Cell-Based Immunization Combined with Single-Round Cell Panning Enables Discovery of PSMA-Targeting Nanobodies from Phage Display Libraries

Abstract

There is a strong need for nanobodies that target novel cancer-associated antigens to advance radioligand imaging and antibody-based therapeutics. In this study, we investigated whether non-targeted llama immunization using tumor cells, combined with non-targeted phage-display panning of human cell lines, could yield nanobodies specific to Prostate-Specific Membrane Antigen (PSMA). Nanobody selection using both classical three-round PSMA negative–positive panning and single-round panning of cell lines (positive or negative) for PSMA showed clear enrichment for PSMA binders in both strategies. Using shRNA knockdown, flow cytometry, cell-ELISA, immunohistochemistry and structural modeling and docking, we confirmed the PSMA-targeting of selected nanobodies. Two distinct epitopes were predicted to be bound by nanobodies PSMANb9 and A7 (JVZ-007), and this was corroborated by epitope competition assays. These findings support the feasibility of non-targeted immunization and panning strategies for isolating antigen-targeting cancer nanobodies.

1. Introduction

The 2018 Nobel-prize-winning phage display technology was first introduced by George Smith in 1985 for peptide display [1] and later exploited by McCafferty et al. to display antibody fragments [2]. The most commonly used antibody fragments include the single-chain variable fragments (scFvs) and variable heavy domain of heavy-chain–only antibodies (VHH, also known as nanobodies). ScFvs are engineered fusion proteins of the variable regions of both heavy and light chains, while nanobodies are naturally occurring, single-domain antibody (sdAb) fragments derived from camelid heavy-chain antibodies [3]. Libraries of nanobodies are typically generated by immunization of llamas with the antigen of interest, after which multiple rounds of pannings take place against the immobilized single target protein [4]. Next-generation sequencing (NGS) of >1 × 10⁵ selected sdAbs replaced the classical Sanger sequencing of <1000 individually picked clones, which had a major impact on insights into enrichment and binding characteristics of clonally related sequences [5].

The conventional approach for targeted nanobody screening typically involves purified antigen proteins for animal immunization and follow-up validation, which may fail to capture epitopes due to a lack of post-translational modifications or conformational variations in natural cell membrane environments. In contrast to targeted selections, we define non-targeted as a discovery strategy where immunization and selection is performed with complex biological mixtures (e.g., cells) and nanobodies against many different targets can be identified, without predefining a single molecular antigen. In our unbiased search for cancer nanobodies, we questioned whether non-targeted llama immunization using whole tumor cells, combined with non-targeted phage panning across a diverse panel of cell lines, could yield nanobodies recognizing a broad repertoire of distinct cell-surface antigens. If so, one can perform a single round of panning against a large collection of cell lines and identify nanobodies binding to a specific subset of samples that uniquely express a protein of interest. Such non-targeted panning cannot easily be performed as a classical multi-round negative–positive panning since the negative samples are different for each target.

As our target of interest, we selected Prostate-Specific Membrane Antigen (PSMA), a type II transmembrane glycoprotein, also known as glutamate carbopeptidase II (GPCII). PSMA is highly expressed in normal prostate epithelial cells and, additionally, in the brain, small intestine, salivary glands and kidneys [6]. In most primary prostate cancers (PCa), and particularly in advanced metastatic lesions, PSMA is overexpressed and is a well-known target for nuclear imaging [7,8,9] and radioligand therapeutics [10,11]. In particular, PSMA PET (Positron Emission Tomography) tracers (e.g., 68Ga-PSMA-HBED-CC (PSMA-11), 68Ga-PSMA-617, and 68Ga-PSMA-I&T) changed the clinical diagnostics of metastasized PCa.

We, and others, have previously identified nanobodies targeting PSMA and tested their efficacy in imaging as mono- and bivalent nanobodies and in therapeutics as bispecifics and CAR-T fusions [12,13,14,15,16,17,18,19]. The clinical application of nanobodies is strengthened by their unique structural features, including their small size (~15 kDa), enhanced tissue penetration and high stability [20]. Technology-wise, the full nanobody sequence (110–140 amino acids) can be determined for millions of clones in a straightforward NGS run. The increasing availability of nanobody sequences, together with their target proteins, has accelerated advances in nanobody structural modeling and docking studies [21,22]. The next steps towards AI-driven target prediction and de novo nanobody design are being taken and will strongly benefit from large panning databases [23,24].

In this study, we compared the panning protocols of nanobodies from three phage display libraries to identify PSMA-binding nanobodies. A single panning round against a series of human cell lines was compared to classical three-round negative–positive targeted panning. Candidate nanobodies exhibiting PSMA-associated binding were examined in cell-binding assays and further validated by the immunostaining of human normal and cancer tissues. Among these, the specificity for PSMA of PSMANb9 and A7 (JVZ-007) [12] was confirmed in PSMA knockdown LNCaP cells. In silico analysis of nanobody–PSMA interactions revealed two distinct binding sites on the extracellular domain of PSMA, and this was confirmed by in vitro epitope competition assays.

2. Materials and Methods

Details of the various protocols can be found in the Supplementary Information.

2.1. Llama Immunization and Nanobody-Library Construction

Three llamas were immunized with four PCa cell lines (LNCaP, PC346C, VCaP, and MDAPCa2b) and a combination of freshly frozen PCa and bladder cancer (BlCa) cell lines, from which three nanobody-phage libraries were constructed. Characteristics of the three different libraries are presented in

Table 1. Immunizations and characteristics of the three nanobody-phage display libraries.

