Much smarter and faster ligand discovery: Iterative, rational optimization of screening and follow-up libraries for the XChem fragment approach

更智能、更快速的配体发现：通过从3D结构数据中进行深度学习，对XChem片段方法的筛选和后续库进行迭代、合理优化基于片段的药物设计（FBDD）现在已成为早期药物发现的一种强大方法，并且在其他方法失败的困难目标方面取得了成功。这是由于片段（约为药物的三分之一大小）与靶标结合的可能性更高，与药物大小的配体相比，相互作用更有效。基于片段的配体开发方法虽然在经验丰富的组织中已经建立并相对强大，但尚未实现其真正的变革潜力，即使定制和有效的配体廉价（<1万英镑）和快速（数周）地广泛获得。即使是最近非常重要的旨在扩大获取的公共和商业投资，包括Diamond的XChem设施或Enamine的真实的廉价化合物，也没有从根本上改变游戏：开发一种有效的配体仍然需要花费约50万英镑。该项目的重点是在实验的一开始就变得更加聪明：找出如何确保起始化合物尽可能产生快速发展所需的所有信息，尽可能少的实验工作。这种雄心并不新鲜，但新鲜的是，（5年的XChem数据）（已经超过150个实验），以及新的深度学习方法来理解蛋白质-配体相互作用，新的合成空间描述也即将进入视野。到整个DPhil项目结束时，我们设想实现（a）更好的筛选化合物的一般集合;（B）用于为任何给定靶选择最佳筛选化合物组的方法，即使先前没有解决配体结合结构;（c）用于选择最佳下一组化合物的算法，以允许从非常小的初始筛选进行有效迭代;以及（d）允许对于充分研究的蛋白质类别完全绕过筛选步骤。该项目同样基于统计部和钻石光源，并且福尔斯属于EPSRC生物信息学研究领域。

Much smarter and faster ligand discovery: Iterative, rational optimization of screening and follow-up libraries for the XChem fragment approach with deep learning from 3D structural dataFragment-based drug design (FBDD) is now well-established as a powerful approach to early-stage drug discovery, and has a track record of success for difficult targets where other methods have failed. This is due to the higher likelihood that a fragment (~third size of drug) will bind to the target, with more efficient interactions compared to a drug-sized ligand. These weakly binding fragments are then linked together to create a potent drug.Fragment-based approaches to ligand development, though well-established and comparatively powerful in experienced organisations, have yet to achieve their true transformative potential of making bespoke and potent ligands widely accessible cheaply (<£10k) and quickly (weeks). Even very significant recent public and commercial investments aimed at widening access, including Diamond's XChem facility or Enamine's REAL cheaply available compounds, have not fundamentally changed the game: developing a potent ligand still costs ~£0.5m. The focus of this project is to get much smarter at the very outset of the experiment: to figure out how to ensure the starting compounds are as likely as possible to yield all the information necessary to progress rapidly to potency, with as little experimental work as possible. The ambition is not new, but what is new are a vast trove (5 years' worth) of XChem data (already >150 experiments), and new Deep Learning approaches to understanding protein-ligand interactions, with new descriptions of synthetic space also coming into view.By the end of the full DPhil project, we envisage achieving (a) a far better general set of screening compounds; (b) an approach for selecting the optimal set of screening compounds for any given target, even if no ligand-bound structure has been solved previously; (c) algorithms to select the best next set of compounds, to allow effective iterations from very small initial screens; and (d) allow bypassing the screening step entirely for well-studied classes of protein. This project is based equally in the Department of Statistics and at Diamond Light Source, and falls within the EPSRC Biological Informatics research area.