Improved protein complex prediction with AlphaFold-multimer by denoising the MSA profile.

Area of Science:

• Computational biology
• Structural biology
• Bioinformatics

Background:

• Protein complex structure prediction is crucial for understanding biological functions.
• While AlphaFold2 and AlphaFold-multimer (AFM) have advanced the field, accurate prediction remains a challenge for many protein dimers.
• Existing methods struggle with complex structures, necessitating novel approaches.

Purpose of the Study:

• To improve the accuracy of protein complex structure prediction using AlphaFold-multimer (AFM).
• To develop a protocol that enhances AFM predictions by optimizing the multiple sequence alignment (MSA) representation.
• To address limitations in current protein complex prediction for difficult targets.

Main Methods:

• Implemented a gradient descent approach to learn biases in the MSA representation within the AFM network.
• Developed the AFProfile protocol to guide AFM predictions.
• Evaluated the method on challenging targets from CASP15 and a dataset of 487 complexes where AFM previously failed.

Main Results:

• AFProfile increased the average MMscore to 0.76 from 0.63 for AFM on seven difficult CASP15 targets.
• Achieved a 33% success rate (MMscore > 0.75) on 487 challenging protein complexes where AFM alone failed.
• Demonstrated significant improvement in predicting the structure of protein complexes, especially for difficult cases.

Conclusions:

• AFProfile effectively enhances protein complex structure prediction accuracy by leveraging MSA information.
• The protocol offers a novel way to direct predictions towards a specific target function.
• Gradient descent over the MSA shows promise for various protein structure prediction tasks.