(403f) Data-Driven Discovery of Polymeric Vehicles for Gene Editing:Serendipity-Inspired Design Directions.

Synthetic materials such as cationic polymers are poised to replace engineered viruses as vehicles for therapeutic nucleic acids. Polymeric materials offer a vast design space where desired material properties can be engineered by deft optimization of chemical composition and polymer architecture. However, rational polymer design is impeded by the “curse of dimensionality” since systematic examination of numerous variables such as composition, architecture, length and formulation parameters is paramount. To accelerate polymeric vector discovery, we need to develop models that predict the impact of key design attributes such as phase behavior, basicity, surface charge, nucleic acid binding, and hydrophobicity on the ultimate biological responses observed. In this work, we eschew low-throughput hypothesis-driven design of polymers in favor of a data-driven approach that exploits combinatorial polymer libraries and high-throughput experimentation. Thereafter, we employ statistical learning to extract the molecular attributes of the hit polymers that promoted gene editing efficiency and apply this knowledge to provide experimental guidance on designing subsequent polymer libraries.

Drawing inspiration from high-throughput experimental approaches used in the pharmaceutical industry to screen drug libraries for therapeutic activity, we employed parallel polymer synthesis, formulation and well plate based biological assays for rapid screening of gene editing efficiency We synthesized a chemically diverse library of copolymers combining different ratios of 1) cationic monomers bearing amines spanning a broad range of basicity and 2) neutral monomers of varying hydrophilicity. Our combinatorially designed library allows for systematic investigation of the effect of amine basicity by studying variations in polymer pKa resulting from the use of 4 cationic monomers, in ratios of 100, 75, 50 and 25 %. Subsequent to parallel polymer synthesis, extensive physicochemical characterization was completed using automated tools -: composition and molecular weight analysis, pK_a, polyplex size distribution, binding assays, and ζ-potential measurements were acquired in high-throughput modes to generate a rich dataset. Polymers were complexed with ribonucleoprotein (RNP) payloads and gene editing was quantified by estimating the proportion of mCherry positive cells. To resolve the trade-off between sensitivity and experimental throughput, we employed image cytometry and developed an image processing algorithm to quantify mCherry expression in a robust and automated fashion from a bank of images arising from 200 unique formulations. At the end of the high-throughput screening campaign, we obtained a high-performing hit polymer that outperformed state-of-the-art synthetic transfection reagents.

Having identified a hit formulation from our library, the challenge was to: 1) unravel the relationship between polymer attributes and gene editing efficiency. 2) build on these structure-activity relationships to guide the design of future polymer libraries that will yield a higher “hit rate”. In order to derive these predictive relationships, we sought to understand how 10 polymer descriptors influenced biological performance. We turned to principal component analysis (PCA), to deal with the dimensionality challenge created by our dataset. Through PCA, we concluded that a single polymer descriptor cannot be used in isolation to guide the design of future libraries. Rather, complex non-linear relationships between several molecular attributes were responsible for editing performance. Discarding preconceived notions of how various chemical functional groups will influence delivery, we screened a large chemically diverse polymer library to discover design guidelines that do not conform to traditional heuristics as well as a promising hit polymer, which may not have been accessible through hypothesis-testing. If statistical learning and automated experimental workflows are applied in tandem, we can overcome challenges originating from the complexity of structure-function relationships governing polymeric gene delivery.