A versatile machine learning image recognition software for automating synchrotron Macromolecular Beamlines

大分子结晶学是研究生物学中最重要的分子机器-蛋白质-的最常用技术之一，因为它允许我们确定这些分子的3D结构并推断它们的功能。这与解决从遗传疾病、癌症和与人类病原体作斗争等各种人类疾病特别相关，并取得了已被证明的成果。这项技术还被用于农业和粮食研究领域，如开发新型除草剂或抗旱作物，以应对当前气候变化的影响。最后，能量存储和电池技术最近也受益于晶体同步加速器，帮助我们这个时代的关键制造和清洁增长挑战。结晶学被从学术到行业药学公司的大量研究人员使用。这些研究人员经常将样品送到大型研究机构，如同步加速器，然后远程收集X射线衍射数据，或使用全自动系统。随着同步加速器技术的最新进展，瓶颈已经从缺乏同步加速器X射线的强度或探测器技术的速度转移到硬件和软件上，这些硬件和软件通过将样品居中并为数据收集做好准备，使样品对X射线可见。对单个晶体进行数据采集通常不到10秒，但所有其他任务都会使每个样品的时间达到~2分钟。人工智能的最新进展创造了图像分析的范式转变。在英国以外的同步加速器设施中，已经有一些原型使用人工智能来提高这些基本任务的速度和可靠性。我们建议使用其中一个经过验证的原型，并将其进一步开发用于样品定心、同步辐射X射线光束线诊断和机器人碰撞风险缓解。这将对英国国家同步加速器-钻石光源(DLS)的MX光束线非常有利。许多DLS姊妹机构可以从人工智能的应用中受益，但缺乏从头开始实现正常工作的人工智能代码的“诀窍”。该项目旨在将这项技术引入英国，同时也促进了人工智能在世界各地的大分子晶体光束线中的使用。首先，将法国国家同步加速器-Soleil训练的用于样品保持器和样品识别的神经网络整合到一个易于访问的模块中，以便在全球任何同步加速器使用，这将是巨大的好处。然后，该系统将通过利用我们先前图像的不同同步加速器数据库进行扩展，这些数据库将用于培训更先进的模型。保罗·谢勒研究所即将关闭的Swiss光源(SLS)创造了一个机会，他们的员工可以进行协作，他们计划的休假计划与我们的项目愿景非常一致。最后，该项目将对钻石2计划升级的路线图有很大帮助。

Macromolecular Crystallography is one of the most used techniques for the study of the most important molecular machines in biology - Proteins - as it allows us to determine the 3D structure of these molecules and infer their function. This is particularly relevant to and has proven results in addressing human diseases ranging from genetic disorders, cancers and fighting of human pathogens. This technique is also used in agricultural and food research areas like the development of novel herbicides or drought resistant crops to address current impacts of climate change. Finally, energy storage and battery technologies have also more recently benefited from crystallography synchrotron instruments helping key manufacturing and clean growth challenges of our era. Crystallography is used by a huge range of researchers from academic to industry pharmacological companies. These researchers often send their samples to large research facilities, like synchrotrons, and then collect X-ray diffraction data remotely or use fully automated systems. With recent advances in synchrotron technology the bottlenecks have moved from the lack of intensity of the synchrotron X-rays or the speed of the detector technology to the hardware and software that makes the sample visible to X-rays by centering the sample and preparing it for data collection. A data collection on a single crystal usually takes less than 10 seconds but all the other tasks bring the time per sample to ~2 minutes. Recent advances in AI have created a paradigm shift in image analysis. There are already a few prototypes in synchrotron facilities outside of the UK using AI to improve the speed and reliability of these essential tasks. We propose to use one of the proven prototypes and further develop it for sample centring, synchrotron X-ray beamline diagnostics, and robot collision risk mitigation. This will be extremely beneficial for the MX beamlines at the UK national Synchrotron - Diamond Light Source (DLS). Many DLS sister facilities can benefit from the application of AI but lack the "know-how" to implement working AI code from scratch. This project aims to bring the technology to the UK but also facilitate the usage of AI in macromolecular crystallography beamlines across the world. Starting by integrating the French national Synchrotron - SOLEIL - trained neural network for sample holder and sample identification into an easily accessible module for use at any synchrotron worldwide would be of huge benefit. This system will then be extended by leveraging our different synchrotron databases of prior images that will be used to train even more advanced models. The coming SwissLight Source (SLS) shutdown at the Paul Scherrer Institute creates an opportunity where their staff are available for collaborations and their planned sabbatical program aligns strongly with our project vision. Finally, this project would help significantly with the roadmap for the Diamond 2 planned upgrade.