TRIANNI humanized mice (TRIANNI, 2016) were immunized over a 12-week period using Adjuplex™ adjuvant and 3T3 cells engineered to express full length mesothelin. Splenocytes were isolated, and hybridomas generated by electrofusion (Greenfield, 2014). Hybridomas were cultured in microtiter plates and secreted antibodies were identified by flow cytometry using both HEK 293 and MV-4-11 cells expressing the full length MSLN and counter-screening with parental controls. Polyclonal positives were subcloned using a ClonePix II™ and validated mesothelin binding antibody secretion by flow cytometry.

We employed the Daedalus human cell line expression platform for the production and purification of all reported proteins, using methods described previously (Bandaranayake et al., 2011). The expression system utilizes suspension adapted HEK293 FreeStyle™ cells (ThermoFisher catalog #R79007) and lentiviral transduction to generate cell lines that secrete proteins at high levels. The lentiviral vector contains a cis-linked fluorescent protein reporter driven by an internal ribosome entry site (IRES) that allows for tracking of relative protein expression levels. All mammalian proteins described in Supplementary Table S5 were purified directly from conditioned media using HisTrap FF Crude columns (GE catalog #17528601) and subsequently polished on a Superose 6 10/300 GL SEC column (GE #17517201) using an AKTA pure 25 instrument. For crystallization, MSLN fragments were produced from HEK293 FreeStyle™ cells treated with kifunensine (Chang et al., 2007) to reduce N-glycan heterogeneity and overall amount. Antibody Fab fragments were produced by papain cleavage of IgGs and the Fab purified on a MabSelect protein A column; antibody scFvs of 1A12 and anetumab were engineered with a flexible (GGGS)₄ linker. The amatuximab scFv was produced with a thrombin cleavable linker incorporating an internal HisTag for purification.

Surface plasmon resonance (SPR) experiments were performed at 25°C on a Biacore T100 instrument (Cytiva) using a running buffer of 10 mM HEPES, pH 7.4, 150 mM NaCl, 3 mM EDTA, 0.05% surfactant P20 with 0.1 mg/mL bovine serum albumin. Goat anti-human IgG, Fcγ fragment specific antibody (Jackson ImmunoResearch: 109005098) was amine coupled to 2 flow cells of a Series S CM4 chip (∼2,800 RUs). Each antibody (1 μg/mL 1A12 or 0.3 μg/mL Amatuximab) was injected at 10 μL/min over 1 flow cell of immobilized anti-human IgG Fcγ for 35, 105, 35 or 30 s to capture 57.1 ± 0.8, 96 ± 1, 82 ± 2 or 81.9 ± 0.6 RUs of MDT846/984, MDT846/985, MDT1119/1121 or Amatuximab, respectively. MSLN-Avi was prepared by cleaving MDT536 with tobacco etch virus protease. Purified MSLN-Avi was run as a concentration series at 50 μL/min over both the captured antibody and anti-human IgG Fcγ alone surfaces. MSLN-Avi concentrations (serial 2-fold dilutions starting at 1 μM for the 1A12 constructs and 25 nM for Amatuximab) were run in duplicate, randomized, and included a buffer blank every 4th injection. MSLN-Avi was injected for 5 min and allowed to dissociate for 7 min from 1A12 to 3 min from Amatuximab. In addition, triplicate 50 nM samples of MSLN-Avi alternated with buffer blanks were injected over Amatuximab and allowed to dissociate for 2 h. The CM4 chip was regenerated with 10 mM glycine, pH 1.5 at 50 μL/min for 30 s, or 45 s for MDT1119/1121, and antibody recaptured prior to each MSLN-Avi injection. Data was double referenced and analyzed in BiaEval 2.0.4 with a 1:1 kinetic binding model.

Cancer cell lines OVCAR-3 and SKOV-3 (ATCC #HTB-161 and #HTB-77) were lentivirally transduced to express a near-infrared fluorescent reporter. Similarly, the NOMO-1 line (DSMZ #ACC-542) was marked with a gene encoding GFP. Cryopreserved healthy, unstimulated human donor PBMCs were obtained from Bloodworks NW. On the day of assay preparation, T-cells were isolated using a CD3⁺ magnetic negative selection kit (Stemcell Technologies #17951) according to the manufacturer’s instructions. All further assay preparation and sample collection was automated using a Microlab Starlet liquid handler (Hamilton). Target cells were co-cultured with purified T-cells at a 5:1 E:T ratio. Test molecules were added with all conditions performed in triplicate. Cells were incubated at 37°C for 48 H.

