In 1979 I joined the Mathematische Applikationen section of Ciba-Geigy's Wissenschaftliches Rechenzentrum in Basel. I was mainly active in providing support to the toxicology functions in both Ciba-Geigy's agrochemical and pharmaceutical businesses, but had the space and time to investigate the use of Bayesian methods broadly within Ciba-Geigy's various businesses. In this talk I will look at some of the statistical applications that I was involved in from determining the safe level of Urethane in kirsch for the Bundesamt für Gesundheit in Bern, to providing evidence that batches of a pesticide was no more toxic than Ciba-Geigy's published data as claimed by one national authority. I will also address approaches to the use of historical data, a topic of continuing interest today, and the appropriate experimental units in one class of toxicology experiments which has links to cluster randomised trials.

Dose-response (or concentration-response, time-response) analyses are an integral part of toxicological research. Often, the goal is to find the lowest condition, where a (significant) change of the response in comparison to a negative control can be observed. For this, both observation-based methods (e.g. Dunnett-test, LOEC) or model-based methods (e.g. ED-values, BMD) are used. For viability assays, parametric modelling with sigmoidal models is well-established. In recent methodological research, parametric modelling and calculation of alert doses based on this parametric modelling has been adressed also for gene expression data

In this talk a review of dose-response analyses published in 2021 in three major toxicological journals is presented. Dose-response analyses from published figures were included when at least four concentrations were measured, where the control is also counted.The review was performed in terms of the biological background (the type of assay, the type of exposition), in terms of the design (the number of considered conditions and the actual condition values, as well as the sample sizes) and in terms of the statistical analysis (the display, the analysis goal, the used methods for testing/modelling and the alert dose of interest). The results from this review are presented comprehensively.

Supported by the findings from the review, a comprehensible guidance is provided, which aspects to consider when designing and analysing dose-response analyses in toxicological research.

In systemic toxicity studies replacement of concurrent control animals by so-called Virtual Control Groups (VCGs) may reduce the use of animals thus contributing to the 3R's principle of animal experimentation: Replacement, Reduction, and Refinement.

Within the Innovative Medicine Initiatives project eTRANSAFE (enhancing TRANSlational SAFEty Assessment through Integrative Knowledge Management)¹ the VCG subgroup was formed tackling the current major obstacles: (i) collection, curation and sharing of suitable sets of historical control data from preclinical toxicity studies, (ii) investigation of methodologies for derivation of ViCoGs from historical control data and performance testing in statistical analysis and (iii) to reach out for regulatory advice to gain acceptance of this concept as early as possible².

This talk introduces the ViCoG database with currently (FEB 2023) more than 60 000 animals' data donated from Bayer AG, Merck KGaA, Novartis Pharma AG, F. Hoffmann - La Roche AG, Sanofi Aventis GmbH and Fraunhofer Gesellschaft ITEM. The SEND (Standard for Exchange of Non-clinical Data)³ domains Demographics, Organ Measurements, Clinical Observations, Macroscopic Findings, Body Weights, Laboratory Measurements and Microscopic Findings are covered and are populated with between >65.000 and up to >257.000 records per domain. All partners contribute with their expertise to this unique cross-industry resource for advancing both VCG-method development, data reuse and rethinking statistical analysis frameworks in toxicity studies in general.

A proof of concept based on 67 toxicology four-week studies revealed overall good agreement between original test results and reanalysis using VCGs for 19 clinical chemistry parameters.

Emphasizing SEND regarding data-formatting and provision of required experimental information, a case study on the effect of a hidden confounder illustrates the special role of understanding sources of variability within the data⁴.

Strategies to address the impact of the hidden confounder are presented leading ultimately to a proposal of dedicated control-chart approaches fostering provision of required experimental information for fine-tuned data-selection in a VCG-framework.

In an outlook the impact of FDA’s recent changes allowing promotion of drug-candidates to human trials after either animal or nonanimal tests, on the potential role of VCGs in nonanimal testing situations will be considered.

[1] www.etransafe.eu/the-project/about/

[2] Steger-Hartmann, T., Kreuchwig, A., Vaas, L., Wichard, J., Bringezu, F., Amberg, A., Muster, W., Pognan, F. and Barber, C. (2020) Introducing the concept of virtual control groups into preclinical toxicology testing, ALTEX - Alternatives to animal experimentation, 37(3), pp. 343–349. doi: 10.14573/altex.2001311.

[3] www.cdisc.org/standards/foundational/send

[4] Gurjanov, A., Steger-Hartmann, T., Kreuchwig, A., and Vaas, L.A.I.(2023) Hurdles and Signposts on the Road to Virtual Control Groups -A Case study illustrating the Influence of Anesthesia Protocols on Electrolyte Levels in Rats, Front. Pharmacol. (under review)

In several pre-clinical or toxicological applications, it is of interest to verify, if an actual observation is in line with historical knowledge. A strategy for this verification is the application of prediction intervals, that should cover an actual (or future) observation with a predefined probability.

For example, dichotomous endpoints such as tumor incidences obtained in longterm carcinogenicity assays lead to overdispersed binomial data. Counted observations such as the number of reverant bacteria colonies in the Ames assay can be modelled to be overdispersed Poisson. Continuous data (e.g. body weights or immunogenicity reactions) often involves one or several nested random effects, when several historical studies are jointly assessed.

Hence, prediction intervals for overdispersed binomial and Poisson data are implemented in the R package predint. Furthermore, prediction intervals computed based on random effects models, that allow for different random effect structures in the historical as well as in the actual data, are available via predint.

The use of the proposed prediction intervals will be demonstrated based on real life data.

Scheduling a production plan for a drug product line comes with many challenges. The planner has to account for the available equipment, working times of operators, and scheduled maintenance operations. Other constraints are related to the products being manufactured, as each product has a specific format and process duration. In addition, there can also be unplanned events which require a total redesign of the schedule.

In order to facilitate the planning process, we have developed a data-driven simulation tool that accounts for all of these constraints and provides explicit planning schedules that can be easily geared to the needs of each product line. The simulation can identify ideal production plans for different priorities, e.g., schedules that maximize the overall equipment effectiveness (OEE), minimize the end-to-end lead time, or even decrease the amount of manual labor between the change over of different products. In addition, this simulation can use historical data to evaluate the chances of successfully running a given production plan and, in consequence, identify potential bottlenecks. It can also help process managers to quickly adapt and optimally assign tasks especially during unplanned events.

We have successfully scaled and applied our method to multiple lines at Roche. The results will be shared and lessons learned will be discussed. The simulation tool helps increase the asset utilization and production capacity, which ultimately enables more drug products to be delivered to our patients.