"If I'd only looked at scores, I would have missed it."
Lina showed a relieved expression.
"What did you find?" Sena showed interest.
"Interaction fingerprints. A method to represent compound-protein interaction patterns as bit strings."
Akira added explanation. "For example, whether hydrogen bonding with Tyr123, expressed as 0 or 1."
"Binary code?"
"Yes. Record the presence or absence of interactions for all potentially interacting residues."
Lina showed the screen. "This is the fingerprint of an active compound."
A number sequence continued: "1011001010110..."
"I don't... understand," Sena was confused.
"Looking at each one makes sense. The first 1 is a salt bridge with Asp47. The next 0 is no interaction with Glu51."
Akira displayed another compound's fingerprint.
"Comparing with this..."
"Similar!" Sena noticed. "About 80 percent match."
"Correct. High fingerprint similarity means similar binding modes."
Lina continued. "And if binding modes are similar, activity tends to be similar too."
"So pattern similarity is more important than score values?"
"In some cases, yes," Akira nodded. "Scoring functions predict absolute values, but with large errors."
"But interaction patterns become more stable indicators."
Sena wrote in her notebook. "Pattern recognition..."
Lina showed a new analysis. "Look at this. I clustered fingerprints of 30 compounds."
A dendrogram appeared.
"Active compounds gather in the same cluster," Sena observed.
"Yes. Compounds with similar interaction patterns have similar activity."
Akira pointed to another cluster. "Conversely, this cluster is all inactive."
"The interaction pattern is decisively different somehow?"
Lina zoomed in. "All lack interaction with Arg88."
"So interaction with Arg88 is essential for activity?"
"As a hypothesis, yes. But we need experimental confirmation."
Sena got excited. "From fingerprints, we can identify important interactions!"
Akira said calmly, "But caution is needed. Correlation and causation are different."
"What do you mean?"
"Even if interaction with Arg88 seems essential, it might actually be another factor. For example, structural constraints needed to achieve that arrangement."
Lina agreed. "So fingerprints are tools for hypothesis generation. Not proof."
Sena understood. "But they tell us what to investigate experimentally."
"Exactly. Can be used for efficient experimental design."
Lina showed another example. "This is an interesting case. Fingerprints are 90 percent identical, but activity differs by 100-fold."
"Why?"
Akira analyzed. "The remaining 10 percent difference must be decisive."
"Which interaction?"
Lina highlighted. "This part. Presence or absence of hydrophobic interaction with Met142."
"From just one difference?"
"Yes. So not all interactions are equal. Weighting is necessary."
Akira proposed, "We can learn the importance of each bit with machine learning."
"Like Random Forest?"
"That's good too. Gradient Boosting also outputs feature importance."
Sena asked, "But if data is scarce?"
"Risk of overfitting," Lina admitted. "That's why cross-validation is important."
Akira offered another perspective. "Don't forget physicochemical interpretation. Don't blindly trust machine learning results."
Lina nodded. "Right. Fingerprints are tools to help human understanding."
Sena summarized. "The truth told by interaction fingerprints is..."
"Compound activity is determined by the combination pattern of individual interactions," Akira continued.
"And by understanding that pattern, we can design better compounds," Lina concluded.
Sena felt a strange beauty in how complex intermolecular interactions could be represented by simple bit strings.
"Now I can see which interactions to optimize next."
"Good perspective," Lina smiled. "Fingerprints are a compass for drug design."
The three continued searching for the next move from the sea of patterns.