Computational modeling and experimental validation of the interaction between tumor biomarker mesothelin and an engineered targeting protein with therapeutic activity

Abstract

Mesothelin (MSLN) is a cell surface glycoprotein overexpressed in many solid tumors, which is known to interact with cancer antigen CA125/MUC16, promoting cancer cell adhesion and metastasis. MSLN has been used as a target of multiple antibody-based therapeutic strategies, but their efficacy remains limited, potentially due to inherent pharmacokinetics conferred by the structure of antibodies (~150 kDa). To provide an alternative targeting molecule, we engineered a small scaffold protein derived from the tenth domain of human fibronectin type III (Fn3, 12.8 kDa) to bind MSLN with nanomolar affinity as a theranostic agent for MSLN-positive cancers. In this study, we explored the Fn3-MSLN interaction site through computational modeling and experimentally validated the model through domain-level and fine epitope mapping. Fn3-MSLN binding was predicted by a consensus approach, comparing multiple protein-protein docking software, the deep-learning-based algorithm AlphaFold3, and performing molecular dynamics (MD) simulations. To validate the prediction, full-length MSLN, single MSLN domains, or combinations of domains were expressed on the yeast surface, and Fn3 binding to displayed MSLN domains was measured by flow cytometry. The employed algorithms predicted two distinct binding modes for Fn3. Overall, experimental data agreed with our in silico prediction resulting from the AlphaFold3 model, confirming that MSLN domains B and C are predominantly involved in the interaction.

1. INTRODUCTION

Mesothelin (MSLN) is a surface glycoprotein overexpressed in numerous malignancies, including mesothelioma, ovarian, pancreatic, and lung cancers.^1,2 MSLN is anchored on the cell surface by the C-terminus through a glycosylphosphatidylinositol (GPI) anchor. MSLN derives from the cleavage of a 68 kDa precursor (628 amino acids), encoded by the MSLN gene.³ The cleavage occurs at arginine 295 (Arg295) and leads to two peptides: soluble megakaryocyte potentiating factor (MPF, 31 kDa) released into the bloodstream and membrane-bound MSLN (40 kDa).¹ MSLN can also be shed in blood, pleural and peritoneal fluid, in its soluble form, which is referred to as soluble mesothelin-related peptide (SMRP).⁴ In the mature form, membrane-bound MSLN has three predicted N-glycosylation sites (Asn388, Asn488, and Asn515).⁵ Despite its biological role remaining unclear,⁶ MSLN has been proposed as a promising cancer biomarker for diagnosis and therapy since it is aberrantly expressed in many solid tumors and has low or no expression in healthy tissues.^7–10 Furthermore, MSLN was shown to bind mucin 16 (CA125/MUC16), an antigen overexpressed in various cancers.¹¹ MSLN-MUC16 binding was found to promote cell adhesion and metastasis, revealing the potential tumorigenic role of MSLN.^12–15 Therefore, there is sustained interest in MSLN as a target for diagnosis and therapy. In the last decade, multiple antibody-based approaches targeting MSLN were developed and tested in clinical trials.^16–19

Recently, a crystal structure of MSLN was published by Zhan et al. (2023), showing that the MSLN membrane distal region (N-terminus) is rigid, whereas the C-terminus, bound to the cell membrane, is more flexible.²⁰ The N-terminal region has been observed to bind CA125/MUC16 and a chimeric monoclonal antibody (mAb), known as MORAb-009 or amatuximab, that has been used to target MSLN for immunotherapy.^21–24 The C-terminal region is closer to the cleavage site that leads to SMRP and is recognized by mAb 15B6, which was produced in recent studies to avoid being released from the cell during the process of MSLN shedding and SMRP release.^18,20

Overall, clinical trials studying antibody-based strategies targeting MSLN for therapy resulted in poor efficacy, attributed to immunogenicity and an “on target/off tumor” effect due to binding circulating SMRP, and trials have yet to result in approved MSLN-targeting approaches.^10,24,25 Since MSLN-positive cancers are generally solid tumors, the antibodies’ large size could limit the ligand penetration into the tumor, lowering the antibody efficacy.²⁶ It has been observed that the anti-tumor efficiency of antibody-based strategies is increased in blood cancers, where the tumor cells, thus the target, are easily accessible.²⁶

To provide an alternative targeting molecule for MSLN, we recently reported engineered MSLN-binding proteins based on the tenth type III domain of human fibronectin (Fn3).^27,28 The Fn3 structure (~13 kDa, 98 amino acids) consists of a hydrophobic core that confers stability to the protein with a beta-sandwich structure, and three loops, which are the variable regions, amenable to diversification.^29,30 In previous studies, Fn3 has been engineered to successfully recognize various targets, and preclinical and clinical studies have shown the potential of Fn3-based molecules for therapeutic and molecular imaging applications.^31–37 We previously observed that the MSLN-binding Fn3 variants internalize and promote apoptosis in MSLN-positive ovarian cancer cells.^27,28 However, we have not previously had knowledge regarding the protein-protein binding interface between engineered Fn3 variants and MSLN.

