Immunogenicity Publication Bias and Its Consequences for Predictive Models: A Call for Transparent Reporting

ABSTRACT

Understanding immunogenicity is crucial to improving therapeutic protein development. By comparing Phase I to III antidrug antibody (ADA) incidence data of Roche‐internal and approved monoclonal antibodies, we demonstrate a bias toward lower ADA incidence in published data, partly because ADA data from early trials—often discontinued for reasons related to high immunogenicity—are rarely published. Through an empirical model, we show how this bias affects ADA incidence time‐course predictions, underscoring the need for cross‐industry transparent data reporting.

1. Introduction

Immunogenicity lowers therapeutic protein technical success rates [1]. However, gaps and biases—especially unpublished immunogenicity data of early‐phase trials—impede antidrug antibody (ADA) incidence modeling and clinical impact prediction. We examine disparities between in‐house and published ADA data and their effect on model‐based predictions:

• Publication Bias: We assess whether published ADA rates for monoclonal antibodies (mAbs) reflect typical rates during development by comparing data from approved mAbs and Roche internal Phase I‐III studies. This phase‐wise analysis addresses the gap of assessing the impact of clinical trial phase on ADA production noted by Galle et al. [2].

• Benchmark Immunogenicity Model: Using insights from publication bias, ADA incidence data, and animal data, we build a model that predicts ADA positivity over time. This simple, data‐derived model represents a typical mAb in Phase I–III trials. It can be refined with mAb‐, cohort‐, or study‐specific covariates.

• Limitations and Future Directions: Our model highlights challenges in validating in silico and in vitro immunogenicity models due to publication bias and context‐of‐use constraints.

2. Methods

To examine the influence of ADA publication bias on model‐based ADA‐positivity rate predictions, we developed an empirical model. This was initially built on Roche‐internal subject‐level animal study ADA profiles, then translated to humans using published and Roche‐internal clinical phase‐separated cumulative ADA‐positivity rate data. The final human models built on data from approved versus Roche‐internal Phase I‐III mAbs simulated the impact of underlying data set choice on ADA‐positivity rates over time.

2.1. Nonhuman Primate Data

We used subject‐level ADA time courses from 33 cynomolgus monkey studies testing one compound each. Study durations ranged from 15 to 386 days, typically including control and multiple treatment groups varying by dose, dosing frequency, and administration route (subcutaneous or intravenous). Covariates included treatment group, compound, co‐treatment, subject count per time point (addressing irregular sampling), and four in silico paratope quality scores, see Supplement S1 for details.

2.2. Human ADA Positivity Rate Data

Cumulative ADA‐positivity rates came from published and Roche‐internal datasets. The published set (Figure 1, gray line), compiled by Vaisman‐Mentesh et al. [3], includes 65 mAbs approved 1997–2020 by the Food and Drug Administration (FDA) (and mostly also the European Medicines Agency), with data from drug labels (32 compounds) and journal articles (33 compounds). Reported intervals were summarized by geometric mean; if an extra value was given, the final value was the arithmetic mean of that and the geometric mean. Reported 0% or < 1% incidences were replaced with exact values from FDA labels; if unavailable, 0% was substituted with 0.001% to enable logit‐normal fitting. The Roche‐internal dataset (Figure 1, blue lines) included cumulative ADA‐positivity rates from the latest‐phase study per mAb, covering immunology, neuroscience, oncology, respiratory, and cardiometabolic indications (2014–2024): 22 Phase I, 8 Phase II, and 15 Phase III studies.

FIGURE 1.

Distribution of ADA‐positives ratio for approved mAbs (gray solid line) and for phase‐separated in‐house compounds (3 shades of blue, one panel per phase). The figure highlights two key behaviors: (1) Approved mAbs exhibit a distribution skewed toward low ADA rates compared to in‐house Phase I and II data. (2) In‐house compounds display distinct patterns across clinical phases. Early‐phase compounds exhibit broader distributions, reflecting variability in ADA rates in the early stages of development. In contrast, Phase III compounds show a distribution closer to marketed mAbs.

2.3. Empirical Model of ADA Incidence Dynamics in Non‐Human Primates (NHPs)

We described the ADA‐positivity rate time course using a nonlinear mixed‐effects ( NLME ) model [4]. The final model in Equation (1), selected as best‐performing among several built on NHP data (Supplement S1), corresponds to a closed form solution of a two‐state Markov Model (MM) and its ordinary differential equations (ODEs). It captures ADA positivity dynamics across treatment groups in a single NHP study and is parameterized by four key variables:

• y _max: Maximum ADA‐positive incidence.

