SBIR Phase I: Highly resource-efficient protein engineering using machine learning

这个小企业创新研究（SBIR）第一阶段项目的更广泛的影响/商业潜力是改善、加速和减轻不同行业的蛋白质工程成本，包括工业生物催化剂、生物制造、食品技术和治疗学。今天，后期蛋白质工程代表了一个主要的时间、劳动力和财政瓶颈。由于现实世界的翻译是后期开发的重点，因此分析更能反映其最终用途应用，因此必然需要更多的时间、劳动力和资本。这就排除了在这一阶段进行筛选的许多变异。在这些开发的后期阶段失败是代价高昂的，并且通常是由于早期高通量筛管测试条件的环境参数变化造成的。基于最小数据但在最终使用条件下具有高功能可能性的蛋白质变异的准确预测是一个关键的未满足需求。拟议的项目将展示利用机器学习模型的可行性，该模型经过原始蛋白质序列、诱变数据集和自然序列-功能对的训练，以预测感兴趣蛋白质（POI）的高功能变体，而不需要特定于所选POI和应用的序列-功能数据集。这种被称为零射击学习的方法迄今尚未应用于蛋白质工程。为了实现这一目标，将使用来自公共和私人数据库以及突变数据集的近50亿个精心策划的未标记蛋白质序列来训练大规模的语言模型。然后，可以将这个通用知识模型与特定于应用程序的顶层模型融合在一起，该顶层模型来源于自然序列（不同于POI）及其自然环境的参数。这种训练被假设为向模型灌输一个概念，即在一般意义上，以及在特定的环境条件下（例如，高温，高盐度等），哪个序列特征可以改善蛋白质的功能。为了证明这种方法的可行性和实用性，该模型将用于虚拟定向进化实验，以优化两种治疗相关的酶，优化其在非原生环境中的功能，并在体外评估其功能。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

The broader impact/commercial potential of this Small Business Innovation Research (SBIR) Phase I project is to improve, accelerate, and alleviate costs of protein engineering across diverse industries including industrial biocatalysts, biomanufacturing, food technology, and therapeutics. Today, late-stage protein engineering represents a major time, labor, and financial bottleneck. Since real-world translation is the focus of late-stage development, assays are more reflective of their end-use application and therefore necessarily require more time, labor, and capital. This precludes many variants from being screened at this stage. Failure at these late stages of development is costly, and often results from a change in environmental parameters from test conditions in early high throughput screens. Accurate prediction of protein variants based on minimal data but with high likelihood of function under end-use conditions is a critical unmet need.The proposed project will demonstrate the feasibility of leveraging a machine learning model, trained on raw protein sequences, mutagenesis datasets and natural sequence- function pairs, to predict highly functional variants of a protein of interest (POI) without sequence-function datasets specific to the selected POI and application. Such an approach, known as zero-shot learning, has not been applied to protein engineering to date. To achieve this, a large-scale language model will be trained with almost 5 billion curated unlabeled protein sequences from public and private databases and a collection of mutagenesis datasets. This general knowledge model can then be fused with an application-specific top model derived from natural sequences (distinct from the POI) paired with parameters of their natural environments. This training is hypothesized to imbue the model with a notion of which sequence features improve protein function in a general sense, and under particular environmental conditions (e.g., high temperature, high salinity, etc.). To demonstrate the feasibility and utility of this approach, the model will be used in virtual directed evolution experiments to optimize two therapeutically relevant enzymes, optimized for function in non-native environments, and assessed for this function in vitro.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.