A detailed understanding of the molecular mechanism regulating protein-protein interactions requires an atomistic-level knowledge of the adducts formed between the binding partners. This information is usually gained by solving the three-dimensional structure of protein-protein complexes using experimental techniques, such as X-ray crystallography or nuclear magnetic resonance spectroscopy. However, such knowledge is often challenging to obtain for large biomolecular complexes. Therefore, in recent years, molecular modeling and simulations have contributed significantly to unraveling the molecular interactions occurring between biological macromolecules.³⁸ More recently, deep-learning-based algorithms, such as AlphaFold3, have proved to accurately predict the three-dimensional structure of proteins and protein-protein complexes, revolutionizing the structural biology field.³⁹

In silico methods, such as protein-protein docking, can predict the native (minimum energy) binding pose and the corresponding free energy involved. Moreover, they are useful in complementing the experimental techniques for determining their complex adducts and identifying interactions and key ‘hot spot’ residues, which are essential for drug discovery.⁴⁰ However, reliable predictions remain challenging to obtain due to the substantial conformational space that these algorithms have to sample exhaustively and the limited accuracy of the employed scoring functions necessary to rank the binding poses.⁴¹ Performing μs-long force field-based MD simulations to relax the adducts predicted by docking simulations can further improve the accurate identification of native binding poses of protein-protein complexes. These could also be followed by the molecular mechanics-Poisson Boltzmann surface area (MM-PBSA) method to estimate their binding free energy (ΔG_b) and refine the relative ranking of the docking poses.⁴²

Throughout this present work, we report results with an Fn3 variant that was identified from five rounds of directed evolution to enhance affinity for MSLN, which we will refer to here as Fn3_5. Here, using a combination of protein-protein docking, machine-learning-based algorithms, and MD simulations, we report key structural information for the Fn3_5/MSLN complex and, thus, its binding interface. Next, we validate our in silico predictions through domain-level epitope mapping and yeast surface display. We determined the predicted Fn3_5-MSLN binding site and the most contributing interacting residues. Investigating the MSLN/Fn3_5 binding interaction can provide a foundation for understanding the biological effect of Fn3_5 on MSLN-expressing tumor cells and inform future rational engineering through point mutations to further improve the MSLN-Fn3 binding affinity.

2. RESULTS

2.1. In silico modeling of the Fn3_5/MSLN complex

To identify the putative three-dimensional structure of the Fn3_5/MSLN complex, we performed blind protein−protein docking calculations. First, the structure of Fn3_5 was obtained via homology modeling (see Methods section) and was subsequently relaxed by performing a 1 μs-long MD simulation (Figure S1). Due to the known limitations of docking algorithms in correctly predicting the native protein−protein binding interface, we have employed a consensus docking strategy, which has been successfully performed for similar systems.⁴³ In particular, we selected three well-known protein-protein docking algorithms: ClusPro⁴⁴, HDOCK⁴⁵, and pyDock⁴⁶. Remarkably, all three algorithms predicted a similar binding region for Fn3_5, located at the N-terminal part of MSLN (Figure 1A). This binding interface overlaps with the epitope of the MORAb-009 antibody (Figure S2), which was recently solved in its complex with MSLN.²⁰

Figure 1. Top-ranking Fn3_5/MSLN binding models.

(A) Superposition of the top-ranking models obtained with ClusPro (pink), HDOCK (red), pyDock (cyan) software, and the AlphaFold server (blue). The mesothelin (MSLN) structure was divided into four domains, as reported in previous studies.²⁰ (B) Binding free-energy (ΔG_b, kcal/mol) of the MSLN/Fn3_5 complexes obtained from all the MD trajectories. The disordered C-terminal tail of Fn3_5 is the result of a hydrophilic portion of the Fn3_5 sequence that supports expression and solubility of engineered Fn3 variants.

While this work was under review, the new version of the deep-learning method for the prediction of three-dimensional biomolecular structures, AlphaFold3, was published and was made easily accessible to the scientific community through the AlphaFold Server (https://alphafoldserver.com/).³⁹ Since AlphaFold3 performance is reported to be improved significantly compared to previous versions, particularly for protein-protein interactions, we have compared our consensus predictions obtained using canonical protein-protein docking algorithms and this new methodology. Interestingly, we obtained a new binding pose different from those predicted by ClusPro, HDOCK, and pyDock (Figure 1A). In particular, in the AlphaFold3 model, Fn3_5 interacts mainly with the B and C domains of MSLN.

Next, to relax the obtained models and refine the binding poses, we have performed μs-long MD simulations on the best poses obtained. All the investigated models rapidly reached structural convergence, with the one obtained by AlphaFold3 showing very small fluctuations (Figure S3). Then, we evaluated the free energy of binding (ΔG_b) using the molecular mechanics-Poisson Boltzmann surface area (MM-PBSA) method.⁴⁷ The AlphaFold3 model showed the strongest binding (Figure 1B), while the other three related poses show lower, although similar, binding energies, with the ClusPro model showing the lowest value. Due to the general consensus shown by the three docking algorithms, we decided to further analyze the ClusPro model along with the one obtained using AlphaFold3. Thus, we dissected the residues that mainly contributed to the binding at the protein-protein interface between MSLN and Fn3_5 by applying a per-residue decomposition to the three replicas of the ClusPro and AlphaFold3 models (in Table 1, the lowest energy model is shown, while in Tables S1 and S2, the per-residue decomposition of all three replicas is reported).

Table 1. Per-residue decomposition of binding free-energy (ΔG_b) of the ClusPro and AlphaFold3 complexes.

These were obtained from the lowest energy replica of the ClusPro and AlphaFold3 docking poses and reported together with their per-residue decomposition analysis. Residues stabilizing the binding pose are marked in orange (positive ΔG_b), yellow (ΔG_b between −1.01 and −2.00 kcal/mol), light green (ΔG_b between −2.01 and −3.00 kcal/mol), and dark green (ΔG_b greater than −3.01 kcal/mol).

