Creating Novel Activated Factor XI Inhibitors through Fragment Based Lead Generation and Structure Aided Drug Design

Activated factor XI (FXIa) inhibitors are anticipated to combine anticoagulant and profibrinolytic effects with a low bleeding risk. This motivated a structure aided fragment based lead generation campaign to create novel FXIa inhibitor leads. A virtual screen, based on docking experiments, was performed to generate a FXIa targeted fragment library for an NMR screen that resulted in the identification of fragments binding in the FXIa S1 binding pocket. The neutral 6-chloro-3,4-dihydro-1H-quinolin-2-one and the weakly basic quinolin-2-amine structures are novel FXIa P1 fragments. The expansion of these fragments towards the FXIa prime side binding sites was aided by solving the X-ray structures of reported FXIa inhibitors that we found to bind in the S1-S1’-S2’ FXIa binding pockets. Combining the X-ray structure information from the identified S1 binding 6-chloro-3,4-dihydro-1H-quinolin-2-one fragment and the S1-S1’-S2’ binding reference compounds enabled structure guided linking and expansion work to achieve one of the most potent and selective FXIa inhibitors reported to date, compound 13, with a FXIa IC50 of 1.0 nM. The hydrophilicity and large polar surface area of the potent S1-S1’-S2’ binding FXIa inhibitors compromised permeability. Initial work to expand the 6-chloro-3,4-dihydro-1H-quinolin-2-one fragment towards the prime side to yield molecules with less hydrophilicity shows promise to afford potent, selective and orally bioavailable compounds.


Introduction
A well balanced haemostasis system is important to both minimize blood loss and disturbances of blood flow. Upon injury of the vessel wall, blood is exposed to tissue factor which via a cascade reaction leads to thrombin generation and a fibrin cross-linked clot to mend the injury and stop bleeding. Factor XI (FXI) has an important role in thrombin generation in the amplification phase of the coagulation process. However, over-production of thrombin may lead to excessive clots resulting in thrombosis. Also, high levels of thrombin cause activation of thrombin activated fibrinolysis inhibitor which hinders fibrinolysis. Therefore, decreased levels of thrombin will indirectly increase the rate of fibrinolysis. Inhibition of activated FXI (FXIa) should decrease thrombin generation in the amplification phase, but not in the initiation phase, and thus yield an antithrombotic and profibrinolytic effect with minimal risk of bleeding (see reviews [1][2][3]). Bleeding is a serious concern with current antithrombotic drugs and FXIa inhibitors could address this issue.
The role of FXIa in haemostasis and thrombosis in human has been extensively studied. Human haemophilia C patients who are severely deficient in FXI display reduced incidence of ischemic stroke [4]. Unlike haemophilia A and B patients, who are deficient in FVIII and FIX, respectively, haemophilia C patients seldom experience spontaneous bleeding [5]. The bleeding associated with FXI deficiency usually occurs after trauma or surgery in the tissues with high fibrinolytic activity [6,7]. An increased level of factor XI has been reported as a risk factor for deep venous thrombosis [8,9], myocardial infarction [10] and ischemic stroke [11,12]. There is also much research on the role of FXI in animals. Several studies have demonstrated that FXInull mice are protected against venous and arterial thrombosis without an adverse effect on bleeding time [13][14][15][16][17][18]. Recent reports present similar effects in mice [19] and primates [20] using antisense oligonucleotides to inhibit FXI production [19]. Antibodies against FXI/FXIa have been shown in one study to reduce thrombus growth in the rabbit iliac artery in the presence of repeated balloon injury [21], and in another study to increase endogenous thrombolysis in rabbit about two-fold in comparison to control antibodies [22]. Also, an anti-human antibody, aXIMab, prevented vascular graft occlusion in baboons [23]. In summary, there is ample evidence in support of FXIa as an attractive antithrombotic and profibrinolytic target.
FXIa small molecule inhibitors have not reached the same level of maturity as thrombin and activated factor X (FXa) inhibitors. The thrombin inhibitor dabigatran [24] and the FXa inhibitor rivaroxaban and apixaban [25] are approved anticoagulant drugs in several markets, but adverse bleeding remains an area where improvement is requested. In contrast, inhibitors of FXIa are still in preclinical development. Daiichi Sankyo Co has reported on potent and selective peptidomimetic alpha-ketothiazole arginine based covalent FXIa inhibitors [26,27], and one compound was shown to display similar antithrombotic efficacy as heparin in a rat venous thrombosis model [26]. Similarly, Bristol Myers Squibb (BMS) demonstrated in vivo antithrombotic efficacy in rat models with BMS-262084, a potent and selective beta-lactam arginine that irreversibly inhibits FXIa with an IC 50 of 2.8 nM [28]. Recently, BMS also showed antithrombotic efficacy without increased bleeding in a rabbit model with a reversible selective small molecule FXIa inhibitor [29]. Patent applications from BMS display lists of selective FXIa inhibitors, or dual FXIa and plasma kallikrein inhibitors, with IC 50 values in the low nM range [30][31][32]. These examples encourage further work with the aim of reaching the clinical setting for small molecule FXIa inhibitors.
In-house high throughput screening (HTS) attempts had previously failed to identify viable leads. Therefore, structure aided fragment based lead generation (FBLG) was chosen as a rescue strategy to create new FXIa inhibitor leads. The choice was supported by the feasibility to generate X-ray structures, to perform a fragment screen using nuclear magnetic resonance (NMR) and to apply structure based virtual screening to select a fragment library.
These techniques are suitable in fragment campaigns, as they allow for detection of the very weak binding (dissociation binding constant (K D ) in the mM range) that can be expected for the small molecules screened. Fragment based approaches to drug discovery have during the last years presented many successful examples [45,46], and at least one campaign has lead to a product on the market [46,47]. Fig. 1 depicts the FXIa active site and describes the standard nomenclature according to Schechter and Berger [48] for substrate and ligand binding to this site. P1 and P1' denote the peptide residues N-and C-terminal of the scissile bond, and S1 and S1' denote the corresponding enzyme binding sites.
The aim of the present study was to create novel FXIa inhibitors with potential to be developed into potent, selective and orally bioavailable compounds by targeting the S1-S1'-S2' FXIa binding pockets. In the FXIa FBLG campaign presented here, virtual screening, NMR, SPR and X-ray crystallography together with enzymatic inhibition assays were successfully applied. We report the identification of two novel FXIa S1 binding fragments and the evolution of these into nM inhibitors using X-ray structures to guide the design. Initial fragment expansion work shows promise to afford potent, selective and orally bioavailable compounds, which motivates further optimization work on presented compounds. Depiction of FXIa active site in complex with FIXa substrate residues (from PDB entry 1XXD [82]). According to standard nomenclature, the substrate P1 residue binds the enzyme S1 site, the P1' residue binds the S1' site, and so on. The scissile bond is marked with a red dashed line.

