Immunogenicity of Generic Peptide Impurities: Current Orthogonal Approaches

Abstract

Generic drugs have saved consumers billions of dollars in the United States. The demand for lower-cost and effective drugs, particularly for well-known peptide drugs like Ozempic and Wegovy (brand names for semaglutide), has resulted in a surge of generic drug development to address perceived shortages in the supply of the reference listed drugs (RLD). To address this demand for generics and expedite consumer access to lower-cost generic versions of approved drugs, the U.S. Food and Drug Administration (FDA) has developed an “Abbreviated New Drug Application” (ANDA) pathway that simplifies the generic drug review process and expands access to these much-needed medicines without compromising quality and safety standards. Guidelines for this pathway require sponsors to identify and characterize both process- and product-related impurities in drug formulations that differ in nature or concentration from the RLD. The ANDA pathway devotes specific attention to immunogenicity and recommends the use of orthogonal methods of assessment to demonstrate that a proposed generic drug is immunologically equivalent to its RLD and therefore suitable for submission via the ANDA pathway. In this perspective, we describe several orthogonal methods for immunogenicity risk assessment of generic peptide impurities and contrast these with other methods such as MHC-Associated Peptide Proteomics peptide elution (MAPPs) assays. Given their importance in the generic drug approval pathway, we have submitted the “PANDA^®” immunogenicity risk assessment methods as a ‘model master file’.

Graphical Abstract

Introduction

A Perspective on Orthogonal Immunogenicity Risk Assessment for Generic Drugs

The FDA has released detailed guidelines for the Abbreviated New Drug Application (ANDA) pathway that require sponsors to identify process- and product-related impurities in their drug formulation to ensure safety and efficacy [1]. The ANDA pathway devotes specific attention to immunogenicity and recommends the use of orthogonal assessment methods to demonstrate that a proposed generic drug is equivalent to its reference listed drug (RLD) and therefore suitable for submission via the ANDA pathway. This perspective article will briefly review methods used by our group that were developed during successive projects funded by the Generic Drug User Fee Amendments (GDUFA) research program. The pre-submission requirements for generic peptide drug applications are described in the final ANDA guidance, which refers to glucagon, liraglutide, nesiritide, teriparatide, and teduglutide.

Extensive documentation of the in silico immunogenicity risk assessment methods described in this article has been submitted as a Model Master File (MMF). The FDA’s MMF initiative will streamline regulatory approval of submissions that provide supporting data resulting from quantitative models or modeling platforms [2]. Like Drug Master Files, MMFs describe the specifics of models and results predicting features of drug products, such as pharmacodynamics, and provide sufficient validation data to be considered acceptable for regulatory purposes by FDA experts. In silico tools that are routinely used to assess the immunogenicity of peptide drugs and their impurities, such as those described in this article, are well suited for the MMF initiative and their submission is expected to improve the speed of the ANDA review.

Defining Immunogenicity

Immunogenicity is broadly defined as the ability of a substance (autologous or foreign) to induce immune responses by the adaptive or innate immune system. In the case of peptide drugs and their impurities, immunogenicity is usually an undesirable side effect that may impair the efficacy and safety of peptide drugs. For example, some novel peptide drugs have induced serious systemic allergic or anaphylactic responses in humans. These side effects may have compromised further development of the product [3]. The regulatory requirements introduced by the FDA for the ANDA pathway serve to reduce the likelihood of these types of adverse effects associated with generic drug products.

Relevant to peptide drugs, immunogenicity can be attributed to several product- or process-related factors, such as aggregation, fibril formation, degradation during storage, and impurities introduced during the synthetic production of the active pharmaceutical ingredient (API) [4]. Peptide- (sequence) related impurities caused by these events may drive adaptive immune responses. The potential for peptide-related impurities to induce adaptive immune responses has contributed to a new emphasis on the immunogenicity of the impurities. Specifically, introduction of Human Leukocyte Antigen (HLA)-binding sequences in impurities that were not present in the RLD could drive new immune responses that may interfere with the efficacy of the drug or cause unwanted injection site reactions or other adverse effects.

An example of undesired immunogenicity occurred during the development of Taspoglutide, an investigational GLP-1 receptor agonist intended for the treatment of type 2 diabetes. During Phase 3 clinical trials, 25 participants exposed to Taspoglutide experienced hypersensitivity reactions similar to anaphylaxis (anaphylactoid reactions) [3]. Patients also reported a high incidence of injection site reactions, which led to discomfort and adherence issues. These reactions were thought to be related to immunogenicity, as some participants also developed anti-drug antibodies (ADAs) against the drug. The adverse results raised safety concerns that may have ultimately ended Taspoglutide’s development. Ultimately, in discussions at scientific meetings on immunogenicity, scientists associated with the project attributed the systemic adverse effects to the presence of sequence impurities and, in retrospective studies, were able to link selected HLA-DR of trial participants to the unwanted immune responses. This finding of HLA-restricted responses to the drug supports the idea that T cell-mediated adaptive immune responses to sequence impurities in the drug formulation might have been responsible for unexpected adverse effects [3, 5].