ClusPro
Domains	Mesothelin Residues	Per-residue decomposition	Loops	Fn3_5 residues	Per-residue decomposition
A	Ala341	−1.61 ± 0.54	BC	Ala29	−1.67 ± 0.53
	Pro343	−3.03 ± 0.60		Leu30	−5.94 ± 0.74
	Phe344	−2.59 ± 0.76		Tyr31	−1.67 ± 0.76
	Tyr346	−1.26 ± 0.43	DE	Pro51	−1.95 ± 0.38
B	Tyr374	−3.23 ± 0.79		Arg52	−3.65 ± 1.13
	Glu400	−3.43 ± 1.36		Tyr53	−6.29 ± 0.64
	Lys403	+1.91 ± 1.30		Tyr54	−2.15 ± 0.64
	Gly404	−-2.31 ± 0.79	FG	Arg77	−4.86 ± 1.72
	Glu406	−1.82 ± 1.09
AlphaFold3
Domains	Mesothelin Residues	Per-residue decomposition	Loops	Fn3_5 residues	Per-residue decomposition
B	Glu406	−1.14 ± 1.59	BC	Tyr24	−2.23 ± 1.47
	Met407	−2.93 ± 0.81		Leu30	−3.19 ± 0.69
	Pro409	−1.85 ± 0.72		Ty31	−6.91 ± 1.81
	Ala412	−1.16 ± 0.42	Framework	Phe48	−1.35 ± 0.97
	Ile415	−2.09 ± 0.38		Thr49	−3.03 ± 2.40
C	Ser443	−2.48 ± 0.64		Val50	−1.35 ± 0.42
	Leu444	−1.17 ± 0.47	DE	Pro51	−4.52 ± 0.54
	Ser445	−4.09 ± 0.89		Arg52	−5.97 ± 1.55
	Pro446	−2.20 ± 0.42		Tyr53	−4.66 ± 0.91
	Asp469	+1.80 ± 1.75		Leu55	−2.49 ± 0.56
	Gln472	−1.36 ± 0.48	FG	Val75	−1.20 ± 0.28
				Arg77	−1.56 ± 1.57
				Asn79	−2.34 ± 1.27

In the ClusPro model, we noticed that the residues mostly contributing to the binding are located in domains A and B of MSLN and all three loops BC, DE, and FG of Fn3_5 (Figure 2A). In particular, some residues such as Pro343, Phe344, and Tyr374 of MSLN and Leu30, Tyr31 (loop BC), Pro51, Arg52, Tyr53, Tyr54 (loop DE), and Arg77 (loop FG) are consistently present in all the MD simulations (Table 1 and S1). Our simulation points toward a significant involvement of the loop DE of Fn3_5 in the binding. A closer view of the protein-protein binding interface is shown in Figure 2A. Here, we noticed the presence of persistent salt bridges between Arg52 and Arg77 of Fn3_5 and Glu400 and Glu406 of MSLN (Figure 2A). π-π stacking interactions performed by Phe344 and Tyr374 of MSLN and Tyr53 and Tyr54 of Fn3_5 also strongly contribute to the binding. Some of these residues are also part of the epitope recognized by the MORAb-009 antibody,²⁰ which indirectly validates the proposed binding interface. On the other hand, the binding of the other two models (HDOCK and pyDock, Tables S3 and S4) appears to be less specific; due to an increased tilt of Fn3_5, the FG loop is less involved in the binding of MSLN in the HDOCK model and absent in the one obtained with pyDock. Also, in these cases, we observe a significant contribution of Loop DE of Fn3_5, especially in the pyDock model, which accounts for almost all the interactions. However, all of these simulations indicate the MSLN domains A and B as the most involved in Fn3_5 binding.

Figure 2. Residues of Fn3_5 most significantly contributing to binding.

The binding interface between Fn3_5 and MSLN in the A) ClusPro and B) AlphaFold3 binding poses are highlighted. In the insets are reported the residues most significantly contributing to the ΔG_b. Salt bridges are highlighted using dashed blue lines. Fn3_5 is shown as pink/blue ribbons.

In the AlphaFold3 model, we observed that the MSLN Domain C is strongly involved in the binding with Fn3_5 (Figure 2B). In particular, Pro409, Ser443, Ser445, and Pro446 residues of MSLN and Leu30, Tyr31 (loop BC), Thr49 (framework region), Arg52, Tyr53, and Tyr54 (loop DE) of Fn3_5 are contributing to the binding in all MD simulations (Table 1 and S2).

We also noticed that the most involved Fn3 loop in the binding is DE, as observed in the ClusPro model, and in particular, Arg52 and Tyr53. A closer inspection of the binding interface shows that Arg52 is also involved in a salt bridge with Glu406 of MSLN. Another stabilizing hydrogen bond is formed by Ser445 of MSLN, and also the overall binding is stabilized by hydrophobic interactions of Pro409, Ile415, and Pro446 of MSLN and Tyr31, Pro51, and Tyr53 of Fn3_5 (Figure 2B). Therefore, this model, which has shown the lowest binding energy (Figure 1B), underlines a clear contribution of both domain B and C of MSLN in the binding with Fn3_5, also resulting in the lowest calculated binding energy.

The MSLN structure is composed of a combination of HEAT and ARM motifs, which generally mediate protein-protein interactions by furnishing unique mechanical properties.⁴⁸ To investigate if Fn3_5 binding would influence the MSLN internal dynamics, we have performed a principal component analysis (PCA, Figure S4) using the MD trajectories obtained from the ClusPro and AlphaFold3 models. We have observed in the MSLN structure, and in both systems, a breathing motion where the N- and C- termini move with respect to a hinge centered on domain C. By comparing this movement with an MD simulation performed on MSLN alone, we have noticed that this motion is not affected by the presence of Fn3_5. Finally, we have observed a local rigidification of the MSLN region implicated in Fn3_5 binding, mostly involving domains A and B for ClusPro and domains B and C for the AlphaFold3 model (Figure S5).

