Visualization for AlphaFold3 Results

Hanziwww commented 1 day ago

Hi,

In response to the growing number of issues and comments regarding result visualization in AlphaFold3, I wanted to share an update and a potential solution. I have added Visualization tools to the latest version afusion package. This module focuses specifically on improving the ease and accessibility of visualizing AlphaFold3 predictions.

1. Visualization Module

2. Visualization GUI Interface

3. Online Usability

To make this even more accessible, I have deployed an online app where users can upload their AlphaFold3 result files and visualize them in real time.
This app is fully functional and available at the following link: Usable visualization site (fully usable)

Augustin-Zidek commented 1 day ago

Amazing work, congratulations on shipping this impressively quickly!

A few suggestions / bug reports:

It would be helpful to draw chain boundaries in the PAE plot. See e.g. https://www.uvio.bio/alphafold-outputs/hemoglobin/pae_error_plot.png
The pair metrics would be nice to organise as square matrices with chain IDs as headers for columns and rows, so that it is easier to tell which pair of chains each number is for.
The PAE can get really big, I recommend using dtype=np.float16 in https://github.com/Hanziwww/AlphaFold3-GUI/blob/main/afusion/visualization.py#L31, instead of the default np.float64 (4x less RAM use). Since the precision is just 1 decimal digit, float16 is more than enough. If not possible, use at least np.float32.
Explicitly specify PAE min and max (0.0 -- 31.75) in https://github.com/Hanziwww/AlphaFold3-GUI/blob/main/afusion/visualization.py#L175, otherwise e.g. for very confident prediction that only ranges from 0.0--1.0 you will get skewed colours.

Hanziwww commented 1 day ago

@Augustin-Zidek Thank you so much for your valuable suggestions and taking the time to share these insights! I'll make sure to incorporate your recommendations to improve the code. Your feedback is greatly appreciated : )

Augustin-Zidek commented 1 day ago

My pleasure! :) Closing the issue, but pointing people your way if they don't want to deal with command-line.

google-deepmind / alphafold3