Immunogenicity Risk of Generic Peptides

Despite the potential for adaptive immune responses (ADA, T cell responses) and hypersensitivity reactions, there is a notable lack of immunogenicity information in published clinical trials of peptide drugs. This may be due to the long-held belief that peptides are not immunogenic because of their small size and short half-life. The relatively positive safety profile of many peptide drugs may indeed be related to the use of smaller peptides, unnatural amino acids, and cyclized peptides, contributing to a lack of HLA-binding sequences (a prerequisite for immunogenicity). Thus, the relationship between adverse effects and adaptive immunity has not been a primary focus of peptide drug development. Specifically for generic peptide drug APIs, limited information is available on the immunogenicity profile for most of these products. However, following years of clinical use, immunogenicity has been associated with certain RLDs, and the goal of the ANDA program can be simplified to the following: reduce the likelihood that approved generic drugs are more immunogenic in patients than the RLD.

Many RLDs were developed using recombinant DNA technology, while advances in peptide synthesis have enabled generic drug manufacturers to produce the same drug synthetically. This change in manufacturing may introduce novel impurities in the final product. Therefore, the ANDA guidance now asks sponsors to identify significant differences in impurity profiles between RLD versions of peptide drugs and their generic alternatives, for risks they may pose to safety.

Reflecting the relevance of the concentration of impurities on immunogenicity, impurities that are unique to the generic product and present at a concentration between 0.1% and 0.5% of the API must be evaluated for immunogenic risk potential relative to the API. Impurities present in both the originator and generic products must also be evaluated if the concentration of the impurity in the generic drug product exceeds the concentration in the originator product [1]. Furthermore, the guidance defines the need to use at least two orthogonal methods, independently assessing immunogenicity risk, for the ANDA submission.

In concert with FDA experts [FDA contracts HHSF223018186C and 75F40120C00157], our group has revised methods previously used for protein-sized therapeutics (see, for example, reference [6]) to address the ANDA requirement for orthogonal immunogenicity risk assessments of generic peptides. This perspective describes those orthogonal methods, that include: (1) in silico screening and (2) an independent in vitro T cell assay comparing naïve T cell responses to the API and impurities using blood samples from a diverse naïve donor population. An in vitro HLA binding assay that evaluates the increased binding potential of new impurities is an additional means of assessment. The FDA specifically mentions their concern whether the impurities “contain sequences that have an increased affinity for major histocompatibility complex (MHC), known as T-cell epitopes” in their final guidance [1]. Animal models, where MHC may differ from human MHC, are not suitable to assess T cell-mediated immune responses for ANDA applications.

Peptide Impurities and Their Sources

Solid Phase Peptide Synthesis (SPPS) is the gold standard for synthesis of peptides 5–50 amino acids in length. It involves a cyclic process in which amino acids are linked individually through coupling reactions to form the desired peptide. During the coupling steps, the C-terminus of amino acid sequences are anchored to a solid support matrix and peptides are synthesized from C-terminus to N-terminus. Peptide impurities may be introduced during any given cycle of elongation, including during further processing or due to degradation of the peptide. Collectively, these peptide-related impurities may include amino acid insertions or duplications, amino acid deletions, truncation of the N- or C-terminus, incorporation of D-stereoisomers, or amino acid side chain modifications (including deamidation, oxidation, reduction, or incomplete removal of protecting groups).

Adaptive Immune Response to Peptide Drugs

Central to the immune response to peptide drugs are CD4 + T helper cells. A critical step in T cell activation is recognition of a linear peptide bound to HLA molecules on antigen presenting cells (APCs) through their T cell receptor (TCR) (Fig. 1). The (linear) amino acid sequence presented by APCs via HLA and recognized by a TCR is known as the T cell epitope. Once activated, effector CD4 + T cells provide essential cytokine support to other members of the immune system, including B cells and cytotoxic T cells. In the case of B cell help, peptides presented by B cells via their HLA following uptake and processing are recognized by antigen-specific CD4 + T cells. This interaction activates antigen-presenting B cells, leading to the development of memory B cells and the production of ADAs. T cell help, provided by HLA Class II (DR)-bearing CD4 + T follicular helper cells, is a critical contributor to B cell development and production of plasma and memory B cells [7].

Fig. 1.

Generic peptides can be processed and presented by APCs to two types of T cells: T effector (T Eff) and regulatory T (Treg) cells. Depending on the phenotype of the T cell, an antibody response to the drug is activated and plasma cells and memory cells are generated, or, if Treg cells are activated in the lymphoid follicle, the antibody response is muted or suppressed. The overall response to a peptide impurity depends on the HLA of the subject and the balance between Treg and T effector responses.

When considering forces affecting the development of immunogenicity, it is important to remember that more than one type of CD4 + T cell is present in the lymphoid follicle where antibodies are produced. Both T follicular helper cells (Tfh), which target foreign antigens, and T follicular regulatory cells (Tfr), which mostly target self-antigens and mediate tolerance, are present in lymphoid follicles [8]. The purpose of regulatory T cells that have been trained in the thymus is to suppress immune responses to self-proteins, including some peptides that contain sequences identical to natural hormones or growth factors (Fig. 1).