At 24 H, supernatant was collected to screen for cytokine release. At 48 H, cytotoxicity was measured by obtaining target cell counts. For the adherent lines, OVCAR-3 and SKOV-3, cells were imaged using an ImageXpress Nano and positive fluorescence was analyzed using the accompanying MetaXpress software (Molecular Devices). For the suspension line, NOMO-1, cells were collected, mixed with 1x DAPI, and run on an iQue Screener Plus (Sartorius) and live target cells were identified by forward/side scatter distribution, negativity for DAPI, and positive fluorescence. Raw cell counts were exported to Microsoft Excel and percent cytotoxicity was calculated utilizing the following formula: %   C y t o t o x i c i t y = M e a n   c o u n t   o f   l i v e   i R F P + c e l l s   n o   d r u g − c o u n t   o f   l i v e   i R F P + c e l l s   w /   d r u g M e a n   c o u n t   o f   l i v e   i R F P + c e l l s   n o   d r u g

Fitted curves were generated by applying a Sigmoidal, 4PL, x is log (concentration) model using Prism (GraphPad).

Cytokine quantification was performed using the iQue Qbeads^® PlexScreen kit (Sartorius) as per the manufacturer’s protocol. Briefly, supernatants and assay kit cytokine standard were incubated with capture beads specific to IFNγ and IL-2. The samples were incubated for 2 H at room temperature with shaking (1,250 rpm). The beads were washed and incubated with prepared detection reagents to detect bound cytokines. Supernatant and reagent addition was performed on the Microlab Starlet liquid handler. The data was collected on the iQue Screen Plus, measuring fluorescence intensity in the BL2 channel. Cytokine concentrations were interpolated from cytokine standards and the resultant data was fitted using the Sigmoidal, 4PL, x is log (concentration) model in Prism.

Crystallization of free MSLN fragments proved infeasible, even after extensive screening. Alternatively, MSLN fragment complexes with combinations of Fabs and/or scFvs derived from amatuximab, anetumab, or 1A12 were isolated by SEC in 12 mM PIPES (pH = 7.2), 150 mM NaCl, 1 mM EDTA and concentrated to ∼10 mg/mL. Diffraction-quality crystals (d_min = 2.6 Å) of the ternary complex of MSLN fragment #1 with the scFvs of anetumab and 1A12 were obtained by vapor diffusion over a well solution of 1.2–1.4 M ammonium sulfate, 100 mM bicine (pH = 9.0), 3% v/v glycerol after multiple rounds of macro-seeding. Diffraction-quality crystals (d_min = 3.3 Å) of the binary complex of kifunensine-treated MSLN fragment #2 with 1A12 Fab were obtained by vapor diffusion over a well solution of 19%–22% w/w polyethylene glycol (M_r = 3,350), 100 mM TRIS (pH = 8.0), plus 250 mM sodium citrate. The highest quality crystals were obtained after multiple rounds of macro-seeding into drops of 1 mL of 3–6 mg/mL protein solution mixed with 1 mL of well solution. Crystals were cryopreserved by the addition of 15%–20% v/v glycerol to the well solution. Diffraction-quality crystals (d_min = 4.3 Å) of the ternary complex of kifunensine-treated MSLN fragment #2, the amatuximab scFv, and the 1A12 Fab were obtained by vapor diffusion over a well solution of 1.3–1.5 M ammonium sulfate, 50 mM TRIS (pH = 8.0), plus 14 mM ZnCl₂. Diffraction data were collected at the Advanced Light Source (Berkeley, CA) beamlines 5.0.1 or 5.0.2 and integrated and scaled with HKL-2000 (Otwinowski and Minor, 1997). Initial phases for the MSLN fragment #2/1A12 Fab complex structure were determined by molecular replacement using Phaser (McCoy et al., 2007) as implemented in the CCP4 software suite (Winn et al., 2011) with 4F3F.pdb (Ma et al., 2012) (MSLN NTD/amatuximab Fab complex) as the search model, though preliminary electron density maps only supported building of a partial model. This partial model was used as a search model to generate initial phases by molecular replacement for the ternary MSLN fragment #2/amatuximab scFv/1A12 Fab and MSLN fragment #1/anetumab scFv/1A12 scFv complex structures. To complete model building, the RoseTTAFold MSLN model was used as a guide template iteratively across the three models. Model building and positional refinement were performed with refmac (Murshudov et al., 1997) and COOT (Emsley and Cowtan, 2004), including placing ordered solvent molecules, followed by a final round of TLS refinement (Winn et al., 2001). Residues or side chains that did not exhibit clear 2F_obs-F_calc electron density when contoured at 0.7σ were removed or truncated to the Cβ atom. The quality of the final models was assessed using ProCheck and Molprobity (Supplementary Table S3).

