The emergence of cell-based protein arrays to test for polyspecific off-target binding of antibody therapeutics

ABSTRACT

Specificity profiling is a requirement for monoclonal antibodies (mAbs) and antibody-directed biotherapeutics such as CAR-T cells prior to initiating human trials. However, traditional approaches to assess the specificity of mAbs, primarily tissue cross-reactivity studies, have been unreliable, leading to off-target binding going undetected. Here, we review the emergence of cell-based protein arrays as an alternative and improved assessment of mAb specificity. Cell-based protein arrays assess binding across the full human membrane proteome, ~6,000 membrane proteins each individually expressed in their native structural configuration within live or unfixed cells. Our own profiling indicates a surprisingly high off-target rate across the industry, with 33% of lead candidates displaying off-target binding. Moreover, about 20% of therapeutic mAbs in clinical development and currently on the market display off-target binding. Case studies and off-target rates at different phases of biotherapeutic drug approval suggest that off-target binding is likely a major cause of adverse events and drug attrition.

Introduction

The development of monoclonal antibodies (mAbs) has opened new avenues for precision medicine, offering targeted and effective treatment options with reduced side effects. Over 250 mAbs entered clinical trials in 2022,¹ and this number continues to increase due to the high demand of biologics and the advancement of new antibody-directed modalities such as chimeric antigen receptors (CARs).²

MAbs are usually isolated using a target antigen and so have been presumed to recognize only that target. As such, off-target binding has been an understudied aspect of the antibody discovery process. In this review, we discuss off-target cross-reactivity due to “polyspecific” binding, also described as complementarity-determining region (CDR)-specific off-target binding. This is in contrast to “polyreactivity”, i.e., ‘sticky’ antibodies that bind to undefined proteins,³ which is a well-known developability risk.^4,5 Polyspecific off-target binding has the potential to cause unintended cellular toxicity, adverse events, and clinical trial failures. Emerging case studies demonstrate that a number of mAbs have had developability and safety issues as a result of polyspecific off-target binding,^6–11 and our own data suggest a surprisingly high polyspecificity rate for therapeutic mAbs currently in development and on the market. The presumption that every mAb confers absolute specificity is simply not accurate.

Testing for specificity is a critical part of the safety assessment of all antibody-based therapeutics, including mAbs, Fabs, single-chain variable fragments (scFvs), and VHH nanobodies. Moreover, absolute specificity is vital for therapeutics designed to kill the target cell, such as CAR-T cell therapies, antibody-drug conjugates (ADCs), and bispecifics. Current US Food and Drug Administration (FDA) and International Conference on Harmonization (ICH) guidance for biotherapeutic development^12,13 describe the use of in vitro tissue cross-reactivity (TCR) studies to assess specificity. However, most toxicologists no longer trust TCR to predict safety,¹⁴ and the TCR assay has not undergone major improvements since its introduction over 40 y ago.¹⁵

Recently, “cell-based protein arrays” have emerged as an alternative solution for evaluating specificity. FDA and ICH biotherapeutic guidance documents state that “other technologies can be employed in place of IHC techniques to demonstrate target/binding site distribution”¹² and that “appropriate newer technologies should be employed as they become available and validated”.¹³ In 2018, the FDA advised that TCR studies are not necessary for certain life-threatening oncology indications.¹⁶ In January 2024, the FDA issued a new guidance document for CAR-T cell therapeutic development¹⁷ recognizing that “unintended targeting of other antigens expressed on healthy/normal tissue is a safety concern that may be evaluated using in vitro and/or in vivo studies” and for the first time delineated specific alternative methods beyond TCR studies, notably including “protein arrays”. Here, we review the technology behind cell-based protein arrays, best practices for using them, and data gathered to date that suggest their ability to improve drug safety.

Cell-based protein arrays to evaluate specificity and predict safety

In 2001, we and others first created cell-based protein arrays by reverse-transfecting arrayed expression plasmids into eukaryotic cell lines.^18,19 The resulting array contains cells expressing thousands of proteins across the human proteome. Target binding is determined by high-throughput flow-cytometry on unfixed cells²⁰ (Figure 1) or immunofluorescence (IF) on fixed cells.²¹ Importantly, each mAb interaction is tested against each individual protein target, so the identity of the protein can be readily determined. There are currently two commercially available cell-based protein array platforms, the Membrane Proteome Array™ (Integral Molecular, Figure 1) and the Retrogenix® Human Cell Microarray (Charles River Labs).

Figure 1.