One driver of regulatory T cell response in the lymphoid follicle may be the API in the peptide drugs. Many of these drugs are derived from self-proteins and contain HLA-binding sequences that can be presented and recognized by natural regulatory (thymus-derived) T cells that are present in the lymph node (Tfr cells). Thus, each assessment method must take into consideration the potential for tolerance when estimating the immunogenicity risk of peptide impurities. Depending on the phenotype of the T helper cell that responds to the sequence, the immune response to a peptide impurity may be promoted (in the case of Tfh recognition) or suppressed (in the case of Tfr recognition).

The rationale for considering ‘self-ness’ in generic peptides is that the HLA-binding regions of these human-derived products are likely to have been presented in the thymus. T cells trained on these sequences have likely been removed (thymic deletion) during thymic education or may recognize sequences with a tolerogenic (T regulatory) response. Immune responses to these HLA-binding sequences are therefore unlikely. However, changes in the sequence of the API, due to modifications resulting from peptide synthesis, may result in the formation of entirely new HLA-binding sequences that are not ‘self’. This is usually due to the introduction of new ‘anchor’ amino acids that can bind to the HLA binding pocket, or modifications of TCR-facing residues. For example, the insertion of an amino acid, as illustrated in Fig. 2, can remove an HLA binding motif (scenario 1, low risk) or introduce a new one (scenario 2, high risk). Occasionally, removal (scenario 1 in Fig. 2) may eliminate a tolerogenic signal that is recognized by circulating regulatory T cells. In this case, the development of a new effector epitope and loss of a Treg epitope would have a significant impact on the immunogenicity of the drug [9].

Fig. 2.

Impact of impurities on HLA binding frames. Impurities can impact immunogenicity by altering HLA binding motifs present in a sequence. In this example, the duplication of the amino acid residue in position 2 shifts downstream amino acid residues, which could lead to one of two scenarios. The insertion of an amino acid can remove epitope content by shifting amino acid residues that are more conducive to binding downstream, lowering immunogenicity risk (Scenario 1). Alternatively, the insertion of an amino acid can introduce new epitope content by shifting amino acid residues more conducive to binding into binding positions, increasing immunogenicity risk (Scenario 2). The EpiMatrix tool can be used to predict which scenario will occur when an impurity is introduced.

Many examples of peptide impurities are described in the literature. For example, a recent survey of five manufacturers of the liraglutide API identified several impurities by LC–MS (Table I) in the manufactured products [4]; however, further characterization of T cell responses was not performed. Many of the impurities involved deletions, insertions, or multiple simultaneous modifications. Table I provides an initial assessment of the individual liraglutide impurities using in silico analysis for T cell epitopes (EpiMatrix) and for self-ness (JanusMatrix) (see details in the next sections). Note that many impurity sequences do not include a change to the core HLA binding sequence (FTSDVSSYL, shaded gray). Thus, the overall immunogenicity does not change dramatically compared to the parent API except when a new ‘promiscuous’ binding region (also shaded gray) is introduced by the deletion of a single amino acid (−31Trp) or by the deletion of two amino acids (−19TyrLeu). Initial in silico analyses can be independently confirmed by performing orthogonal methods, such as T cell assays or HLA binding assays, to further assess the immunogenicity risk of the identified impurities.

Table I.

In Silico Assessment of Common Peptide Additions and Deletions Observed in Synthetically Produced Liraglutide, Illustrating the in Silico Scores That Contribute to the Overall Immune Risk Assessment

In silico tools such as EpiMatrix (EMX) and JanusMatrix (JMX) can be used to categorize impurities based on whether they are expected to exhibit increased immunogenic potential relative to the API

*Red letters or dashes indicate insertions or deletions, while blue Z indicates the palmitic acid group at position 26 in liraglutide (low-affinity binding residue; for further details on side chain substitutions, see Mattei et al., ref [10])

^‡EMX Score changes are classified as either an increase, a decrease, or a slight (sl.) increase or decrease relative to the score of the API

Current Orthogonal Approaches for Assessing Immunogenicity

In general, when performing risk assessment for the ANDA pathway, an ‘orthogonal’ assay would provide immunogenicity risk assessment based on different scientific principles or mechanisms, resulting in complementary information that independently confirms the previous method’s results. For example, in vitro T cell assays using human peripheral blood mononuclear cells (PBMC) can be used together with in silico methods to provide complementary data on peptide immunogenicity. Alternatively, in silico tools can be used in combination with Major Histocompatibility Complex-II (MHC-II)-Associated Peptide Proteomics (MAPPs) or Class II HLA binding assays to fulfill the orthogonal method requirement.

With funding from FDA contracts HHSF223018186C and 75F40120C00157, our group tested these in silico and in vitro methods for immunogenicity risk assessment of common generic peptides: salmon calcitonin [11], teriparatide [9], liraglutide, and semaglutide. Both in silico and in vitro methods are used in the “PANDA^®” (Peptide ANDA) screening to assess the immunogenic risk of peptides and their impurities (described in an overview publication, reference [12]). Standardizing our approach has resulted in a calibrated set of orthogonal methods tailored to meet current guidelines. Orthogonal methods used by our group to support ANDA are described in the next section.

