Deep probabilistic models for analysing complex DNA structures in high-resolution atomic force microscopy images.

罗莎琳德富兰克林建立DNA原子结构的开创性工作，巩固了我们对“生命分子”的理解。基因组DNA向细胞核中的压缩导致显著的拓扑应力和卷曲、扭曲和打结的DNA结构的形成，其影响细胞活力，从DNA复制到癌症和感染中的治疗剂的活性具有分支。了解这些复杂的DNA结构如何影响DNA加工的挑战从根本上受到可用工具的限制。高分辨率原子力显微镜（AFM）的独特之处在于，它能够以纳米分辨率提供有关液体中DNA结构、功能和动力学的定量信息，无需标记或平均[1]，然而，到目前为止，这些数据集的分析依赖于有经验的显微镜专家的眼睛[2]。尽管AFM生成的数据集越来越大，但自动分析和/或机器学习技术并没有得到常规应用。机器学习已经推动了我们对生物现象的理解（例如AlphaFold）。使用人工神经网络的深度学习已被应用于使用相邻显微镜（特别是其分辨率革命中的cryo-EM）产生的数据集，以解决以前无法解决的生物学问题。高斯过程（GPs）是另一种重要的机器学习技术，在数据不太丰富的情况下很有用，并且更多地了解正在建模的系统的行为（例如DNA力学）。我们建议使用这些和类似的技术相结合，适应和改进它们在复杂的生物AFM数据集的分析。

Rosalind Franklin's pioneering work to establish the atomic structure of DNA has underpinned much of our understanding of the 'molecule of life'. The compaction of genomic DNA into the nucleus results in significant topological stress and the formation of coiled, twisted and knotted DNA structures which impact cell viability, with ramifications from DNA replication to the activity of therapeutic agents in cancer and infection. The challenge of understanding how these complex DNA structures influence DNA processing has been fundamentally limited by the tools available.High-resolution atomic force microscopy (AFM) is unique in its ability to provide quantitative information on DNA structure, function and kinetics in liquid with nanometre resolution without labelling or averaging [1], however the analysis of these datasets has until now relied on the eye of an experienced microscopist [2]. Despite the increasing size of datasets generated by AFM, automated analysis and/or machine learning techniques are not routinely applied. Machine learning has driven step changes in our understanding of biological phenomena (e.g. AlphaFold). Deep learning using artificial neural networks has been applied to datasets produced with adjacent microscopies (notably cryo-EM in its resolution revolution), to solve previously inaccessible biological problems. Gaussian processes (GPs) are another important machine learning technique, useful in situations where data is less abundant and more is known about the behaviour of the system being modelled (e.g. DNA mechanics). We propose to use a combination of these and similar techniques, adapting and improving them in analysis of complex bio-AFM datasets.