Fig. 2: D-I-TASSER modeling results on 500 hard nonredundant domains.
From: Deep-learning-based single-domain and multidomain protein structure prediction with D-I-TASSER
a–c, TM scores of the first-rank models built by D-I-TASSER versus those of I-TASSER (a), C-I-TASSER (b) and AlphaFold2 (c). d, TM score comparisons of I-TASSER with different deep learning potentials and Alphafold2 versions, where ‘I-TASSER + DeepPotential + AttentionPotential + AlphaFold2 distances’ is equivalent to D-I-TASSER. The height of the histogram indicates the mean value, and the error bar depicts s.d. e, Structure superposition of the best LOMETS template (PDB ID: 4cvhA) over the target structure (PDB ID: 3fpiA). f, Structure superposition of the first D-I-TASSER model with the target structure. g, Comparison of inter-residue distance map predicted from deep learning models (upper triangle) and the distance map calculated from the target structure (lower triangle) for PDB ID: 3fpiA. h, Trajectory of TM scores and MAEm during the REMC cycles of the replica that starts with template PDB ID: 4cvhA. The structures are decoy models taken from different simulation steps. i, Structure superposition of the AlphaFold2 model over the target structure (PDB ID: 4jgnA). j, Structure superposition of the D-I-TASSER model with the target structure (PDB ID: 4jgnA). k–m, Comparisons of inter-residue distance map from the target structure (lower triangle) for PDB ID: 4jgnA versus the predicted distance maps (lower triangle) by standard AlphaFold2 (k), AlphaFold2 with DeepMSA2 MSA (l) and D-I-TASSER assembly (m).