Specificity testing of MAbs using cell-based protein arrays. Each location in a cell-based protein array contains a plasmid clone encoding a defined protein, typically a membrane protein. Cells in each well express this protein in its native conformation. Once expressed, the entire array can be screened by flow cytometry on live or unfixed cells with a selected mAb to quantify its binding to each protein target. MAbs that are reactive with a target protein (histogram in blue) can clearly be distinguished from non-reactive proteins (histograms in gray). The data for all ~ 6,000 membrane protein binding interactions are plotted to clearly visualize on- and off-target binding across the membrane proteome.

Cell-based protein arrays differentiate from traditional protein arrays in that the targets remain expressed in eukaryotic cells (typically human cells), while they are evaluated for binding. In traditional spotted protein arrays, recombinant proteins are purified from cell lysates before they are spotted on a solid surface (slides or membranes) to test the binding of the molecule.^22–24 Proteins are typically expressed in bacteria or yeast for highest production, but platforms that use human HEK cells are also available. Traditional protein arrays are available that cover up to 80% of the human proteome and have been used as a research tool to profile auto-antibody targets.²⁵ However, this approach of spotting proteins can give rise to unnatural surface interactions and denaturation of proteins, and spotted protein arrays are particularly problematic for multi-pass membrane proteins that require a lipid bilayer to maintain their native structure. Cell-based protein arrays overcome the issue of protein misfolding by keeping each protein intact in its native conformation within the cell and so are much more relevant for testing for specificity.

Cell-based protein arrays have primarily focused on cell surface membrane proteins because of the role of these proteins in directing therapeutics to the correct cell type, as any off-target binding can lead to cellular toxicity, particularly for biotherapeutics designed to kill cancer cells. Binding to secreted proteins can also be tested but typically does not pose as significant a safety risk, and other assays to test binding and clearance in human plasma are commonly performed during preclinical development (e.g., plasma protein binding assays). Cell-based protein arrays can be used to test the specificity of all antibody-based biotherapeutic modalities, including IgG, IgM, scFv, VH/VHH nanobodies, ADCs, and bispecifics, as well as other protein modalities. CAR-based cell therapies or other non-IgG formats can also be evaluated using the targeting moiety in an Fc fusion or epitope-tagged format, as indicated by FDA guidance allowing testing of the “antigen recognition domain used to confer target specificity”.¹⁷

Best practices in using cell-based protein arrays

There are a number of factors to consider when using cell-based protein array technologies to ensure reliable results acceptable to regulatory agencies (summarized in Table 1).

Table 1.

Best practices in cell-based protein arrays.

Criteria	Best Practice
Library Composition	Comprehensive library of unique membrane proteins, heterocomplexes, GPI-linked proteins, viral membrane proteins
Expression Validation	Epitope tags to confirm full-length expression of each protein in the array
Native Screening	Live or unfixed cells during screening to preserve the native conformation of each protein
Cell Type(s)	Human cell types (e.g., HEK-293) for native expression, avian or other cell types to reduce background if needed
MAb Concentration	20 μg/mL, or highest possible based on each mAb’s background, to detect all possible off-targets with the highest sensitivity
Isotype Controls	Isotype-matched molecule containing identical Fc and constant domains, produced using same methodology and from same cell type as test molecule
Quantitative Analysis	Quantitative data that enables statistical analysis of hits and thresholds to meet regulatory rigor requirements

Library composition and validation

The FDA’s Points to Consider in the Manufacturing and Testing of Monoclonal Antibody Products for Human Use¹³ provides a guide for inclusion of targets that need to be assessed for off-target binding. This document identifies 34 normal adult human tissues, covering the major organs of the human body, including the gastrointestinal, circulatory, musculoskeletal, endocrine, respiratory, reproductive, immune, and nervous systems. Olfactory tissues are not included in the panel recommended by the FDA. From the named tissues, bulk RNA sequencing data and available topology servers,²⁶ as well as databases such as UniProt, identify the membrane proteins expressed within each tissue, which is the basis for the libraries used for cell-based protein arrays. The canonical isoform of the gene coding for each protein should be used because the canonical isoform usually codes for the full target protein, while isoforms most often delete introns. In addition to monomeric proteins, some membrane proteins only express or fold properly as part of an obligate heteromer, and thus heterocomplexes should also be included. Glycosylphosphatidylinositol (GPI)-linked proteins are expressed on the cell surface and so should also be included even though they lack a traditional transmembrane (TM) domain. Viral envelope proteins can also be useful because of their importance in human health and biotherapeutic development. While membrane proteins expressed in adult human tissues are the primary targets of interest, some membrane proteins are expressed only during development (placental, fetal, neonatal, or juvenile stages). The inclusion of placental and fetal proteins can provide a basis for testing these proteins without the need for procuring fetal tissues. To ensure consistent results and avoid false-negatives, all target proteins should have some kind of measure of expression, such as an epitope tag or co-expressed reporter protein.

