution of anti-HA mouse primary antibody (Fisher Scientific) for 30 min at room temperature. Following washing with matriptase assay buffer, samples were labeled with secondary antibodies to measure yeast expression (1:100 dilution of anti-mouse phycoerythrin (Invitrogen)) and matriptase binding (1:100 dilution of anti-His fluorescein isothiocyanate (FITC) (Bethyl Laboratories)) in 50 l of matriptase assay buffer. Samples were incubated at 4°C for 15 min, washed, and maintained on ice until loading onto a flow cytometer for analysis (BD Accuri) or library sorting (BD Aria II). Data analysis and three-parameter curve fits were performed with GraphPad Prism version 6 software. Sorted cells were recovered in SD-CAA liquid medium and incubated at 30°C overnight or until reaching an OD of 4 -8. Yeast surface protein expression was then induced by culturing cells in SG-CAA media at 20°C overnight. The initial library (ϳ5 ϫ 10 7 yeast) was screened twice in the presence of 10 nM matriptase, and then isolated yeast were lysed for DNA extraction (Zymoprep, Fisher Scientific). This DNA was subjected to additional mutagenesis (see above) before retransformation into yeast. This library was screened seven times under equilibrium sorting conditions with 10 nM matriptase. Parallel screening for binding against secondary reagents alone was used to reduce false positives. The final pool of isolated yeast was lysed for DNA extraction (Zymoprep) and transformed into DH10B electrocompetent Escherichia coli cells for plasmid amplification and sequencing (Sequetech, Molecular Cloning Laboratories).

Protease inhibition assay
First, 0.05 nM matriptase (R&D Systems) was added to soluble inhibitor (final concentrations ranging from 0 to 50 nM) containing a 100-l final matriptase assay buffer volume. Solu-ble matriptase substrate (1 M; Boc-QAR-AMC) (R&D Systems) was added to initiate the reaction. Matriptase inhibition assays were carried out on yeast surface-displayed proteins using the same matriptase and substrate conditions except yeast were first counted (ranging from 10 to 10 5 yeast cells/ sample) and incubated with matriptase for at least 1 h prior to addition of the substrate in a 100-l final matriptase assay buffer volume. Additional protease inhibition assays were carried out using 0.5 nM urokinase, trypsin 3, kallikrein 4, and hepsin with a 1 M concentration of each enzyme-specific substrate, Z-GGR-AMC, Mca-RPKPVE-Nval-WRK(Dnp)-NH 2 , Boc-VPR-AMC, and Boc-QRR-AMC, respectively. All enzymes and substrates purchased from R&D Systems were assumed 100% active; buffers and assay conditions were prepared as each R&D Systems protocol describes with 100-l final volume reactions. Soluble inhibitor (final concentrations ranging from 0 to 150 nM excluding inhibitor-depleting conditions) was added initially to each enzyme containing its respective assay buffer, then the reaction was initiated at least 30 min later by addition of substrate, and fluorescent output over time at 380 nm/460 nm (matriptase, urokinase, kallikrein 4, and hepsin) and 320 nm/405 nm (trypsin 3) was measured. Protease activation of fluorescent substrate (relative fluorescent units (RFU)/s) was measured for at least 30 min using a 96-well clearbottom black plate (Fisher Scientific) and a kinetic microplate reader (Synergy H4, BioTek), corrected for background, and then converted to initial relative velocity, v/v 0 . Relative velocities were plotted against inhibitor concentration, and apparent inhibition constants (K i app ) were determined by fitting each curve to the Morrison binding equation (Equation 1) as described previously (38,40,59) using GraphPad Prism version 6 software. Inhibition constants (K i ) were then calculated using

