This is a test version of Biostars. For the public version, visit https://www.biostars.org.
AutoDock Vina Gives Different Results From Different Simulations Using Same Parameters, and Different From Experimental Results

Hello everyone,

When I run multiple Autodock Vina independent simulations (at exhaustiveness of 100) on the SAME protein and the SAME ligand using the exact SAME parameters but with different random seeds I get different results. Each line here is the highest binding affinity of a separate docking simulation. You can see that every time I re-dock the same protein-ligand I get large differences in "the best" binding affinity (-3 to -37):

                 mode |   affinity | dist from best mode
                      | (kcal/mol) | rmsd l.b.| rmsd u.b.
                 -----+------------+----------+----------
Simulation 1        1       -10.35          0          0      random seed: 1615206230
Simulation 2        1       -10.46          0          0      random seed: -273847999
Simulation 3        1        -10.8          0          0      random seed: 2071239951
Simulation 4        1       -11.76          0          0      random seed: 1725032250
Simulation 5        1       -1.665          0          0      random seed: -172889668
Simulation 6        1       -17.23          0          0      random seed: 1076344612
Simulation 7        1       -2.826          0          0      random seed: -16921503
Simulation 8        1       -3.042          0          0      random seed: -366290784
Simulation 9        1        -3.26          0          0      random seed: -1254578750
Simulation 10       1       -3.584          0          0      random seed: -1014984749
Simulation 11       1       -3.614          0          0      random seed: 1532520213
Simulation 12       1       -3.619          0          0      random seed: 1182472761
Simulation 13       1       -3.666          0          0      random seed: -1188328766
Simulation 14       1       -36.89          0          0      random seed: 1542674337
Simulation 15       1       -37.26          0          0      random seed: 607682916

  • The search space is not on the surface, it is actually a tube within a trimer (image here), you can see that the search box volume is small (1,400 angstrom^2).

  • If I perform 250 separate docking simulations each at 32 exhaustiveness I get these distribution results for 3 molecules (results here).

  • f we take the mode of the distribution curves we get:

MOL6 = -14.6 Kcal/mol while the paper reported -10.97 Kcal/mol (Kd = 9e-9 M)

MOL7 = -10.8 Kcal/mol while the paper reported -10.75 Kcal/mol (Kd = 1.3e-8 M)

MOL8 = -11.3 Kcal/mol while the paper reported -6.35 Kcal/mol (Kd = 2.2e-5 M)

The paper's experimental measurements were done using the Biacore. I used the following formula to convert from the paper's Kd to dG [dG=RTlnKd]:

import math
def affinity(Kd, R=0.00198720425864083, T=298):
    dG = R*T*math.log(Kd)
    print(f'dG = {dG} Kcal/mol')

Questions:

  1. How do we interpret these data? what does it mean that the computed affinity is different from the experimental affinity? is it possible to consolidate the computed affinity with the experimentally determined affinity?

  2. If I want to run a virtual screen to find molecules that bind with a stronger affinity than MOL6, MOL7, or MOL8, which baseline should I use, the computed affinity or the experimental affinity?

screening vina docking virtual affinity binding autodock discovery drug
• 1,238 views • link
updated by Matic Broz • Matic Broz • written by ac.research • ac.research

1 answer

This is expected to some extent, but the size of the spread you show is a warning sign that the setup/search may not be well behaved.

Vina uses a stochastic search, so different random seeds can give different poses and scores. Small differences are normal. Very large swings, such as -3 to -37 kcal/mol for the same receptor/ligand/box, usually mean you should inspect the poses rather than treat the top score as a stable affinity estimate.

A few checks I would do:

  1. Compare the top poses from different seeds, not just the scores. Are the very negative poses in the same physical pocket, or are they buried in an unrealistic region, overlapping protein atoms, or exploiting a narrow cavity?

  2. Cluster the poses by RMSD or interaction pattern. If the same pose family appears repeatedly and has similar scores, that is more meaningful than the single best score from one seed.

  3. Increase exhaustiveness and check convergence, but do not assume high exhaustiveness alone fixes a bad search space. A small or awkwardly shaped tunnel/cavity can still be difficult for the search.

  4. Redock a ligand with a known experimental pose, if available. If Vina cannot recover the known pose under your preparation/grid settings, the protocol is not reliable enough for prospective ranking.

  5. Recheck receptor and ligand preparation: protonation state, tautomer, formal charges, missing atoms, cofactors/metals, waters that mediate binding, and whether the receptor conformation matches the experimental binding state.

For the comparison to experiment: Vina scores are not experimental binding free energies. They are approximate scoring-function values that are often useful for rough ranking within a consistent protocol, but they are not expected to reproduce Biacore-derived dG values compound by compound. Your Kd-to-dG conversion is fine in principle, but the docking score and experimental dG are not measuring the same model of the system.

For virtual screening, I would not use the experimental affinity as a direct score cutoff inside Vina. I would use the known ligands as controls: dock MOL6/MOL7/MOL8 with the same protocol, check whether their poses are plausible, then rank new compounds relative to those computed controls and prioritize compounds that have both better or comparable scores and chemically plausible poses/interactions. If MOL8 is experimentally weak but scores well computationally, that is useful evidence that this protocol may not discriminate your series reliably.

If setup/file handling is the blocker, ProteinIQ's Vina page may help; I work on ProteinIQ and this is one of the workflows we support. But the main scientific point here is that repeated runs and experimental disagreement should be interpreted through pose consistency, controls, and receptor/ligand preparation rather than score alone.

• 0 views • link
written by Matic Broz • Matic Broz

Log in to answer this question.

More posts like this