The coordinates for the final models have been deposited in the RCSB PDB (Berman et al., 2000), accession codes: 8CZ8 (MSLN fragment #1/anetumab scFv/1A12 scFv complex), 8CYH (MSLN fragment #2/1A12 Fab complex), and 8CXC (MSLN fragment #1/anetumab scFv/1A12 scFv complex).

In order to target distinct regions of the MSLN ectodomain with novel antibody-based reagents, murine 3T3 cells were engineered to surface express a chimeric murine/human MSLN molecule, where the membrane-proximal ectodomain sequence of murine MSLN (residues 464-625, based on UniProt: Q61468) was swapped with the corresponding region of human MSLN (residues 470-630, based on UniProt: Q13421-1). The logic was that murine protein sequences would not be immunogenic in mice focusing induced humoral responses on the human membrane-proximal region (Supplementary Figure S1A). TRIANNI mice, which produce chimeric antibodies from a full repertoire of human heavy and light chain variable domains, while maintaining mouse constant domains, were immunized with the engineered murine 3T3 cells and hybridomas were isolated. Sequencing mesothelin-positive monoclonal hybridomas yielded 22 paired heavy (V_H) and light (V_L) chain variable domain sequences of which 18 belonged to a single sequence family (Supplementary Figure S1B), where the V_H domain was derived from the IGHV3-23*01 germline, defining a single clonotype: sequence identities across these 18 CDRH3 and CDRL3 segments were >80% and >78%, respectively. There were few predicted deamidation or isomerization liabilities (Yan et al., 2018) and no non-standard cysteines or glycosylation sites, particularly in the CDRs. The clonotype archetype, clone 1A12, as well as other related clones, were reformatted by replacing the murine constant regions with human constant regions (aglycosylated IgG1) and recombinantly expressed using FreeStyle 293F cells (Bandaranayake et al., 2011). The binding of 1A12, as an IgG1, to MSLN+ and MSLN-knockout cell lines was evaluated by flow cytometry in head-to-head comparisons with amatuximab, showing comparable sensitivity and specificity (Figure 1A). The interaction parameters of 1A12 and amatuximab for recombinant MSLN were quantitated by surface plasmon resonance (SPR) biosensor analyses, showing univalent K _D s of 64.4 ± 0.1 nM (ligand: captured 1A12 IgG, analyte: MSLN ectodomain; Figure 1B) and 112 ± 1 pM (ligand: captured amatuximab IgG, analyte: MSLN ectodomain; Figure 1C). The reported univalent K _D of amatuximab for MSLN is 1.5 nM (Hassan et al., 2007), and of anetumab for MSLN is 10 nM (Golfier et al., 2014).

In order to confirm that the desired targeting of 1A12 was achieved and to compare epitopes, three co-crystal structures with recombinant MSLN ectodomain fragments [residues 434-590 (#1), and 296-605 (#2)] were determined at d_mins ranging from 2.6 to 4.3 Å (Figure 2; Supplementary Table S3). MSLN fragments were produced in human cell lines using the Daedalus protein expression system (Bandaranayake et al., 2011). Using combinations of either Fab or scFv fragments of 1A12, amatuximab, and anetumab as co-crystallization “chaperones” enabled these structure determinations, including a fragment of MSLN (#2) spanning the entire ectodomain, revealing its overall fold. Antibody Fab fragments were produced by papain cleavage of IgGs and purified on a MabSelect™ protein A column; antibody scFvs were engineered with a flexible (GGGS)₄ linker and either purified separately or as co-expressed MSLN ectodomain complexes. Even in the lowest resolution structure, the final electron density maps were cleanly resolvable, ultimately allowing modeling of the entire structure, including N-acetylglucosamine residues at the two N-glycan sites in MSLN (N388, N523), except for a prominent disordered loop (residues 405-414). This disordered loop spans an eight amino acid insertion (QAPRRPLP) only present in human MSLN isoform 1 (UniProt: Q13421-1) among human isoforms and Pan troglodytes isoform 4 (A0A2J8J880) among primate MSLN sequences, though a similar insert of matching length does appear in a few more distantly related primate orthologs.

