Scoring Functions and Binding Affinity

Every pose that AutoDock Vina reports comes with a single number in kcal/mol called the predicted binding affinity. That number is produced by Vina's empirical scoring function, a weighted sum of physics-inspired terms whose weights were fitted against a training set of protein-ligand complexes with known structures and measured affinities. Knowing what feeds into it tells you exactly how much to trust it.

Anatomy of the Score

The score adds up a handful of pairwise atom-atom terms, each evaluated on the surface distance between a ligand atom and a receptor atom rather than the raw center-to-center distance. Two Gaussian terms reward close but non-clashing contact and stand in for attractive van der Waals forces, while a repulsion term penalizes atoms that overlap. A hydrophobic term rewards nonpolar atoms resting against each other, and a hydrogen-bond term rewards donor-acceptor pairs sitting at bonding distance.

# Vina intermolecular score: weighted sum over atom pairs
# d = surface distance = r - Ri - Rj  (r = center distance, R = atom radii)

c_inter =  -0.0356 * Gauss1(d)       # attractive vdW, sharp well
           -0.00516 * Gauss2(d)      # attractive vdW, broad well
           +0.840  * Repulsion(d)    # steric clash penalty (raises score)
           -0.0351 * Hydrophobic(d)  # nonpolar contact bonus
           -0.587  * HBond(d)        # hydrogen-bond bonus

# torsional entropy penalty divides the sum by a rotatable-bond term
affinity = c_inter / (1 + 0.0585 * N_rot)   # reported in kcal/mol

Because the van der Waals, hydrophobic, and hydrogen-bond weights are negative, favorable interactions pull the total downward, so a more negative affinity means stronger predicted binding. The intermolecular sum is then divided by a term that grows with the number of active rotatable bonds, which is Vina's torsional entropy penalty: a floppy ligand must freeze several bonds to bind, and that costs it. This is why an otherwise well-fitting pose of a very flexible molecule can still land at a modest score.

Reading the Vina Log

# Vina 1.2 removed the --log flag, so send stdout to a file with tee
vina \
  --receptor receptor.pdbqt \
  --ligand ligand.pdbqt \
  --center_x 15.0 --center_y 53.5 --center_z 12.0 \
  --size_x 22 --size_y 22 --size_z 22 \
  --exhaustiveness 8 \
  --out docked.pdbqt | tee vina_log.txt

AutoDock Vina v1.2.5

Scoring function : vina
Rigid receptor: receptor.pdbqt
Ligand: ligand.pdbqt
Grid center: X 15 Y 53.5 Z 12
Grid size  : X 22 Y 22 Z 22
Grid space : 0.375
Exhaustiveness: 8
CPU: 0
Verbosity: 1

Computing Vina grid ... done.
Performing docking (random seed: 1274908391) ...
0%   10   20   30   40   50   60   70   80   90   100%
|----|----|----|----|----|----|----|----|----|----|
***************************************************

mode |   affinity | dist from best mode
     | (kcal/mol) | rmsd l.b.| rmsd u.b.
-----+------------+----------+----------
   1       -9.234          0          0
   2       -8.719      2.015      3.128
   3       -8.507      1.876      2.442
   4       -7.892      3.045      5.211
   5       -7.634      2.333      3.874

Vina sorts poses from best to worst, so mode 1 always carries the lowest, most negative affinity, and the affinity column is the one you read. As a rough guide, scores near -6 kcal/mol are weak, -7 to -9 is typical for a genuine hit, and values past -10 are strong but worth double-checking for grid-box or protonation artifacts. Treat the gap between mode 1 and the runner-up as a hint of how confident the search is in its top pose.

# Pull just mode number and affinity out of the log
awk '/^-----/{seen=1; next} seen && $1 ~ /^[0-9]+$/ {print "pose "$1"  "$2" kcal/mol"}' vina_log.txt

pose 1  -9.234 kcal/mol
pose 2  -8.719 kcal/mol
pose 3  -8.507 kcal/mol
pose 4  -7.892 kcal/mol
pose 5  -7.634 kcal/mol