Fragment library selection
The structural characteristics of the target protein binding site and the quality of the initial library of fragments to be screened are two critical determinants for the success of an FBLG campaign. The most prominent and probable fragment binding site of FXIa is the S1 pocket. A key feature of the S1 pocket is Asp189 (trypsin numbering) deep down in the cavity and most known ligands contain a strongly basic group that forms ionic interactions with this residue [26,27,49]. However, it has been shown that the basicity is not required [50,51], and since strongly basic P1 groups often pose high hurdles in achieving adequate oral bioavailability, we steered our fragment selection towards non-basic or weakly basic compounds. Importantly, the S1 pocket in FXIa harbors an alanine in position 190, a feature shared with thrombin and FXa whereas the corresponding residue in activated factor IX (FIXa) and activated factor VIIa is serine. The Ala190 provides a pocket where a neutral P1 is likely to be more easily accommodated than when this residue is a serine. Apart from S1, cavity analyses using the Schrödinger Sitemap tool [52] suggested that favorable ligand-protein interactions could be picked up towards the S1'-S2' sites in FXIa. Also, comparisons of primary sequences and structural features around the S1'-S2' sites for thrombin, FXa and FXIa suggested opportunities for achieving selective FXIa inhibitors. Thus, the project set out to primarily target the S1-S1'-S2' binding pockets.
The initial fragment selection focused on targeting the S1 site. A total of about 65000 fragments from the AstraZeneca screening collection with MW < 250 g/mol were docked using the Glide software from Schrödinger [53][54][55] to the FXIa X-ray structure (PDB id 1ZSJ), with the N-(7-Carbamimidoyl-naphthalen-1-yl)-3-hydroxy-2-methyl-benzamide ligand and all water molecules removed. Docked fragments with a halogen and/or nitrogen interacting with S1 pocket residues were then manually inspected for a selection of about 1200 structures. The same 65000 fragments were also docked into an in-house structure of FXIa catalytic domain (CD) in complex with compound 1, see Fig. 2, with the ligand and waters removed. From these docking experiments, an additional 600 structures were selected based on Glide score and chemical tractability. In total, 1800 structures were submitted to experimental screening.

Screening cascade
Ligand detected one dimensional (1D) NMR spectroscopy [56][57][58] was used as the primary fragment screening tool, followed by SPR and enzymatic assays to determine K D and IC 50 values, respectively. Finally, structures of the FXIa catalytic domain, containing the four mutations described by Jin et al., [59] (herein referred to as FXIa CD) in complex with ligands were solved by X-ray crystallography.
Ligand detected 1D NMR. Compounds were screened at 100 μM in mixtures of six. Weakly binding fragments were identified from intensity changes through measurements on three samples viz. (1) mixtures of fragments at 100 μM in buffer; (2) with addition of 5 μM FXIa CD; (3) with subsequent addition of 10 μM of the high affinity inhibitor compound 2 (see Fig. 2). The signal from a compound binding with low affinity to FXIa CD decreased upon addition of protein and regained upon addition of a high affinity reference compound. Comparison with reference spectra of individual compounds enabled identification of the binding fragment in a mixture. Compound 2 binds to the S1-S2-S3-S4 pockets of FXIa CD with an IC 50 of 110 nM, and is expected to displace any fragment binding weakly to the S1 site as well as to the nonprime side of the binding pocket, i.e. S2-S3-S4. As a positive control, the S1 binding ligand para-methyl-benzamidine (pMeBza), compound 3, was used, and in Fig. 3 the doublet at 7.66 ppm is shown.