• b: Baseline ADA‐positive ratio.

• t _lag: Latency period before the onset of post‐baseline ADA positivity.

• k _np: Transition rate from ADA‐negative to ADA‐positive states in the MM.

$$y=y_{\max }-\left(y_{\max }-b\right)·\mathit{exp}\left(-\frac{k_{\mathrm{np}}}{y_{\max }}·\left(t-t_{\mathrm{lag}}\right)\right)$$ (1)

2.4. Human Translation of the Empirical Model and ADA Time‐Course Simulations

Three model parameters for the human ADA model were derived from NHP data: Baseline ADA positivity (b) was sampled from a logit‐normal distribution (0, 1); onset time (t _lag) and transition rate (k _np) were both sampled from lognormal distributions. Their mean and standard deviations were based on the NHP study population values and standard deviation of random effects, respectively. y _max, representing peak ADA incidence, was based on human data to capture species‐specific immunogenicity incidence: A logit‐normal distribution (0, 1) was fitted to published and in‐house cumulative ADA data. For in‐house data, this was repeated per phase.

We then performed 10,000 simulations, each representing a single compound, using y _max sampled from (1) the published‐data distribution and (2) phase‐specific in‐house distributions. Results are shown in Figure 2; phase‐merged outcomes in Supplement S2. The source code underlying the simulations is available in the repository cited in the Code Availability statement. These simulations allowed us to directly compare the predicted immunogenicity dynamics based on different data sources, as detailed below.

FIGURE 2.

Simulated ADA‐positive ratio over time based on y _max values sampled from logit‐normal distributions fitted to published (magenta, last panel) and in‐house cumulative ADA data (blue, first three panels, one panel per phase). The median trajectory is represented by the solid line, while shaded regions indicate percentile ranges of the simulated distributions (10%–90%, 25%–75%, and 40%–60%, from lightest to darkest, respectively). The simulations illustrate how ADA‐positive ratios evolve and vary between phases of in‐house compounds and approved mAbs. Early‐phase compounds (Phases I and II in Panels 1 and 2, respectively) exhibit broader distributions, with higher medians and wider percentile ranges, suggesting greater uncertainty and higher immunogenicity potential. Using Phase III compounds (Panel 3), the simulated ADA‐positive ratios converge toward lower medians with reduced variability, aligning more closely with the profiles simulated from approved mAbs (Panel 4), which display comparatively skewed distributions toward low ADA‐positive ratios.

3. Results and Discussion

Figure 1 shows that ADA‐positivity ratio distributions for approved and internal Phase III mAbs skew lower than those of internal Phase I and II compounds. Early‐phase compounds show broader distributions—reflecting higher ADA‐positivity rate variability in early‐phase development—while internal Phase III‐compound distributions resemble those of marketed mAbs. This likely reflects that late‐stage failures from immunogenicity are comparatively rare; efficacy is usually their main driver [5]. Minor differences include a trimodal, rather than monomodal, distribution for internal Phase III compounds, possibly due to the small sample size of 15 compounds. The local density peaks may reflect inclusion of all, not just approved, Phase III compounds—some possibly terminated (partially) due to immunogenicity. This observation supports our core concern: High‐ADA compounds often fail early [5], and non‐publication of such ADA data skews public datasets toward low immunogenicity. Moreover, it could reflect that some of the marketed compounds were assessed for ADA pre‐2014 and the FDA‐required Phase III‐assay sensitivity increased in 2009 and 2014. Together, these phase‐specific trends highlight the importance of considering development stages, study time, assay information and dataset completeness and size when interpreting immunogenicity data.

Figure 2 illustrates how the above biases affect model predictions. Particularly, as shown in the two rightmost panels, setting y _max using only published ADA data yields a low predicted median maximal incidence of 2.7%—similar to internal Phase III‐based predictions. Both the marketed mAb‐based simulations and those based on in‐house Phase I‐III data start from a median incidence of 0.019% at time 0. By day 100, the median predicted incidence for the approved mAb‐based simulations is 3.3%, compared with 4.3%, 3.5%, and 2.9% for simulations based on in‐house Phase I, II, and III datasets, respectively. Predictive models like ours are most useful in early therapeutic development. For such early use, the (publicly unavailable) Phase I in‐house dataset is the most appropriate dataset available to us to base y _max on for the typical, nonspecific mAb that we currently simulate.