Library	Immunization of Lama Glama	# of Boosts After First Immunization	Independent Clones	Reference
L1P4	VCaP, PC346C, LNCaP, MDAPCa2b cell lines	3	3 × 109	[12]
LUPCa1	Pool of cell fractions isolated from 12 prostate tumors and 7 bladder tumors, from different patients	2	1.2 × 109
LUPCa2		2	0.6 × 109

.

2.2. Screening and Selection of Nanobodies Using Phage Display Technology

To produce phage display libraries, each nanobody library was infected with M13 K07ΔpIII hyperphages (Progen Biotechnik, Heidelberg, Germany). To screen PSMA-binding nanobodies, the full L1P4 library or the mix of three libraries were used in a 0.13:0.52:0.35 ratio and screened involved three biopanning strategies: three rounds of negative–positive panning (3R-NegPos) against B16-WT and B16-PSMA, one round of single-cell panning (1R-SC) against B16-WT and B16-PSMA and one round of biopanning against PSMA-positive and PSMA-negative cancer cell lines (1R-cell line panning). The biopanning protocol was adapted from Chatalic et al. [12]. For 3R-NegPos, each round consisted of a negative selection against B16-WT cells using 100× the L1P4 library followed by a positive selection against B16-PSMA cells. In the 1R-SC strategy, one individual positive selection was performed against B16-WT and B16-PSMA, respectively. For the 1R-cell line panning, one single positive selection was performed on each cell line, separately.

To select B16-PSMA specific nanobodies, the conventional colony-picking method was used, as described previously [12]. For the NGS analysis of the selected phages at the end of each panning round, the nanobody cDNA from each nanobody-phage sample corresponding to an individual cell line from a single selection round was amplified and sequenced using the Illumina MiSeq 2 × 300 bp protocol to a 1-mln paired read-depth per sample (Illumina MiSeq Reagent Kits v3, San Diego, CA, USA). For each sample and panning round, a specific barcode (index) was present in the PCR amplification primers to allow multiplexing of samples in each MiSeq run. NGS data was processed using an in-house developed pipeline, as described in the Supplementary Information.

2.3. Selection of PSMA-Binding Nanobodies from Nanobody Database

To select PSMA-binding nanobodies, we defined certain criteria within each biopanning strategy. For 3R-NegPos, we first determined the ratio between selection rounds (SR) 2/1 and 3/1 with respect to the frequency of each unique nanobody sequence and selected candidate nanobodies for which both the SR2/SR1 and SR3/1 was greater than 5, while the value in SR3 was set to greater than 1000 (is 0.001% of scaled 100 million). The read-count threshold and fold ratio were conservatively selected above the bend of the ratio–rank curve to enrich for nanobodies beyond the background/noise (Supplementary Figure S2). For 1R-SC, the ratio between B16-PSMA and B16-WT should be greater than 5, and the value of the average B16-PSMA greater than 1000. For the 1R-cell line panning, the average level of nanobody binding to PSMApos and PSMAneg cells was determined where nanobodies with a PSMApos/PSMAneg ratio greater than 5, XMRVpos/XMRVneg ratio less than 3 and average PSMApos greater than 1000 were selected as top candidates.

The DNA sequence of each selected nanobody was retrieved from sequencing data to produce nanobody-phages, which were further characterized.

2.4. Hierarchical Clustering and Circular Dendrogram of Nanobody CDR3 Sequences

Nanobody CDR3 sequences were clustered based on pairwise Levenshtein distances using average-linkage hierarchical clustering. The resulting dendrogram was displayed in a circular configuration using R (igraph, ggraph, and stringdist). Branches were colored by dominant library—L1P4 (coral red), LUPCa1 (green), and LUPCa2 (blue)—with progressively darker shades indicating deeper hierarchical levels. Selected nanobodies were indicated by black outer nodes, whereas unselected ones were represented by light gray.

2.5. Whole-Cell ELISA

PSMA binding of the selected nanobodies was assessed by cell-ELISA. In the preliminary screening and validation phases we utilized the nanobody-phages instead of pure nanobody proteins, due to the high efficiency/throughput, replication of phage-based detection, higher signal intensity and cost-effectiveness of the phage display format. The experiment was performed using two PSMA-positive cell lines, LNCaP and B16-PSMA, along with two PSMA-negative cell lines, DU145 and B16-WT. Nanobody binding was detected using mouse anti-M13 antibody as the primary antibody followed by goat anti-mouse HRP secondary antibody and OPD substrate. Antibody information is provided in Supplementary Table S3.

2.6. Cell Immunohistochemistry (IHC)

To investigate the specific binding of nanobodies, we created a cell array where 20 cell lines were coated onto a glass cavity slide (Marienfeld, Lauda-Königshofen, Germany) and stained with nanobodies. Briefly, cells were blocked with 1 mL 1% BSA-PBS and then stained with 5 × 10¹⁰ PFU/slide of nanobody-phages. Nanobody binding was detected using mouse anti-M13 antibody followed by goat anti-mouse HRP secondary antibody and 3,3′-Diaminobenzidine (DAB) substrate.