SPR.
Compounds that displayed binding to FXIa CD in the NMR screen were followed up by K D determinations using SPR. The SPR data were then used to rank and select compounds for X-ray crystallography experiments. The SPR method was based on an inhibition in solution assay in which compound 4, immobilized on the biosensor surface, served as a probe for the FXIa catalytic site. The advantages of inhibition in solution assay as compared to the more conventional direct binding assays where the protein is immobilized include (1) substantially increased sensitivity, (2) generic and rapid assay development and (3) immediate verification of competitive binding. Fig. 4 illustrates the design of compound 4 based on the crystal structure of FXIa in complex with compound 1. Since any ligand that binds to the S1 pocket competes with compound 4 for binding to FXIa CD, it is possible to derive the K D for that ligand [60]. The initial binding rate of the protein to compound 4, K on , was used to determine the percentage of free protein in solution which changed by varying the concentration of the competing ligand. The free protein concentration was plotted against the logarithm of the ligand concentration and a sigmoidal dose-response curve-fit model was applied to determine the K D -value.
Enzymatic assay. All fragments that showed competitive binding in the NMR assay (fragment hit compounds) and all compounds resulting from expansions of fragments were screened in an enzymatic assay to determine their inhibitory activity against FXIa. Selected compounds were also screened against other serine proteases to evaluate their selectivity profiles. All protease activities were determined by chromogenic assays monitoring the release of para-nitroanilide at 405 nm.
X-ray crystallography. Selected fragments and expanded compounds that had been confirmed to be binding to the active site by either NMR or SPR were attempted for structure determination by X-ray crystallography. To improve crystallisability, the quadruple mutant (Ser75Ala, Lys78Ala, Thr97Ala and Cys123Ser), first reported by Jin et al., [59] was generated. The mutated residues in this construct (FXIa CD) are located at least 16 Å from the catalytic residue Ser195 and the alterations are therefore not anticipated to impact ligand binding in the active site. The selected fragments were either co-crystallized with FXIa CD, or the protein crystals were soaked in a solution containing the ligand. The crystals typically diffracted to around 2 Å on an in-house X-ray source. Fragments were considered for chemical expansion if their binding mode could be determined by X-ray crystallography.

Screening results
NMR screening of the 1800 structures selected by virtual screening resulted in 13 neutral or weakly basic fragment hits (calculated pKa < 9.0, Advanced Chemistry Development, Inc. ACD/LabspKaDB (version 6.0) software [61]). These hits were FXIa S1 binding fragments as demonstrated by pMeBza competition in NMR experiments. About 600 near neighbours to the original 13 hits were selected and screened by NMR to yield a total of 50 hits. Several of the fragment hit compounds were only weakly basic, such as aminopyridines and aminoquinolines, and neutral compounds, such as chloroquinolinones and the known p-chlorophenyltetrazoles [32], were also identified. Two prioritized fragment hits, 6-chloro-3,4-dihydro-1Hquinolin-2-one (5) and 4-methylquinolin-6-amine (6) are shown in Fig. 5. To our knowledge, 5 is a novel serine protease P1 structure, and 6 is here described for the first time to bind FXIa. The crystal structures of compounds 5 and 6 in complex with FXIa were solved to 2.0 and 1.7 Å resolution, respectively, and the fragments could be unambiguously placed in the electron density maps (Fig. 5). The two fragments display distinct binding modes as depicted in Fig. 5. These binding modes were predicted from the GLIDE dockings of compounds 12 and 19 in the virtual screen. After alignment of the proteins using the SSM function in Coot [62], a root mean square deviation (r.m.s.d.) of 0.44 Å for the compound 12 docking pose and crystal structure was calculated using heavy atom positions only. For compound 6, the binding mode was predicted using the docking pose of compound 19 (r.m.s.d. of 0.61 Å including all heavy atoms except for the amino group in the 6-position).
Fragment 5 displayed a FXIa IC 50 of 140 μM. The chloro atom is directed almost perpendicular to the aromatic ring of Tyr228 and is located at a distance of 3.8 Å to the closest carbon atom of the phenyl ring. This chloro atom replaces a highly conserved water molecule (see Fig. 5B) and its position is similar to the chloro atom in chloro-phenyl containing P1s of FXIa, thrombin and FXa inhibitor [63,64]. The halogen bonding between electron rich groups including pi systems and the σ-hole of the halogen with a clear directionality has recently received attention and has been reviewed in several articles [65][66][67][68]. The chloroquinolinone compound 5 picks up an additional interaction with FXIa, i.e. the quinolinone NH group forms a 2.7 Å hydrogen bond with the backbone carbonyl of Gly218. Notably, the chloroquinolinone is slightly more potent than benzamidine, compound 7 (see Fig. 2), despite the fact that the former is neutral and does not form polar interactions with Asp189. This suggests that the chloroquinolinone P1 is a promising FXIa fragment inhibitor, with potential to be expanded and developed into an orally active drug.
The aminoquinoline fragment 6 shown in Fig. 5 binds to the S1 pocket with an IC 50 of 160 μM. The calculated pKa 7.9 (ACD/LabspKaDB [61]), is supported by an experimental pKa value of 7.1 from the expanded analogue, compound 8. The amino group of the ligand interacts with the carboxyl group in Asp189 through a 3.1 Å hydrogen bond, and the quinoline nitrogen forms a 2.9 Å hydrogen bond with the carbonyl group of Gly218 (Fig. 5). Both the quinolinone and the aminoquinoline fragments form hydrogen bonds to Gly218, even though their overall binding geometries in the S1 pocket are quite different. The conserved water molecule stacking against Tyr228 is bound in the S1 pocket close to the fragment with a distance of 3.6 Å between the ligand amine nitrogen and the water oxygen.

