Short Story ◈ Drug Design

Behind the Docking Score

Understanding what computational chemistry scoring functions evaluate and what they overlook.

  • #docking
  • #scoring function
  • #binding affinity
  • #computational chemistry

"The docking score was good, but it didn't work at all in experiments..."

Sena reported to Lina with a discouraged expression.

"Happens often," Lina answered while looking at her laptop. "Scores aren't omnipotent."

"But minus 10 kcal/mol... I thought it was a really good number."

Akira sat next to her. "Do you understand what the scoring function calculates?"

"Binding strength... right?"

"That's the ultimate goal. But in reality, it's approximation upon approximation," Lina began explaining.

The screen displayed the scoring function equation.

"Van der Waals forces, electrostatic interactions, hydrogen bonds, hydrophobic effects... each has a coefficient."

"What about these coefficients?" Sena asked.

"Determined by fitting from known complex data," Akira answered. "In other words, empirical parameters."

"Not perfect...?"

"Far from it," Lina said frankly. "Entropy loss, solvation free energy, induced fit... many factors aren't calculated."

Akira supplemented. "Especially conformational entropy loss upon binding is difficult. The more flexible the molecule, the greater the cost of losing degrees of freedom."

"So flexible molecules are disadvantaged?"

"They're hard to reflect in scores. So rigid molecules are sometimes overestimated."

Lina opened another screen. "This is the same complex calculated with different scoring functions."

"The values are completely different..."

"Different functions emphasize different things. One emphasizes electrostatic interactions, another emphasizes hydrophobic effects."

Akira said, "So it's wise to evaluate with multiple functions. If all give good scores, reliability increases."

"But," Sena pondered, "how much correlation is there with experimental values?"

Lina displayed a graph. "In good cases, correlation coefficient around 0.7. In bad cases, below 0.3."

"Not very high..."

"It can be used for ranking," Akira explained. "Even if absolute values aren't reliable, relative superiority is often discernible."

"So a score of -10 versus -7 means -10 is more likely better, roughly?"

"That's the level of resolution," Lina acknowledged. "So you need to select multiple top candidates for experiments."

Sena took notes. "Can't judge by score alone."

"Visual confirmation is also important," Akira rotated the molecule on screen. "Is this binding mode really reasonable?"

"The hydrogen bond angle looks strange..."

"Scoring functions often simplify angle dependence. So unnatural structures can get good scores."

Lina showed another example. "Here a hydrophilic group is stuck in a hydrophobic pocket."

"Obviously wrong..."

"But the score isn't bad. Because the hydrogen bond term contributes greatly."

Akira summarized. "A score is a single number. But binding quality is multidimensional. One number can't fully represent it."

"So what should we do?" Sena asked.

"Use scores as reference only," Lina answered. "Look at the structure visually and judge if it's chemically reasonable."

"Look at interaction patterns," Akira continued. "Are there important hydrogen bonds, is hydrophobic matching good, is there steric hindrance."

"And compare multiple poses," Lina added. "Docking doesn't give just one solution. Look at the top several and search for common patterns."

Sena began to understand. "Scores are tools, not answers..."

"Exactly," Akira smiled. "Knowing the tool's limitations and using it is the professional way."

Lina switched screens. "There are more accurate calculations. Molecular dynamics simulations, free energy calculations..."

"Are those accurate?"

"More accurate, but computational cost is orders of magnitude higher. One complex can take days."

"So first screen with docking, then do precision calculations only on promising candidates," Akira explained.

"Narrowing down in stages."

"Yes. At each stage, choose methods with appropriate precision. That's efficient drug discovery."

Outside the window, clouds drifted by. Perfect prediction is difficult. But by using imperfect tools wisely, you can approach the truth. Computational chemists need this sense of balance.

"Next, let's study free energy perturbation methods," Lina suggested.

"Sounds difficult..."

"It is. But understanding the principles will help you see score limitations even better."

Sena nodded with mixed feelings of anticipation and anxiety. The world of computation seemed even deeper.