Transferable deep generative modeling of intrinsically disordered protein conformations.

Area of Science:

• Protein dynamics and intrinsically disordered proteins (IDPs).
• Computational structural biology.
• Machine learning in bioinformatics.

Background:

• Intrinsically disordered proteins (IDPs) lack stable 3D structures, necessitating advanced methods for conformational ensemble elucidation.
• Molecular simulations are crucial for IDP structural analysis but are computationally expensive.
• Existing machine learning models for IDP ensembles show limited transferability to novel sequences and conformations.

Approach:

• Developed idpSAM, a novel latent diffusion model utilizing transformer neural networks.
• Integrated an autoencoder for protein geometry representation and a diffusion model for conformation sampling.
• Trained idpSAM on extensive simulation data of disordered protein regions using the ABSINTH implicit solvent model.

Key Points:

• idpSAM demonstrates high transferability, accurately modeling 3D structural ensembles for test sequences dissimilar to the training set.
• The model's expressiveness and stable training contribute to its reliable performance.
• Highlights the potential for generating comprehensive conformational ensembles from limited simulation data.

Conclusions:

• idpSAM represents a significant advancement in transferable machine learning-based protein ensemble modeling.
• Emphasizes the critical role of training set size and diversity for model generalization.
• Offers an efficient alternative to resource-intensive molecular simulations for IDP structural studies.