Screening unfixed cells that retain protein conformation

MAbs should be tested throughout the study, particularly during screening, on live or unfixed cells. Fixation often alters the conformation of multi-spanning membrane proteins and the entire point of screening on cell-based protein arrays is to retain the native conformation of the protein as it exists within humans. As an example, an antibody against CD19 (a common CAR-T target) was tested on our own Membrane Proteome Array under fixed and non-fixed conditions and identified an off-target binding interaction only in the native unfixed condition (Figure 2). Binding to this off-target protein was confirmed in validation studies, demonstrating that fixation of cells can significantly mask epitopes and cause important off-target binding to be missed during screening.

Figure 2.

Cell fixation can mask epitopes. An anti-CD19 MAb was screened on Integral Molecular’s Membrane Proteome Array (MPA) using both unfixed and fixed cells in parallel. Both conditions identified the target, but the fixed cells failed to identify an important off-target interaction. Figure adapted from.²⁷

Low background cell types

Human HEK-293 cells are often preferred in cell-based protein arrays because of their ease of transfection, high expression levels, and often low background reactivity. The use of human cells ensures that each protein folds in its native conformation and provides native post-translational modifications, glycosylation, native partner proteins, and multimerization (if applicable).²⁸ Alternative cell types, such as avian QT6 cells, can be used to avoid any background reactivity problems caused by target proteins endogenously expressed in HEK-293 cells.

Assay sensitivity, isotype controls, and quantitative analysis

Screens should be designed to be as sensitive as possible to detect potential off-targets, using high concentrations of the test antibody (typically 2–20 μg/mL), over-expressed target proteins, unfixed cells, and low thresholds of detection (to identify even weak off-targets). Validation of initial hits serves as an integral verification step to identify real interactions and eliminate any screen artifacts or false-positives. An isotype control that is precisely matched to the same format and production process as the test molecule can be used during validation to differentiate protein interactions mediated by Fc or carbohydrates (i.e., endogenous Fc receptors and lectins). Such binding may or may not affect safety, but there are well-characterized strategies to mitigate such binding.²⁹ Finally, data from cell-based protein arrays should be quantitative in nature to allow for rigorous statistical interpretation and analysis that meet regulatory requirements (e.g., three standard deviations (SD) above average reactivity across the array).

If best practices are followed, the overall result from using a cell-based protein array should be a refined list of targets to which the test antibody binds. Binding data across every protein in the array should be shown and quantified to avoid misinterpretation and meet regulatory expectations. Since their inception, subsequent versions of cell-based protein array platforms have undergone improvements in library quality, screening conditions, and detection mechanisms in an effort to reduce false-positive and false-negative results. From an in vitro screening perspective, false-negatives that could cause adverse events in patients are far more important to avoid than false-positives, which are quickly eliminated using validation assays or by adjusting thresholds of detection.

Off-target risk assessment to interpret cell-based protein array results

Cell-based protein arrays can identify specific off-target proteins that allow for a focused investigation into potential safety liabilities. Such investigations are particularly important when cell-based protein array data are used for investigational new drug (IND) submissions, as off-target binding does not always lead to adverse events. A thorough off-target risk analysis can determine whether a lead molecule warrants reconsideration or if further progression through development and IND is still appropriate. Relevant considerations for off-target binding include statistical confirmation of off-target binding, the relative strength of the off-target interaction, the epitope location and accessibility of the off-target, and therapeutic modality (summarized in Figure 3).

Figure 3.

Risk assessment for off-targets identified from a cell-based protein array study. Evaluation of risk associated with an off-target should consider target binding (relative affinity, epitope location), accessibility (cellular and tissue), and potential for therapeutic toxicity (MOA, agonism, dosing).

