Collaborative Research: MFB: Integrating Deep Learning and High-throughput Experimentation to Rapidly Navigate Protein Fitness Landscapes for Non-native Enzyme Catalysis

理解蛋白质结构和功能之间的关系仍然是一个重大挑战。这一知识将有助于药物设计、回收和化学品生产。这个项目的目的是学习如何创造出促进自然界中的反应的蛋白质。人工智能将解释实验产生的数据。将对两类酶进行修饰，以促进新的反应。为了帮助STEM劳动力多样化，将为对蛋白质设计感兴趣的学生提供机器学习研讨会。这种方法既模拟了预期活动，也模拟了预测的不确定性。训练深度学习模型是数据密集型的。采用变压器结构的卷积神经网络(CNN)将使用模拟的序列函数数据进行预训练。模拟数据将使用Rosetta生成。预先训练的CNN将用使用组合密码子突变(CCM)产生的实验数据进行改进。将使用GFP表达、基于FACS的筛选和下一代DNA测序来监测单个细菌细胞中的酶活性，以确定相应的氨基酸序列。当存在多个细胞时，生物传感器筛选可能会受到串扰的影响。罗梅罗实验室开发的皮升规模的微滴筛选技术将被用来避免这个问题。将开发一种模拟退火算法来随机搜索序列位置和退化密码子，以寻找具有预期批量BO目标的高值文库。此外，还将设计和实施一个使用基于抽样的推理来估计密码子最佳组合的概率计划。该项目由化学、生物工程、环境和运输系统部门(CBET)、化学部门(CHE)和信息和智能系统部门(IIS)联合支持。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Understanding the relationship between protein structure and function remains a major challenge. This knowledge would benefit drug design, recycling, and chemical production. This project is designed to learn how to create proteins that will facilitate reactions seen in nature. Artificial intelligence will interpret the data generated by experiments. Two classes of enzymes will be modified to facilitate novel reactions. To help diversify the STEM workforce, workshops in machine learning will be offered to students interested in protein design. Summer research opportunities will be offered to high school and undergraduate students traditionally underrepresented in STEM fields.In this project, protein engineering is treated as a Bayesian optimization problem, with the objective to explore sequence space for improved specific activity. This approach models both the expected activity and the uncertainty of the prediction made. Training deep learning models is data intensive. A convolution neural net (CNN) using transformer architecture will use simulated sequence-function data to pretrain. The simulated data will be generated using Rosetta. Pretrained CNN will be refined with experimental data generated using combinatorial codon mutagenesis (CCM). Enzyme activity in single bacterial cells will be monitored using GFP expression, FACS-based screening, and next-generation DNA sequencing to determine the corresponding amino acid sequences. Biosensor screening can suffer from crosstalk when multiple cells are present. A picoliter-scale microdroplet screening technology developed in the Romero lab will be utilized to avoid this issue. A simulated annealing algorithm to randomly search over sequence positions and degenerate codons for libraries with high values for the expected batch BO objective will be developed. In addition, a probabilistic program using sampling-based inference to estimate the optimal combination of codons will be designed and implemented.This project is jointly supported by the Division of Chemical, Bioengineering, Environmental and Transport Systems (CBET), the Division of Chemistry (CHE), and the Division of Information and Intelligent Systems (IIS).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.