Structure aided fragment expansion
In parallel with the fragment hit identification work, compounds 9-11 from patent applications from Bristol Myers Squibb [30,31] were synthesized and characterized. The compounds are potent FXIa and plasma prekallikrein inhibitors and selective against e.g. thrombin and FXa as seen in Fig. 6. The crystal structures of 9-11 in complex with FXIa CD were not available in the public domain and were therefore solved by X-ray crystallography. Given our aim to target the prime side, it was exciting to note the S1-S1'-S2' binding mode for all three compounds, as shown for compound 9 in Fig. 7. Superposition of the crystal structures showed that the chloro-phenyl group of compound 9 overlays with the corresponding group of the chloroquinolinone fragment (compound 5), and the chloro-phenyltetrazole group binds as anticipated from previously reported thrombin structures in complex with ligands carrying the same P1 [63,64]. The P1-P1' amide group, extended from the P1 via an ethyl linker, interacts with the active site Ser195 where the carbonyl oxygen forms hydrogen bonds with the amide nitrogen of Ser195 and a central tightly bound water molecule (Fig. 7). The benzyl group acts as P1' and binds in the S1' cavity and the amine of the amide group linking P1' and P2' forms a hydrogen bond with the central water molecule that is not only located within hydrogen bond distances to both amide groups of the inhibitor, but also to the hydroxyl of Ser195 and to the backbone carbonyl of Leu39. The carbonyl from the P1-P1' amide is positioned in the oxyanion hole with hydrogen bonds to the main chain nitrogens of Gly193 and Ser195. The P2' hydroxyquinolinone moiety interacts with His38 and Tyr143 in the S2' pocket. Two glycerol molecules, originating from the cryo solution from the crystal freezing step, were bound in close vicinity to the ligand, one stacking against the hydroxyquinolinone in the S2' region and the second was bound close to the benzyl moiety (not shown). However, since the S2' substituent of compound 11 shows a highly similar pose (not shown), the glycerol molecules do not appear to alter the binding mode. It should be noted that after completion of the experimental work presented here, the structures of two FXIa inhibitors with closely analogous P1'-P2' moieties that show similar binding modes have been published [51].
The crystal structures of compounds 9-11 in complex with FXIa CD gave a clear handle on directed expansion of fragments towards the prime side and allowed linking exercises. Fragment 5 was selected as the primary target for linking/expansion work. Disregarding overall physicochemical properties in general and ADME properties in particular, the first key question addressed was if the novel chloroquinolinone P1 could be linked to the P1'-P2' parts of compound 9 to yield highly potent and selective FXIa inhibitors. Fig. 8 outlines the structural basis and outcome of this exercise. The X-ray structures of the FXIa CD complexes with the chloroquinolinone methylated in the 4-position (compound 12) and compound 9 were aligned in order to guide linking. Based on X-ray structure information and docking studies, we deduced that the hydrogen bond acceptor in the defined space as exemplified by the P1-P1' amide carbonyl oxygen in compound 9 (see Fig. 7) was favorable and thus judged critical to maintain in any designed new compounds. Another key feature is a hydrogen donor moiety, exemplified by the P1'-P2' amide nitrogen that interacts with a central water molecule that is highly conserved in position (see Fig. 7). Thus, linkers were designed to preserve the experimentally determined positions of both the P1 fragment, the P1-P1' amide carbonyl oxygen and the P1'-P2' amide nitrogen, without introducing steric strain. A range of possible linker sizes and geometries were conceived by visual inspection of the overlay of available protein-ligand crystal structures, and subsequently tested and prioritized by docking experiments of linked full length compounds using GOLD [69] and GLIDE [53][54][55]. The most promising compounds with unstrained geometry, maintaining the linked moieties in place, were chosen for synthesis. No estimate for the free energy of binding was used, but the approach rather selected compounds for synthesis based on good predicted geometry, leaving the actual binding affinity to be determined experimentally. The first linker chosen to connect compound 12 and the P1'-P2' parts of compound 9 resulted in compound 13, which binds FXIa CD with a K D of 0.5 nM and is one of the most potent FXIa inhibitors (IC 50 = 1.0 nM) reported to date.
The potency SAR observed in this first expansion/linking work is summarized in Fig. 9. The FXIa IC 50 of the chloroquinolinone 5 P1 binder is 140 μM with a ligand efficiency [70] (LE) of 0.44. The P1-P1' compound 14 increasing in MW from 180 to 398 resulted in only minor improvement in potency with an IC 50 of 33 μM and a drop in LE to 0.22. Adding on the P2' part boosted the potency more than four orders of magnitude to an IC 50 of 1 nM and a LE of 0.32 (compound 13). The isolated P1'-P2' moiety, compound 15, showed a poor binding affinity (K D = 1.4 mM), and the N,N-dimethylated analogue, compound 16, confirmed S1'-S2' binding with a nearly perfect overlap with the identical P1'-P2' groups of compound 13 (X-ray data not shown). The SAR outlined above can be regarded as an example of superadditivity energetics achieved by fragment linking, similarly as recently described in detail by Nazaré et al. for FXa inhibitors [71]. Rigid body entropy loss is expected to be similar for the individual fragments (here, fragments 5 and 15) as for the final fragments linked molecule (here, compound 13). This rigid body entropy loss has been estimated as 15-20 kJ/mol at 298 K by Murray and Verdonk [72]. Thus, superadditivity with respect to the starting fragments binding affinities, in the range of three orders of magnitude, is expected when fragments are optimally linked. Adapting the following equation from Nazaré et al. [71], to the present example yields: ΔG link = ΔG compound 13 -ΔG fragment 5 -ΔG fragment 15 = -53.1 kJ/mol + 23.7 kJ/mol + 16.3 kJ/mol = -13.1 kJ/mol. This linker contribution of -13.1 kJ/mol corresponds to an improvement by about 2.3 orders of magnitude relative to the individual fragment affinities, and agrees well with the estimated rigid body entropy loss of 15-20 kJ/mol [72]. This example shows the reward from superadditive affinity gain by joining two weakly binding fragments with a near optimal linker. Fragment expansion/linking, based on compound 9, with different P1 fragments are summarized in Fig. 2. It is worth noting that the 6-methyl-3,4-dihydro-1H-quinolin-2-one P1 fragment 17 was not identified as a hit in NMR but after the expansion that resulted in compound 18 yielded a FXIa IC 50 of 10 nM, emphasizing that fragments with very weak binding affinities can provide useful starting points for expansion. The expansion/linking of the aminoquinoline fragment 6 with the P1'-P2' part of compound 9 resulted in compound 8 with an IC 50 of 120 nM. The choice of the 5-position of 6 as site for initial expansion of the aminoquinoline was based on synthetic feasibility considerations. However, analyses of the aligned structures suggested that analogous expansion/linking from the 4-position might be more favorable, i.e. expanding from the methyl position in compound 19 (an equipotent analogue to compound 6).
Analogous expansion/linking work was also performed based on compound 10 with amide bonds reversed as compared to compound 9, see data in Fig The structure aided expansion/linking work described above is briefly summarized as follows: Both the X-ray structures and docking studies presented here showed that the hydrogen bond acceptor and donor positions highlighted in Fig. 7 are clearly defined in space and thus key features to utilize in modeling. The novel FXIa P1s, chloroquinolinone, dihydroquinolinone and aminoquinoline, can upon expansion yield potent and selective P1-P1'-P2' compounds. The 4-position in chloroquinolinones and dihydroquinolinones is a favorable point for expansion. Amides and reverse amides are both acceptable, but with compounds synthesized so far, the amide nitrogen closer to the P1 part yield more potent binders.
Figs. 12-16 describe the synthetic steps to arrive at remaining screening compounds.