Relative binding strength

Cell-based protein arrays are designed to be screened at the highest sensitivity levels to reveal any low-level reactivity with off-targets. Low, but real, reactivity can be caused by low affinity of the interaction or low expression levels of the target, and rigorous statistical assessment of off-target binding compared to positive and negative controls can differentiate real interactions from typical variation. Antibody titration studies using flow cytometry or biosensor measurements can differentiate the underlying strength of binding interaction by measuring Bmax, EC₅₀, and K_D values that provide insight into whether off-target engagement is meaningful relative to binding of the actual target. For example, low-affinity binding to an off-target known to be expressed at low levels in tissues may have minimal impact on the therapeutic window of the drug compared to a high affinity off-target interaction. The measurement of relative binding strength also complies with FDA guidance to measure “affinity/avidity for the target antigen/cells to evaluate the potential for on-target/off-tumor and off-target toxicities”.¹⁷

Accessibility of epitope

Accessibility of the off-target binding site within cells and tissues is also a critical consideration. Cell-based protein arrays are performed on permeabilized cells to maximize sensitivity, so intracellular binding interactions are routinely detected. Many cellular membrane proteins are normally expressed intracellularly, on the nucleus, Golgi, endoplasmic reticulum, and other intracellular membranous structures. Even if the off-target protein is expressed on the cell surface, the binding epitope may be intracellular, particularly since many membrane proteins have large intracellular domains. Flow cytometry can be used to determine epitope location by comparing binding under non-permeabilized vs. permeabilized conditions. Intracellular binding location can also be visualized by high-resolution immunofluorescence. Off-target proteins may also be inaccessible depending on their location within human tissues. For example, proteins solely expressed in the brain are not usually accessible to biologics. Tissue expression profiles available through online databases (e.g., Human Protein Atlas, GTEx, Expression Atlas, Bgee) can be used to identify which tissues express the known off-target to interpret its accessibility in vivo.

Therapeutic modality

Finally, any off-target binding must be analyzed in the context of therapeutic mechanism of action (MOA), drug dosing, and target biology. A therapeutic that does not function by killing a cell likely has a lower risk than a CAR-T, ADC, or bispecific designed to kill a cell that happens to express an off-target protein. Similarly, antagonist or agonist mAbs that are designed to carefully modulate the function of a target protein are not likely to similarly alter the function of an off-target protein. In addition, a therapeutic given at a very low dose may have minimal effect on an off-target, depending on the relative binding to the off-target and its localization. Additional preclinical safety evaluations, such as functional studies using primary cells or cell lines known to express the off-target, may be necessary to understand the complete risk of off-target cell killing.^30,31

Prevalence of off-target binding

Off-target related issues have traditionally been the single largest cause of failed preclinical drug programs. A retrospective analysis of the AstraZeneca pipeline determined that 62% of all of their preclinical failures were caused by off-target binding that caused safety problems.³² However, most of their pipeline at the time focused on small molecules, and a similar study of biologics has not yet been published. To estimate the frequency of off-target binding by candidate antibody therapeutics, we conducted a retrospective analysis of the last ~250 samples that completed a Membrane Proteome Array (MPA) study at Integral Molecular. These samples primarily represent lead candidates at biopharmaceutical companies throughout the industry. A total of 83 out of 254 evaluated samples (32.7%) displayed polyspecific off-target binding (Figure 4a). Of the polyspecific antibodies tested, about half had a single off-target, while the other half had two or more off-targets (Figure 4b). Off-target binding in each case was detected during MPA screening of the entire library of ~6,000 native membrane proteins, as well as in follow-up validation studies, so the 33% reflects true off-target binding interactions with high confidence. These results do not include common non-CDR-mediated interactions that are tracked separately, such as binding to Fc receptors or lectins, which would make the off-target rate even higher if counted. Notably, off-target binding was almost always to completely unpredictable membrane proteins that demonstrate no significant sequence homology to the intended target (most mAbs screened on the MPA have already been tested against related family members). This result suggests that polyspecificity is more common than previously thought and highlights the need for comprehensive specificity screening outside of related family members. These data also suggest that cell-based protein arrays are especially useful during the lead selection phase of development, when candidates with potential toxicity issues can be identified and rectified with relatively little impact on a drug program. For example, we tested two lead candidate mAbs that bound to the same GFRα4 target, but one of the mAbs demonstrated polyspecificity with strong binding to a completely unrelated off-target protein.³³ When biodistribution studies in mice were conducted on these same mAbs, both localized to the thyroid (the location of the primary target), but the polyspecific mAb also localized to another organ (where the off-target protein is expressed), thereby prioritizing the highly specific mAb for clinical development.

Figure 4.

