Seven years of ISOLDE: Lessons learned and paths forward - Tristan Croll скачать в хорошем качестве

Seven years of ISOLDE: Lessons learned and paths forward - Tristan Croll

Starting with its first beta release in 2018, ISOLDE has been designed explicitly to ease the challenges unique to working with less-than-ideal crystallographic or cryo-EM data, in the regime where density combined with traditional bonded restraints are insufficient for reliable model building (i.e. resolutions below about 2-2.5 Å. The more complete physical representation provided by molecular dynamics (MD) force fields offers clear advantages here, at the cost of dramatically higher computational complexity. Once an insurmountable barrier to the dream of interactive MD, this is now easily handled by modern consumer GPUs and is the core engine underlying ISOLDE. Building on this base, I added a number of key innovations aimed at reducing the number of rebuild/refine/validate cycles needed for a typical model – in particular (a) providing live geometry validation markup on the model representation, and (b) for crystallographic models, performing background, parallelised structure factor calculations to constantly update the maps as the model evolves. In addition ISOLDE includes its own take on well-trodden ideas such as the use of reference-model restraints to help with particularly challenging density. In this talk, I will discuss the current state of ISOLDE and its day-to-day applications, and some of the lessons I have learned during its development. I will also explore some of the challenges and limitations introduced by its reliance on the ageing AMBER force field, and some of the exciting progress made towards much more general approaches to MD parameterisation. Finally, I will consider the extraordinary challenges arising from the impending data deluge as the rate of information coming from cryo-EM experiments begins to overwhelm the capacity of humans to perform gold-standard model building, and the increasingly urgent need for next-generation validation approaches reliable enough to allow users to confidently skip over well-modelled regions.