DeepPath: overcoming data scarcity for protein transition pathway prediction using physics-based deep learning.

Area of Science:

• Computational Biology
• Structural Biology
• Biophysics

Background:

• Protein structural dynamics are vital for function.
• Current deep learning methods often provide static protein snapshots, limiting dynamic insights.
• Capturing protein dynamics typically requires simulations or experiments.

Purpose of the Study:

• To introduce DeepPath, a novel physics-guided deep learning framework.
• To enable rapid prediction of realistic protein transition pathways at atomistic resolution.
• To overcome limitations of static predictions in current deep learning models.

Main Methods:

• Developed DeepPath, a physics-guided deep learning framework.
• Employed generative active learning (GAL) for iterative prediction refinement.
• Utilized molecular mechanical force fields as oracles to guide pathway generation.

Main Results:

• Successfully predicted protein transition pathways for AdK, SHP2, CdiB H1, and BAM-complex.
• Reproduced key transient interactions from previous studies.
• Identified a novel intermediate for BAM-complex gating, aligning with experimental data (TM-score = 0.91).

Conclusions:

• DeepPath offers rapid, atomistic resolution prediction of protein transition pathways.
• Generative active learning (GAL) shows significant potential for advancing protein structure prediction.
• The framework accurately models complex protein dynamics and conformational changes.