Deep learning-based data analysis methods have gained considerable attention in all fields of science over the last decade. In recent years, this trend has reached the single-molecule community. In this review, we will survey significant contributions of the application of deep learning in single-molecule imaging experiments. Additionally, we will describe the historical events that led to the development of modern deep learning methods, summarize the fundamental concepts of deep learning, and highlight the importance of proper data composition for accurate, unbiased results.

The mammalian glycocalyx is a heavily glycosylated extramembrane compartment found on nearly every cell. Despite its relevance in both health and disease, studies of the glycocalyx remain hampered by a paucity of methods to spatially classify its components. We combine metabolic labeling, bioorthogonal chemistry, and super-resolution localization microscopy to image two constituents of cell-surface glycans, N-acetylgalactosamine (GalNAc) and sialic acid, with 10-20 nm precision in 2D and 3D. This approach enables two measurements: glycocalyx height and the distribution of individual sugars distal from the membrane. These measurements show that the glycocalyx exhibits nanoscale architecture, on both cell lines and primary human tumor cells. Additionally, we observe enhanced glycocalyx height in response to epithelial-tomesenchymal transition and to oncogenic KRAS activation. In the latter case, we trace increased height to an effector gene, GALNT7. These data highlight the power of advanced imaging methods to provide molecular and functional insights into glycocalyx biology.

Background fluorescence, especially when it exhibits undesired spatial features, is a primary factor for reduced image quality in optical microscopy. Structured background is particularly detrimental when analyzing single-molecule images for 3D localization microscopy or single-molecule tracking. Here, we introduce BGnet, a deep neural network with a U-net-type architecture, as a general method to rapidly estimate the background underlying the image of a point source with excellent accuracy, even when point spread function (PSF) engineering is in use to create complex PSF shapes. We trained BGnet to extract the background from images of various PSFs and show that the identification is accurate for a wide range of different interfering background structures constructed from many spatial frequencies. Furthermore, we demonstrate that the obtained background-corrected PSF images, both for simulated and experimental data, lead to a substantial improvement in localization precision. Finally, we verify that structured background estimation with BGnet results in higher quality of super-resolution reconstructions of biological structures. Significance:A main factor that degrades image quality in fluorescence microscopy is unwanted background fluorescence. Background is almost never uniform, especially in complex samples. Rather, background usually exhibits some structure, making it very difficult to distinguish from the signal of interest. Due to this challenging problem, background fluorescence is often assumed to be uniform even though it is not. This assumption leads to deteriorated image quality, e.g. in localization-based superresolution microscopy, and to the introduction of uncharacterized biases. To overcome this challenge, we developed a general framework rooted in deep learning to accurately and rapidly estimate arbitrarily structured background using several distinct image shapes for the single emitter. Proper background estimation allows for critical performance improvement of various optical microscopy methods. Main Text:

We present a systematic isotope labeling study of the protein G mutant NuG2b as a step toward the production of reliable, structurally stable, experimental standards for amide I infrared spectroscopic simulations. By introducing isotope enriched amino acids into a minimal growth medium during bacterial expression, we induce uniform labeling of the amide bonds following specific amino acids, avoiding the need for chemical peptide synthesis. We use experimental data to test several common amide I frequency maps and explore the influence of various factors on map performance. Comparison of the predicted absorption frequencies for the four maps tested with empirical assignments to our experimental spectra yields a root-mean-square error of 6-12 cm(-1), with outliers of at least 12 cm(-1) in all models. This means that the predictions may be useful for predicting general trends such as changes in hydrogen bonding configuration; however, for finer structural constraints or absolute frequency assignments, the models are unreliable. The results indicate the need for careful testing of existing literature maps and shed light on possible next steps for the development of quantitative spectral maps.