Fig. 4: Supervised DNN.
From: Determining structures of RNA conformers using AFM and deep neural networks
a, True versus estimated r.m.s.d. (r = 0.95) for the training-validation set. Inset, histogram of the residuals (true r.m.s.d. – predicted r.m.s.d.) with σ = 2.6 Å. b, Loss function versus training epoch for the training, training-validation and validation sets. The loss was evaluated after the end of each training epoch. c–g, Tests using the full trajectories from BM1 (c), BM2 (d), BM3 (e), BM4 (f) and BM5 (g) (excluding training and training-validation data), the mean square error (MSE) for the cohort of best models are highlighted in a blue dashed circle in the lower left corners. Approximate number of trajectory models: BM1, 10 million; BM2, 11 million; BM3, 7 million; BM4, 2 million; BM5, 13 million. h, Violin plots comparing the true r.m.s.d. values of the models from the UML-selected cohort with n = 7,450, 11,927, 10,867, 93 and 12,771 structure models, and the top 10 models from DNN (n = 10) for each benchmark. Maximum, mean and minimum values of the plots for UML and DNN, respectively: BM1 (12.0, 7.4, 5.7) and (7.8, 7.3, 6.6); BM2 (10.8, 6.5, 4.1) and (6.1, 5.5, 4.1); BM3 (7.2, 2.9, 1.8) and (3.2, 2.7, 2.3); BM4 (5.5, 4.8, 3.9) and (5.3, 4.8, 4.4); BM5 (9.3, 7.2, 5.8) and (7.4, 6.5, 6.2). i, Uncertainty of estimated accuracy, provided as root-mean-square error (r.m.s.e.), across random samples of all tests set of benchmarks. Dots indicate the average r.m.s.e. and vertical lines are s.d. for five bins (range indicated by horizontal lines) of estimated accuracy: 3–5 Å, 5–7 Å, 7–10 Å, 10–20 Å and 20–30 Å, n = 5.4 million total independent structures.