Method 1: Computational Prediction Models

The ISPRI platform (Immunogenicity Screening and Protein Re-engineering Interface, ISPRI) is used for comprehensive in silico immunogenicity risk assessment of biologics. ISPRI contains multiple immunoinformatics algorithms such as EpiMatrix and JanusMatrix, described below, that also enable comprehensive immunogenicity analysis of peptide drugs. A range of applications for the tools comprising ISPRI have been published, including for preclinical risk assessment of monoclonal antibodies, bispecific antibodies, therapeutic cytokines, peptide drugs, and vaccines [6, 10–16]. These algorithms and their tabular and graphical outputs have been submitted as a MMF for the ANDA pathway.

EpiMatrix and JanusMatrix: Defining Two types of Epitopes and Two Phenotypes of T Cell Response

EpiMatrix

This algorithm analyzes the immunogenic potential of the amino acid sequence of a given protein or peptide by parsing the sequence into overlapping 9-mer frames. These 9-mers are assessed for binding potential to a set of HLA-DR “supertype” alleles, including DRB1*01:01, DRB1*03:01, DRB1*04:01, DRB1*07:01, DRB1*08:01, DRB1*09:01, DRB1*11:01, DRB1*13:01, and DRB1*15:01. HLA-DR is the most important allele for immunogenicity risk assessment for peptide and protein drugs [6]. Together, these alleles cover more than 95% of the HLA-DRB1 alleles expressed in the global population [17]. Allele-specific assessments for each 9-mer frame are summarized to calculate an overall EpiMatrix Score for a sequence, which represents the epitope content relative to a random standard. The scores are centered on zero, with negative scores indicating a scarcity of putative epitopes and positive scores indicating an abundance of epitopes compared to random expectation and, consequently, an elevated immunogenic risk [10].

EpiMatrix predictions are routinely compared to experimentally determined binding affinities using assays detailed in the next section. In a recent benchmark analysis, a set of 251 peptides prospectively selected by EpiMatrix were tested in HLA binding assays to determine their affinity (IC50) to DRB1*01:01, DRB1*03:01, DRB1*04:01, DRB1*07:01, DRB1*08:01, DRB1*11:01, DRB1*13:01, and DRB1*15:01. A parallel benchmark was performed using the latest version of the best-in-class public T cell epitope prediction algorithm, NetMHCIIpan 4.3 [18]. The average predicted accuracy was 74% for EpiMatrix, compared to 57% for NetMHCIIpan 4.3 (Fig. 3). Average F1 scores, providing a more balanced assessment of both precision and recall, were 0.84 for EpiMatrix and 0.63 for NetMHCIIpan 4.3. These observations highlight a close alignment between EpiMatrix results and observed binding to HLA, and superior prediction capabilities compared to industry-leading tools.

Fig. 3.

EpiMatrix and NetMHCIIpan 4.3 benchmark analysis. IC50 values were experimentally determined for a set of 251 peptides against a panel of 8 HLA alleles. Observed results for each HLA were compared to predictions from EpiMatrix and NetMHCIIpan 4.3 using default cutoffs. (A) Average accuracy (+ SD) across 8 HLAs. (B) Average F1 scores (+ SD) across 8 HLAs. F1 scores range from 0 (worst predictor) to 1 (perfect predictor). ** p-value < 0.01, Paired t test.

JanusMatrix

The JanusMatrix algorithm [13, 19] is applied to identify T cell epitopes that are conserved with epitopes found in the human proteome (“self”). JanusMatrix evaluates each 9-mer frame in a peptide sequence for conservation with similar, but not exactly matched, peptides in the human genome. Matching is not required for amino acids that act as HLA binding anchors, as long as the HLA binding likelihood is similar to the original 9-mer. However, the comparison peptides must have exactly matched TCR-facing contours, which may be cross-conserved with other self-peptides. Higher numbers of JanusMatrix-matched human sequences indicate greater potential for tolerance, as T cells trained on self are likely to have been deleted during thymic selection, rendered anergic, or may be actively tolerogenic in the periphery. JanusMatrix Scores above 3 are considered to have high homology with endogenous proteins and have a higher chance of being tolerated by the human immune system when compared to predicted epitopes with lower JanusMatrix Scores.

Both the EpiMatrix and JanusMatrix algorithms are applied to evaluate immunogenic risk of generic peptides and their impurities in the PANDA^® program. First, the EpiMatrix algorithm identifies putative T cell epitopes present in the API and impurities. Second, the JanusMatrix algorithm is used to compare the T cell epitope content of the API and impurities to self-epitopes (from the human proteome). The Immunogenicity Scale (Fig. 4) ranks proteins and peptides based on HLA-DR-restricted T cell epitope content. The Immunogenicity Quadrant plot (Fig. 5) adds an examination of cross-conservation with existing peptides in the human proteome to the evaluation and compares the overall predicted epitope content (EpiMatrix Score) and humanness of the epitope content (JanusMatrix Score) of the API and related impurities to known benchmarks.

Fig. 4.