2.7. Immunohistochemistry Staining of Frozen Tissue Microarrays and Frozen Prostate Tissue Sections

Tissue microarray (TMA) slides representing 20 human normal tissues in duplicate were purchased from Biochain (Newark, CA, USA). Frozen human prostate normal and cancer tissues were obtained from the Erasmus MC Tissue Biobank (Rotterdam, The Netherlands), sectioned with 5–10 um thickness, attached to Starfrost Adhesive slides (Deltalab, Rubí, Spain) and frozen to −80 °C. Immunohistochemistry (IHC)-staining of tissues was performed according to the protocol described above. Slides were scanned by PARTS (Erasmus MC Pathology and Research and Trial Service, Rotterdam, The Netherlands) using Nanozoomer 2.0 HT digital slide scanner (Hamamatsu Photonics K.K, Herrsching am Ammersee, Germany) and images were visualized and analyzed using NDP.view2 (Version 2.9.25) software.

2.8. Lentiviral Transduction of shRNA for PSMA Knockdown

HEK293T cells were transfected using calcium phosphate transfection kit with a combination of 5 different PSMA shRNAs to produce lentiviruses (Mission shRNA libraries Sigma-Aldrich, Erasmus MC Center for Biomics, Rotterdam, The Netherlands). A GFP-Luciferase fusion construct (M21) and non-targeting human gene shRNA (shNC) served as positive and negative controls, respectively. The virus containing supernatant was collected 48 h post-transfection and used to infect LNCaP cells.

2.9. Western-Blot Analysis of PSMA Expression

Cells were lysed using RIPA buffer (50 mM Tris-HCl, 150 mM NaCl, 1% Triton X-100, 0.25% Sodium deoxycholate, 0.1% SDS, 2 mM EDTA) and total protein was extracted. Five µg of each protein sample including wild type (WT), knockdown (shPSMA) and shRNA negative control (shNC) LNCaP cells along with DU145, B16-WT and B16-PSMA were resolved on a 10% SDS-PAGE gel and transferred to a PVDF membrane. PSMA was detected using rabbit monoclonal anti-PSMA antibody while anti-β-actin antibody and anti-human GAPDH antibody were used as controls. Protein bands were detected using goat-anti-rabbit HRP or goat-anti-mouse HRP followed by BM chemiluminescence blotting substrate.

2.10. Binding of Nanobodies to LNCaP Cells in a Flow Cytometry Assay

Binding of nanobody-phages to wild type (WT) and PSMA knockdown (shPSMA) LNCaP cells was investigated in a flow cytometry assay. For detection, A7 and PSMANb9 nanobody-phages were directly labeled with 0.25 mM ATTO488 NHS-ester dye (ATTO-TEC, Siegen, Germany). Mouse monoclonal anti-PSMA antibody and goat-anti-mouse IgG cross-adsorbed secondary antibody Alexa Fluor 647 (Thermo Fisher Scientific, Waltham, MA, USA) were used to detect PSMA expression in LNCaP-WT and shPSMA LNCaP cells.

2.11. Nanobody Protein Production

A7 and PSMANb9 nanobodies were cloned into the pHEN vector (Twist Bioscience, South San Francisco, CA, USA) and proteins expressed, including a C-terminal Myc-tag and 6xHis tag using E. coli HB2151 cells as host. Nanobody proteins were purified on a Ni-NTA agarose column using standard 25–500 mM imidazole gradient.

2.12. In Silico Modeling and Docking of Nanobody–PSMA Complex

The 3D structure of human PSMA ectodomain (PDB ID: 1Z8L) was processed using BIOVIA Discovery Studio Visualizer (Dassault Systèmes BIOVIA, San Diego, CA, USA) to remove water molecules and unbound atoms. The 3D structures of candidate nanobodies-A3, A7, PSMANb5, PSMANb6, PSMANb9 and PSMANb11, were predicted using Robetta (http://robetta.bakerlab.org, accessed on 1 April 2024) and SWISS-MODEL server (https://swissmodel.expasy.org, accessed on 1 April 2024), both of which leverage homology modeling method [25]. To generate more accurate models, particularly within CDR3 regions, we used RoseTTAFold, a software tool that employs deep learning methods for accurate prediction of protein structures [25]. The predicted models were downloaded and evaluated for their overall quality (Supplementary Information).

Nanobody models and PSMA ectodomain were prepared for docking simulation using Dock Prep tool in UCSF Chimera (Resource for Biocomputing, Visualization, and Informatics, University of California, San Francisco, CA, USA) followed by docking simulation of nanobody–PSMA monomer using the “easy interface” of HADDOCK 2.4 (High Ambiguity Driven protein–protein DOCKing) web server (https://wenmr.science.uu.nl/haddock2.4, accessed on 30 April 2024). The generated models were clustered and ranked based on their HADDOCK score. To compare the PSMA epitopes bound by PSMA-617 and our selected nanobodies, we also docked PSMA617 peptide to PSMA monomer using PyRx 0.8 [26] and Autodock 4.2 (The Scripps Research Institute, La Jolla, CA, USA) [27]. Molecular interactions were visualized and analyzed using Python Molecular Graphics (PyMOL, Version 2.3.4; Schrödinger, LLC, New York, NY, USA) and LigPlot+ (Version 2.2.9; European Bioinformatics Institute, Cambridge, UK) software [28]. The stability of nanobody–PSMA complex models was evaluated using GROMACS 2022 package (GROningen MAchine for Chemical Simulations, Groningen, The Netherlands) and OPLS-AA force field (Optimized Potential for Liquid Simulations-All Atom, Yale University, New Haven, CT, USA) [29].