Compound characterization
It is important to achieve potent FXIa inhibitors with high selectivity vs. other serine proteases in order to minimize risk for adverse effects, particularly with regard to bleeding, which limits the utility of current antithrombotic drugs, e.g. dabigatran and rivaroxaban. Selectivity data for FXa, thrombin, trypsin, plasma kallikrein (PKK), FIXa, tissue plasminogen activator (tPA) and plasmin were therefore generated and are summarized in Fig. 6. Assessment of fragment selectivity is limited both due to the low potencies of the fragments and that the highest compound concentration used in the assay was 99 μM. The P1' expanded chloroquinolinone, compound 14, with a FXIa IC 50  Regarding species selectivity, it was noted that compound 13 was about 30000-fold less potent against rat FXIa than against human FXIa. We ascribed this discrepancy primarily due to the replacement of Ala190 in human FXIa by Thr190 in rat, which makes the S1 pocket more narrow and hydrophilic and thus disfavor binding of the neutral chloroquinolinone P1. In contrast, benzamidine containing FXIa inhibitors show similar FXIa potencies between human and rat. This finding has implications on choice of species for in vivo studies and highlights the importance of doing species analyses early in projects.
The next stage of the lead identification work focused on a more balanced approach to address not only potency and selectivity but also physicochemical and ADME properties. As seen in Fig. 17, compound 13 displays high polar surface area (PSA), low permeability, high efflux and high in vivo clearance. The high rat in vivo clearance despite the low/moderate in vitro intrinsic clearance in rat hepatocytes may suggest active transport into bile and/or urine in vivo. We hypothesized that, relative to compound 13, improved permeability and decreased efflux  should increase bioavailability and lower clearance in vivo. Initial ADME exploration targeted minor modifications of compound 13 to first methylate the P1-P1' amide nitrogen, as in compound 25, and subsequently adding a second methyl group on the P1'-P2' amide nitrogen, yielding compound 26. Minor improvements in permeability and efflux were observed with compound 25, and compound 26 provided no further benefit. From the structure of FXIa in complex with compound 11, the methyl group on the P1-P1' amide in compound 25 is predicted to point out towards the solution, whereas the methyl group on the P1'-P2' in compound 26 is predicted to disturb important hydrogen bond interactions with the tightly bound water molecule shown in Fig. 7 and would therefore be detrimental for potency. The IC 50 values of <0.005 and 1.7 μM for compound 25 and 26, respectively, support this notion. After the methylation exercise, it was decided to keep the methylated P1-P1' amide as first step expansion of the P1 fragment and search for further expansion replacing the P1'-P2' amide with less hydrophilic structures. In this context, 13 compounds available in our corporate compound collection with structures containing the N-methyl-N-[(2-oxo-1H-quinolin-4-yl)methyl]formamide scaffold were screened. Compound 27 with a FXIa IC 50 of 99 μM was identified, which displayed high permeability (see Fig. 17), but still showed relatively high efflux. The low potency of compound 27 was improved by approximately two orders of magnitude upon addition of the chloro atom to the aminoquinolinone P1 to afford compound 28 with an IC 50 of 0.82 μM. The ADME properties of compound 28 with high permeability (Caco2 A to B = 19 Á 10 -6 cm/s), moderate rat in vivo clearance (18 mL/min/kg) and a measured oral bioavailability in rat of 27% show promise towards the goal of developing an oral drug. Potency improvements of this lead with maintained favorable ADME properties remain to be shown. However, comparing the potencies of the P1-P1' compounds 28 and 14 (IC 50 values of 0.82 and 33 μM, respectively), and realizing the potency boost feasible by adding a P2' moiety (compare compounds 14 and 13 in Fig. 9), suggest exciting opportunities for development of a potent, selective and oral FXIa inhibitor drug.

Summary and Conclusion
A virtual screen based on docking experiments was performed to select a fragment library, and NMR was used as the primary screen with minimal risk of false positives. SPR and enzymatic screening quantified the affinities of the NMR hits, and X-ray structures provided the basis for fragment expansion and linking design.
The structure aided fragment based lead generation strategy adopted in this work successfully identified two novel FXIa P1 fragments that were expanded towards the S1'-S2' pockets to yield potent and selective inhibitors. The chloroquinolinone is neutral and the aminoquinoline is only weakly basic and should provide opportunities to be developed into bioavailable full length FXIa inhibitors without relying on a prodrug approach. It remains to be explored whether the FXIa chloroquinolinone and aminoquinoline P1s will have more general utility as serine protease P1s. The very potent and selective compound 13 gives promise for further optimization. The future challenge will be to improve the physicochemical properties of the chloroquinolinone and aminoquinoline P1 based compounds presented here to optimize the combined potency, selectivity, bioavailability and clearance profiles. Compound 28 indicates