EpiMatrix and JanusMatrix analysis helps to determine the phenotype of CD4 T cell response. EpiMatrix scores are generated from a matrix-based algorithm designed & trained based on careful curation of vast amounts of in vitro HLA binding, ligand elution (MAPPs), and T cell assay data. From these HLA-specific matrices, for any given sequence, the tool can predict HLA ligands or T cell epitopes across HLA families (HLA DR supertypes) and ultimately generate cumulative scores that reflect the peptide’s binding to HLA DR, the sum of which can be used to estimate the peptide’s immunogenic potential.

Fig. 5.

Quadrant plot. This quadrant plot shows the five generic peptides that are the focus of the ANDA guidance and selected liraglutide impurities from Table I. Much of the potential immunogenicity of each peptide is concentrated in one 9-mer frame, referred to as an EpiBar. The EpiBar of salmon calcitonin is entirely foreign, locating it in the upper left ‘higher risk’ quadrant. Teriparatide also has an EpiBar, however the EpiBar is ‘highly human’ based on cross-conservation of TCR facing amino acid side chains with the human genome, thus is located in the high-scoring, but highly human quadrant. Controls include Hepatitis C NS3 peptide epitope, a well-known Tetanus Toxin epitope, Influenza hemagglutinin (HA) epitope, Human CLIP (hCLIP), Tregitope (treg epitopes) 167, a peptide from EBV. See reference [11] for more information.

The Quadrant Plot

The quadrant plot incorporates EpiMatrix scores and JanusMatrix scores. Based on these parameters, protein and peptide sequences, as well as peptide impurities, can be plotted in one of four quadrants (Fig. 5), enabling an overall assessment of their immunogenic potential compared to well-established standards. The figure also highlights the two dimensional analysis of generic peptide APIs impurities (from Table I), using EpiMatrix and JanusMatrix.

Method 2: In Vitro Assays

HLA Binding Assay

Class II HLA Binding Assay

HLA binding assays are used to validate promiscuous binding across a panel of HLA-DRB1 supertype alleles, providing an independent (orthogonal) measure of binding that can be compared to predicted (in silico) binding assessments. Our group has developed a seven-point in vitro peptide-HLA binding assay that evaluates the binding of predicted epitopes to HLA over a range of seven concentrations, allowing the generation of an IC50 value for each peptide (Fig. 6) [12].

Fig. 6.

Examples of results from HLA binding assays performed at EpiVax demonstrating the binding affinity of peptides to HLA molecules, categorized as high, moderate, weak, negligible, or non-binder. Bar graphs are color-coded according to the respective binding affinity category, with the threshold for each category indicated in the lower right schematic. The absence of a test peptide is indicated by a red bar. The sequences of the peptides are not shown due to client confidentiality.

Soluble HLA binding assays can be performed for the most common HLA-DRB1 alleles, covering > 80% of the HLA-DR subtypes in the human population. This assay was adapted for commercial use from a publication by Steere et al. [20] and provides an indirect and comparative measure of peptide-MHC affinity. As shown in Fig. 6, a biotinylated, HLA-specific high binding control peptide, and soluble HLA molecules are incubated along with an unlabeled experimental peptide. The experimental peptide is mixed with the binding reagents at seven concentrations. The following day, the HLA-peptide complexes are captured on an enzyme-linked immunosorbent assay (ELISA) plate coated with a pan anti-human HLA-DR antibody. Time-resolved fluorescence measurement of bound labeled control peptide is then assessed.

By using a high affinity peptide as the control, a dose–response curve can be generated to calculate the half-maximal concentration at which the test peptide can displace the control peptide (IC50). The HLA binding assays have been used with high accuracy in two FDA-sponsored generic drug research programs. For four alleles, the binding assay demonstrated 100% accuracy. The overall accuracy for the seven alleles evaluated in vitro was 71%.

Naïve CD4 + T Cell Assay

T cell assay protocols have also been developed in the PANDA^® program to assess the differentiation of naïve T cells into effector T cells by antigen stimulation with peptide drugs and impurities. To improve assay sensitivity, the number of naïve CD4 + T cells exposed to peptides during primary stimulation is adjusted to 2.0–2.5 million naïve T cells per control or test peptide. IFN-gamma (Th1) and IL-5 (Th2) cytokine secretion is measured by FluoroSpot assay, allowing for an assessment of the magnitude and quality of the effector T cells that drive the immune response (Table II).

Table II.

Sample Naïve T cell Assay Results

A previous iteration of the Naïve T cell assay has been used under both FDA funded projects to evaluate the predicted immunogenicity of peptide impurities. Recent changes to the assay, as described here, have been made to bring the Naïve assay in alignment with current agency requests.

While the Naïve T cell assay is widely used for detecting immunogenicity of generic peptide impurities due to its relative sensitivity, it requires the use of sufficient human-derived, antigen-inexperienced T cells that have a broad spectrum of immune reactivity without biases introduced by pre-existing immune memory. The assay requires freshly isolated, healthy donor cells to obtain the most accurate assessment. This is due to the reported poor viability of dendritic cells and regulatory T cells following freeze–thaw [21, 22]. In addition, each in vitro assay may reflect the responses of T cells with different T cell receptor repertoires due to the immune history of the donors, which introduces variability and can impact reproducibility. The complexity of the assay, including the need for fresh blood samples and extensive cell culture and monitoring, can make it costly, especially when risk assessment involves large sets of impurities.