Off-target reactivity of antibodies in lead selection stage. A. Specificity testing of 254 antibodies on Integral Molecular’s Membrane Proteome Array found that 83 (32.7%) mAbs tested demonstrate validated off-target binding. Validated binding to an off-target is confirmed after screening by comparing mean fluorescent intensity (MFI) binding signal to the off-target vs that of the negative control (mock transfected cells), each run in quadruplicate by flow cytometry on unfixed cells. B. The number of off-targets bound by the 83 mAbs is shown.

As our data suggests that polyspecificity is more common than previously appreciated among lead candidate MAbs, we also wanted to determine whether polyspecificity is prevalent among mAbs that have gone into humans. To test this, we produced biosimilars of clinical-stage, FDA-approved, and withdrawn mAbs³⁴ and screened them on the MPA. In this dataset, we screened a total of 83 biosimilars and found polyspecificity among mAbs at all stages of clinical development. We identified off-target interactions for 18.1% of the mAbs overall. Compared to the 32.7% polyspecificity rate of lead candidate mAbs, this suggests that off-target binding is a significant risk contributing to attrition during the drug development process. Interestingly, when we further categorized the mAbs by their clinical status, we found a slightly higher off-target rate for mAbs that were withdrawn (22.2%) or in Phase 2 or Phase 3 (20.0%) compared to that of approved MAbs (15.0%) (Table 2). While more studies are needed, the data indicate that mAbs still in clinical development or withdrawn may have a higher rate of polyspecificity compared to approved mAbs, which suggests that off-target binding may be a significant cause of failure during drug development.

Table 2.

Rate of off-target binding in clinical mAbs. Screening of 83 clinical mAb biosimilars suggests that off-target binding may be a significant risk that contributes to the withdrawal of marketed therapeutics and their failure during clinical trials.

	Clinical mAbs screened on the MPA
	All (83)	Approved (40)	Phase2/3 (25)	Withdrawn (18)
Clean	68	34	20	14
Off-target	15 (18.1%)	6 (15.0%)	5 (20.0%)	4 (22.2%)

In screening these clinical mAbs, we identified several cases where off-target binding may explain adverse events. For example, one mAb currently in clinical trials demonstrated off-target binding to a widely expressed protein that would be predicted to result in adverse events due to mAb-mediated toxicity (the mAb is armed with a toxic payload) (Figure 5a). Another mAb was withdrawn from clinical trials due to severe patient adverse events and demonstrated off-target binding to a completely unrelated membrane protein even better than it bound its intended target (Figure 5b). In another example, we tested an FDA-approved mAb that showed significant off-target reactivity to an unrelated protein (Figure 5c). This biotherapeutic has very significant side effects (much greater than other clinical mAbs that bind to the same target), suggesting that this off-target binding could potentially explain the drug’s safety profile in patients. Overall, these data show that polyspecificity is currently found in mAbs at all stages of clinical development and even in FDA-approved mAbs on the market today. Additional studies are ongoing to understand the off-target effects of these mAbs and of other mAbs currently on the market or in clinical trials.

Figure 5.

Off-targets identified in late-stage therapeutic mAbs. Integral Molecular’s MPA was used to test the specificity of a) a clinical-stage mAb intended to kill cancer cells, b) a mAb withdrawn after clinical trials, and c) an FDA approved mAb currently on the market. The correct target was identified for all mAbs, but all three mAbs also showed off-target binding that may explain their adverse events and reasons for withdrawal. The off-target protein identified for each are unrelated membrane proteins (i.e., not to immediate protein family members, FcRs, or lectins). Figure adapted from (27).

Mechanisms of polyspecificity

Most off-target proteins that we identify are unrelated to the intended target based on the primary amino acid sequence. Three mechanisms causing off-target binding to unrelated proteins have been reported to date. The first of these is molecular mimicry, where a local structural region with just a few critical epitope residues is mirrored in a completely unrelated protein. We identified such a scenario in a previous study of a panel of antibodies against SLC2A4 (GLUT4), a 12-TM insulin sensitive glucose transporter.²⁰ We found one mAb that demonstrated low-level reactivity to the unrelated signaling protein Notch1. This was unexpected, as SLC2A4 and Notch1 are completely unrelated structurally (12-TM vs 1-TM) and share less than 7% sequence identity. High-resolution shotgun mutagenesis epitope mapping localized the mAb’s epitope residues to a loop-constrained ‘LGXXGP’ sequence motif on SLC2A4. Despite being completely unrelated in sequence and structure, the exact same LGXXGP motif was shared on a disulfide-constrained loop of Notch1 (Figure 6), explaining how this mAb could bind to both proteins.

Figure 6.

