In this blog post, I would like to take a recently published scientific article by Andris et al. [i], as an opportunity to give deeper insight into modeling of antibody-drug conjugates (ADCs). This class of therapeutics has gained increasing importance over the years. Already in 2018, eight of the 33 antibody therapeutics in late-stage clinical development for cancer were ADCs [ii]. As the name reveals, they are based on monoclonal antibodies (mAbs). For the application in targeted cancer therapy, the mAb is conjugated to a small-molecule drug.
During the conjugation reaction, different species are formed. Some antibody molecules will stay unmodified, others will have 1 or 2 drug molecules attached to them, but also 10-drug ADCs can occur. The resulting drug-to-antibody ratio (DAR) is a typical quality attribute of the drug product and must be controlled. According to the recent publication of Andris and Hubbuch high-DAR species (DAR 9-10) exhibit different pharmacokinetics, efficacy, and tolerability of the product compared to components with less drug molecules attached (DAR 2-6). After the conjugation reaction, the DAR is typically adjusted by hydrophobic interaction chromatography (HIC). The species with different degrees of conjugation can be separated well with HIC as the small-molecule drugs are very hydrophobic. The separation of species with the same DAR but different conjugation sites is naturally more difficult.
The first published model for ADCs on HIC
In their paper, Andris and Hubbuch combined a previously developed mechanistic model for the conjugation step with a newly developed model for DAR-adjustment of an ADC baaed on an AstraZeneca mAb – with excellent results. Practically, the model development followed the typical ChromX workflow. First the system and column are characterized with tracer experiments (dextran and acetone). Thereafter, a set of bind/elute experiments using salt gradients and steps are used for model calibration. Gradients lead to a good identifiability of the model parameters describing the binding equilibrium while steps allow to get insights on the mass transfer and binding kinetics. The latter are especially important for HIC as binding kinetics are typically slow in comparison to e.g. ion-exchange chromatography.
Andris and Hubbuch combine a previously developed mechanistic model for the conjugation step with a newly developed model for DAR-adjustment of an ADC based on an AstraZeneca mAb – with excellent results: presenting the first ever published model for ADCs on HIC.
Process optimization with ChromX
The adsorption model used was the HIC isotherm proposed by Mollerup in 2008 [iii]. This model is readily available in ChromX and works very well when binding is rather strong. The symmetric peaks in the gradient runs indicate that the load challenge was rather low in this study. In consequence, Andris and Hubbuch identified only the relevant model parameters for this regime. For high load densities, the model would have to be extended.
After model building, Andris and Hubbuch used ChromX to identify optimized process conditions. In comparison to the earlier process set-point, the ChromX-optimized conditions achieved a higher yield of the target DAR-2 component in shorter time and with higher concentration, all while maintaining purity. Yield and DAR of these runs were predicted by the model with relative errors between 1% and 4%.
To try the optimization yourself, as we did in figure 1, please download our ChromX file Andris_2020.cmx, which includes all model parameters from the paper and mimics Runs 1-4 and 8-9.
Implementing Quality-by-design in ADC development
Andris and Hubbuch conclude that their modeling approach “could be effectively applied to process development and optimization and could support the implementation of QbD in ADC development by yielding process knowledge and facilitating a more robust realization of critical quality attributes like the DAR.” I fully agree with them. To learn more about the use of models in the Quality-by-Design paradigm, please take a look at the description on process characterization on our website.
Application to other ADCs and HIC steps
GoSilico has modeled quite a few industrial HIC steps, including DAR-adjustment of ADCs. As almost all our projects are in active (pre-)clinical development, these models, unfortunately, never get published. What we can share is that our ADC models also worked under high-load conditions, and for varying pH values. To this, we extended Mollerup’s HIC model similarly to GoSilico’s HIC model [iv] as presented at Recovery Conference 2018.
Getting a bit more into the technical details, here’s an example of a 4-drug ADC species calibrated for an industrial HIC step. I built this model myself from 5 experiments of a pre-existing DoE for an industrial partner. While I cannot share parameter values, the following 95% confidence intervals indicate well identifiable parameters under practical conditions:
- Binding kinetics ± 1.6 %
- Binding equilibrium ± 1.9 %
- pH-dependency of equilibrium ± 3.3 %
- Ks ± 0.4 %
- Stoichiometric number ± 0.6%
- Saturation concentration = ± 1.4%
We have also modeled industrial HIC steps outside the antibody field successfully, with protein sizes down to single digit kilo Daltons (kDa). Some model protein data can be found in the poster presented at Recovery Conference 2018. Meanwhile, GoSilico’s HIC model has also proven itself in further industrial projects. It is particularly superior to Mollerup’s model under weak binding conditions, which are frequently observed for flow-through or weak partitioning processes.
If you want to model your own HIC process or in case of general questions regarding ADC or HIC modeling, please contact me directly via my author page.
[i] Andris et al. (2020): Modeling of hydrophobic interaction chromatography for the separation of antibody-drug conjugates and its application towards quality by design, Journal of Biotechnology, Volume 317, 2020, pp. 48-58.