The MSLN ectodomain is composed of a series of short amphipathic helices that wind in a coiled fashion to produce a slightly bent, elongated structure (Figures 2A–C), extending the compact, right-handed α-superhelical fold observed in the preceding structure of the MSLN NTD (Ma et al., 2012). Each helix is oriented to position its non-polar face towards the central axis resulting in the packing of an extended hydrophobic core. The ectodomain contains two disulfide bonds, with the first (C302-C326) in the NTD securing the initial helical packing and the second (C450-C476) bridging between superhelical turns near the middle of the ectodomain. This disulfide constrains the packing of several surrounding helices resulting in the solvent-exposure of W466 and defining a hinge point that is the largest element of overall flexibility observed in the MSLN ectodomain (∼19°) when the two full length MSLN ectodomain were superimposed on their C-terminal ends (residues 476-587; Figure 2D).

One goal of MSLN structural analyses was to advance its functional annotation by identifying structurally related proteins with potentially informative, related functions. However, the overall fold of the extracellular domain of MSLN is fairly distinct, based on structural similarity searches using the DALI server (Holm, 2020). The closest match was the NTD of Saccharomyces cerevisiae Rif1 [5NW5.pdb (Mattarocci et al., 2017)], but the structural similarity was weak, with a Z-score of 4.9, obvious also by visual inspection (Figure 2E), and there was no obvious sequence similarity. Functionally, Rif1 is an intracellular DNA binding protein involved in double-strand break processing, telomere capping, replication initiation, silencing, and rDNA stability, providing no useful information about MSLN extracellular domain function/s. However, despite the relative uniqueness of the MSLN structure as determined experimentally, state-of-the-art structure prediction algorithms performed extremely well, even in detail. We compared the MSLN crystal structures with its predicted structure, available either through the AlphaFold protein structure database (Jumper et al., 2021; Varadi et al., 2022) (rmsd = 0.80 Å, residues 449-589) or as a one-side blinded prediction using RoseTTAFold (Baek et al., 2021) (rmsd = 1.63 Å; residues 448-589; Supplementary Figure S2). While slight deviations from experiment were noted throughout both predictions, both algorithms correctly identified the 405-414 disordered loop. Indeed, the accuracy of these results was such that at least one of the crystallographer co-authors decided to check the status of their retirement plan.

These three antibody/MSLN co-crystal structures enabled high-resolution mapping of the epitopes of anetumab and 1A12 and the relationship between these and that of amatuximab (Supplementary Figure S3A; Supplementary Tables S3, S4). The amatuximab scFv was used in combination with the 1A12 Fab to generate one of the two full length MSLN co-crystal structures which recapitulated all the salient details of the previously reported co-crystal structure with the isolated MSLN NTD. Anetumab binds near the middle of the ectodomain and 1A12, as intended, binds near the C-terminus. All three antibodies have distinct epitopes which do not cross-block, though all three preferentially bind to relatively negatively-charged patches on MSLN, which overall displays a patchwork of positive/negative surface charge (Supplementary Figure S3B). At the antibody/MSLN interfaces, amatuximab buries ∼890 Å² of solvent-accessible surface and makes 6-8 hydrogen bonds and two salt bridges; anetumab buries ∼825 Å² making 6-8 hydrogen bonds and 2-3 salt bridges; and 1A12 buries ∼600 Å² and interacts primarily through hydrophobic contacts plus 4-6 hydrogen bonds. The reduced footprint of 1A12 likely contributes to its reduced affinity for MSLN: ∼6-fold weaker than anetumab and ∼40-fold weaker than amatuximab.