2.13. Epitope Binding Competition Assay

A titration assay was conducted to determine the nanobody protein concentration sufficient to fully block the PSMA-binding sites on the cell surface. LNCaP and B16-PSMA cells were incubated with varying concentrations of A7 or PSMANb9 nanobody proteins and binding was detected using a mouse monoclonal anti-Myc tag antibody followed by a goat anti-mouse Alexa Fluor 647 secondary antibody. Similar titration was performed using varying concentrations of ATTO488-labeled A7 or PSMANb9 nanobody-phages. For competition, LNCaP cells or B16-PSMA were first incubated with A7 or PSMANb9 proteins and, without washing, this was followed by adding ATTO488-labeled PSMANb9 or A7 nanobody-phages.

2.14. Statistics

For hierarchical clustering and visualization and statistical analyses we used GraphPad Prism 9.0 (San Diego, CA, USA). The hypergeometric test was used for determining p-values of overlap in the Venn diagrams. The nonparametric Mann–Whitney test was applied for the cell-ELISA.

3. Results

3.1. Schematic Workflow of the Panning and Selection Strategy

In our search for an undirected cell-based panning approach to identify novel PSMA nanobodies, we compared two panning strategies: (i) the classical targeted panning against the wild type mouse melanoma B16 (B16-WT) cell line and human PSMA-expressing B16 (B16-PSMA) cell line and (ii) panning against a panel of PSMA-positive and PSMA-negative human cell lines (Supplementary Figure S1).

For the nanobody-phage panning, we included three different libraries. L1P4 was previously generated by immunization of a llama with the prostate cancer (PCa) cell line suspensions of (PSMA-positive) VCaP, PC346C, LNCaP and MDAPCa2b [12]. The LUPCa1 and LUPCa2 libraries were created by llama immunization with cell fractions from a collection of freshly frozen PCa and bladder cancer (BlCa) patient tumors (Section 2 Table 1).

3.2. PSMA-Targeted L1P4 and Mixed Library Panning Using B16-WT and B16-PSMA Cell Lines

For the targeted panning against B16-WT and B16-PSMA using the L1P4 library, we compared two different panning procedures: (i) the classical three rounds of negative–positive panning (3R-NegPos) and (ii) one-round single-cell line panning (1R-SC).

The 3R-NegPos selection consisted of three rounds, starting with the 100× L1P4 library against B16-WT cells, followed by incubation of the unbound fraction with B16-PSMA cells to isolate bound nanobody-phages. Upon next-generation sequencing (NGS) and database building, we selected 462 nanobody clusters with different CDR3s that were >5-times more abundant in rounds two and three as compared to round one, with a minimal read count of 0.001% in round three of the total number of reads. We applied these thresholds to identify nanobody clusters showing a pronounced enrichment to the background/noise in a ratio–rank curve (Supplementary Figure S2).

To select PSMA-targeting nanobodies without negative selection, we performed a one-round single-cell line panning (1R-SC) using B16-WT and B16-PSMA cells. We selected 341 nanobody clusters that were >5-times more abundant in the B16-PSMA panning as compared to the B16-WT and had a read count of >0.001% in the B16-PSMA of the total number of reads.

The overlap between the 462 3R-NegPos nanobody clusters and the 341 1R-SC clusters was only 6 (p < 1.7 × 10⁻⁴) (Figure 1A). These six common nanobody clusters were strongly enriched during 3R-NegPos panning from 0.001% of the reads in the L1P4 library to 3.0%, 15.7% and 29.9% of the reads in rounds one, two and three, respectively. In the single-round B16-PSMA panning, the six PSMA nanobody clusters accounted for only 0.41% of the total number of reads sequenced. Previously, using 3R-NegPos panning, the A3, A5 (JVZ-005) and A7 (JVZ-007) PSMA nanobodies were identified via colony-picking [12]. This corresponds to our NGS identification as these three nanobodies overlap with the six highly abundant clusters. The other three abundant nanobody clusters are all highly similar to the A3 CDR3 sequence and are expected to bind the same PSMA epitope.

Next to the panning using the L1P4 library, we tested whether a mix of L1P4 with the LUPCa1 and LUPCa2 libraries could reproduce the 1R-SC using the B16-WT and B16-PSMA pannings. The three libraries were mixed and based on a vector barcode, and the origin of each individual nanobody upon sequencing was determined.

Using an identical 1R-SC panning protocol and nanobody selection criteria, 507 nanobody clusters were selected as higher binders in B16-PSMA cells as compared to the B16-WT. Of these 507 nanobodies, 125 were from L1P4, of which 68 overlap with the 341 nanobody clusters identified using the L1P4 library mentioned above (p < 1 × 10⁻¹⁵⁰). All six nanobody clusters that overlapped with the 3R-NegPos are present in the list (Figure 1A).

3.3. Single-Round Panning Against a Collection of Human Cell Lines

Next, we performed one round of panning using the L1P4 library against 40 different (cancer) cell lines and the mix of three libraries (L1P4 with LUPCa1 and LUPCa2) against 12 cell lines. The cell lines were grouped into PSMA-positive (PSMApos) and PSMA-negative (PSMAneg) cell lines based on the level of FOLH1 mRNA (encoding the PSMA protein). We selected 498 nanobody clusters with, on average, more than five-fold higher binding to PSMApos cell lines and a minimal read count of 0.001% of the average count of the PSMApos cells. Comparison with the 3R-NegPos and 1R-SC selections against B16-PSMA, identified an overlap of 52 nanobody clusters, of which 5 intersect with all groups and 47 with 1R-SC (Figure 1B,C; Supplementary Table S1).