2.2. Domain-level epitope mapping system for MSLN binding established using yeast surface display

To experimentally validate the proposed binding interaction between Fn3_5 and MSLN, we designed a set of yeast-surface displayed constructs composed of combinations of MSLN domains (Figure 3A), based on the published MSLN crystal structure.²⁰ MSLN was divided into four domains, following the naming convention of domains A, B, C, and D from Zhan et al.²⁰ We refer to the full length domain as ABCD. Yeast surface display has previously been used as a method to map the binding domains on EGFR for a selection of anti-EGFR antibodies, by displaying fragments of EGFR.⁴⁹ Because of the differences in yeast and human glycosylation patterns, with yeast exhibiting hypermannosylation that could sterically hinder binding interactions, we removed the three N-linked glycosylation sites of MSLN, by mutating asparagine to glutamine in the gene encoding the full-length shed MSLN, at the following three sites: Asn388Gln (in domain B), Asn488Gln (in domain C), Asn515Gln (in domain D). We confirmed that each of the 10 domain combinations were successfully expressed on the surface of yeast with an antibody that recognizes a c-myc tag at the C-terminus of each displayed construct (Figure 3B). A non-expressing population of yeast is characteristic of all yeast surface displayed proteins, and provides an internal negative control for these epitope mapping experiments.

Figure 3. Independent domains of MSLN can be expressed on the yeast surface.

(A) Design of MSLN domains for epitope mapping using yeast surface display. (B) All domain combinations expressed well on the surface of yeast, detected by an antibody to the C-terminal c-myc epitope tag. A non-expressing population of yeast is characteristic of yeast surface display. Expression tests were conducted at least four times for each construct. Representative expression data shown, with all constructs expressed the same day in freshly prepared media. Percent expressing population is shown for each histogram, and median fluorescence intensity is indicated for each positive population.

We then determined the domain of MSLN bound by a commercially available anti-MSLN antibody. The epitope of MSLN that the antibody was raised against was previously not reported. We simultaneously labeled each MSLN construct with the anti-MSLN antibody and an anti-c-myc antibody, to separate those yeast that were expressing from those that were not. Our data showed that domain C was necessary and sufficient for binding the anti-MSLN antibody (Figure 4, top). According to the manufacturer website, in addition to applications in flow cytometry, the antibody was suitable for western blots, suggesting that the antibody recognizes a linear epitope on MSLN, which would be recognizable regardless of whether the displayed MSLN was properly folded.

Figure 4. Validation of domain-level epitope mapping of MSLN binding.

Expression of each domain on the yeast surface is shown on the x-axis, as measured in flow cytometry by a C-terminal c-myc epitope tag for each construct. The construct being tested is labeled in the upper left corner of each plot. The y-axis shows the detection of binding to either an anti-MSLN antibody (top) or a fragment of MUC16 (bottom). The anti-MSLN antibody recognizes an epitope in domain C. MUC16 binds domain A, as previously reported.²¹

Due to the eukaryotic protein quality control machinery of S. cerevisiae, the expression of a protein in yeast is often properly folded, although correct fold must be experimentally confirmed, as there are also examples of misfolded, expressed domains in yeast surface display.^50,51 Towards validating that the MSLN surface-displayed fragments were properly folded, we incubated the displayed MSLN constructs with a soluble domain of MUC16, which is reported to predominantly bind MSLN domain A, with minor contributions from domain B.²¹ We co-incubated MUC16 with an antibody that recognizes the C-terminal c-myc epitope tag in yeast surface display, to identify those yeast that were expressing each construct. For the expressing population of yeast, the presence of domain A was necessary and sufficient for binding MUC16 (Figure 4, bottom), confirming proper fold for domain A of each construct. The strongest binding signal was observed for domains ABCD and ABC compared to AB and isolated domain A. This data suggests that while the domains can be independently expressed, each domain may contribute to stabilizing the overall fold of MSLN, and therefore each domain may play an indirect role in enhancing the binding of MSLN with targeting molecules.

2.3. Binding assays support in silico modeling of Fn3_5/MSLN interaction

After establishing the domain-level epitope mapping strategy using subdomains of MSLN displayed on the surface of yeast, we then measured the dissociation constant (K_D) of Fn3_5 and the full-length MSLN construct ABCD using an equilibrium titration binding assay and flow cytometry. Using human cancer cell lines with high levels of expression of MSLN on their surface, we have measured a K_D = 11 +/− 4 nM (unpublished, manuscript under review). For the human cell line measurement, the MSLN is full length, encompassing domains ABCD and an additional short C-terminal fragment with a glycosylphosphatidylinositol anchor to the mammalian cell membrane. The MSLN expressed in cancer cell lines also has intact N-linked glycosylation. For the yeast-displayed ABCD and Fn3_5, we measured K_D = 36 +/− 13 nM (mean ± standard deviation, ten replicates) (Figure 5A), consistent with the interaction between Fn3_5 and human cancer cells with MSLN, further validating our approach to mapping the binding domains for Fn3_5.

Figure 5. Fn3_5 binding interaction mediated by a combination of domains A, B, C, and D of MSLN.

