When researchers choose compounds to test, they often compare scores from computational methods. In this example, we compared ten candidate compounds. The candidate ranked first in the initial assessment fell to sixth in the next calculation. The candidate ranked eighth moved to first. The candidate list stayed the same; the evaluation method and input structures changed.
These are Nadoo Bio calculation records from October 22, 2026. We first selected candidates with matching chemical features, then scored specific binding poses. Comparing the two candidates whose positions changed, together with their scores before and after docking, helps explain what each calculation evaluates.
Matching features and finding poses
A protein can have a site where a small molecule binds. Chemical features that may contribute to binding, and their arrangement in space, can be distilled into a model called a pharmacophore. PharmacoNet evaluates how well a compound’s features match this model. In this run, we selected the ten highest-scoring candidates. [1]
We then used Vina to search for the compound’s position and orientation at a particular site on the protein, as well as its internal conformation. This search is docking, and a resulting arrangement is a binding pose. The workflow moves from matching features to evaluating a specific arrangement of atoms. [2]
Finally, we supplied the structures produced by Vina to PIGNet2. PIGNet2 takes a protein and compound structure and predicts binding affinity—how strongly the two molecules bind. Its output here is a predicted score, not an experimentally measured binding strength. [3]
Figure 1. An explanatory diagram connecting the inputs and outputs of this workflow. The protein outline simplifies the coordinates of chain A in the public structure 6OIM. It does not depict a candidate’s docking result. [4]
Higher PharmacoNet scores are preferred, whereas lower scores are preferred in this Vina and PIGNet2 comparison. We did not combine the three values into an overall score. The methods evaluate different things, and even though Vina and PIGNet2 use the same energy unit, their numerical magnitudes cannot be compared directly.
First to sixth, eighth to first
This example used the public structure 6OIM. Only the ten candidates selected with PharmacoNet were evaluated with Vina and PIGNet2. The rankings below therefore apply to these ten selected candidates.
Figure 2. Lines connect identical candidate IDs. The IDs for 37632291 (blue) and 307754643 (orange) appear on both sides. Vertical position represents rank within each tool, not binding strength.
Candidate 37632291 had the highest PharmacoNet score, 159.8. In Vina, however, its score was −5.54, placing it sixth. Candidate 307754643 ranked eighth in PharmacoNet with 90.4 but first in Vina with −6.22. Its post-docking PIGNet2 score, −4.1, was also the lowest.
This comparison does not include experimental measurements obtained under matching conditions. We therefore cannot tell whether 307754643, the computational leader, actually binds best.
Vina and PIGNet2 selected the same top candidate, but PIGNet2 evaluated structures produced by Vina. This is agreement between two assessments that share structural input.
One candidate, scores of +64.4 and −4.1
The records also include PIGNet2 scores for structures before docking. Candidate 307754643 scored +64.4 before docking and −4.1 afterward. Candidate 37632291 changed from +62.4 to −2.6. All ten candidates had positive scores before docking and negative scores afterward.
Figure 3. Scores for each candidate’s inputs before and after docking, in kcal/mol. These are not experimentally measured energy changes. Blue points show pre-docking scores; orange points show post-docking scores.
Why can the score change so much when the compound ID stays the same? A structure-based model does not evaluate a compound by its name alone. Its position and orientation relative to the protein, and its internal conformation, are also part of the input. How hydrogen atoms and charges are handled during file preparation can affect the result too.
In this run, we docked the structures, converted the files and then rescored them. Because both steps changed, we cannot isolate the entire score difference as the effect of molecular placement alone. What this example shows is that scores for the same compound changed substantially after its input structure and preparation procedure changed.
The official PIGNet2 instructions recommend structures optimized with the docking tool SMINA and evaluation using four models. We used Vina structures, so our procedure differs from that recommendation. [3]
What the two candidates leave us to investigate
The two candidates call for different follow-up comparisons. For 37632291, the structural question is why its advantage in feature matching did not carry through to its post-docking rank. For 307754643, the question is whether similar poses and rankings persist when search settings change or calculations are repeated.
Keeping only PharmacoNet’s top candidate would have excluded 307754643. Carrying ten candidates through docking revealed that it led both subsequent assessments. How many candidates survive the initial selection determines which ones can be compared later. Candidate 307754643 illustrates that effect here. Which candidate performs better experimentally remains unknown.
Values and conditions used in the comparison
Candidate ID | PharmacoNet | Vina | PIGNet2 after docking |
|---|---|---|---|
37632291 | 159.8 | −5.54 | −2.6 |
62719435 | 121.2 | −5.28 | −3.9 |
307710026 | 115.2 | −5.88 | −2.9 |
217657246 | 104.1 | −4.95 | −3.0 |
1893000282 | 97.2 | −5.92 | −3.6 |
144895264 | 95.7 | −5.87 | −4.0 |
146251011 | 90.7 | −5.94 | −3.3 |
307754643 | 90.4 | −6.22 | −4.1 |
1952375463 | 90.1 | −5.53 | −2.6 |
2209318782 | 84.5 | −5.48 | −3.2 |
Nadoo Bio calculation records, October 22, 2026. Vina and PIGNet2 values are in kcal/mol. IDs are identifiers from the execution files. PIGNet2 values are shown to one decimal place; candidates with identical displayed values are not assigned distinct fine-grained ranks here.
We used Vina 1.2.3. For PIGNet2, we selected each candidate’s lowest score within each of four models and averaged those four values. The lowest-scoring pose may differ between models, so this is distinct from averaging four evaluations of one identical pose. Hydrogen and charge preparation was also modified for the execution environment. We therefore do not present these absolute values as a reproduction of the official example. This is a case study of ten candidates selected for one structure, not a benchmark of the tools’ general performance.
Our earlier analysis of public KDDF projects examined the share of research devoted to finding early candidates. This example shows how the set of candidates worth examining can change as additional calculations are applied after that initial selection.
References
PharmacoNet developers. Official PharmacoNet README. Feature models, virtual screening and input guidance. Checked October 26, 2026.
AutoDock Vina developers. Basic docking. Official user documentation; documentation source in the official repository checked October 26, 2026.
PIGNet2 developers. Official PIGNet2 README. Recommended inputs, four models and output columns. Checked October 26, 2026.
RCSB PDB. Structure record 6OIM. Public experimental structure. PDB coordinates checked October 26, 2026; used for the protein outline in Figure 1.