Figure 1.
Performance of rigid-body docking.
(a) Justification for decoy set size. The top 20 CCD (▵) or RMSD (▿) ranks for the N = 47 complexes in the benchmark set are plotted against the size of the decoy set generated using molecules in the bound conformation. (b,c,d) The performance for the five docking trials using different combinations of bound and unbound conformations: bound protein/bound DNA (•, B/B), bound, but with rebuilt side chains (▪, S/B), unbound/bound (▿, U/B), bound/unbound (▵, B/U), and unbound/unbound (○, U/U). A total of 105 docked conformations are generated for each complex. (b) Fraction of the complexes having an RMSD of the 20th best decoy better than the abscissa (e.g., 50 percent of the benchmark can be rebuilt to within 10 Å using the unbound forms as marked with the two intersecting grey lines). (c) Same as (b) but for fraction of native contacts. The higher the score the better; docking would have the value 1.0. For 80 percent of the benchmark, decoys have at least 30% of the native contacts (intersecting grey lines). (d) Same as (b) but for the MCC. The true positives (TP), false positives (FP) and false negatives (FN) are computed by comparing the protein residues in contact with DNA in the crystal structure compared to a docked model. Higher values are better; a perfect docking would score1.0. As much as 90 percent of the benchmark decoys have MCC greater than 0.6.
Figure 2.
Protein and DNA representation and the Chemical Context Profile.
(a) Interacting centers for nucleic acids. For each of the nucleic acids, A:adenine, C:cytosine, G:guanine, U:uracil and T:thymine, the three moieties used to describe a nucleotide are highlighted by shaded disks. Each nucleotide has three interacting centers; one in the phosphate group and one in each of the grooves, major and minor. The nucleic acids are paired in the canonical Watson-Crick configuration to expose the positions of the two double-helical grooves. The thick curved lines represent the major grooves (over the base pairs), while the thin ones the minor grooves (under the base pairs). (b) Comparison of the CCP magnitude with the loss of accessible surface area upon docking using complex 1A0A with a 103 poses decoy set. The area loss is computed as the area of the complex minus the area of the isolated protein and DNA, using the msms computer program (ref). The squared Pearson correlation coefficient is 0.6. (c,d,e,f) CCD versus RMSD for 105 decoys. The CCD correlates with the RMSD when the RMSD values are low (e.g. <10 Å). (c) The non palindromic DNA (2IRF). (d) The DNA sequence (1ZME) has two palindromic regions at both ends (e) The DNA (1Z9C) has a palindromic sequence. (f) Expanded version of (c,d,e) near the origin into the zero closed triangles, 2IRF; open circles, 1ZME; and closed boxes, 1Z9C. The Pearson's R-Squared values for linear fits are provided.
Table 1.
Scoring function components optimized in this work.
Figure 3.
Performance of various scoring functions for identifying DNA binding.
The scoring functions are statistical contacts (a,b), statistical distance-dependent (c,d) and the one derived here (e,f, bottom row). Three decoy sets are used, known DNA-binding proteins (•), proteins with isoelectric points (pI) lower than 7 (▿) and greater than 7 (▵). Left panels (a,c,e) illustrate how the protein-DNA complexes are scored in relation to the pI of the proteins. Right panels (b,d,f) illustrate how the scores of the three decoy sets overlap with one another: a perfect scoring function would be able to systematically score authentic DNA-binding proteins from those that do not bind to DNA (no overlap). The blue region highlights the overlap between the known DNA-binding proteins with those proteins that have a pI greater than 7. The tick marks at the top of the plots indicate the scores of the known DNA-binding proteins, while those at the bottom are for the two other protein sets.
Table 2.
Performance of various scoring functions.
Figure 4.
Performance of various scoring functions for identifying the native binding pose.
The scoring functions are random (▪), statistical contacts (▾), statistical distance-dependent (▴) and our work (•). Two decoys sets are used for the evaluation: both protein and DNA in the bound form (closed figures) or both in the unbound form (opened figures). Left panes (ace) are results with the native DNA sequences, while the right panes (b,d,f) features “flipped” sequences. For each decoy set in the benchmark, the decoys are scored by the specified scoring functions (rows), then sorted from best to worst: lowest RMSD in top row, highest fraction of native contacts in middle row, and highest protein MCC in the bottom row, reporting the 5th value in the top 50 score. (a,b) DNA RMSD. (c,d) Fraction of native contacts. (e,f) Protein MCC. The true positives (TP), false positives (FP) and false negatives (FN) are computed by comparing the protein residues in contact with DNA in the crystal structure compared to a redocked model.
Figure 5.
Protein DNA-binding potential.
A protein's DNA-binding potential is revealed using one of the test moieties. Each test site is represented by a ball, whose color varies blue (favorable) to red (unfavorable). The first three proteins are known to bind DNA, while the last two are not known to bind DNA. For DNA-binding proteins, a graph tracks the potential for each type of moiety along the actual DNA coordinates. The blue curve tracks the 5′ strand of the DNA (tagged with a star), while the green curve tracks the 3′ strand. A point marks each base pair step, and a black horizontal line each 5 base pair steps. Since the 5′ phosphate is absent for both the 5′ and 3′ strands, the base pair step index starts at two. The potential is more favorable as the curve is more to the left. The x axis is scaled to show the relative change along the DNA molecule.