Computational design of multi-state proteins using molecular simulations and machine learning

计算设计从头蛋白质的能力是未来合成生物学和生物材料发现进展的关键技术。然而，目前的从头设计方法往往会产生缺乏功能性的僵硬、超稳定的褶皱。为了设计真正具有功能的从头蛋白质，设计具有多种构象状态的蛋白质是至关重要的。这将使蛋白质能够设计为催化剂、生物传感器或生物开关，以响应环境变化，如pH或温度的变化。1这种努力对当前蛋白质设计软件中的能量功能具有挑战性，因为它需要设计结构不同但能量相似的折叠。在原子水平上描述蛋白质动力学和溶剂化效应的严格的分子动力学模拟方法已经显示出这种“动态设计”问题的潜力。2然而，当前的分子动力学方法太慢和太复杂，不能常规地应用于蛋白质设计。这个博士项目将验证一种跨学科的方法，该方法利用新的机器学习算法来显著加快从头开始蛋白质设计问题的分子动力学模拟的速度。在整个项目中，我们将解决一系列在蛋白质设计中至关重要的问题，从优化环肽的二级结构偏好3到调整新螺旋束的首选齐聚状态。4由于现有的大量实验数据和经过验证的原子建模的能量函数，这些系统非常适合数据驱动/物理驱动的组合分子建模方法。这种多态系统的精确建模将为设计用于合成生物学应用的多功能构件和发现新的生物材料开辟新的途径。通过设计可切换螺旋束，对设计方法进行了实验验证。设计将在实验室中使用分子生物学或固相多肽合成来产生，并将在体外使用一系列生物物理技术进行表征。

The ability to computationally design de novo proteins is a key technology that underpins future progress in synthetic biology and biomaterials discovery. However current de novo design methods tend to produce rigid, hyperstable folds that lack functionality. To progress towards truly functional de novo protein design it is crucial to design proteins that adopt multiple conformational states. This would enable the design of protein as catalysts, or biosensors or bioswitches that respond to environmental changes such as a variation in pH or temperature.1Such an endeavour is challenging for the energy functions in current protein-design software because it requires designing folds that adopt structurally distinct yet energetically similar states. Rigorous molecular dynamics simulation methods that describe protein dynamics and solvation effects at the atomistic level have shown potential for such 'dynamic design' problems.2 However current molecular dynamics methods are too slow and complex for routine application to protein design.This PhD project will validate an interdisciplinary approach that leverages new classes of machine learning algorithms to substantially accelerate the speed of molecular dynamics simulations for de novo protein design problems. Throughout the project we will tackle a range of problems of fundamental importance in protein design, ranging from optimising secondary structure preferences of cyclic peptides,3 to tuning the preferred oligomerisation states of de novo-helical bundles.4 Such systems are well suited to a combined data-driven/physics-driven molecular modelling approach owing to a large amount of existing experimental data, and validated energy functions for atomistic modelling. The accurate modelling of such multi-state systems will open new routes for the design of versatile building blocks for synthetic biology applications, and for the discovery of new biomaterials. Experimental validation of the design method will be performed by designing switchable -helical bundles. Designs will be produced in the lab using either molecular biology or solid phase peptide synthesis and will be characterised in vitro using a range of biophysical techniques.