KD2 variant cloning and library construction
The natural KD2 domain comprises amino acids Cys-375 to Cys-425 (GenBank TM accession number AY358969.1). KD2 variant constructs Arg-Cys-Arg (graft 1) and Arg-Cys-Arg-Gly (graft 2) were created using PCR, and gene products were cloned into the pTMY yeast display vector described previously (52) using NheI and MluI restriction sites. Plasmids were transformed into the yeast strain EBY100 using electroporation, expanded in SD-CAA medium at 30°C, and then induced for surface expression using SG-CAA medium at 20°C (45). All yeast-displayed proteins were cloned and prepared in this manner. Yeast expression levels were measured using an anti-HA epitope tag antibody (1:250 dilution of anti-HA mouse primary antibody (Fisher Scientific)) and an anti-HAI-1 antibody (1:100 dilution of rabbit anti-HAI-1 primary antibody (Fisher Scien-Engineering a potent matriptase inhibitor tific)) followed by analysis using flow cytometry (BD Accuri). The KD2 graft 2 library was constructed using error-prone PCR as reported previously (46). In short, variable concentrations of Mn 2ϩ (0.075 or 0.15 mM) were added to reactions to alter the mutational frequency (final average, 2 amino acid mutations/ gene), whereas elevated ratios of dCTP and dTTP nucleotides were added to account for mutational bias. Error-prone PCR was carried out using a low-fidelity Taq polymerase and the following primers: forward primer 5Ј-gctatcttcgctgctttgc-3Ј and reverse primer 5Ј-tgtcagttcctgcaagtcttctt-3Ј. Final PCR products were amplified under high-fidelity conditions (without Mn 2ϩ , equal dNTP ratios, and Phusion polymerase) and then transformed, along with digested pTMY plasmid, into EBY100 competent cells as described previously (60). Cell samples were collected post-transformation, serially diluted, and plated on SD-CAA plates to quantify a transformation efficiency of 5 ϫ 10 7 cells. Cells were then expanded in SD-CAA medium at 30°C, and 5 ϫ 10 8 cells were induced for surface expression in SG-CAA medium at 20°C in preparation for library sorting. Additional rounds of mutagenesis followed the same method as above (46), but the concentration of Mn 2ϩ was increased to 0.3 or 0.4 mM to increase the mutational frequency to 2-5 amino acid mutations per gene. It also appears the wildtype KD1 gene (Cys-250 to Val-303; Gen-Bank accession number AY358969.1) was present during PCR and transformation, allowing recombination of genetic regions of KD1 and KD2 to generate clone 33. Final PCR products were then amplified as before and transformed into EBY100 with an estimated transformation efficiency of 1 ϫ 10 7 cells. Yeast cells were expanded and induced for expression prior to sorting as described above.
Cancer cell binding assay 5 ϫ 10 5 cancer cells were resuspended in cold 1ϫ PBS with 1 mg/ml bovine serum albumin (0.1% BPBS) (50 l to a 10-ml final volume) containing soluble inhibitor (50 pM to 1 M serial dilution) or a 1:100 dilution of soluble human matriptase antibody (Fisher Scientific). Cell solutions were incubated at 4°C for at least 3 h (for single point binding assays) or overnight (for binding curve assays). Cells were then washed with 1 ml of cold 0.01% BPBS, and protein binding was measured using 1:100 anti-mouse phycoerythrin (Invitrogen). Following incubation at 4°C for 15 min, cells were washed with 1 ml of cold 0.01% BPBS and analyzed using flow cytometry (BD Accuri). Mean cell binding (mean RFU) was quantified from at least 10,000 cell events and corrected for controls incubated with antibodies alone. Values were then plotted against inhibitor concentration and normalized to saturating conditions, and K d app values were calculated by fitting graphs to a four-parameter equation (GraphPad version 6).

Pro-HGF activation assay
A gradient of soluble inhibitors (0 -5 nM, serial dilution) was incubated with 0.05 nM soluble matriptase in 10 l of matriptase assay buffer for 30 min at room temperature. Soluble pro-HGF (125 nM) was then added to the solutions, and the reactions were incubated for an additional 2 h at room temperature. Reaction products were then boiled for 10 min at 95°C in the presence of loading dye and reducing agent, then loaded onto a 12% SDS-polyacrylamide gel (GenScript), and subjected to electrophoresis for protein fragment separation. Protein bands were then transferred to a nitrocellulose membrane and probed with a primary anti-HGF ␣-chain antibody (Abcam) followed by an anti-rabbit horseradish peroxidase (HRP) antibody (Fisher Scientific). Protein presence was then detected and imaged using SuperSignal West Femto HRP substrate (Fisher Scientific).

MDCK cell migration/scatter assay
3 ϫ 10 3 MDCK cells were seeded in 96-well plates in 100 l of 2% serum-supplemented medium. Plates were cultured for 24 h in a humidified tissue culture incubator at 37°C in a 5% CO 2 atmosphere. Following incubation, cells were washed twice with warm 1ϫ PBS, and 99 l of serum-free medium was added. Reaction products from the pro-HGF activation assay described above were mixed with a 50 nM final concentration of respective inhibitors to quench the activation of HGF by matriptase. Inhibitor was not added to the "pro-HGF with matriptase" or "pro-HGF alone" controls to allow uninterrupted matriptase activity. All reactions were then diluted 1:10 in matriptase assay buffer (R&D Systems), and then 1 l of each reaction was added to separate wells, all in duplicate. Serumfree medium only served as the untreated negative control. Plates were cultured for 24 h in a humidified incubator at 37°C in a 5% CO 2 atmosphere. Following incubation, cells were stained with crystal violet and then imaged at 10ϫ magnification for qualitative assessment of cell migration in response to active HGF.

Cancer cell activity assay
1 ϫ 10 5 cancer cells were seeded in a 96-well clear-bottom black plate (Fisher Scientific) in 100 l of 2% serumsupplemented medium. Plates were cultured for 24 h in a humidified incubator at 37°C in a 5% CO 2 atmosphere. Following incubation, cells were washed twice with warm 1ϫ PBS, and serum-free medium was added (for a 100-l final reaction volume). Soluble inhibitors (0 -250 nM serial dilution) were then immediately added, and samples were incubated at 37°C in 5% CO 2 for 1 h. A 100 M final matriptase substrate (Boc-QAR-AMC) concentration was then added to initiate the reaction, and fluorescence at 380 nm/460 nm was measured once per hour for 5 h using a microplate reader (Synergy H4). The reactions were incubated at 37°C in 5% CO 2 between reads. The obtained matriptase activity rates (RFU/h) were then normalized to conditions lacking inhibitor and plotted against inhibitor concentration tested. Graphs were fitted using a three-parameter equation (GraphPad version 6), and IC 50 values were calculated.