One widely employed approach for developing antibodies targeting tumor antigens into therapeutics is to engineer bispecific reagents, where one antibody specificity is directed against the tumor antigen and the other antibody specificity is directed against a T cell-associated antigen, often the CD3 signaling moiety of the T cell holoreceptor (Arvedson et al., 2022). When infused, these reagents recruit and activate otherwise non-tumor specific, bulk T cells, generating potent anti-tumor responses. 1A12 was engineered into two different bispecific architectures: “IgG-scFv” (Figure 3A) and “scFv-Fc” (Figure 3B). Engineered protein and antibody sequences are appended in Supplementary Table S5 and were produced using the Daedalus protein expression system in human cell lines (Bandaranayake et al., 2011). The 1A12 IgG-scFv was initially tested with four anti-CD3 scFv moieties derived from the parental antibodies OKT3 (Kjer-Nielsen et al., 2004), UCHT1 (Arnett et al., 2004), MICRO194 (Kischel et al., 2009), and ADI26906 (Walker Laura et al., 2020; Liu et al., 2023). All four expressed well using the FreeStyle™ 293F mammalian tissue culture system and could be purified to homogenous, monodisperse species stable even after a freeze-thaw cycle (Supplementary Figure S4). 1A12 was also used to engineer an scFv-Fc construct with OKT3. The 1A12/OKT3 IgG-scFv and scFv-Fc constructs retained binding to MSLN, with univalent K _D s for soluble MSLN of 63.0 ± 0.1 nM (ligand: captured 1A12 IgG-scFv, analyte: MSLN ectodomain) and 103 ± 1 nM (ligand: captured 1A12 scFv-Fc, analyte: MSLN ectodomain), respectively (Figures 3C, D), within two-fold of the parent, un-engineered 1A12 IgG.

Focusing on the OKT3 fusions because of its long history and superior product quality in the IgG-scFv format, both IgG-scFv (“IgG^1A12-scFv^OKT3”) and scFv-Fc (“scFv-Fc^1A12/OKT3”) constructs demonstrated in vitro functional T cell engagement and activation in the nanomolar-to-picomolar range against the MSLN+, human acute monocytic leukemia cell line NOMO-1 in redirected cytotoxicity and cytokine interferon-γ (IFN-γ) and interleukin 2 (IL-2) release assays (Figures 3E, F). Because of its greater potency, reasonably a product of the greater inherent valency of a bi-bivalent IgG-scFv over that of a bi-monovalent scFv-Fc, IgG^1A12-scFv^OKT3 was advanced for head-to-head testing with the analogous constructs incorporating amatuximab (alternate anti-MSLN moiety) or ADI26906 (alternate anti-CD3 moiety): IgG^AMA-scFv^OKT3, IgG^1A12-scFv^ADI, and IgG^AMA-scFv^ADI. IgG-scFv constructs with 1A12-based MSLN-targeting moieties (IgG^1A12-scFv) were roughly a half-log more potent than IgG^AMA-scFv constructs, with either OKT3 anti-CD3 moieties (IgG-scFv^OKT3; Figure 4A) or ADI26906 moieties (IgG-scFv^ADI; Figure 4B) in in vitro redirected cytotoxicity assays against OVCAR-3 cells. The improved potency of both IgG^1A12-scFv constructs over amatuximab-based bispecifics was reasonably the result of targeting an epitope closer to the tumor cell surface, tightening the T cell synapse, as has been previously observed (Bluemel et al., 2010; Tang et al., 2013a).

It is well established that SMRPs have the potential to confound antibody-based therapeutics by preventing the therapies from reaching the tumor cell surface (Zhang et al., 2007; Liu et al., 2022). To assess how soluble MSLN may affect the potency of T-cell engaging therapies we performed in vitro killing assays with SKOV-3 cells in the presence of a high concentration of recombinant MSLN. We compared the potency of the IgG^1A12-scFv^ADI with that of a trispecific T-cell engager (known as HPN536) with a novel V_HH-V_HH-scFv format (referred to as a TriTAC). The TriTAC binds human MSLN, albumin and CD3 with monovalent affinities (K_Ds of .21 nM, 6.3 nM, and 6.6 nM respectively) and has reached clinical trials (Molloy et al., 2021). Both IgG^1A12-scFv (Figure 4C ) and the TriTAC molecule (Figure 4D) kill SKOV-3 cells with high potency (EC50s of 6.6 pM and 3.6 pM respectively). To mimic physiologic MSLN levels we performed the cytotoxicity experiment in the presence of 1ug/mL (∼30 nM) recombinant MSLN (Zhang et al., 2012). Both molecules were inhibited by the soluble ectodomain but the fold-change in potency was much improved (Figures 4C, D), with the bivalent molecule only showing a three-fold reduction in potency (compared to 60-fold for the V_HH-V_HH-scFv). This result suggests that bivalent binding of the IgG-scFv may have advantages over a monomeric T cell engager when soluble antigen sinks are a concern.