Protein expression and purification
Protein expression and purification of the catalytic domain (residues 370-607) was carried out as described by Jin et al. [59]. To improve crystallisability this protein contains the four mutations: S434A, K437A, T475A and C482S (S75A, K78A, T115A, C123S, trypsin numbering). This protein was used for crystallisation, NMR and SPR measurements. Structures with associated FXIa potencies and selected physicochemical and ADME parameters. a logD measured by liquid chromatography. b Caco2 A to B permeability measured at pH 6.5. c Efflux ratio determined in Caco2 cells by dividing P app B to A with P app A to B permeabilities measured at pH 7.4 at side A. d Rat intrinsic clearance determined from measurements in rat hepatocytes. e Rat in vivo clearance. f Caco2 A to B permeability measured at pH 7.4 at side A. NMR screening NMR samples were prepared in aqueous buffer containing 50 mM deuterated TRIS, pH 7.4, 3 mM NaN 3 , and 10% D 2 O. FXIa CD stock solution was buffer exchanged and concentrated to approximately 100 μM (2.7 mg/mL) (UV absorption ε 280 = 60390 cm -1 ) using 10 kDa cut-off Millipore filter centrifugation tubes. Control experiments with pMeBza and compound 2 were done on all protein preparations (Fig. 3). Stock solutions of test compounds were prepared in randomised mixtures of six by pooling 10 μL aliquots of 10 mM DMSO solutions from the AZ compound library in microtiter plates followed by vacuum evaporation of the solvent. The solid compound mixtures were then redissolved in 20 μL DMSO-d 6 yielding 5 mM compound stock solutions. Samples for retesting of individual compounds were prepared from original compound library solution in regular DMSO, necessitating the simultaneous suppression of solvent and DMSO peaks (see below). Samples were prepared "just in time" using a Genesis liquid handling robot (TECAN, CH) and transferred to the NMR spectrometer via a SampleRail system (Bruker Biospin, DE). Mixtures were screened at 100 μM compound concentration in aqueous buffer with 5 μM FXIa CD added in the second step. At the third step, 10 μM inhibitor (compound 2) was added.
NMR data were collected at 20 o C on a Bruker Avance III 600 MHz NMR spectrometer equipped with a cryo probe head. 1D 1 H T 1ρ NMR spectra [56] were recorded with 256 scans, 3.2 s interscan delay, 180 ms spin lock time and 64 kHz spin lock field. Solvent suppression was achieved by applying an 8 ms excitation sculpting sequence [58] before acquisition. For samples prepared from non-deuterated DMSO stock solutions, simultaneous suppression of the solvent and DMSO peaks was achieved by replacing the selective 180 degree flip back pulses with bilinear pulses created with the Pulse tool module in TopSpin (Bruker, CH). Spectra were analyzed in TopSpin and compounds were considered hits if they displayed significant signal amplitude changes upon addition of protein and inhibitor.

FXIa surface plasmon resonance assay
The tethered free amine ligand 4 was covalently linked to the dextran layer on the CM5 sensor chip (GE Healthcare/Biacore) using the Amine Coupling Kit (Healthcare/Biacore) and performed according to the manufacturer's recommendations. Carboxyl groups on the dextran layer were activated by injecting a 1:1 mixture of 0.5 M N-ethyl-N' (3-dimethylaminopropyl)carbodiimide and 0.5 M N-hydroxysuccinimide at 10 μL/min for 7 min. The activated carboxyl groups on the control channels were blocked by injecting a 1 M ethanolamine solution for 7 min at 10 μL/min. The tethered ligand was coupled to the enzyme binding channels by injecting a 100 μM solution of the compound (10 mM 4-(2-hydroxyethyl)-1-piperazineethanesulfonic acid, pH 7.4, 0.15 M NaCl, 0.005% (v/v) polysorbate 20, 2% DMSO) over the chip surface at 10 μL/min for 1 min. The remaining activated groups were blocked with 1 M ethanolamine. Compounds were tested in a concentration response curve. The compounds were serially diluted 1:3 in DMSO to form a 10-point concentration response curve with 5 mM for the highest concentration. The FXIa CD was added to the plate at a final concentration of 50 nM and the mixture was incubated for 15 min at RT. The plate was then sealed and placed in the Biacore 3000 instrument. Samples were injected at 20 μL/min for 2 min. The signal generated on the control channel was subtracted from the signal of the channel with the tethered ligand. The value used to calculate percent inhibition of binding was the slope of the response measured 8 s after the start of the injection. The surface of the chip was regenerated by injecting 0.5% SDS for a period of 36 s.

Crystallisation and data collection
In order to obtain ligand complexes with compounds 5, 6, 8, 10, 11, 12 and 13, co-crystallisation was used. FXIa (% 1 mg/ml in 20 mM Tris pH 7.4, 200 mM NaCl) was mixed with respective compound to a concentration of 1-5 mM and incubated at room temperature for around 1 h. The complexes were then concentrated to about 20 mg/ml. Crystallisations was set up at 20°C using  Table 1.
For obtaining ligand complexes with compounds 1 and 9, powder soaks to preformed FXIa-benzamidine crystals were performed. FXIa-benzamidine crystals were obtained by mixing diluted FXIa (% 1 mg/mL in 20 mM Tris pH 7.4, 200 mM NaCl) with benzamidine to a concentration of 10 mM and incubating at room temperature for 1 h. The complex was concentrated to % 20 mg/mL and crystallisation set up with the precipitate solution of 1.  Table 1.