Cancer cell invasion assay
24-well 8-m pore inserts (Corning) were coated with 50 g of Matrigel (Corning) following the manufacturer's protocols and inserted into a 24-well plate.1 ϫ 10 4 cancer cells in 0.5 ml of 2% serum-supplemented medium were seeded onto the Matrigel-coated inserts, and 1 ϫ 10 4 HEK cells in 0.75 ml of 2% serum-supplemented medium were then seeded onto the Engineering a potent matriptase inhibitor lower chamber; a 100 nM final concentration of soluble inhibitor was added to the lower chamber immediately. The 24-well plates were cultured for 48 h in a humidified tissue culture incubator at 37°C in a 5% CO 2 atmosphere. Following incubation, non-invading cells were removed, and invaded cells were stained using crystal violet and then imaged at three random center fields of view at 10ϫ magnification. Images were analyzed using ImageJ and adjusted for brightness-contrast, cell borders were defined by watershed command; and cell numbers were quantified with the threshold particle setting Ͻ500 pixels excluded to reduce artifacts.

Recombinant protein production
HAI-1 comprises amino acids Met-1 to Glu-449 (GenBank accession number AY358969.1), KD1 comprises amino acids Cys-250 to Val-303 (GenBank accession number AY358969.1), and pro-HGF comprises amino acids Met-1 to Ser-728 (HGF isoform 3, GenBank accession number NP_001010932). DNA encoding the open reading frame of the HAI-1 monomer, pro-HGF, and Fc fusion constructs was cloned into the pCEP4 mammalian expression plasmid (Invitrogen). Genes were cloned into the pCEP4 vector using NotI and HindIII restriction sites and included a C-terminal hexahistidine tag (pro-HGF and monomer inhibitors), or a mouse IgG2a Fc domain was genetically linked using NotI and XhoI restriction sites. pCEP4 vectors were amplified and transfected into adherent HEK cells using Lipofectamine 2000 (Fisher Scientific). Transfected HEK cells were selected using 400 g/ml hygromycin B (Fisher Scientific) and cultured in DMEM containing 10% FBS in a humidified incubator at 37°C in 5% CO 2 . Selected HEK cells were then expanded in T-225 culture flasks (Fisher Scientific) until reaching 70% confluence, recombinant protein expression was initiated by the addition of 25 ml of serum-free DMEM, and protein expression occurred for at least 1 week. The KD1 monomer was cloned into the pPIC9 yeast expression plasmid, transformed into the yeast strain P. pastoris, and expressed using reagents, media, and protocols exactly as described previously (52). Protein-containing supernatants from HEK cells and P. pastoris were purified by nickel-nitrilotriacetic acid metal-chelating chromatography and eluted using 500 mM imidazole (for monomer inhibitors and pro-HGF containing a hexahistidine tag) or by Protein A affinity chromatography and eluted using pH 5 citrate elution buffer (for inhibitors containing a Fc fusion). Eluted protein samples were then buffer-exchanged into 1ϫ PBS or 1ϫ PBS with 500 mM NaCl (pro-HGF only). Each protein was additionally purified by sizeexclusion chromatography (Superdex 75 10/300 GL and Superdex 200 Increase 10/300 GL, GE Healthcare). Purified protein was then characterized using SDS-polyacrylamide gel electrophoresis (SDS-PAGE), and concentrations were quantified by UV-visible absorbance (280 nm) and the following extinction coefficients: HAI-Fc fusion variants, 179,810 cm Ϫ1 M Ϫ1 ; HAI monomer, 57,100 cm Ϫ1 M Ϫ1 ; KD1 monomer, 11,835 cm Ϫ1 M Ϫ1 ; pro-HGF, 149,180 cm Ϫ1 M Ϫ1 . Protein yields typically ranged from 1 to 22 mg/ml. Purified proteins were stored in 1ϫ PBS (inhibitors) or 1ϫ PBS with 500 mM NaCl (pro-HGF) at 4°C and tested within 3 weeks or flash frozen with 0.01% Tween 80 for long-term storage at Ϫ80°C.
Author contributions-A. C. M. conceived and coordinated the study, performed and analyzed (or assisted others with) the experiments in each figure, prepared artwork, and wrote the paper. D. K. helped perform and analyze the experiments shown in Tables 1 and  2, Fig. 4B, and Fig. S5 and helped produce and test constructs in Fig.  3. S. A. H. helped design and optimize the experiments shown in Fig.  2, A-C, and helped produce and test constructs in Fig. 3. R. A. P. S. helped design, perform, and analyze the experiments shown in Figs.