(A) Equilibrium titration binding assay of Fn3_5 with yeast surface displayed MSLN domain ABCD. K_D = 36 ± 13 nM (mean ± standard deviation of ten replicates, representative curve shown). (B) Domain mapping of all MSLN domain constructs with 100 nM and 400 nM Fn3_5 co-incubated with an anti-c-myc tag antibody detecting domain expression. Binding was detected to ABCD, ABC, BCD, and BC, without conclusive evidence of binding for other subdomains.

We then assessed the binding of Fn3_5 to each displayed MSLN construct, measuring at both 100 and 400 nM of protein (Figure 5B). We chose 100 nM as a near-saturating amount for binding to the ABCD construct. Since our in silico calculations predict that Fn3_5 interacts with both domains A and B, in the case of the ClusPro model or with both domain B and C, in the case of the AlphaFold3 model, we hypothesized that the higher 400 nM concentration would be needed to identify binding to MSLN fragments that did not encompass the entire binding interface. At the 100 nM concentration, the following MSLN fragments demonstrated binding to Fn3_5: ABCD, ABC, and BCD. We also observed moderate binding to domain BC, but did not observe detectable binding for other constructs. For the 400 nM concentration of Fn3_5, we saw similar binding signals comparing ABCD, ABC, and BCD, suggesting they were near saturating concentrations of Fn3_5. We also observed an increase in binding signal for domain BC, consistent with 100 nM and 400 nM both being at sub-saturating levels for BC. We did not observe any conclusive binding for other domains displayed on the surface of yeast. A full titration equilibrium binding assay with domain AB further confirmed that there was no measurable binding to domain AB when expressed alone on the surface of yeast (Figure S6). It is also possible that there is binding between Fn3_5 and domains A, B, or C alone, but that the interaction is sufficiently weak not to be measurable using flow cytometry at concentrations up to 400 nM of Fn3_5. These data are in agreement with the model obtained by AlphaFold3, which predicts an interaction of the B and C domains of MSLN with Fn3_5. This also agrees with the relative rankings of our computational predictions, showing a higher ΔG_b for the AlphaFold3 model compared with the other models.

2.4. Interacting residues predicted from AlphaFold3 model validated via MSLN point mutagenesis and binding assays

To validate potential interacting residues identified via our in silico approach on the ClusPro and AlphaFold3 models, we have selected some of the amino acids mostly contributing to binding (Table 1), mutated them into alanine (Ala), and repeated the titration binding assays between the resulting MSLN mutants and Fn3_5. For the ClusPro model, the Glu400Ala mutation was selected, since in our simulations Glu400 is involved in a persistent salt bridge with Arg52 of Fn3_5, resulting in a stabilization of the Fn3_5/MSLN complex. On the other hand, for the AlphaFold3 model, we have chosen to mutate the residue mostly contributing to the binding, Ser445Ala, as highlighted by our per-residue decomposition analysis (Table 1). Then we also selected an amino acid slightly contributing to the stabilization of both the ClusPro and AlphaFold3 models by establishing salt bridges with arginine residues of Fn3_5 (Figure 2), selecting Glu406Ala. Finally, a double mutant, Glu406Ala/Ser445Ala, was also constructed for testing the AlphaFold3 model.

The four full-length MSLN variants expressed on the surface of yeast and bound MUC16, confirming that domain A was properly folded (Figure 6A). Titration binding assays of Fn3_5 with the MSLN variants were repeated in at least triplicate, and a mean K_D and standard deviation determined (Figure 6B). To assess statistical significance of the differences in dissociation constants, comparisons of the mean K_D values of at least three replicates were analyzed using a two-tailed t-test. Resulting p-values are given in Table S5. Mutation of the key residue for the ClusPro model, Glu400Ala, did not diminish binding of Fn3_5 compared to the ABCD wildtype construct (WT: K_D=36 ± 13 nM vs Glu400Ala: K_D=37 ± 10 nM, p-value = 0.91). Meanwhile, the individual mutants from the AlphaFold3 model both had decreased Fn3_5 binding compared to ABCD wildtype for Glu406Ala (K_D=62 ± 16 nM, p-value = 0.0143) and Ser445Ala (K_D=117 ± 19 nM, p-value < .0001), with an even greater decrease in binding for the double mutant Glu406Ala/Ser445Ala (K_D=164 ± 40 nM, p < .0001). Combined with the domain-level epitope mapping data, these experiments support the AlphaFold3 docking pose.

Figure 6. Fine epitope mapping of Fn3_5 binding to MSLN.

(A) Three residues in ABCD MSLN were chosen for mutation to Ala based on the AlphaFold3 or ClusPro models, generating three single mutants and one double mutant. All four MSLN variants expressed on the surface of yeast and bound MUC16. (B) Titration binding assays revealed that Glu406 and Ser445 were important for binding, supporting the AlphaFold3 model. Meanwhile, the Glu400Ala mutant corresponding to the ClusPro model binding interface maintained Fn3_5 binding that was not statistically different compared to wildtype ABCD. Representative binding curves of at least three replicates conducted on different days are shown, and mean and standard deviation reported for all replicates. Statistical analysis of differences comparing ABCD WT and mutants are given in the text and in Table S5.

3. DISCUSSION

In this study, we reported the computational modeling and experimental validation of the interaction between cancer antigen MSLN and a high-affinity MSLN-binding Fn3 variant. Prior to this work, Fn3 variants were engineered to bind tumor cell surface protein MSLN, and the Fn3 proteins were shown to be internalized and have a targeted therapeutic effect on MSLN-positive cancer cell lines.^27,28 However, there was little understanding of the molecular interaction between Fn3 variants and MSLN that enabled the observed internalization and cell killing.

