Accelerated discovery of synthetic polymers for ribonucleoprotein delivery through the integration of active learning, machine learning, and polymer science

项目摘要/摘要 像CRISPR/Cas9这样的基因编辑系统作为研究工具迅速流行起来，并保持着 有可能治愈一系列不同的遗传疾病。然而，有效、安全和有效的交付仍然是 对于治疗性翻译和应用于难以体外培养的细胞类型来说，这是一个巨大的挑战。 理想情况下，完整的Cas9蛋白将与其引导RNA(SgRNA)一起作为纯化的核糖核蛋白(RNP)输送， 而不是编码Cas9的信使核糖核酸或质粒，以最大限度地减少脱靶效应。病毒载体(例如AAVs)不能 交付如此大的货物由于其船体大小有限，这在以下方面表现出额外的挑战 免疫原性、成本和可制造性。幸运的是，合成聚合物--在 核酸递送和作为生物材料--最近显示出作为体内递送sgRNA的载体的前景。 Cas9 RNPs。然而，目前还没有一致的设计原则来改善新型合成聚合物的性能 可以开发出递送效率、组织特异性和安全性。有太多的聚合物结构 彻底测试或通过特别实验，因此对聚合物设计、合成、 需要进行评估，以确定有前途的候选人。这项提议提出了一个发现的框架。 通过贝叶斯实验设计的功能聚合物。基于实验训练的机器学习模型 结果将作为实验的替代品，以便虚拟地筛选一个巨大的潜在信息库 聚合物候选者。将通过贝叶斯优化算法选择候选聚合物，以 平衡对未知化学空间的探索和对已知结构的开发，以有效地提供RNPs。 目标1将涉及(A)合成各种可生物降解的聚酯脲胺(PEUA)文库，(B) 使用模型荧光报告基因敲入/敲除试验评估它们的功能性能，a 细胞活力分析和代谢活性分析，以及(C)机器学习的开发和验证 了解聚合物结构/组成与这些多重性能之间的定量关系的模型 指标。目标2将涉及(A)计算可综合利用的多联体的化学空间，以及(B) 基于机器学习模型的贝叶斯优化框架的开发与应用 从目标1开始，通过迭代轮次指导从列举的空间中选择候选聚合物 实验。拟议工作的结果将是一个综合工具，结合了机器学习和 聚合物科学不偏不倚地探索广阔的生物材料设计空间，通过 开发高效、安全的RNP基因编辑载体，超越现有的商业 聚合物车辆解决方案。

PROJECT SUMMARY/ABSTRACT Gene editing systems such as CRISPR/Cas9 have rapidly grown in popularity as research tools and hold the potential to cure a diverse set of genetic disorders. However, effective, safe, and effective delivery remains a significant challenge for therapeutic translation and for application to cell types that are difficult to culture ex vivo. Ideally, intact Cas9 protein would be delivered with its guide RNA (sgRNA) as a purified ribonucleoprotein (RNP), as opposed to Cas9-encoding mRNA or plasmids, to minimize off-target effects. Viral vectors (e.g., AAVs) cannot deliver such large cargo due to their limited capsid size, which exhibit additional challenges with respect to immunogenicity, cost, and manufacturability. Fortunately, synthetic polymers--widely studied in the context of nucleic acid delivery and as biomaterials--have recently shown promise as vehicles for in vivo delivery of sgRNA- Cas9 RNPs. However, there are no consistent design principles by which novel synthetic polymers with improved delivery efficiency, tissue specificity, and safety can be developed. There are far too many polymer structures to test exhaustively or through ad hoc experimentation, so a systematic approach to polymer design, synthesis, and evaluation is required to identify promising candidates. This proposal presents a framework for the discovery of functional polymers through Bayesian experimental design. Machine learning models trained on experimental outcomes will serve as surrogates for experimentation in order to virtually screen a massive library of potential polymer candidates. Polymer candidates will be selected algorithmically through Bayesian Optimization to balance exploration of unknown chemical space and exploitation of structures known to effectively deliver RNPs. Aim 1 will involve (a) the synthesis of a diverse library of biodegradable poly(ester urea amines) (PEUAs), (b) the evaluation of their functional performance using a model fluorescent reporter knock-in/knock-out assay, a cell viability assay, and a metabolic activity assay, and (c) the development and validation of a machine learning model to learn a quantitative relationship between polymer structure/composition and these multiple performance metrics. Aim 2 will involve (a) the enumeration of the chemical space of synthetically accessible PEUAs, and (b) the development and application of a Bayesian Optimization framework leveraging the machine learning model from Aim 1 to guide the selection of candidate polymers from the enumerated space through iterative rounds of experimentation. The outcome of the proposed work will be an integrated tool combining machine learning and polymer science for the unbiased exploration of a broad biomaterial design space, validated through the development of effective and safe RNP delivery vehicles for gene editing that outperform existing commercial polymeric vehicle solutions.