Epitope mimicry explains the unexpected cross-reactivity of a mAb against SLC2A4 (GLUT4) with Notch1. The MPA identified cross-reactivity of an SLC2A4 mAb with the unrelated signaling protein Notch1. Shotgun mutagenesis epitope mapping studies localized the binding site on SLC2A4 to critical residues L61, G65, and P66.²⁰ The epitope structure is visualized on SLC2A4 (left), and the exact same sequence motif is found on a disulfide-constrained loop in the EGF-like domain of Notch1 (right).

A second mechanism underlying polyspecificity is CDR plasticity, a phenomenon in which restructuring of the antibody paratope enables the antibody CDRs to adapt to more than one antigen. This mechanism requires considerable CDR conformational flexibility but has been documented.⁶ Finally, a third mechanism of polyspecificity can occur through differential engagement of the VL and VH CDRs. In such cases, a single antibody can have multiple (potentially overlapping) functional paratopes, each specific for a different target. This mechanism also does not appear to be very common but has been documented.³⁵ Overall, all these mechanisms of polyspecificity involve binding mediated by either single amino acids or small conformational changes of the paratope or target protein. Therefore, it is criticalto evaluate specificity under native conditions (on a cell membrane, without fixatives or drying of tissues/cells) to avoid any alterations to secondary structures that can change these interactions.

In our own antibody engineering experience, we have observed off-targets appear (or disappear) with even single amino acid changes, for example with our CLDN6 antibodies now in clinical trials.³⁶ As lead candidate antibodies typically undergo a number of modifications during the development process, such as humanization, affinity maturation, and developability engineering, it is important to test the final candidate selected for clinical studies.

Traditional assays used to evaluate specificity: tissue cross-reactivity studies

TCR studies were first recommended for specificity screens by the FDA in 1983 guidance documents, and their use has subsequently been updated in the most recent document versions.^12,13 TCR studies involve screening >30 human tissues from three different donors to identify unexpected binding, whether it be off-target binding or unknown sites of on-target binding. Binding is determined by immunohistochemical (IHC) staining, and staining patterns are evaluated by a licensed pathologist (reviewed in detail in Leach, et al.¹⁵). Despite its broad use, TCR studies to assess antigen specificity have numerous and significant limitations.

The primary limitation of TCR studies is that the molecular target of binding is never known; specificity issues may be identified, but the specific cause cannot be determined. Moreover, on-target tissue reactivity may provide false reassurance if an off-target protein is expressed within the same tissue or the off-target protein is expressed at low levels. Natural tissue sources display high variability for endogenously expressed and disease-relevant proteins. Most tissue samples used in TCR are fixed (using various fixatives) or frozen (e.g., snap frozen, desiccated) and placed onto glass slides, potentially altering the conformation and native structure of proteins (especially membrane proteins that are highly sensitive to fixatives and staining conditions).^37–39 Finally, TCR results are scored by specialized pathologists, and so are based on subjective observations and interpretations, which are not statistically quantitative and can lead to precision issues.^40,41

Several reviews^5,14,42 and case studies^15,43 illustrate the limitations of using TCR studies for predicting in vivo toxicity and safety. A detailed review on TCR by Leach et al. presents case studies across the industry that show examples of the uses, interpretations, and limitations of TCR studies.¹⁵ Examples are presented where antibodies demonstrated in vivo toxicity even though the TCR study showed only expected staining patterns. In contrast, other antibodies with clear off-target TCR tissue staining patterns demonstrated no in vivo toxicity. Strikingly, surveys of toxicologists at pharmaceutical and biotechnology companies now report that TCR results do not influence the development strategy of 90–95% of biologics.¹⁴ This suggests that the vast majority of companies believe that TCR results are not predictive of in vivo toxicity and are not in actuality used anymore to make any critical decisions about safety.^14,42

Similar to TCR studies, tissue microarray assays (TMAs) are sometimes used to understand target tissue expression profiles. TMAs can consist of up to 1,000 tissue samples on a slide, and binding is determined by IHC. However, TMAs also do not identify the molecular target of binding and are not recommended by the FDA for IND filings because they represent very limited tissue cross sections (<2 mm) that do not faithfully represent all elements of the originating tissue.¹⁵

Advantages of cell-based protein arrays

Compared to TCR studies, cell-based protein arrays provide numerous advantages for specificity testing (summarized in Figure 7).

Figure 7.

Comparison of cell-based protein arrays to TCR studies. Cell-based protein arrays and TCR studies differ in the methods, conditions, and quality of results generated.

