"Design proteins from scratch."
Mikhail rotated the three-dimensional structure on screen.
"From scratch?" Eiji was surprised. "Proteins that don't exist in nature?"
"Yes. De novo protein design."
Lina watched with interest. "How?"
"First, decide the desired function. For example, binding a specific molecule, or having catalytic activity."
"Yeah."
"Next, design the structure that realizes that function."
"From structure?"
"Reversed. Function → structure → sequence. Think in this order."
Eiji tilted his head. "Normally it's sequence → structure → function, right?"
"Traditional biology is. But design is reversed."
Lina understood. "An engineering approach."
"Exactly. Decide what you want to make, then draw the blueprint to realize it."
Mikhail showed a concrete example. A small protein, about 50 amino acids.
"This was designed to recognize a specific peptide sequence."
"How?"
"First, determine the binding interface structure. Hydrogen bonding patterns, placement of hydrophobic contacts."
"Like a pharmacophore," Sena entered and said.
"Yes. Protein design is similar to ligand design."
"But proteins are large."
"So computation is hard. The sequence space is enormous."
Lina asked. "20 amino acids, 50 residues, so 20 to the 50th power possibilities?"
"Yes. Astronomical. Impossible to try all."
"So what do you do?"
Mikhail explained. "Use machine learning, especially deep learning."
"What model?"
"Variational autoencoders or diffusion models. Learning from known protein structures to generate new structures."
Eiji supplemented. "Like the reverse of AlphaFold."
"Close. AlphaFold predicts structure from sequence. Protein design predicts sequence from structure."
"But how do you confirm the predicted sequence really takes that structure?"
"That's where we use AlphaFold. Input the designed sequence into AlphaFold to predict structure."
Lina's eyes lit up. "A loop of design and validation."
"Yes. Design → predict → evaluate → redesign. Repeat this."
Sena asked. "Do you actually make it and confirm?"
"Ultimately yes. Express in E. coli and confirm structure with X-ray or cryo-EM."
"Does it work?"
Mikhail smiled. "Recently success rate is increasing. Especially tools like RosettaFold and ProteinMPNN are powerful."
Eiji presented another perspective. "There's also modifying existing proteins, right?"
"Yes. Not completely de novo, but template-based design."
"What's the difference?"
"Start from a known protein and modify only specific parts. Higher safety."
Lina gave an example. "Like antibody affinity maturation."
"Exactly. Optimize only CDR regions. Keep the framework."
Sena showed interest. "Can you design antibodies too?"
"You can. Recent research is progressing on computationally designing antibodies that recognize antigens from antigen structures."
"Amazing."
Mikhail looked serious. "But there are challenges."
"What?"
"Function. Even if the structure is correct, it doesn't necessarily have the expected function."
"Why?"
"Because dynamics are important. Proteins aren't static. They fluctuate."
Lina supplemented. "Crystal structure is just one snapshot."
"Yes. So molecular dynamics simulations are also needed."
Eiji said, "Computational costs keep rising."
"Yes. But with computer advances and algorithm improvements, it's becoming practical."
Sena said thoughtfully, "Proteins becoming the designed."
"The era of making life's parts ourselves," Mikhail said.
"But," Lina warned, "we mustn't forget ethical issues."
"Right. The more powerful the technology, the more carefully we must use it."
Mikhail nodded. "That's why we clarify objectives and prioritize safety."
The moment proteins are designed. That's the beginning of a new chapter in life sciences.