Fast DNA-PAINT imaging using a deep neural network

Abstract

DNA points accumulation for imaging in nanoscale topography (DNA-PAINT) is a super-resolution technique with relatively easy-to-implement multi-target imaging. However, image acquisition is slow as sufficient statistical data has to be generated from spatio-temporally isolated single emitters. Here, we train the neural network (NN) DeepSTORM to predict fluorophore positions from high emitter density DNA-PAINT data. This achieves image acquisition in one minute. We demonstrate multi-colour super-resolution imaging of structure-conserved semi-thin neuronal tissue and imaging of large samples. This improvement can be integrated into any single-molecule imaging modality to enable fast single-molecule super-resolution microscopy.

Multi-color super-resolution microscopy accelerated by a neural network

“…Experimental details on sample preparation, image acquisition and analysis are reported in detail in Narayanasamy et al 12 .…”

“…SMLM frames were acquired using multi-dimensional acquisition (MDA) mode in µManager 2.0 16 . DNA-PAINT imaging was conducted with DNA-labeled antibodies and using fluorophore-labeled imager strand concentrations of of 20 pM (training data), 0.2 -0.5 nM (ground truth) and 5 -20 nM (testing data) (for more information, we refer to 12 ). SMLM images were reconstructed with Picasso 10 .…”

“…Training took 35 min with ColabPro. Image analysis was conducted with Fiji 19 , multi-scale SSIM 20 , SQUIRREL 21 and HAWKMAN 22 ; full details are available in Narayanasamy et al 12 . To determine the quality of DeepSTORM predicted images, ground truth images were generated from low density emitter data (0.5 nM imager strands) for 10,000 frames corresponding to 25 minutes imaging time.…”

“…Multi-target imaging in such tissue samples was reported previously using antibody exchange 32 , or DNA barcoding (DNA-PAINT) 13 . By implementing the DeepSTORM NN into the SMLM imaging workflow, the acquisition time for single-molecule raw data could be substantially reduced up to ~20-fold, and down to one minute per data set 12 . Using Exchange PAINT 14 and a trained model, multi-color super-resolution images were predicted (bottom) and compared to ground truth images (top) for glial fibrillary acidic proteins (GFAP; red) and neurofilament M (NF-M; cyan; Figure 2A) which shows the distribution of astrocytes and neurons in the neuronal tissue section, α-tubulin fibrillar (Figure 2B) and bundled axonal structures (red; Figure 2C) with mitochondria labeled with TOM20 (cyan), and the synaptic active zone scaffold proteins Homer (red) and Bassoon (cyan; Figure 2D).…”

“…Here, we report fast single-molecule super-resolution microscopy by utilizing a neural network (NN), DeepSTORM 11 , to predict super-resolution SMLM images from high-density DNA-PAINT data 12 . First, the neural network was trained with sparse emitter density DNA-PAINT data.…”