Fig 1.
The 3-stage hierarchical structure modeling protocol of DConStruct.
Fig 2.
TM-score distributions of the reconstructed models on 150 FRAGFOLD soluble protein domains using (A) FT-COMAR, CONFOLD, and DConStruct for true Cα–Cα contact maps; and (B) CONFOLD and DConStruct for true Cβ–Cβ contact maps.
Table 1.
Reconstruction performance of soluble proteins on 150 FRAGFOLD domains for true Cα-Cα contact maps.
The mean TM-score of top predicted models are reported. Values in bold represent the best performance.
Table 2.
Reconstruction performance of soluble proteins on 150 FRAGFOLD domains for true Cβ-Cβ contact maps.
The mean TM-score of top predicted models are reported. Values in bold represent the best performance.
Fig 3.
Superimpositions between the reconstructed models (rainbow) and the corresponding experimental structures (gray) for two soluble proteins (A) PDB ID 1gmi and chain A; and (B) PDB ID 1a3a and chain A; using DConStruct and CONFOLD for true Cα-Cα contact maps at 8, 10, and 12Å thresholds.
Fig 4.
Superimpositions between the reconstruction models (rainbow) and its native structure (gray) for two soluble proteins (A) PDB ID 1gmi and chain A; and (B) PDB ID 1a3a and chain A; generated by DConStruct with hybrid interaction maps at tri-level thresholding.
Table 3.
Folding performance of CASP free modeling protein targets on 40 CASP12 and CASP13 free modeling target domains.
Values in bold represent the best performance.
Fig 5.
Ribbon diagrams of 3D models for the four CASP FM targets: T0898-D1, T0870-D1, T0968s2-D1, and T0957s2-D1; predicted by DMPfold and DConStruct along with the experimental structures.
All molecules are rainbow colored blue to red from the N- to C-termini. Models are optimally superimposed on the experimental structures, and then separated by translations along the horizontal direction.
Fig 6.
Ribbon diagrams of 3D models for the three CASP FM large targets: T0864-D1, T0904-D1, and T0969-D1; predicted by the top CASP human and server predictors, trRosetta distance-only modeling, and DConStruct using trRosetta-predicted distance maps along with the experimental structures.
A7D prediction is relevant only for CASP13 target T0969-D1. All molecules are rainbow colored blue to red from the N- to C-termini. Models are optimally superimposed on the experimental structures, and then separated by translations along the horizontal direction.
Table 4.
Folding performance of DConStruct using trRosetta-predicted distance maps compared to top CASP predictors and trRosetta distance-only modeling on 40 CASP12 and CASP13 free modeling target domains.
Table 5.
Folding performance of 510 MPs using Xu’s deep transfer learning (DTL) with CNS and DConStruct.
Values in bold represent the best performance.
Fig 7.
Ribbon diagrams of 3D models of MPs predicted using Xu’s DTL with CNS and DConStruct (using DMPfold- and trRosetta-predicted distance maps) along with the experimental structures for five protein targets: PDB ID 4xu4 and chain A, PDB ID 5doq and chain B, PDB ID 4g1u and chain A, PDB ID 4m64 and chain A, and PDB ID 1o5w and chain A.
All molecules are rainbow-colored blue to red from the N- to C-termini. Models are optimally superimposed on the experimental structures and then separated by translations along the horizontal direction.
Fig 8.
Stagewise TM-score distributions of the reconstructed models on 15 proteins from the EVfold set.
Stage-by-stage (A) density plots and (B) violin plots are shown with the means indicated using the unfilled circles, the medians indicated using the horizontal yellow lines, and the interquartile ranges indicated using the vertical black strips.
Table 6.
Stage-by-stage 3D reconstruction results for 15 protein targets in EVfold dataset using true Cβ-Cβ contact maps at 8Å threshold and true secondary structuresa.
Fig 9.
Visualization of effect of the three stages adopted in DConStruct for a representative protein target from EVfold dataset (PDB ID 5p21 and chain A).
Fig 10.
Stagewise distributions of the pseudo-covalent bond lengths for the protein target (PDB ID 5p21 and chain A).
The inset magnifies the stage 1 and stage 2 distributions around the ideal value of 3.8Å.
Fig 11.
Stagewise distributions of the virtual bond and dihedral angles for residues in (A-B) alpha helices and (C-D) beta strands for the protein target (PDB ID 5p21 and chain A).
Table 7.
TM-score gain Δ2 (TM-score of stage 2 –TM-score of stage 1) and Δ3 (TM-score of stage 3 –TM-score of stage 2) after performing stage-by-stage 3D reconstruction for 15 protein targets in EVfold dataset using true Cβ-Cβ contact maps at 8, 10, and 12Å thresholds and true secondary structures.