Alternative Methods: Mass Spectrometry

Major Histocompatibility Complex-II (MHC-II)-Associated Peptide Proteomics (MAPPs) is a mass spectrometry-based approach for the identification and relative quantification of naturally processed and presented MHC-II-associated peptides. The major value of this assay lies in its coverage of three important aspects of antigen presentation: (1) antigen processing, (2) antigen binding, and (3) peptide presentation on the surface of antigen presenting cells, all three of which are important steps in activating CD4 + T cells eliciting a patient’s immune response. During preclinical development, this assay can be an important component of the immunogenicity risk assessment toolkit.

MAPPs is conducted as follows: Monocytes are isolated from PBMCs and differentiated into dendritic cells (DCs). Then, immature DCs are matured by incubation with peptides or proteins. Following this incubation period, mature DCs are lysed and the released peptide-loaded HLA molecules are collected by immunoprecipitation. The presented peptides are eluted from the HLA molecules and analyzed by mass spectrometry. The eluted peptide sequences are scanned against a database of known antigens. Large amounts of eluted peptide data are available in public databases online (for example in the HLA Ligandome database).

However, this approach has its drawbacks. It requires expensive instrumentation and significant instrumental expertise, meaning that this type of assay is usually outsourced by most biotech companies. Furthermore, HLA allele coverage may be limited by the number of donors that are tested, and cost increases with each additional donor. Compilation of results from enough donors, at significant cost, should be sufficient for regulatory filing support; however, the distribution of HLA should be similar to the target population and ‘positive’ standards that confirm assay reliability have not yet been developed. Furthermore, the identification of epitopes based solely on peptide sequences does not guarantee immunogenicity, as it lacks the functional context of T cell recognition and does not account for factors like MHC binding affinity or T cell receptor interaction and the generation of regulatory T cell responses. Thus, MAPPs needs to be complemented by functional assays for comprehensive immune risk assessment.

Comparison with ISPRI

We have performed a comparison between ISPRI and MAPPs to determine whether we should include MAPPs in our own PANDA^® program. Fortunately, a large amount of MAPPs data is currently available in public databases such as the HLA Ligandome, the Immune Epitope Database (IEDB), and others, making it possible to assess the accuracy of in silico methods as compared to MAPPs. A retrospective comparison of HLA ligands identified in published datasets of peptides eluted from HLA molecules in MAPPs assays was recently compared with data generated for the same peptides using the ISPRI toolkit [10]. For studies where donor HLA haplotypes were available, eluted peptides were assessed with EpiMatrix using HLA allele-specific predictions. We observed that 83% of the eluted peptides contained a predicted 9-mer binding core restricted to the appropriate HLA haplotype, highlighting a good alignment between EpiMatrix prediction and MAPPs data.

In addition, we screened ISPRI-derived T cell epitope clusters from a benchmark set of 24 monoclonal antibodies against the records in IEDB to determine their rate of detection in published elution studies. In this analysis, 135 of the 150 T cell epitope clusters (90%) could be related to previously eluted peptides cataloged by IEDB. Of the remaining 15 predicted clusters, 14 overlapped antibody CDRs, reducing the likelihood that they were observed in eluted peptide datasets in IEDB.

In a separate study, we assessed the alignment between in silico and in vitro results for a single monoclonal antibody natalizumab, and the gene-editing protein Cas9, for which detailed donor-peptide-HLA haplotype data were available. We found that 100% of natalizumab (MAPPs) peptides eluted from donor HLA (n = 2) were predicted ligands (using ISPRI), whereas 82% of Cas9 MAPPs peptides were predicted ligands (using ISPRI) for their donor’s HLA DR alleles (n = 18).

Overall, it appears that there is significantly positive alignment between MAPPs results and in silico results generated using EpiMatrix, suggesting that in silico analysis may be a rapid and more cost effective means of assessing immunogenicity risk potential than MAPPs. The addition of a comparison to self-peptides to the PANDA^® program, using JanusMatrix enhances the utility of EpiMatrix-JanusMatrix combined in silico results for generic drug risk assessment.

Case Studies and Applications of Orthogonal Approaches

Five generic peptides are listed in the ANDA guidance, even though the general principles appear to be applied more broadly. They are glucagon, liraglutide, nesiritide, teriparatide, and teduglutide. We have provided two case studies (salmon calcitonin and teriparatide) that illustrate the principles described in the model that can be addressed using the orthogonal methods described here. The same approach can be applied to other generic peptides listed in the ANDA guidance.

Salmon Calcitonin (SCT), a short peptide with 50% homology to human calcitonin, is FDA approved to treat postmenopausal osteoporosis, Paget’s disease, and hypercalcemia. Drug-neutralizing antibodies are found in over 60% of antibody-positive patients, while cross-reactive antibodies capable of binding to human calcitonin are not detected. A detailed assessment of SCT has been published [11]. EpiMatrix analysis identified a promiscuous epitope in the C-terminal half of SCT that binds across multiple HLA alleles and is not found in human calcitonin. The region was previously identified as antigenic, suggesting an important role in the immunogenicity of the drug.

