(M1530-09-56) Rational Design of Topical Semi-Solids: An All-Virtual, Inverse Design Process

Purpose: Many aspects of topical semi-solid formulation design are done on a trial-and-error basis, rather than using rational technical concepts that may be able to inform the design workflow. This can result in excessively long product development timelines. We introduce a virtual approach to topical formulation development combining scientific first principles and AI/ML algorithms to predict formulation attributes in-silico. The workflow consists of building a virtual input formulation which is used to predict formulation microstructure, rheological/flow properties, and other relevant attributes. The approach can also be extended to be generic to numerous dosage forms. We demonstrate the workflow by developing an emulsion-based formulation benchmarked to a reference listed drug using ZoomLab® and show Q3 sameness1 within experimental uncertainty using only a single benchtop-scale attempt. Evidently a virtual approach to formulation design can substantially reduce product development timelines by minimizing trial-and-error.

Methods: The virtual workflow, shown schematically in Figure 1, consists of several virtual applications in ZoomLab®. Each app in the workflow targets one aspect of either formulation/processing design or prediction of formulation Q3 attributes. The workflow begins with a comprehensive Q3 characterization of the reference listed drug (RLD), using rotational and oscillatory rheology, density, and optical microscopy. Advanced microstructural analysis utilizes digiM’s I2S machine-learning workflow. The second step combines this data with the RLD label information to construct a series of candidate formulations with similar flow properties to the RLD as predicted by the ZoomLab® ‘Property Predictor’ app. The ‘Property Predictor’ module is powered by statistical modeling of Q3 attributes of at least 60 formulations using JMP and identifies correlations between a formulations rheological properties and Q1/Q2 attributes. The analysis was limited to contain terms in existing emulsion viscosity models.² Finally, candidates are further screened using ZoomLab®’s ‘emulsion stability calculator’, based on hydrophile-lipophile difference (HLD) concepts to filter potentially unstable formulations. The remaining candidates were formulated and subjected to an identical testing regimen as the RLD.

Results: A comprehensive characterization of the RLD was conducted first. Microstructural analysis conducted using I2S machine learning algorithms revealed an average d50 of 5.8±0.8μm and an emulsified phase fraction of 23±5%; these values, as well as measured rheological properties such as zero-shear viscosity, yield stress, and viscoelastic properties, were used as input parameters for the virtual formulation design. After screening, the virtual workflow outputs a single candidate formulation for RLD benchmarking (Figure 1). Select Q3 attributes of a 0.1L benchtop-scale batch are compared to the ZoomLab® predicted values and the RLD values in Figure 2. There is good agreement between the measured and predicted values, with most properties within ~20% of the RLD. The formulation aspects with the largest discrepancies (e.g., storage and loss moduli) were used in combination with the ‘Property Predictor’ app to refine the input formulation. However, this worsened Q3 sameness compared to the original candidate and indicated the original test formulation is a better match. We also formulated 5 of the candidate formulations rejected for predicted differences in flow attributes as a means of verifying the predictive capabilities of the ZoomLab® property predictor app—the difference between predicted and measured values was within 15% for each formulation, indicating the app’s accuracy. The microstructures of the two formulations are also in good agreement with similar droplet diameters and emulsified phase contents. We then manufactured a pilot-scale 1L batch of the formulation containing the API to investigate the workflow utility for scale-up. This necessitated a change in processing equipment; the ‘emulsion stability calculator’ app, which also estimates microstructural features of emulsions, was used to successfully accommodate this change. The effects of API addition on the formulation stability were captured by treating it as a polar oil within the HLD framework.³ This analysis provided insights on reformulation efforts needed avoid an unstable formulation. This was accomplished by reducing the Kolliphor® SML 20 content from 0.5 to 0.25 wt/wt%. The pilot-scale batch properties are also well-predicted by the property predictor module and are in good agreement with those of the RLD. This workflow can be extended to other dosage forms through additional experiments to correlate formulation Q1/Q2/Q3 attributes. We repeated the statistical analysis used to build the ‘Property Predictor’ app for PEG-based ointments and find that their flow properties can also be reliably predicted (Figure 3). However, such tools applied to more complex topical dosage forms may require additional experimental verification.

Conclusion: We developed and tested a series of virtual formulation tools to allow for the rational design of topical products of various dosage forms, based on the input formulation and processing parameters. The microstructure and Q3 flow attributes of emulsion-based formulations are well-predicted by the employed machine-learning tools in the ‘Property Predictor’ app, and their physical stability can be optimized using the ‘Emulsion Stability Calculator’ app. Such stepwise virtual workflows enable the rational, performance-based design of topical formulations and represents a promising avenue to reduce product development timelines.

References: [1]: United States FDA, "Physicochemical and Structural (Q3) Characterization of Topical Drug Products Submitted in ANDAs", 2022
[2]: Pal, “Novel Viscosity Equations for emulsions of two immiscible liquids”, Journal of Rheology, 2001
[3]: Ghayour & Acosta, "Characterizing the Oil-Like and Surfactant-Like Behavior or Polar Oils", Langmuir, 2019

The proposed virtual workflow for properties-based formulation design. The workflow inputs quantitative microstructural and flow information on a reference listed drug, and predicts physical stability, droplet sizes, and other Q3 attributes of a user-specified 'virtual' formulation

Visual, rheological, and microstructural characterization of the RLD and the candidate formulation identified through in-silico design, manufactured using both benchtop-scale and pilot-scale equipment. The dashed lines in the rheological comparison chart are the values predicted by our application of machine learning to an existing dataset of topical formulations. Both candidate formulations are similar to the RLD in each aspect.

Extension of the machine-learning application to predict rheological attributes of other topical dosages forms. Here, a Poly(ethylene glycol) (PEG) ointment candidate formulation was benchmarked to a RLD containing the same excipients. The candidate formulation and RLD have identical properties within experimental uncertainty.