Applying machine learning to understand photoprotection: how do triazine-based UV-filters really work?

摘要：紫外线辐射（UVR）对生命有深远的影响，如人类皮肤癌和植物光合机制的破坏。该项目将研究提供具有光保护特性的天然分子的基本机制，使它们能够吸收UVR并将其无害地作为热量消散。我们将利用机器学习势能面的力量来加速基于时间依赖密度泛函理论的电子结构计算，这是准确的，但对于动力学来说太慢了。这将使在复杂的溶剂环境中的性能，如激发态寿命的计算。背景：太阳光谱中的紫外线辐射（UVR）对人类、动物和植物的生命有着深远的影响：它当然是众所周知的皮肤癌的原因，虽然阳光是植物光合作用的必要条件，但UVR也会破坏光合作用机制，增加对入侵病原体的敏感性。这项研究的重点是研究为自然产生的分子提供光保护特性的基本机制。实现这样的模拟需要从头算动力学的显著加速，这可以通过使用基于DFT数据训练的神经网络的机器学习势能面来实现。沃里克正在开发的ESTEEM代码是一个Python包，用于溶剂中分子激发态势能面的机器学习：我们可以学会以足够快的从头算精度来执行动力学，以生成数百或数千个长轨道-至少在量子力学时间尺度上是这样。我们将使用这些动力学模型，通过基于电子结构的动力学模拟补充时间分辨光谱学实验，研究UVR与商业防晒制剂中发现的三嗪基紫外线过滤剂的相互作用。这种结合是建立这些分子中能量耗散机制的严格结构-动力学-功能图的关键（例如Tinosorb S中酚OH和三嗪N之间的分子内质子转移）以及它们对基态恢复的保真度，使得它们可用于许多吸收/恢复循环。分子结构的改性;和在UV过滤剂上的共混物（与润肤剂的混合物）。这也将使我们能够在尽可能接近实际使用的条件下，对这些紫外线的光化学进行前所未有的了解。重要的是，这种“结构-动力学-功能”方法可以使我们能够为下一代UVFs建立“创新设计规则”，以解决配方科学中的工业挑战，例如：（1）增加防晒系数（SPF），（2）增加临界波长（> 370 nm），以及（3）在暴露2小时后增加高达90%的光稳定性。

Summary:Ultraviolet radiation (UVR) has far-reaching consequences on life such as skin cancer in humans and damage to photosynthetic machinery in plants. This project will study the fundamental mechanisms that provide naturally-occurring molecules with photoprotective properties, allowing them to absorb UVR and dissipate it harmlessly as heat. We will harness the power of Machine-Learned Potential Energy Surfaces to accelerate electronic structure calculations based on Time- Dependent Density Functional Theory, which are accurate but far too slow for dynamics. This will enable calculations of properties such as excited state lifetimes in a complex solvent environment. One main target of this will be triazine-based UV filters found in commercial sunscreen formulations.Background:Ultraviolet radiation (UVR) within the solar spectrum has far-reaching consequences on human, animal and plant life: it is of course a well-known cause of skin cancers, and while sunlight is necessary for photosynthesis in plants, UVR also damages photosynthetic machinery and increases susceptibility to invading pathogens. The proposed research focuses on studying the fundamental mechanisms that provide naturally-occurring molecules with photoprotective properties.Enabling such simulations requires dramatic acceleration of ab initio dynamics, which is possible through the use of Machine-Learned Potential Energy Surfaces based on neural networks trained on DFT data. The ESTEEM code under development at Warwick is a Python package for machine learning of potential energy surfaces of excited states of molecules in solvents: we can learn to perform dynamics with ab initio accuracy fast enough to generate hundreds or thousands of trajectories that are long - at least on quantum mechanical timescales (100s of picoseconds).We will use these dynamics models to study UVR interaction with triazine-based UV filters found in commercial sunscreen formulations by electronic structure-based dynamical simulations complementing time-resolved spectroscopy experiments. This combination holds the key to establishing a rigorous structure-dynamics-function picture of the energy dissipation mechanisms operating in these molecules (for example intramolecular proton transfer between phenolic OH and triazine N in Tinosorb S) as well as their fidelity for ground state recovery so that they are available for numerous absorption/recovery cycles.Such studies will allow us to quantify the effects of: solvent; pH; modification of molecular structure; and blend (mixture with emollient) on the UV filters. This will also enable us to garner an unprecedented understanding of the photochemistry of these UVFs, in as-close to real-use conditions as possible. Importantly, such a 'structure-dynamics-function' approach may enable us to establish 'innovative design rules' for next generation UVFs which tackle industrial challenges in formulation science such as: (1) increased sun protection factor (SPF), (2) increased critical wavelength (> 370 nm), and (3) increased photostability of up to 90% after 2 hours of exposure.