PLM-GAN: A Large-Scale Protein Loop Modeling Using pix2pix GAN.

Area of Science:

• Computational Biology
• Structural Biology
• Artificial Intelligence in Biochemistry

Background:

• Protein tertiary structure is crucial for function, but missing regions pose design challenges.
• Accurate protein structure prediction is vital for understanding biological processes and developing therapeutics.
• Existing methods struggle with reconstructing complex missing segments in protein structures.

Purpose of the Study:

• To develop novel deep learning models for reconstructing missing regions in protein tertiary structures.
• To enhance protein design capabilities, including loop modeling and interface prediction.
• To improve the accuracy and efficiency of protein structure modeling.

Main Methods:

• Implemented pix2pix generative adversarial network (GAN) for generating and inpainting protein distance matrices.
• Developed PLM-GAN by integrating residual blocks into the pix2pix GAN's U-Net architecture.
• Introduced a novel "missing to real regions loss" (L_MTR) function and a unique distance matrix pairing strategy.
• Extended reconstruction capabilities to segments up to 30 amino acids and increased protein length by 128 amino acids.

Main Results:

• pix2pix GAN and PLM-GAN demonstrated proficiency in generating and inpainting protein distance matrices.
• The novel L_MTR loss function and pairing strategy significantly improved model performance.
• Models successfully reconstructed missing protein regions, including longer segments.
• Evaluations on natural protein datasets (4ZCB, 3FJB, 2REZ) yielded promising experimental results.

Conclusions:

• pix2pix GAN and PLM-GAN offer effective solutions for protein structure reconstruction challenges.
• These models represent significant advancements in protein design and engineering.
• The developed methods have the potential to accelerate research in protein-protein interactions and drug design.