To unravel the putative three-dimensional structure of the Fn3_5/MSLN complex along with the interface’s residues mostly contributing to the binding, we have applied a combination of computational techniques. Notably, while the protein-protein docking algorithms ClusPro, HDOCK, and pyDock point towards a general consensus on the binding site, which is close to the N-terminus of MSLN, the newly developed deep-leaning based algorithm AlphaFold3 suggests a different binding pose, in which Fn3_5 interacts with the B and C domains of MSLN (Figure 1A). All the obtained models were refined using μs-long MD simulations, followed by an evaluation of the binding free energy. This approach allowed us to obtain a relative energetic ranking and a per-residue description of the interactions (Table 1). Finally, the experimental data are in agreement with the model obtained using AlphaFold3, showing that MSLN domains B and C are the most important for interacting with Fn3_5.

MSLN has three N-linked glycosylation sites (Asn388, Asn488, and Asn515). There is experimental evidence that the glycan patterns can vary across cancer cell lines and patient samples, with some data indicating that differential glycosylation may influence MSLN binding to MUC16.^5,11,20,52 In the present study, the computational prediction did not incorporate the MSLN glycosylation data. Since the sites were mutated for the validation step, MSLN expressed on the yeast surface was not glycosylated; thus, the epitope mapping revealed that Fn3_5 can bind MSLN without glycans. The model supports this experimental data, in that the Fn3_5-MSLN binding site was predicted to be far from the MSLN glycosylation sites. Therefore, Fn3_5-MSLN interaction does not seem dependent on MSLN glycosylation (Figure S7). However, since one of the glycosylation sites (Asn388) is located in domain B (Tyr360-Lys427), in future analyses, we could explore if the glycosylation state of MSLN could stabilize the MSLN subdomain structures and/or Fn3_5-MSLN binding.

In this work, the MSLN-Fn3_5 binding prediction was carried out using the structure of MSLN in its shed form (residues Glu296-Gly580), not including the C-terminal fragment. Therefore, data about the membrane-proximal region of MSLN are not included in the in silico analysis. Additionally, neither the engineering process nor the experimental validation (MSLN and the subdomains expressed on the yeast surface) incorporates the C-terminus. Thus, our results highlighted that Fn3_5 can bind MSLN with high affinity, despite the disordered C-terminus, suggesting that this region is not critical for Fn3_5-MSLN complex formation. In recent years, the membrane-proximal region of MSLN has been a target for other studies, in which antibodies or CAR-T cells were produced to bind the C-terminus of MSLN.^18,53,54 Targeting this disordered C-terminal fragment seems important for CAR-T cell and antibody-based strategies but may be less critical for MSLN-binding Fn3s that were reported to be quickly internalized.²⁷ Targeting the unshed portion of MSLN was proposed as an alternative to avoid the “off-tumor” effect due to SMRP in a patient’s bloodstream that acted like a decoy for previous antibody-based strategies, which resulted in low anti-tumor efficacy.^10,24 Therefore, when testing Fn3_5 in further research, with in vivo models or in patients, the SMRP concentration in the bloodstream should be considered, adjusting the Fn3_5 dose to avoid or limit the effect that “off-tumor” MSLN binding may have on reducing tumor binding.

In conclusion, we have identified a binding interface between Fn3_5 and MSLN by a combination of computational modeling and experimental validation. This study supports the reported consensus that the AlphaFold3 algorithm greatly improves the prediction of the three-dimensional structure of protein-protein complexes, which, in our case, outperforms existing protein-protein docking methods. In this regard, AlphaFold3 seems to bypass the well-known problems that plague docking calculations, such as the rigid model receptor, which is intrinsically overcome by the new methodology. Thus, the identification of the MSLN-Fn3_5 binding interface can serve as a foundation for the design of novel Fn3 variants to enable an even stronger binding affinity toward MSLN. Efforts to target MSLN for cancer diagnosis and therapy are ongoing,^55–58 and these results can inform the further development of candidate anti-MSLN molecules for diagnostic and therapeutic purposes. Additionally, the methods reported are also relevant for studying the interaction of other targeting macromolecules and their receptors.

4. MATERIALS AND METHODS

4.1. Model building

The structure of mesothelin was obtained from the protein data bank (pdb ID 8CX3).²⁰ The Fn3 variants described by Sirois et al. to bind MSLN were further matured as a fifth-generation library through yeast surface display and directed evolution, using our previously reported methods;^27,28 details will be reported separately (unpublished, manuscript under review). The structure of variant Fn3_5 was modeled using the SWISS-MODEL server,⁵⁹ starting as a template from the 6XAY pdb. The model was then relaxed with a 1 μs-long MD simulation, and the most representative structure was extracted using a cluster analysis for the protein-protein docking calculations.

4.2. Docking calculations

We performed blind protein–protein docking calculations to identify the three-dimensional structure of the MSLN/Fn3_5 complex. Due to the limitations of docking algorithms in correctly predicting the native protein–protein binding interface, we have employed a consensus docking strategy, which has been successfully performed for similar systems.⁴³ We used programs based on different search algorithms and scoring functions to verify if the predicted binding poses were reproducible. In particular, we confronted several docking Web servers such as ClusPro,⁴⁴ HDOCK,⁴⁴ and pyDock⁴⁶ using the suggested parameters and without specifying any constraints on interacting residues. These all share search algorithms based on fast-Fourier transform (FFT) methods for grid matching and similar scoring functions, mainly based on desolvation and electrostatic contributions, with the exception of HDOCK, which, instead, employs an iterative knowledge-based scoring function. Moreover, AlphaFold3 was used to predict the 3D structure of the MSLN/Fn3_5 complex. The generated model reported an overall Predicted Local Distance Difference Test (pLDDT) greater than 90, indicating high confidence in the accuracy of the structured regions. The overall quality of the model, as indicated by the high pLDDT values, suggests that the well-structured regions are reliable for further analysis. Then, we selected the highest-ranking model from each used algorithm for performing MD simulations.