HLA binding studies confirmed binding to multiple Class II HLA alleles for the original SCT sequence and impurities retaining the predicted epitope. T cell assays demonstrated the immunogenic potential of these sequences. In addition, SCT impurities with predicted low T cell epitope content showed low immunogenicity in these assays [11].

Teriparatide (TPT) is a peptide derived from the N-terminal 34 amino acids of human parathyroid hormone (PTH) that stimulates bone formation for the treatment of osteoporosis. Teriparatide has shown a low immunogenicity profile in clinical trials, with a small number of patients developing anti-drug antibodies, none of which exhibit drug-neutralizing activity. A detailed assessment of TPT is in preparation for publication.

Using EpiMatrix, we identified a region within the N-terminus of the TPT API that is capable of binding multiple Class II HLA alleles, suggesting potential immunogenic activity. Additional analysis using JanusMatrix revealed that this epitope region shares a high degree of homology with several self-epitopes found in proteins such as beta-tubulin. This combination of broad HLA binding and high degree of humanness defined the immunogenicity risk as low, and the peptide was classified as potentially tolerogenic.

An in vitro bystander suppression assay, the Tetanus Toxoid Bystander Suppression Assay (TTBSA), was used to confirm T regulatory epitopes predicted in silico [23]. This assay confirmed that the TPT API epitope, identified in silico as tolerogenic, can suppress CD4 + T memory cell responses. Interestingly, a single amino acid modification to this tolerogenic epitope is sufficient to break tolerance and dramatically increase the immunogenicity of the peptide [9].

This case study highlights the need to assess both potential immunogenicity based on T cell epitope content and homology with the human proteome to evaluate APIs and their impurities. In silico and in vitro results were complementary. MAPPs assays would not have provided differentiation between the EpiBar in salmon calcitonin (immunogenic) and a similar epitope in teriparatide (non-immunogenic). Thus, the PANDA^® approach, which employs in silico prediction as one orthogonal method for generic peptide immunogenicity risk assessment, followed by Naïve CD4 + T cell assays, may provide more accurate assessment than MAPPs alone.

Limitations of Current Approaches

In silico

In silico methods have become integral to preclinical evaluation of immunogenicity when identifying T cell epitopes that can drive immune responses to peptide drug products. One concern regarding web-based public tools such as IEDB’s NetMHCIIpan is that proprietary information might be disclosed or intercepted by interested parties. ISPRI, the website used to perform in silico analysis by our group, operates in a secure-access environment, entirely protected so proprietary information remains secure.

An additional limitation is that T cell epitope prediction models rely heavily on training datasets derived from experimental data, which can introduce significant biases and reduce predictive power, especially for peptides with novel or rare sequences. EpiVax has developed an approach to curate input data in order to develop more accurate predictive tools. The complex biology of immune recognition—encompassing peptide processing, HLA binding, TCR engagement, and the impact of the immune microenvironment—is challenging to capture in a computational model. As a result, in silico predictions are likely to yield some false positives (predicting immunogenicity where none exists) and false negatives (missing truly immunogenic peptides, which supports the FDA rationale for using orthogonal approaches).

In vitro

HLA binding assays are cell-free and provide only a relative measure of affinity which cannot account for natural processing. Naïve CD4 + T cell assays tend to provide a “snapshot” of immune potential in a cohort of donors limited to one window of time in the evolution of a natural immune responses. However, when both in silico analysis and the results of in vitro assays are compared to clinical data for reference-listed products, similar immunogenicity risk profiles are typically equivalent.

MAPPs

One major issue for MAPPs methods is HLA coverage of donors (as discussed above). The HLA alleles in a donor panel may not represent the full spectrum of HLA types in the human population and could skew observed results. For studies involving MAPPs, if broad HLA coverage is not included in the study population, certain peptides may not be identified as potential epitopes, especially if they are restricted to rare HLA variants. Therefore, it is important to consider broad HLA-DR coverage, with multiple donors per allele, when selecting PBMC for a given assay. Furthermore, use of MAPPs alone without an analysis of the potential phenotype of the putative T cell epitope may lead to higher risk rankings than would be observed using in silico tools and T cell assays.

Additional Caveats Related to In Vitro Assays

T cell receptor (TCR) diversity within a sample of peripheral blood T cells is another challenge, as a variety of factors influence the frequency of individual TCRs within a sample, such as age or previous exposure. In addition, standard blood draws only contain T cells that constitute a subset of all circulating TCRs. Therefore, assays such as naive T cell assays may underestimate the immunogenicity potential of impurities when evaluating generic peptide drugs.

Finally, the immune status of donors can significantly impact the outcome of T cell assays. Donors with chronic infections, autoimmune diseases, or immunosuppressive conditions may have altered immune responses, including T cell dysfunction or skewed immune responses due to existing inflammation or immune exhaustion. These factors can reduce the sensitivity of the assay, as the T cells may be less responsive or less capable of mounting a robust response to novel peptides.