Data processing and structure determination
The crystals belong to the spacegroup I23 with one copy of the FXIa CD in the asymmetric unit. The data for the FXIa:compound 1 complex was processed using autoPROC [73] and XDS [74] and scaled with Aimless [75]. All other data was processed with the programs Mosflm [76] and Scala [75,77]. Structure determination by molecular replacement was conducted with the program Molrep [75,78], Phaser [79] or rigid body refinement with Refmac [75,80], using an in house high-resolution crystal structure as search model. Rebuilding of the model and addition of water molecules were done in the program Coot [62]. The structure of the FXIa:compound 1 complex was refined with Refmac and autoBUSTER (BUSTER version 2.11.5. Cambridge, United Kingdom: Global Phasing Ltd) and the final protein model comprised residues 16-130 and 133-244 (chymotrypsin numbering). Compound 1 was well defined in the electron density, except for the benzyl group which was modelled in two different orientations, the predominant orientation is depicted in Fig. 4. All other datasets were refined using Refmac [80] and all protein residues, 16-245, were included in the final models. The ligands were placed in the F o F c difference maps during the last steps of refinement. Details from data collection, refinement and the final models are summarised in Table 1. The atomic coordinates and structure factors have been deposited in the protein data bank with accession codes 4CR5, 4CR9, 4CRA, 4CRB, 4CRC, 4CRD, 4CRE, 4CRF and 4CRG.

Enzymatic assays
Once the compounds were synthesized and characterized, enzymatic assays were performed to evaluate the inhibitory activity of the resulting compounds against FXIa and other serine proteases. All protease activities were determined by chromogenic assays monitoring the release of para-nitroanilide at 405 nm. IC 50 values for the compounds were determined using equation: Q denotes the assay readout, V denotes the uninhibited velocity of the enzyme, [I] denotes the inhibitor concentration, n denotes the slope and IC 50 is the inhibitor concentration at V/2. The IC 50 values were converted into K i values for competitive inhibitors using the equation:  Thrombin. Human recombinant thrombin was prepared in house. Inhibition by compounds was measured using 300 μM pyroGlu-Pro-Arg-pNAÁHCl (Km = 280 μM) as substrate in 50 mM Tris/HCl pH 7.4, 100 mM NaCl, 0.2% BSA (w/v), 1% DMSO (v/v).
FIXa. Human FIXa was from Enzyme Research Laboratories, Swansea, UK. Inhibition by compounds was measured using 500 μM CH 3

Chromatographic determination of lipophilicity
The chromatographic data was obtained utilizing a Waters Aquity UPLC chromatographic system running a gradient from 5-95% acetonitrile at pH7.4. The stationary phase was a Waters BEH C18 2.1x50 mm. Standardized retention data were obtained by normalization against a set of compounds with precisely determined (octanol/water) logD-values. Lipophilicity values for samples were derived from a corresponding calibration curve.

Caco2 permeability and efflux
A monolayer of Caco2 cells, cultured on semi-permeable polycarbonate surfaces, was used to study the permeability in the apical to basolateral direction. The process was atomized by a robotic Tecan EVO platform (Männerdorf, Switzerland) and 24-transwell plates from Costar (Cambridge, MA, USA). HBSS buffer pH 7.4 was dispensed to the basal side of the monolayer. The assay was initiated by adding the test substrate (10 μmol/l in HBSS buffer pH 6.5) to the apical side of the monolayer. Samples for quantitative analysis by LC/MS/MS were withdrawn before the addition of the test substrate and at 45 and 120 min post addition of the test substrate. During incubation the transwell plates were placed in a shaking incubator at 37°C between sampling. The peak areas were exported to Excel where Papp-values and recoveries were calculated. Twenty-two reference compounds with known human oral bioavailability were used to establish a correlation curve with P app versus the fraction of the oral dose absorbed (f a ).
Efflux was studied in the same Caco2 cells with the modification that HBSS buffer pH 7.4 was used on both sides of the monolayer. The permeability was studied in the apical to basolateral direction (P app AB) as well as in the basolateral to apical direction and (P app BA). Efflux ratios were derived from the following equation: P app BA/P app AB.

Rat hepatocyte metabolic stability assay
Rat hepatocyte metabolic stability was determined in accordance with the method described by Jacobson et al. [81]. Cryo preserved hepatocytes at a concentration of 10 6 viable cells/mL were used. After thawing, hepatocytes were incubated for 10 min to warm to 37°C and test compounds, dissolved in acetonitrile, were added to give a final concentration of 1 μM. At 0.5, 5, 15, 30, 45, 60, 80, 100 and 120 min, the incubation system was mixed and 20 μL aliquots were transferred at each time point to wells in a separate plate filled with 80 μL acetonitrile to stop the reaction. The quenching plate was then vortexed followed by centrifugation, and supernatants were analyzed by LC/MS/MS. Peak areas were determined from extracted ion chromatograms, and the in vitro intrinsic clearance (in vitro CL int , in μL/min/10 6 cells) of parent compound was calculated from the slope in the regression analysis of the natural logarithm of parent concentration vs. time curve.

In vivo rat PK
Two days prior to dosing, female Sprague Dawley rats were prepared by cannulation of the left carotid artery for blood sampling and by cannulation of the right jugular vein for intravenous administration. The catheters were filled with heparin (100 IU/mL), exteriorized at the nape of the neck and sealed. The surgery was performed under isoflurane (Forene 1 , Abbott) anesthesia. After surgery the rats were housed individually and had free access to food and water. About 16 h prior to dosing the animals were deprived of food, and fasted until 4 h after dosing. The rats had free access to drinking water during the experiment. On the experiment day, the test item formulation was administered orally by gavage or intravenously in the jugular vein. At pre-defined time points, blood samples of about 0.150 mL were withdrawn from the carotid artery up to 24h after dosing. A total of 10 samples were withdrawn. The blood samples were collected in heparinized plastic tubes and centrifuged, within 30 min, for five min at 10 000 g and +4°C. The plasma was transferred to a 96 well plate and stored at -20°C until analysis by LC/MS/MS. The studies were approved by the Göteborg Animal Research Ethical Board.