The combined targeted B16-PSMA and undirected cell-line pannings resulted in the identification of 32, 18 and 2 PSMA-binding nanobody clusters from the L1P4, LUPCa1 and LUPCa2 libraries, respectively. From these groups, we selected 11 nanobody clusters (A5, A7, and PSMANb1–3, 5–7, 9–11) based on binding specificity to B16-PSMA and PSMApos cell lines, their panning procedure and library origin (Figure 1B). The most abundant nanobody sequence from each cluster was DNA synthesized, cloned into the pHEN vector and phages were regenerated in TG1 bacteria for nanobody-phage validation. Due to the stringent threshold criteria, the clone-picked A3 nanobody did not fulfil the PSMApos/PSMAneg cell line selection criteria (<0.001% of reads). We still took it along as it was identified by colony-picking previously. In total, 12 nanobody-phages were generated for further validation.

3.4. Circular Dendrogram Illustrating Library-Specific Clustering of Nanobody CDR3 Repertoires

Circular dendrogram analysis was performed to visualize the sequence diversity and inter-library distribution of nanobody CDR3 repertoires after 1R-SC PSMApos cell line panning (n = 498; Supplementary Figure S3A). In the mixed repertoire, distinct sequence families corresponding to the three nanobody libraries were evident. The L1P4 library displays the highest diversity, harboring numerous potential PSMA-targeting nanobodies with low sequence similarity. In contrast, the LUPCa1 library contains a large number of sequences mainly confined within several dominant families, whereas most of the 52 overlapping PSMANbs fell outside these clusters. The LUPCa2 library exhibited, overall, fewer PSMApos cell line nanobody sequences, reflecting inter-animal variability in immune response. In the refined dendrogram of the 52 overlapping clusters (Supplementary Figure S3B), nanobodies from all three libraries were represented, including both unique and family-associated members, collectively illustrating the repertoire’s sequence diversity and representativeness for subsequent experimental validation.

3.5. Validation of Binding Specificity of Selected PSMA Nanobodies

We tested the binding specificity of the 12 selected nanobody-phages in a whole-cell ELISA assay using four cell lines: PSMA-positive cell lines (LNCaP and B16-PSMA) and PSMA-negative cell lines (DU145 and B16-WT) (Figure 2A). Six out of twelve nanobodies (A3, A7, and PSMANb5, 6, 9, and 11) showed significant higher binding in PSMA positive versus negative cell lines. We further examined PSMA-binding specificity of these 6 nanobody-phages as compared to the binding specificity of an IgG PSMA antibody. 20 different PSMApos and PSMAneg cell lines were used in an immunohistochemistry (IHC) assay where nanobody-phages were incubated with cells spotted on a glass slide (Figure 2B; Supplementary Figure S4).

The validation reproduced the panning data well: the Pearson Correlation Coefficient of the IHC scores and panning data of six nanobody-phages was, on average, (±SD) 0.60 (±0.18). As expected, the IgG anti-PSMA antibody bound specifically to PSMA-positive cell lines with a binding level reflecting the known FOLH1 RNA expression levels. The only exception was the VCaP cells having high mRNA levels, but relatively low antibody surface binding. Among the nanobody-phages, A7 and PSMANb9 consistently yielded strong staining across PSMA-positive cell lines, with the most intense staining observed in LNCaP cells. As compared to A3, A5 and PSMANb5, the 6, the 11, the A7 and the PSMANb9 nanobodies only bound (weakly) to a few PSMA-negative cell lines.

3.6. Nanobody-Phages A7 and PSMANb9 Specifically Bind Prostate (Cancer) Tissue

To further investigate the PSMA specificity of A7 and PSMANb9, we stained frozen sections of human PCa and normal adjacent prostate (NAP) tissues. IgG anti-PSMA antibody strongly stained PCa tissues compared to NAP tissues (Figure 3A,E). Similarly, A7 and PSMANb9 nanobody-phages exhibited comparable stronger staining patterns in PCa (Figure 3B,C) versus NAP, although with weaker intensity (Figure 3F,G).

Staining of human normal-frozen-tissue microarrays (TMA) (Supplementary Figure S5) showed weak staining with the phage without a nanobody (NoNb), particularly in liver and small intestine tissues, indicating non-specific interactions of phage proteins with tissue proteins. In concordance, the positive control anti-CD9 antibody and anti-CD9 nanobody-phage (H6) (Supplementary Figure S6) bound strongly across most tissues (www.proteinatlas.org/ENSG00000010278-CD9/tissue, accessed on 1 April 2024). The anti-PSMA IgG antibody strongly bound to small intestine tissue and to a lesser extent to salivary gland, kidney, and spleen tissue, which was in accordance with the levels of PSMA expression known for these tissues (www.proteinatlas.org/ENSG00000086205-FOLH1/tissue, accessed on 1 April 2024). Similarly, PSMANb9 and A7 nanobody-phages demonstrated binding to small intestine, salivary gland and kidney tissues, with an overall weaker binding across other normal tissues, which could be attributed to the non-specific interactions of phage coat proteins.

3.7. A7 and PSMANb9 Nanobody-Phages Bind PSMA Protein on the Surface of PSMA-Positive Cell Lines

To show that nanobody-phages A7 and PSMANb9 interact with the PSMA protein on the surface of PSMA-positive cell lines, PSMA expression was knocked down in LNCaP cells using shRNA technology, as confirmed by Western-blot analysis, where a significant knockdown of PSMA was observed in LNCaP-shPSMA cells compared to LNCaP-WT and LNCaP-shNC (negative control shRNA) cells (Figure 4A). An uncropped WB Figure can be found in Supplementary Figure S12.