Across the industry, several methods are employed to measure T cell responses to generic drug peptide impurities, many of which were not described in this perspective. These include dendritic cell-pulsed T cell assays [24, 25], assays using whole PBMCs [26], CD8 + T cell depletion and whole blood assays [27–29]. These methods may also include a range of readouts such as T cell proliferation by CFSE dilution and/or measurement of cytokine expression by Fluorospot, bead array, flow cytometry, or RT-PCR [12, 24, 30].

Concerns may arise at the time of FDA review if the results from orthogonal methods of immunogenicity risk assessment are not consistent (in terms of in silico vs. in vitro, or peptide binding vs. T cell response) or if different providers deliver different results for the same API and/or peptide impurities. As a result, reviewers may rightfully question whether the methods used in the risk assessment are adequate for determining true immunogenicity risk. Standardized, industry-wide assay controls vetted in laboratories performing different types of naïve CD4 + T cell assays would go a long way to help reviewers understand the accuracy and appropriateness of these diverse assay methods. Design and development of such controls is the subject of a new FDA-funded project which began late 2024.

Future Directions and Emerging Technologies

Advances in Computational Immunology

Advances in artificial intelligence (AI) and machine learning (ML) are rapidly improving predictive capabilities of in silico tools. As newer and larger datasets emerge, the ISPRI toolkit can be used to uncover new patterns and improve immunogenicity risk assessment accuracy. For example, MAPPs data, now publicly accessible on databases such as IEDB, can be leveraged with AI/ML techniques to advance the performance of in silico methodologies. AI/ML approaches may also be able to capture patterns and integrate multifactorial data, such as peptide processing, HLA polymorphisms, and TCR repertoire diversity. Eventually, deep learning models may be able to refine predictions by learning non-linear relationships not easily modeled by traditional algorithms.

While in silico methods will never be entirely foolproof due to the inherent variability of the immune system, the continuous improvement of AI and ML models, along with what we call HI (Human Intelligence), suggests that immunogenicity risk assessments for generic peptides will continue to improve.

Model Master Files

The in silico methods from the PANDA^® model have been submitted as an MMF to the FDA, promoting standardization of such techniques and accelerating drug access for patients. The PANDA^® MMF provides ANDA sponsors the advantage of simplifying their regulatory submission by referencing FDA-reviewed MMFs and omitting lengthy descriptions of “routine” immunogenicity models used to support their application. A PANDA^® MMF will also support FDA reviewers by expediting the approval process of ANDA applications and circumventing the need to re-review identical methods across multiple drug sponsors. Several ANDA applications have already been approved via the ANDA pathway, using the PANDA^® approach described here (including for vasopressin, teriparatide, and salmon calcitonin).

Conclusion

Currently there is no standardized method or readout to evaluate T cell immunogenicity for generic peptide drug ANDAs. Each generic drug sponsor is requested to provide sufficient details about their immunogenicity risk assessment assays to allow regulatory reviewers to determine if the methods are sufficient. At least two orthogonal methods for the immunogenicity risk assessment are expected. While the current model may be suitable for ANDA filings for the five generic drugs listed in the ANDA guidance, the method may also apply to other peptide drugs that are not yet approved for the ANDA pathway. For example, many generic drug manufactures are already preparing to submit ANDA filings for semaglutide, in anticipation of the expiration of patents.

Abbreviations

ADA

Anti-drug antibody

ANDA

Abbreviated new drug application

APC

Antigen presenting cell

API

Active pharmaceutical ingredient

GDUFA

Generic drug user fee amendments

HLA

Human leukocyte antigen

LC-MS

Liquid chromatography-mass spectrometry

MAPPs

MHC-II-associated peptide proteomics

MHC

Major histocompatibility complex

MMF

Model master file

PBMC

Peripheral blood mononuclear cell

RLD

Reference listed drug

SPPS

Solid phase peptide synthesis

TCR

T cell receptor

Author Contributions

ADG: Funding acquisition, Conceptualization, Writing – original draft, Data review, Validation, Visualization, Writing – review & editing and Revisions. AM: Writing – original draft, Data Analysis and Visualization. Writing – review & editing. Jennifer Calderi: Data Analysis and Visualization, Writing – review & editing. BG: Project administration, Data Visualization, Writing – review & editing. SL: Data Analysis and Visualization, Writing – review & editing. MM: Data Analysis and Visualization. Writing – review & editing. WDM: Methodology, Supervision, Writing – review & editing. GR: Methodology, Writing – original draft, Supervision, Visualization, Writing – review & editing.

Funding

Several of the studies reviewed here (and published in detail elsewhere) were performed under FDA contracts HHSF223018186C and 75F40120C00157 in concert with FDA experts from OGD and the OPQR. Funding for the activities related to the development of this manuscript was provided by internal funds from EpiVax.

Declarations

Conflict of Interest

ADG and WDM are senior officers and majority shareholders of EpiVax, Inc, a privately owned immunoinformatics and vaccine design company. ADG, BG, AM, WDM, JC, BR, SL, CB, GR, and MM are employees of EpiVax, Inc. Due to this relationship with EpiVax, these authors acknowledge that there is a potential conflict of interest inherent in the publication of this manuscript, and attest that the work contained in this research report is free of any bias that might be associated with the commercial goals of the company.