Chemistry
Chemicals and solvents from commercially available sources was purchased and used without further purification. 1 H NMR spectra were recorded on a Bruker biospin GmbH 400, 500 or 600 MHz spectrometer. Chemical shifts are reported in parts-per-million (δ) relative to DMSO-d 6 at 2.50 ppm, CDCl 3 at 7.26 ppm or CD 3 OD at 3.31 ppm as an internal standard. High-resolution mass spectrometry (HRMS) analysis was performed using a Waters XEVO-qTOF instrument (Waters, Milford, USA) with MassLynx 4.1 software. If necessary, the purity was determined by ultra high performance liquid chromatography (U(H)PLC). Purity of all final compounds was 95% or higher at 210 nm. The instrument was a Waters U(H)PLC Acquity system (Binary pump, Sample Manager, Column Compartment and UV-PDA detector). The column was a Waters Acquity BEH C18, 1.7 μm particle size (100 mm × 2.1 mm).
Heating of reactions with microwaves was performed on a Biotage Initiator 2.45 GHz, 400W microwave for the times indicated. Compounds 1, 2, 4 and 27 were obtained from the AstraZeneca corporate compound collection, compounds 3, 5, 6, 7, 12, 17, and 19 were commercially available and compounds 9 [31], 10 [31] and 11 [30] were synthesized according to literature. Experimental and spectroscopic details for all other non-commercially available compounds are given below or are available in S1 File.
In general, optimizations of the reaction procedures were not performed and yields are given for isolated materials. (4). Compound 4 was synthesized in an 8 step sequence from methyl 6-cyano-2naphthoate as described in the S1 File. 1 1H).
5-Methylquinolin-2-amine (8D). A mixture of 2-chloro-5-methylquinoline 8C (0.484 g, 2.72 mmol), Xantphos (0.126 g, 0.220 mmol), tris(dibenzylideneacetone)dipalladium (0.100 g, 0.110 mmol) and potassium tert-butoxide (0.917 g, 8.17 mmol) in Toulene (15 mL) was heated at 100°C for 1 h using a microwave reactor. The reaction was quenched with water and the phases separated. The aqueous phase was extracted with DCM (3 x 20 mL), the combined organic fractions was passed through a phase separator to remove water and palladium residues. The filtrate was concentrated under reduced pressure, the residue re-dissolved in a mixture of THF (10 mL) and 2M HCl (10 mL), and stirred at room temperature for 2 h. The reaction was concentrated under reduced pressure to give compound 8D as a crude solid (0.427 g, 99%). The crude product was used in the next reaction step without further purifications.
N-(5-Methylquinolin-2-yl)acetamide (8E). A mixture of 5-methylquinolin-2-amine (8D) (0.427 g, 2.72 mmol) and triethylamine (TEA) (3.79 mL, 27.2 mmol) in DCM (10 mL) was stirred at 0°C under an inert atmosphere of nitrogen. Acetyl chloride (1.94 mL, 27.2 mmol) was added drop wise and resulting mixture stirred for 2 h. A large excess methylamine (2M in THF) was then added and resulting mixture stirred for 16 h to cleave di-acetylated products. The reaction was concentrated under reduced pressure and resulting residue filtered through a plug of silica using EtOAc/ heptane (1:3) as the eluent. The crude compound 8E (0.270 g, 50%) was sufficiently pure to be used in the next reaction step without further purifications.
N-(5-(Bromomethyl)quinolin-2-yl)acetamide (8F). A solution of N-(5-methylquinolin-2yl)acetamide 8E (0.270 g, 1.35 mmol), NBS 0.528 g, 2.97 mmol), benzoyl peroxide (0.098 g, 0.400 mmol) was dissolved in (trifluoromethyl)-benzene (6 mL) and the resulting mixture heated at 80°C for 16 h. The reaction proceeded slowly and was heated at reflux for 2 h to achieve complete conversion of the starting material. The reaction was diluted with water and then extracted with DCM (3 x 15 mL). The combined organic fractions was dried (MgSO 4 ), filtered and then concentrated to give compound 8F as a crude pale brown oil (0.377 g, 100%). The crude material was used without further purifications in the next reaction step.
N-(5-(Aminomethyl)quinolin-2-yl)acetamide (8H). The crude (DMF/Ethanol solution, 2 mL:6 mL) solution of N-(5-(azidomethyl)quinolin-2-yl)acetamide 8G was transferred to a 50 mL round bottom flask and palladium hydroxide on carbon (0.050 g, 0.36 mmol) was added. The reaction mixture was stirred under an atmosphere of hydrogen for 16 h. The resulting mixture was filtered through a phase separator to remove palladium residues and then concentrated under reduced pressure to give crude compound 8H (0.291 g, 100%) as a brown oil. The crude oil was used directly in the next reaction step without further purifications.
(E)-Methyl 3-(5-chloro-2-(3-ethoxy-3-oxopropanamido)phenyl)acrylate (20D). To a mixture Ethyl 3-(4-chloro-2-formylphenylamino)-3-oxopropanoate (20C) (3.46 g, 12.8 mmol) in toluene (40 mL) was slowly added methyl 2-(triphenylphosphoranylidene)acetate (5.58 g, 16.7 mmol) and the resulting mixture was refluxed for 2 h. Upon cooling to room temperature a precipitate formed. The precipitate was collected by filtration and dried under reduced pressure. The crude product was purified by silica gel column chromatography using a gradient from 20% to 50% of EtOAc in heptane as the eluent. Fractions containing the product were pooled together and the solvent was removed under reduced pressure to give compound 20D as a white solid (1.29 g, 31%). The material was used in the next reaction step without further purifications.