(1) Select structures from the PDB corresponding to a functional site of interest.
(2) Superimpose the structures with DALI to create a multiple sequence alignment.
(3) Input the multiple sequence alignment to HMMER to train a Hidden Markov Model.
(4) Input the structures to FEATURE to build FEATURE models of the functional site.
(A) Score query sequences with the Hidden Markov Model, filtering out those with Evalue > 1E2.
(B) Simulate retained query structures using OpenMM Zephyr, extracting snapshots to generate structural ensembles.
(C) Score structural ensembles with the FEATURE models, using the scores to guide classification. Prediction results are shown in the table. The first column states the method and dataset (italicized) being evaluated. The second column lists the threshold used for each. The next three columns report the number of thioredoxin active sites found (column 3), number of thioredoxin active sites missed (column 4), and number of sites incorrectly predicted to be thioredoxin active sites (column 5). The last two columns summarize the performance of each method by its fidelity (column 6) and completeness (column 7). Numbers in parentheses reflect inclusion of the training data in the datasets.
See also and .

(A) Multiple sequence alignment of the six thioredoxin training proteins. Columns less than 25% gapped are shown, and universally conserved residues are in red. Sequence conservation scores diagramed below the alignment are gray (score < 7) or colored (score ≥ 7).
(B) Thioredoxin training structure 2FWH by sequence conservation score. Color scheme is identical to (A) with colored residues displayed using a ball-and-stick scheme. The three regions of high conservation correspond to (1) active site and (2–3) core structural sites.
(C) Distribution of BLAST pairwise alignment scores between thioredoxins of the SP-PDB95 dataset. Alignment scores between HMM true positives (TP--TP, blue) and HMM false negatives and true positives (FN--TP, red) are shown. The two distributions are significantly different with a p-value of 9.1E-37 from a two-sample Kolmogorov-Smirnov test. N indicates the number of alignments (E-value < 10 and alignment length ≥ 50) and μ describes the distribution mean.

(A) Schematic of the thioredoxin active site of training structure 2FWH. The environment around each active site cysteine sulfur atom is segmented into six concentric radial shells, each 1.25 Å thick.
(B) Thioredoxin active site of dynamic FEATURE true positive 2DJJ. Shown are the starting PDB conformation (left) and the highest scoring conformation from molecular dynamics (right).
(C) Thioredoxin active sites of dynamic FEATURE false negatives 1WMJ and 2ALB. In (A), (B), and (C), residues near the active site are shown using a ball-and-stick schematic and atoms within 7.5 Å of the active sulfur atoms are colored (C = green, N = blue, O = red, and S = yellow). The conserved cysteines and proline are highlighted by colored covalent bonds. FEATURE scores for the active cysteines are below the panels.
(D) RMSD distribution of conserved active site residues. Standard box plots show RMSD between the static training proteins and the simulation ensembles of the training proteins, 2DJJ, 1WMJ, and 2ALB. Simulation frames are split into high and low FEATURE scores for active sites scoring above or below model thresholds, respectively.
(E) Multi-site thioredoxin FEATURE model fingerprint. The fingerprint shows the physicochemical properties significantly over-represented (abundant, red) and under-represented (deficient, blue) in the local environment of the active cysteines. Color intensity reflects a property’s significance (p-value) with a minimal inclusion threshold of 0.01.

(A) FEATURE score profile for the N-terminal active site cysteine.
(B) FEATURE score profile for the C-terminal active site cysteine. In (A) and (B), FEATURE scores are plotted against simulation frame (gray). The starting scores (blue) and model thresholds (red) are also marked. Simulation frames that score above both model thresholds simultaneously are orange. The highlighted yellow regions correspond to .
(C) Histogram of active cysteine FEATURE scores. Grid intensity indicates the frequency of the score combination, according to the scale on the right. The dotted lines show the FEATURE model thresholds.
(D) Correlation between FEATURE scores and RMSD of conserved active site residues. Active cysteine FEATURE score combinations are plotted against their average RMSD from the static training structures. The RMSD is in angstroms and scales according to the color scale on the right. The dotted lines show the FEATURE model thresholds.
See also .