A more critical distinction than the median ADA incidence highlighted by the simulations is the difference in variability. Simulations using our model built on approved mAb data [3] produced narrower percentile ranges in the simulated distribution than those based on sparser internal Phase I and II data (Figure 2), indicating greater precision in low‐range immunogenicity predictions. Taken together, this calls for caution and action from users and developers of immunogenicity models based largely on published studies and hence mainly (near‐)approved compounds. Applying these during early development phases may underpredict peak ADA incidence with undue confidence. This underscores our call to the pharmaceutical community to publish immunogenicity data from early—potentially terminated—therapeutic protein programs, as well as more detailed data on immunogenicity's impact on efficacy. This would hugely benefit the development of new in silico and in vitro methods as well as the assessment of the predictive ability of existing models [6]. Improved, less biased predictions could enhance technical success rates and reduce unexpected immunogenicity‐related adverse events, ultimately benefiting patients. To illustrate this, consider a thought experiment: What if all projects predicted to exceed 50% ADA incidence before or after Phase I were terminated using a model built on all relevant mAb data? Might the resulting development cost savings outweigh revenue lost from the few compounds that would have succeeded?

4. Considerations and Future Directions

• Data availability for model building: The phase‐specific empirical models relied on sparse Roche‐internal ADA data that may not represent industry‐wide mAbs. Still, they show a clear trend toward lower maximal ADA incidence in later phases—despite longer trials—likely partly due to early attrition from immunogenicity. The published dataset used here includes 65 approved mAbs (1997–2020) [3], which we consider broadly representative of available public data, given the publication bias of favoring late‐phase, low‐ADA studies [7]. A reason to select this pre‐established data collection was that it constitutes a typical easy‐access source used by the community to validate and build in silico and in vitro models. Maintaining an up‐to‐date, unbiased collection of immunogenicity data across therapeutic proteins would support cross‐industry modeling. Publishing (neutralizing) ADA assay details and sampling timing would improve the evaluation of response levels.

• Model simplification: Our simplified model predicts per‐phase ADA incidence across compounds, highlighting ADA publication bias effects on model building and validation. For early‐phase compounds, the variability of the predicted ADA‐positivity ratio is large (cf. Figure 2). This highlights that to enable supporting drug development with study‐centric immunogenicity predictions, the next step would be to incorporate compound‐, study‐, and patient‐specific features, including nonclinical assay results [1, 8].

• Human translation: We based our model on subject‐level animal data, using human data only for maximal ADA incidence due to non‐translatability across species. So we assumed that the structural model, baseline ADA incidence, and time to onset translate from NHP to humans. Ideally, the model would rely entirely on human data. However, data scarcity, high dropout, irregular dosing, and assay variability complicate human‐only models. Cross‐industry sharing may enable this in the future—a notable current example is the Innovative Medicines Initiative (IMI) project ABIRISK consortium [9].

• Prediction beyond immunogenicity incidence: There is a strong need for predictive models supporting a more complete assessment of the impact of immunogenicity on exposure, efficacy and safety [1]. Yet, even advanced efforts [6] are impeded by current data scarcity. We thus welcome the FDA‘s draft guidance [10] on ADA impact labels for safety and efficacy and underline the importance of publishing respective information—also before regulatory drug acceptance—to enable immunogenicity predictions. Ultimately, a more transparent and collaborative approach to sharing immunogenicity data will not only improve predictive models but also accelerate the development of safer, more effective therapies for patients.

Funding

This work was conducted and funded by F. Hoffmann‐LaRoche Ltd.

Conflicts of Interest

L.C.F. and H.P.G. are full‐time employees and shareholders of F. Hoffmann‐La Roche Ltd. Z.H. is a Genentech employee and stockholder of the Roche Group.

Supporting information

Supplement S1. Summary of the model development.

Supplement S2. Phase‐merged in‐house simulations.

Tascedda S., Hu Z., Grimm H. P., and Franssen L. C., “Immunogenicity Publication Bias and Its Consequences for Predictive Models: A Call for Transparent Reporting,” CPT: Pharmacometrics & Systems Pharmacology 15, no. 2 (2026): e70184, 10.1002/psp4.70184.

Data Availability Statement

Code Availability: The code used to perform the simulations and generate the figures presented in this article is publicly available in a GitHub repository (https://github.com/SopTax/IG‐publication‐bias).