We evaluated the binding of ATTO488-labeled nanobody-phages A7 and PSMANb9 to LNCaP-WT and LNCaP-shPSMA cells using flow cytometry. Notably, a sharp reduction in anti-PSMA antibody and nanobody signals was observed when bound to LNCaP-shPSMA cells compared to LNCaP-WT and LNCaP-shNC cells (Figure 4B). Both qualitative Western-blot and quantitative flow cytometry assays consistently demonstrated a clear reduction in PSMA expression.

3.8. Computational Modeling of Nanobody–PSMA Interactions Predicts Two Distinct Epitopes for A7 and PSMANb9

The 3D structures of candidate nanobodies were modeled using two approaches: the homology-based modeling tool SWISS-MODEL [30] and the RoseTTAFold [25] option of the Robetta server, a deep learning-based modeling platform that also incorporates de novo structure prediction for domains lacking sequence homology. The generated models were evaluated comprehensively using ERRAT scores, MolProbity scores, PROCHECK residue statistics, and ProSA-web Z-scores to assess their structural reliability (Supplementary Table S2).

Overall, both Robetta and SWISS-MODEL produced nanobody models of good structural quality. Robetta models generally demonstrated higher ERRAT scores and more favorable ProSA-web Z-scores, reflecting better non-bonded atomic interactions and overall fold quality. PROCHECK analysis showed that the majority of models were well-refined, with more than 90% of residues located within favored and allowed regions of the Ramachandran plot, indicating proper backbone geometry (Supplementary Table S2). ProSA-web results supported the high quality of all models, as the Z-scores (≤−6.03) were consistent with those reported for experimentally determined protein structures.

Despite the strong performance of both tools, Robetta-generated models were selected for further analysis due to their slightly superior scoring metrics. Moreover, Robetta’s ability to integrate de novo prediction methods for regions without suitable templates provides an added advantage in modeling challenging nanobody domains.

The nanobody models were docked to the PSMA ectodomain structure (PDB ID: 1Z8L) using HADDOCK 2.4 to predict the interaction interface between each nanobody and PSMA ectodomain (Figure 5) [31]. The docking generated multiple clusters of models ranked based on the HADDOCK score, which integrates van der Waals, electrostatic, desolvation energies, and restraint violations to evaluate binding affinity and interface quality. The best-ranked complexes exhibited favorable HADDOCK scores ranging from −120 to −83, indicating strong and reliable binding interactions. After obtaining the docked complexes from HADDOCK 2.4, the models were further analyzed using the PDBsum generator (European Bioinformatics Institute, Hinxton, Cambridge, UK) to identify specific interactions at the nanobody-PSMA interface. Based on the detailed analysis of hydrogen bonds (Figure 5C), nanobodies A7, A3, PSMANb6, and PSMANb5 were found to recognize one distinct epitope on the PSMA ectodomain. In contrast, nanobodies PSMANb9 and PSMANb11 targeted a separate and non-overlapping region of PSMA (Figure 5B). The PSMA-617 tracer binds a unique epitope close to the active glutamate carboxypeptidase site of PSMA, consistent with the crystallographic structure of the PSMA-PSMA-617 (PDB ID: 505U) complex (Figure 5A).

3.9. Molecular Docking Simulations Indicate the High Stability of Nanobody–PSMA Complexes

To assess the structural stability of PSMA-A7 and PSMA-PSMANb9 complexes, we performed three times 200 ns production MD (Molecular Dynamics) using GROMACS 2022 and OPLS-AA force field and examined root mean square deviation (RMSD), radius of gyration (Rg) and root mean square fluctuation (RMSF) parameters for each complex (Supplementary Figure S7). Both systems reached equilibrium after approximately 50–80 ns, but the PSMANb9-PSMA complex exhibited lower RMSD values (around 0.2–0.4 nm) compared to A7-PSMA (0.6–0.8 nm), indicating a more stable overall conformation (Supplementary Figure S7A). The radius of gyration (Rg), which reflects the overall molecular size and compactness, remained nearly constant for both complexes throughout the simulation. The slightly lower Rg observed for PSMANb9-PSMA suggests a more compact and stable conformation compared to A7-PSMA (Supplementary Figure S7B). Additionally, analyzing the fluctuations of backbone atoms in each residue showed slight fluctuations for the majority of residues, represented by RMSF values plotted against residue number (Supplementary Figure S7C,D). Residue-based RMSF analysis further revealed that PSMANb9-PSMA showed reduced local flexibility, particularly at the interface regions, supporting the observation of its enhanced structural stability. Together, these results suggest that the PSMANb9 forms a more stable and compact interaction with PSMA than A7.

3.10. PSMA Docking Predictions for Binding of Other Selected Nanobodies from the Top 52 Candidates

We tested whether there might be more than two predicted epitopes on PSMA when we expanded the modeling using the other non-tested 52 PSMA nanobodies identified earlier (Figure 1B). A set of an additional 21 nanobodies with very different CDR3 sequences were selected for Robetta folding and HADDOCK docking. All of these nanobodies were predicted to bind to epitopes bound by A7 or PSMANb9 and no additional epitopes were identified (Supplementary Figure S8).

3.11. A7 and PSMANb9 Nanobodies Do Not Compete for PSMA Binding

To validate the prediction of distinct epitopes on PSMA, we assessed the competitive binding of A7 and PSMANb9 nanobodies to LNCaP and B16-PSMA cell lines using a combination of nanobody proteins and nanobody-phages in a flow cytometry assay. The optimal concentration of the detector ATTO488-conjugated A7 and PSMANb9 nanobody-phages was determined using a concentration series in flow cytometry to LNCaP binding (Supplementary Figure S9). Concentrations of 5 × 10⁹ colony-forming units (CFU) and 1 × 10¹⁰ CFU for ATTO488-labeled A7 and PSMANb9 nanobody-phages were selected, respectively. As unlabeled competitors, PSMANb9 and A7 His-tagged proteins were produced (Supplementary Figure S10). Based on the titration assay with increasing concentration of either A7 or PSMANb9 proteins, we selected the 1 mg/mL concentration to saturate cell-surface PSMA proteins (Supplementary Figure S11).