Target identification

Cell-based protein arrays determine the identity of any cross-reactive, off-target protein, allowing for further investigation into potential safety liabilities. The tissue expression pattern of the entire human proteome has been extensively characterized and published in multiple online databases (e.g., Human Protein Atlas, GTEx, Expression Atlas, Bgee) so that the tissues and cell lines that express any off-targets can be readily identified for subsequent safety investigations. This is in stark contrast to TCR studies where the tissue of binding is identified, but it is difficult or impossible to confirm which specific protein is mediating the positive staining.

Native protein expression

Human proteins within cell-based protein arrays are individually expressed and screened in their native state directly within human cells. This allows proteins to retain their native structural conformation and post-translational modifications. This is especially important for membrane proteins, where the lipid membrane is often critical for native protein folding. Ideally, screens are performed on live or unfixed cells, whereas TCR studies can use fixatives and preservation protocols that alter native protein conformations. IHC on tissue samples is also prone to high background and false positives due to native Fc receptors and endogenous IgG found throughout human tissues.¹⁵

Comprehensive membrane proteome representation

Cell-based protein arrays allow for the identification of specificity across the entire human membrane proteome. Each protein is individually expressed at high levels for maximum detection sensitivity. In contrast, natural tissues and primary cells express variable levels of each protein, with some proteins not expressed at all and others expressed in disease-dependent ways.

Quantitative analysis

Cell-based protein array data obtained by flow cytometry are highly quantitative, allowing for precise thresholds to be determined and target binding values to be statistically compared during analysis to meet the expectations of regulatory agencies. In comparison, IHC or IF staining is scored using qualitative interpretations that are relatively subjective and provide little if any rigorous quantitation, prohibiting statistical analysis or quantified binding thresholds which negatively impact reproducibility.⁴⁴

Compatibility with biologic formats

Specificity screening for IND-enabling studies should ideally be performed on the lead molecule in the same final format as will be used in clinical trials. However, TCR assays are not always compatible with nonstandard antibody formats or CAR-T therapeutics, and low-affinity interactions are difficult to identify by TCR. Often times, a robust IHC assay cannot even be developed.^15,43 Due to the high sensitivity of flow cytometry, cell-based protein arrays are suitable for screening of any format, including IgG, scFv, VH/VHH, bispecific, ADC, and CAR-T. Bispecific antibodies are ideally tested in the final bispecific format, but each arm can also be evaluated individually. Payloads and linkers also have the potential to effect binding, and therefore it is ideal to evaluate the complete extracellular portion of targeting molecules for formats such as ADCs and CAR modalities. For CAR-based therapeutics, the final antigen-binding domain (e.g., scFv-ectodomain-Fc or scFv-Fc fusion compound) can be screened rather than whole cells, as specified in FDA’s CAR-T development guidance document.¹⁷

Throughput and convenience

The high-throughput format of cell-based protein arrays also makes it fast to complete, with a timeline of as little as 4 weeks, compared to ~14 weeks for TCR studies, and multiple test articles can be evaluated simultaneously. TCR studies are also significantly more expensive than cell-based protein arrays, in part due to the need for a trained pathologist to manually score each tissue section.

Comparing cell-based protein array results to TCR

To directly compare cell-based protein arrays to TCR studies, we compared our own MPA data to TCR data obtained from biologics license applications (BLAs) found at drugs@fda (www.fda.gov/drugs/drug-approvals-and-databases/about-drugsfda). Although there is limited access to the raw data in these applications, the summary information is informative. In these case studies, FDA-approved mAbs were reproduced and then screened on the MPA (Figure 8). Discrepancies were found between our findings and the available TCR interpretations.

• In Case Study #1, the antibody targets a plasma membrane protein expressed on lymphocytes. The MPA correctly identified the known target but also identified two off-target membrane proteins that bind even better than the intended target. The TCR summary in the BLA application for this mAb indicated that the TCR staining was consistent with the known expression of the target on lymphocytes. The full TCR report is not publicly available, but the BLA summary suggests that no off-target binding was detected by TCR or further investigated.

• In Case Study #2, the antibody targets a membrane protein expressed both on the plasma membrane and intracellularly in myeloid cells. The MPA correctly identified the known target but also identified an off-target membrane protein. Importantly, the off-target is expressed on the same cell type as the target. The TCR summary in the BLA application for this mAb indicated that the TCR staining was consistent with the known expression of the target and did not mention any off-target staining. In this case, co-expression of the target and off-target protein on the same cell type demonstrates the limited ability of TCR to identify some off-target proteins.

