Automation and Active Learning for the Multi-Objective Optimization of Biopharmaceutical Formulations

Introduction: Over the last forty years, biologics such as monoclonal antibodies (mAbs) have become an increasingly important therapeutic agent in the treatment of numerous diseases, particularly in the areas of immunology and oncology. Between 1986 and 2021, 136 mAb-based treatments have been approved globally, most of which are manufactured as preformulated solutions for subsequent administration to patients. However, bioformulation of complex proteins is a difficult engineering challenge; formulations must be tailored to individual therapies, necessitating time- and material-intensive campaigns to select a combination of excipients to simultaneously optimize an array of design criteria. For example, successful formulations must be biologically active, thermally stable, non-aggregated, non-viscous, non-immunogenic, etc. This often necessitates a combination of excipients to maintain pH (buffers), osmolarity (salts and sugars), viscosity (amino acids), and solubility (surfactants). These many interacting additives complicate formulation design with unintuitive and non-linear structure-activity relationships, thus creating a vast and multidimensional design space that is intrinsically difficult to navigate through and optimize using traditional techniques. To overcome this challenge, we developed a high-throughput discovery pipeline using artificial intelligence / machine learning (AI/ML) to model and predict formulation behavior using Generally Recognized as Safe (GRAS) excipients. This was supported by automation-assisted “on-demand” formulation to produce dozens of uniquely formulated antibody solutions with high reproducibility for downstream evaluation and biophysical characterization. This pipeline was then integrated into an iterative closed-loop cycle of automated Design-Build-Test-Learn (DBTL), where new rounds of experiments are designed by the AI/ML model (Figure 1A).

Materials and

Methods: A seed library of thirty unique formulations was proposed by Latin hypercube sampling to generate representative formulations containing a combination of citric acid, sodium phosphate, sucrose, polysorbate 20, and arginine at concentrations similar to FDA approved therapies on the market. Computer-guided instructions were then implemented by a Hamilton MLSTARlet liquid handling robot to produce this formulation seed library, with bovine IgG (bIgG) serving as a model antibody. High-throughput biophysical characterization was then performed using dynamic light scattering (DLS) and differential scanning fluorometry (DSF). DLS data was collected for each formulation at a range of antibody concentrations, which was used to determine diffusion coefficient (Dc) and diffusion interaction parameters (kD). In parallel, melting temperature (Tm) for each formulation was quantified by DSF using a Boltzmann function to assess thermal stability. These datasets were then used to train two gaussian process regressor (GPR) models to predict Dc and Tm. From these models, three unique Bayesian optimization strategies with differing goals were evaluated: i) maximizing kD, ii) maximizing Tm, or iii) maximizing Hypervolume (HV), a combined multi-objective performance metric of kD and Tm. Each strategy utilized an expected improvement (EI) acquisition function to propose 10 additional formulas (30 total) which collectively formed the library for the next “round” of the DBTL loop. Ultimately, three rounds of DBTL were implemented in order to optimize formulations and train the AI/ML models. All AI/ML methods were implemented in Python 3.10 using the scikit-learn package.

Results, Conclusions, and Discussions: Remarkably, after three rounds of DBTL, the GPR models demonstrated very high accuracies in their predictions of Dc and Tm with calculated R2 of 0.700 and 0.956, respectively, for 90 formulations (Figure 1B). This suggests the model could learn complex biophysical behaviors from relatively small, experimental-scale datasets. One of the key goals of this study was to guide formulation development towards multiple objectives. When a Bayesian optimization strategy optimized hypervolume of both kD and Tm parameters, the resulting formulations demonstrated a significant improvement in antibody Tm, while maintaining kD (Figure 1C). This resulted in significant generation-over-generation improvement for all objectives, demonstrating the capacity for our models to learn from incremental new data and efficiently navigate the formulation landscape. In contrast, when Bayesian optimization attempted to only optimize a single parameter (i.e., kD or Tm), deterioration of the other property was possible (Figure 1D). We also implemented Shapley Additive Explanations (SHAP) analysis on the GPR models to infer the impact of different excipients on mAb behavior and better understand why the model is driving towards certain formulation compositions. For example, SHAP values for the Dc GPR model indicate concentrations of citric acid and sodium phosphate are the most important features, reporting positive and negative correlation with SHAP values, respectively (Figure 1E). This reflects known pH-dependent antibody behavior; as antibodies tend to have weakly basic isoelectric points, acidic conditions result in greater surface charge, leading to larger repulsive forces and higher diffusivities. Likewise, this pH-dependent shift in attractive-versus-repulsive forces is also observed as SHAP values flip from negative to positive with increasing citric acid concentration (Figure 1F). The fact that SHAP implicitly captures this established property indicates that the GPR model is successfully mapping the nuances of the underlying physical phenomena. In conclusion, we successfully demonstrated an active learning campaign could simultaneously optimize multiple parameters of interest in bioformulation – specifically colloidal and thermal stability – as well as map underlying roles of key features within the formulation space. Future work will further broaden the design space with additional excipients and incorporate additional objectives for a more comprehensive formulation optimization challenge.

Acknowledgements (Optional):