After blocking cell-surface PSMA with A7 purified monovalent proteins in either LNCaP or B16-PSMA cells, the binding intensity of ATTO488-labeled A7 nanobody-phages dramatically decreased (Figure 6). This was similar for the PSMANb9 combination, indicating competition between the nanobody protein and corresponding nanobody-phage for the same epitope. In contrast, when cells were blocked either with A7 protein before ATTO488-labeled PSMANb9-phage or PSMANb9 protein before ATTO488 labeled A7-phage, no significant change was observed in the ATTO488 fluorescent signal intensity. This observation revealed that A7 and PSMANb9 bind to two distinct epitopes on the extracellular domain of PSMA, confirming the in silico prediction of two different epitopes of the PSMA protein (Figure 5).

4. Discussion

4.1. Panning Against Collections of Cell Lines Allows for Identification of PSMA Nanobodies

In the present study, we questioned whether non-targeted immunization using tumor cells and non-targeted panning using a large panel of cell lines with diverse protein expression profiles could identify nanobodies targeting PSMA. This approach differs from classical targeted protein-based screening methods. Those typically rely on recombinant antigens with fixed sequences and conformations that may not reflect how cell-surface proteins are naturally presented in a physiological membrane. As a result, epitopes that rely on post-translational modifications (e.g., glycosylation), protein folding, complex formation or isoform variation might be missed. Current understanding of which target modifications are cancer-associated remains incomplete, limiting the efficacy of targeted strategies. In contrast, the cell-based non-targeted panning strategy used in this study enables nanobody screening directly against antigens on natural (cancer) cell surfaces, thereby preserving their native conformation in a physiological molecular environment.

Comparing nanobody binding to 4 PSMApos cell lines with 8 or 13 PSMAneg cell lines, allowed selection of 498 nanobody clusters preferentially binding PSMA-positive cell lines. From these, 10% (n = 52) were also preferentially binding the B16-PSMA cell line over B16-WT, the targeted approach for selecting PSMA nanobodies (Figure 1C). The majority of the nanobodies from the panning of the cell line panel do not seem to bind PSMA, which could be explained by potential binding to cell-surface proteins co-expressed with PSMA such as STEAP1, PSCA, SLC44A4 and SLC39A6 [32]. In addition, any single-round panning is expected to have a high number of non-specific binders that will not be fully eliminated when comparing two groups with limited numbers of cell lines.

The majority of the 52 nanobodies preferentially binding PSMA-positive cell lines were derived from the most diverse nanobody library, L1P4 (Figure 1C). A second explanation for its dominance in PSMA binders is that L1P4 was generated from immunization with an intact and pure mix of AR-positive PCa cell lines of which three have high cell-surface PSMA expression. In contrast, LUPCa1 and 2 were generated from immunization with cell fractions from PCa and BlCa patient tumors containing many more different cell types than purely (PSMA positive) cancer cells. Although the mixing of the libraries for panning is dominated by nanobody clusters from L1P4, novel PSMA-binding nanobodies were identified from LUPCa1 (PSMANb9) and LUPCa2 (PSMANb6).

4.2. Multiple-Rounds Negative–Positive Panning as Compared to Single-Round Panning

Whether the classical three-round negative–positive panning (3R-NegPos) outperforms the one-round single-cell pannings (1R-SC) was tested by comparing 3R-NegPos and 1R-SC using the B16-WT and B16-PSMA cell lines. Only six nanobody clusters overlapped between the 462 and 341 clusters (Figure 1A).

Multiple factors underlie this limited overlap. First, nanobody clusters enriched after 1R-SC were more prevalent in the original library, with an average count of 1145, compared to just 164 for the 462 3R-NegPos clusters. Second, repeated panning rounds likely also amplify nanobodies’ binding to minor surface protein differences between B16-PSMA and B16-WT, other than PSMA. Third, every additional panning round introduces a bias in clone amplification and loss of diversity [33]. Fourth, the sequencing depth, PCR amplification biases and nanobody selection criteria are expected to affect the differences identified [34]. The strong enrichment for six overlapping nanobody clusters during the 3R-NegPos panning (up to 29.9%) resulted in their identification during conventional colony-picking [12]. This would be less likely using 1R-SC, with only up to 0.41% enrichment of the six clusters.

