Autonomous Multimodal Optimization of Surfactant Solutions Utilizing The AFL Platform

Biological formulations such as protein therapies, vaccines, insulin solutions are complex mixtures of a variety of carefully balanced components, for example surfactant stabilizers, anti-microbial agents, buffer salts, and the active biological component. Such solutions are optimized for shelf life, potency, and efficacy of the biologic. Neutron scattering techniques are a critical but costly tool for understanding the solution structure and interplay between the insipient and excipient components. A naïve grid search over each independent component results in an increasingly intractable space due to a span of resource limitations. (experiment and analysis time, cost of reagents, beamline availability, etc.) The autonomous formulation lab (AFL) was developed to conduct an optimized and iterative search of such spaces utilizing neutron scattering. While sensitive to the underlying structure, the neutron scattering is not capable of measuring all the desired properties of the end formulation. Previous work indicates successful exploration through large dimensional spaces but operates only on identifying solution morphology boundaries in the scattering data. A solution that becomes turbid indicates aggregation, a non-starter in biomedical formulations, and can be detected by simple image processing techniques. The combination of both structure information and turbidity can reduce the number of experiments in exploring <i>useful</i> solution composition space, lowering the overall resource cost. Here we present the recent results of combining two sources of information, Gaussian process classification of the neutron scattering patterns and gaussian process regression over the optical turbidity, in active learning campaigns trialing different acquisition functions and search strategies. We benchmark independent source learning with a joint optimization scheme that explore the space via the uncertainty in the classification model while exploiting the expected turbidity of the regression model. This accelerated the exploration and optimization of solutions containing a PS188 surfactant, and two preservative and antimicrobial components benzyl alcohol and phenol in a sodium phosphate buffer.