"AI proposed all of these?"
Sena stared at the screen. Dozens of structures lined up.
"Generative model," Lina explained. "Generates new molecules from learned chemical space."
"Incredible that this is possible..."
Akira was cautious. "But not all are good. Need selection."
Lina executed filtering. "First, screen with Lipinski rules."
Dozens reduced to a dozen.
"Next, synthetic accessibility score."
Further reduced to five.
"Let's look at these five in detail," Akira proposed.
Sena analyzed the first structure. "This has an interesting structure."
"Scaffold I've never seen," Akira showed interest.
Lina ran docking simulation. "Score is good."
"But can we synthesize it?" Akira questioned.
"Let me search synthesis routes," Lina operated.
"6 steps. Feasible."
"Then worth trying."
Sena looked at the second structure. "This resembles existing compounds."
"Bioisostere," Akira pointed out. "Replaced functional group with another of similar properties."
"AI learns known patterns. So it also generates things similar to existing ones," Lina explained.
"But," Akira continued. "Subtle differences can yield big improvements."
Sena asked. "How did the AI learn?"
"Dataset of existing active compounds," Lina answered. "Thousands of structures and activity values."
"Uses them to learn structure-activity relationships."
"But," Akira pointed out. "Patterns not in training data are difficult."
"Extrapolation limits."
Lina admitted. "Yes. So need to combine with domain knowledge."
"AI proposes, chemists evaluate."
Sena looked at the third structure. "This has odd substituents."
Akira analyzed. "Synthesis looks difficult."
"AI doesn't fully understand synthetic difficulty," Lina explained.
"Knows if molecule can exist. But ease of making is different."
"That's why we filter with synthetic accessibility score."
Sena understood. "AI alone is insufficient."
"It's a tool," Akira emphasized. "Powerful but not perfect."
Lina showed another AI function. "Activity prediction model."
Input structure, output activity value.
"IC50 8 nM, predicted."
"Reliable?" Akira was cautious.
"Confidence interval also appears," Lina displayed. "6-12 nM range."
"Wide. But order of magnitude seems right."
Sena asked. "Why the range?"
"Model uncertainty," Lina explained. "Because training data is limited."
"But sufficient for prioritization."
Akira evaluated. "With this prediction, worth synthesizing."
Lina launched another tool. "Retrosynthesis AI exists too."
"From target molecule, back-calculate starting materials and reactions."
Synthesis route displayed on screen.
"4 steps. Synthesizable with known reactions only."
Sena was moved. "AI is this helpful."
"But," Akira reminded. "Ultimately, verify by experiment."
"AI generates hypotheses. Verification is human work."
Lina supplemented. "AI accelerates chemical space exploration."
"Instantly screens space that humans alone couldn't explore in a lifetime."
"But physical laws and chemistry knowledge must be incorporated by humans."
Sena looked at the five candidates. "Which to try first?"
Akira evaluated. "Second and fifth. Extensions of known, lower risk."
"Validate with those first. If successful, challenge novel scaffolds."
Lina agreed. "Stepwise approach. Calibrate AI predictions with experiments."
Sena resolved. "Then I'll synthesize."
"When data accumulates, retrain the model," Lina said.
"Accuracy improves further."
Akira said finally. "AI-human collaboration. That's future drug design."
Sena smiled. "AI illuminates the path, we walk it."
"Exactly," Lina nodded.
"But humans decide the destination," Akira added.
The three gazed at molecules suggested by AI. Possibilities and limits. Understanding both, moving forward. That was the new form of drug discovery.