• In Case Study #3, the antibody targets a membrane protein expressed on both the plasma membrane and intracellular vesicles. While the target is upregulated in certain disease states, it has low and widespread expression in most normal tissues. The MPA screen correctly identified the known target and identified an off-target membrane protein. The TCR summary in the BLA application for this mAb indicated primarily cytoplasmic staining in numerous tissues with some membrane staining, but there was no significant off-target suspicions noted. In this case, where the target has widespread expression at both plasma membrane and intracellular locations, identifying an off-target is more difficult using TCR where the tissues already show positive staining.

Figure 8.

Discrepancies between cell-based protein array results and TCR. FDA-approved mAbs were reproduced and tested for off-target binding on the MPA, and results were compared to available TCR summary data obtained from BLA applications. In several cases, the MPA identified off-targets that were not discussed or identified in the TCR studies. The antibodies tested here on the MPA were biosimilars produced from the published sequence of the antibody variable chains, and so are not identical to the therapeutics formulated by the original manufacturer. Although it is unlikely, discrepancies between manufacturing or Fc components could have an effect on the specificity profile. The dotted line represents 3 SD above the calculated background.

In these examples, any direct attribution of the off-target interaction to safety is not yet known; we have not yet determined if the off-targets identified on the MPA are responsible for any adverse events. However, other examples of mAb off-target binding directly causing toxicity in vivo are now being realized.^9–11 For example, the anti-PD1 antibody camrelizumab unexpectedly caused severe capillary hemangioma in Phase 1 clinical trials.⁴⁵ Retrospective profiling of camrelizumab using a cell-based protein array directly linked these adverse events to off-target binding against VEGFR2.¹¹ Camrelizumab not only bound VEGFR2 but also acted as an agonist to stimulate vascular neogenesis, which in vivo leads to hemangioma. In human trials, rescue therapy with a VEGFR2 antagonist ameliorated the hemangioma effects of camrelizumab, confirming the role of the off-target binding in causing the adverse effect in humans.⁴⁶ CDR mutagenesis was able to abolish binding to VEGFR2 while maintaining a high affinity to PD1, demonstrating that unwanted binding can be fully ameliorated by antibody engineering.¹¹ In another example in non-human primates (NHPs), development of the anti-beta-amyloid antibody ABT-736 was discontinued due to severe toxicity observed in cynomolgus monkeys caused by off-target binding to plasma protein platelet factor 4.⁹ These studies suggest that prior methods used to evaluate specificity were insufficient in identifying the off-target binding events that later caused safety issues in both NHPs and humans.

Conclusions

TCR studies have been the primary specificity assay used since the 1980s. In reality, however, off-target binding is going undetected in many cases, and surveys of toxicologists suggest that few still trust TCR data to predict safety or design in vivo toxicity studies.¹⁴ Instead, cell-based protein arrays are emerging to provide a more accurate, cost-effective, and faster approach to identify off-target binding.^5,14,27

In addition to improved specificity profiling, cell-based protein arrays can also be used to better design animal studies. By identifying on- and off-target proteins, species orthologs of those targets can be readily assessed for binding. The ability to identify safety issues in vitro using cell-based protein arrays has the potential to reduce the number of animals and NHPs currently used for toxicity studies.^43,47 In particular, for mAbs with exogenous targets (e.g., against viral proteins where on-target safety is not a concern), cell-based protein array data showing no off-targets could dramatically reduce the number of animals and NHPs needed for in vivo toxicity studies.

Cell-based protein arrays are designed based on regulatory guidance documents and are now routinely being used to support FDA, European Medicines Agency, and other regulatory submissions for assessing off-target binding that could cause toxicity. Both FDA and ICH biotherapeutic guidance documents allow the use of viable alternatives to TCR to test for off-target binding, with protein arrays now included in FDA’s CAR-T guidance document. In addition, regulatory agencies continue to support the qualification of cell-based protein array technology, such as through the FDA’s ISTAND validation program (MPA is designated Drug Development Tool DDTIST00006 and is certified with ISO9001 quality management systems). Standard inclusion of cell-based protein array data in preclinical studies would lead to improved specificity data, more efficient regulatory review, safer clinical trials, and ultimately more successful therapeutics.

Funding Statement

This work was supported by the National Institute of General Medicine at the National Institutes of Health [grant GM113556 to BJD].

Disclosure statement

Integral Molecular is a biotech company that offers the Membrane Proteome Array as a specificity screening service. D.M.N., C.N., J.T.S., and B.J.D. are current employees and/or shareholders of Integral Molecular.