4.3. Molecular dynamics (MD) simulations

We performed three 1 μs-long replicas of the highest-ranking docking models obtained with all the software. In particular, these were done using the GROMACS 2021.2 software package⁶⁰ and the Amber ff99SB-ILDN force field⁶¹. The models were embedded in a 15 Å layer of TIP3P water molecules. The total charge was neutralized, and additional Na⁺ and Cl⁻ ions were added to achieve a physiological salt concentration of 0.150 M. All the topologies were built with Ambertools 18 ⁶² and were subsequently converted in a GROMACS format using the software parmed. We have used a soft equilibration protocol in all the MD simulations, as described previously.⁶³ The systems initially went through an energy minimization employing a steepest descent algorithm considering a force convergence criterion of 1000 kJ/mol·nm². Next, the models were annealed from 0 to 300 K using a temperature gradient of 50 K every 2 ns, totaling 12 ns. The proteins were subjected to harmonic position restraints with a force constant of 1000 kJ/mol·nm² while water molecules and ions were allowed to move. Once the target temperature of 300 K was reached, 20 ns of NPT simulations were conducted to stabilize the pressure to 1 bar using a Berendsen barostat. The same restraints used in the heating phase were retained in this phase. Temperature control at 300 K was achieved by stochastic velocity rescaling thermostat. Afterward, the barostat was switched to Parrinello-Rahman, and the restraints were restricted only to the backbone atoms. These restraints were then gradually decreased into three consecutive runs of 20 ns each, during which the force constant was set to 1000, 250, and 50 kJ/mol·nm², respectively. Thus, after a long equilibration protocol of ~100 ns, all the position restraints were released, and all production runs were performed for 1000 ns for each model. Finally, productive MD simulations were performed using the isothermal-isobaric ensemble (NPT). The particle mesh Ewald method was used to account for long-range electrostatic interactions with a cutoff of 10 Å, and a LINCS algorithm was used to constrain the bonds involving hydrogen atoms. An integration time step of 2 fs was used in all simulations. To test for the conformational stability of MSLN domains, three replicas of a 3 μs-long MD simulations were carried out.

4.4. MM-PBSA calculations

Binding free energies (ΔG_b) between the MSLN and Fn3_5 proteins were calculated using the molecular mechanics-Poisson Boltzmann surface area (MM-PBSA) method⁴⁷ using the MMPBSA.py program,⁶⁴ as successfully employed in other studies.⁶⁵ Errors are reported as their corresponding standard deviations. MM-PBSA calculations were performed on 100 frames extracted from the MD trajectories. The free energy’s entropy contribution was not considered, as it was suggested that this term does not improve the quality of the results using the MM-PBSA.⁶⁶

4.5. Protein production and purification

MSLN-binding Fn3 variant Fn3_5 was produced as previously described.^27,28 Briefly, Fn3_5 gene was cloned into a pET vector with a C-terminal hexahistidine tag and expressed in BL21(DE3) E. coli. Cultures were grown in LB and induced at 37°C for 3 hours with 0.5 mM isopropyl-b-D-thiogalactopyranoside (IPTG). Cells were resuspended in lysis buffer (35 mM Na₂HPO₄ dibasic, 15 mM NaH₂PO₄ monobasic, 500 mM NaCl, 5 mM CHAPS, 25 mM imidazole, 5% glycerol) supplemented with an EDTA-free protease inhibitor (Pierce), and lysed by freeze-thaw cycles. Soluble fractions were isolated through centrifugation and passed through 0.45 μm and 0.22 μm filters. Then, Fn3 proteins were purified by cobalt affinity chromatography with HisPur cobalt resin (Thermo Fisher). Samples were dialyzed into water, lyophilized, and resuspended with phosphate buffered saline (PBS) to the desired concentration. Fn3 purity was assessed by SDS-PAGE on a BioRad ChemiDoc MP imaging system.

4.6. Cloning and expression of MSLN domains using yeast surface display

Based on the recently reported crystal structure of MSLN,²⁰ the sequence for the full length shed MSLN (residues Glu296 - Gly580) or combinations of the reported domains A, B, C, and D were cloned into the yeast surface display pCT plasmid between the NheI and BamHI restriction sites following established methods.^67,68 The following domain combinations were chosen: single domains A (N-terminal domain, Glu296-Leu359), B (Tyr360-Lys427), C (Asp428-Gly501), D (Gly502-Gly580); two domain combinations AB, BC, CD; three domain combinations ABC, BCD; and full length shed MSLN ABCD. A gene block of the full length ABCD gene was designed (Integrated DNA Technologies) as a PCR template, with the three N-linked glycosylation sites at Asn388, Asn488, and Asn515 removed by mutation of Asn to Gln. Subdomains were cloned using primers designed to add the NheI and BamHI cloning sites. Following subcloning and transformation into XL1-Blue E. coli, sequences were confirmed by standard Sanger sequencing. Plasmids were transformed into EBY100 S. saccharomyces surface display strain following protocols for the Frozen-EZ Yeast Transformation II Kit (Zymo Research). An individual yeast colony with plasmid from an SD-CAA plate was grown overnight in SD-CAA liquid culture at 30°C with aeration, then induced in SG-CAA liquid media at 20°C with aeration for 18 h. Expression of all domains was confirmed by flow cytometry using the mouse 9E10 antibody that recognizes the C-terminal c-myc tag (Fisher Scientific, 13-250-0, 1:50 dilution) and a goat anti-mouse PE conjugate antibody (Sigma, P9670,1:25 dilution). Experiments to assess the ability to express each domain were repeated at least three times for each construct. For residue-level fine epitope mapping studies, additional gene blocks were designed and ordered (Integrated DNA Technologies), with a codon for Ala replacing the codons corresponding to the residues identified for mutation, and the above procedures were followed. The MSLN residues mutated to Ala individually were Glu400, Glu406, or Ser 445. Double mutant Glu406Ala/Ser445Ala was also constructed.

