Transferable deep generative modeling of intrinsically disordered protein conformations.

Area of Science:

• Computational Biology
• Structural Biology
• Machine Learning

Background:

• Intrinsically disordered proteins (IDPs) possess dynamic structures crucial for biological functions.
• Determining IDP conformational ensembles requires integrated computational and experimental approaches.
• Molecular simulations generate structural ensembles but are computationally expensive.

Purpose of the Study:

• To develop a novel, highly transferable machine learning model for intrinsically disordered protein ensemble generation.
• To address the limitations of existing methods in modeling novel sequences and conformations.

Main Methods:

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

Main Results:

• IdpSAM demonstrated high transferability, accurately predicting 3D structural ensembles for test sequences dissimilar to the training set.
• The model effectively captures conformational ensembles even from limited simulation data.
• Achieved stable training and high expressiveness in its neural network architecture.

Conclusions:

• IdpSAM represents a significant advancement in transferable protein ensemble modeling via machine learning.
• The study highlights the critical role of training set size in achieving robust generalization.
• This method offers an efficient alternative to resource-intensive molecular simulations for IDP structural studies.