Fig. 2: Classification of experimental DNA origami datasets.

a Templates of four DNA-origami designs in the digits dataset: the digits 1, 2, and 3 and a 20 nm grid. b Fusion result without classification of 2000 particles (randomly selected from a set of 10,000 images, 2500 per class). This data is imaged separately per class and combined into one dataset prior to the classification. c–f The four classes resulting from the classification of 5000 images (randomly selected from a set of 10,000 images, 2500 per class) containing 1374, 1179, 1214, and 1233 particles, respectively. g Confusion matrix of the classifications c–f with an overall performance of 96.4%. h Fusion result without classification of 2000 particles imaged in one FOV. i–l The four classes resulting from the classification of 5000 particles images in one FOV, containing 1219, 1309, 1278, and 1194 particles, respectively. m Confusion matrix of the classifications i–l, with an overall performance of 96.4%. n Templates of three DNA-origami designs in the letters dataset: letters T, O, and L. o Fusion result without classification of 600 particles (200 per class). This data is imaged separately per class and combined into one dataset prior to the classification. p–s The four classes resulting from the classification of o, containing 207, 122, 176, and 95 particles, respectively. t Average confusion matrix of two independent classifications performed as in p–s with an average performance of 97.3 ± 1.9%, where the class of misfolds, s, is not taken into account. u Fusion result without classification of 800 particles imaged in one FOV. v–z The five classes resulting from the classification of u, containing 170, 238, 130, 139, and 123 particles, respectively. Scale bar of b applies to c–f and h–l. Scale bar of o applies to p–s and u–z.