4.7. Validation of yeast surface displayed MSLN system

All yeast surface display incubations and washes were performed using FACS buffer, composed of PBS with 0.1% bovine serum albumin. To determine the domain on MSLN bound by a rabbit anti-MSLN antibody (Abcam, EPR19025–42), 5 × 10⁵ yeast of each domain to be tested were incubated for 1 h at room temperature and gentle mixing with mouse 9E10 anti-c-myc antibody to measure domain expression and with rabbit anti-MSLN antibody (1:50 dilution) in a total volume of 50 μl. Samples were washed, then incubated for 30 min at 4°C with a goat anti-mouse AF488 conjugate (Thermo Fisher, A11001, 1:200 dilution) and a goat anti-rabbit PE conjugate (Abcam, ab72465, 1:200 dilution) in 25 μl volume. Samples were washed, and immediately analyzed on a Guava flow cytometer. Data were compensated for spectral overlap. Domain-level epitope mapping of MUC16 binding was performed similarly, with alternate antibodies to avoid cross-species reactivity. For each MSLN domain construct, a near-saturating concentration (100 nM) of a MSLN-binding domain of MUC16 with a hexahistidine tag (Acro Biosystems) was incubated with induced yeast during the primary incubation step simultaneously with chicken anti-c-myc antibody (Gallus Immunotech, #ACMYC, 1:330 dilution) for 1 h at room temperature with gentle mixing. Cells were washed, and incubated for 30 min at 4°C with a mouse DyLight 488 anti-6X His tag antibody (Abcam, ab117512, 1:50 dilution) and a goat anti-chicken PE antibody (Abcam, ab72482, 1:50 dilution).

4.8. Domain-level epitope mapping and titration binding assays of Fn3_5 with yeast displayed MSLN constructs

An equilibrium titration binding assay was performed with purified Fn3_5 and the MSLN ABCD construct, or with mutants of ABCD. A range of concentrations of Fn3_5 with hexahistidine tag, from 0.1 nM to 400 nM, was incubated with 5 × 10⁵ yeast displaying an MSLN ABCD variant, co-incubated with chicken anti-c-myc antibody, for 1 h at room temperature with gentle rotation. Yeast were washed, and incubated for 30 min at 4°C with mouse DyLight 488 anti-6X His tag antibody and goat anti-chicken PE antibody. The experiment was performed at least three times for each variant. For each replicate, data were plotted in Kaleidagraph and fit with a sigmoidal curve. The dissociation constant K_D was determined for each replicate as the concentration yielding the half maximal signal, and the mean and standard deviation of all replicates reported. To determine statistical significance of the differences between mean K_D values comparing mutants to ABCD wildtype, a two tailed t-test was conducted, and p-values are reported in Table S5.

Domain-level epitope mapping was performed for each MSLN construct displayed on the surface of yeast, for two concentrations of Fn3_5, at 100 nM or 400 nM of protein. All other experimental conditions and antibodies were as described for the equilibrium titration binding assay. All data were compensated for spectral overlap.

Supplementary Material

SI File 1

Supporting Information File 1 includes:

Figure S1. Homology modeling of Fn3

Figure S2. Superposition with MORAb-009 structure

Figure S3. Root mean square deviation (RMSD)

Figure S4. Principal component analysis (PCA) of MSLN

Figure S5. Root mean square fluctuation (RMSF)

Figure S6. MSLN domains AB on the surface of yeast do not demonstrate detectable binding to Fn3_5

Figure S7. Visualization of glycosylation sites on MSLN

Table S1. Per-residue decomposition of the binding free-energy (ΔG_b) of the ClusPro complex.

Table S2. Per-residue decomposition of the binding free-energy (ΔG_b) of the AlphaFold3 complex.

Table S3. Per-residue decomposition of the binding free-energy (ΔG_b) of the HDOCK complex.

Table S4. Per-residue decomposition of the binding free-energy (ΔG_b) of the pyDock complex.

Table S5. Statistical analysis comparing K_D values of Fn3_5 binding MSLN variants.

SI File 3

Supporting Information File 2 includes: A video reporting the motion of MSLN, as extracted from the molecular dynamic trajectory of the MSLN/Fn3_5 ClusPro complex. Principal component 1 is shown.

SI File 2

Supporting Information File 3 includes: A video reporting the motion of MSLN, as extracted from the molecular dynamic trajectory of the MSLN/Fn3_5 AlphaFold3 complex. Principal component 1 is shown.

SIGNIFICANCE.

A combined experimental and computational approach furnishes the first model of the interaction between cancer biomarker mesothelin and an engineered binding scaffold protein. We identified the region on MLSN that is mainly involved in the binding interface. Gaining atomistic insights could pave the way for developing novel binding variants with stronger affinity, and broadly inform the development of mesothelin-targeting diagnostics and therapies. Further, the methods reported can be used to experimentally validate protein docking models for other biochemical systems.