Fig. 3

Performance analysis of FA-DNN for protein property prediction on three independent test sets. The FA-DNN consists of a fragment library encoder and a protein property predictor for protein property prediction (red bars). As control, a model without the fragment library encoder was trained with the same hyperparameters (blue bars). The performance of the two neural networks were evaluated on all targets of the three test sets. Additionally, performance of simply extraction from fragment libraries was also evaluated (green bars). The mean absolute error (MAE) of torsion angles (ϕ, ψ, θ and τ) and the MAE of C_β − C_β pairwise distances are shown in the left and right subfigures, respectively