4.3. Panning Against Collections of Cell Lines Provides a Means of Selection with Restrictions

The use of a panel of cell lines with diverse protein expression, such as a strong distinction in PSMA expression, allows for the selection of PSMA nanobodies in non-targeted single-round panning. This approach comes with some limitations. First, endogenous cell line protein expression infections can result in differential nanobody binding. The xenotropic murine leukemia virus-related virus (XMRV) [35] is present in cell lines created through mouse engrafting. Except for LNCaP and MSKPCa12D, these unfortunately include all PSMA-positive PCa cell lines (PC346C, MDAPCa2b, 22Rv1, and VCaP) and one PSMA-negative line (PC339C). Based on the binding of the XMRV-recognizing nanobodies, all other cell lines in the panel are XMRV-negative except for the mouse B16-WT and B16-PSMA. Their binding potentially reflects infection by the closely related MelARV or Bxv-1 viruses [36]. During the 3R-NegPos B16-WT and B16-PSMA panning, the XMRV-binding nanobodies were selected out, which was not the case in the 1R-SC cell line pannings for which we corrected the selections, as described in Section 2. The inventory and impact of viral infections in cell cultures are understudied and will distort the selection of cancer nanobodies if not taken into account [37]. Second, to determine PSMA positivity, RNA expression levels were used to separate cell lines into two groups. However, in VCaP cells, high FOLH1 mRNA and PSMA protein (by Western blotting) [38] did not correlate with strong antibody or nanobody binding, suggesting limited PSMA surface exposure. Through inspection of RNA sequencing data, it became clear that this discrepancy is not explained by PSMA splice variants or mutations in VCaP cells. It does highlight that RNA levels do not always reliably represent surface protein abundance [39]. Third, to confidently identify nanobodies, the protein of interest must show distinct expression across a large group of high- and low-expressing cell lines. Within our selection, using 4 PSMApos cell lines and 8 or 13 PSMAneg cell lines, approximately 10% of the 498 nanobody clusters are PSMA binders based on overlay with the B16-PSMA-targeted panning, indicating the limitations in PSMA-specific enrichment.

4.4. Validation and Prediction of Distinct PSMA-Binding Epitopes and Utilization of Docking Software

Of the 12 selected nanobodies tested by cell-ELISA, six did not significantly reproduce the panning data, of which three nanobodies were significant in only one of the two comparisons. This validation of 50% is based on stringent selection criteria and party reflects the comparison of two very different assays: library panning with NGS as compared to a cell-ELISA. In addition, both the ELISA and IHC validation tests were performed using the easy-to-use nanobody-phage format, which resulted in higher background as compared to purified antibodies or nanobodies and could explain some of the variable reproducibility. Computational modeling and docking simulations using the validated nanobodies predicted two distinct PSMA epitopes, which were verified through epitope competition assays between nanobodies A7 and PSMANb9. While homology modeling using Robetta and SWISS-MODEL provided comparable structural predictions, accurate modeling of the diverse CDR3 loop remains challenging due to lack of sufficient sequence homology [40]. Although Robetta uses ab initio modeling for loop regions, no significant difference was observed between quality indicators (Supplementary Table S2). Leveraging deep learning tools like H3-OPT [41] and DeepH3 [41] offers promising alternatives.

HADDOCK-guided docking predicted that the majority of nanobody interactions occur via CDR3, with no evidence of a third PSMA epitope among 27 candidate nanobodies. However, the exact interacting residues should be further verified through standard experimental methods [42,43]. HADDOCK-based docking and molecular dynamics (MD) simulations are valuable computational tools for modeling nanobody–antigen complexes, however, several limitations constrain their predictive accuracy. The performance of HADDOCK depends extensively on the quality of experimental constraints and input structures, which might not adequately represent the conformational flexibility of antigen epitopes or nanobody CDR loops [44,45]. Solvation and electrostatic effects may be underestimated by rigid-body assumptions and simplified scoring functions [23,46]. Similarly, MD simulations are limited by sampling efficiency and computational costs, restricting their ability to explore large conformational transitions [47,48]. Therefore, these computational predictions should be considered qualitative and should, ideally, be validated through experimental approaches such as mutagenesis, X-ray crystallography or SPR [49].

4.5. Future Perspectives

The observation that A7 and PSMANb9 nanobodies bind different epitopes allows for the development of novel biparatopic tracers to increase affinity and internalization due to receptor surface clustering [50]. Such molecules could further expand the exploitation of PSMA in diagnostics, PET/SPECT, intraoperative imaging and therapeutics [51].

Cellular immunization and non-targeted panning against a panel of cell lines could be the basis of identification of cell-surface targets. The classification of cell lines into groups with and without PSMA expression allowed for the selection of nanobodies to potentially bind with this PCa target. At the moment, such a selection alone is not powerful enough to give the almost-certainty of PSMA binding and a targeted approach using B16-WT and B16-PSMA panning is needed to enrich PSMA binders. While PSMA served as an example in this study, the non-targeted method can, theoretically, be applied to other cell-surface antigens as long as the target expression is known in the cell line panel. However, variables such as target abundance, cell-surface accessibility, and level of heterogeneous expression for suitable positive/negative screening may limit the breadth of application for this approach.

Access to an expanding repertoire of nanobody sequences together with their corresponding target proteins facilitates advances in nanobody structural modeling and docking studies [21,22]. The next steps towards AI-driven target prediction and de novo nanobody design are being taken and will strongly benefit from large panning databases [23,24].

5. Conclusions

This study established a non-targeted llama immunization approach using tumor cells, combined with a non-targeted panning method using a panel of (PSMA pos/neg) cell lines as a novel method for identifying PSMA-binding nanobodies. This approach broadens the available discovery toolbox for identifying nanobodies targeting subtypes of (cancer) cells and developing target-specific nanobodies. However, it is still dependent on targeted enrichment strategies to enhance selection efficiency comparable to the traditional protein-based selection methods. Both classical PSMA-directed panning and single-round cell-based selection resulted in substantial enrichment of PSMA-binding clones. Selected nanobodies A7 (JVZ-007) and PSMANb9 were confirmed as top PSMA binders using multiple complementary assays. Epitope analyses revealed that they target distinct epitopes on PSMA. These findings substantiate the utility of non-targeted immunization strategies for generating antigen-binding nanobodies for radioligand imaging and targeted therapeutics.