Use Physicochemical features of AAs as input.

[hypnopump]

It would be useful to use Physicochemical properties beyond Van der Waals radius of AAs as input sucha as:

• surface exposure
• predicted solvent accessibility
• polarity
• isoelectric point
• pairwise potential ...?

[daniel-z-kaplan]

Those all sound like desirable traits to include.

[hypnopump]

Implementation should be straightforward. If you want inspiration you can see the function onehotter_aa from preprocessing/angle_data_preparation_py.ipynb and apply the method to the function wider from models/predicting_distances.ipynb. @Daniel-z-Kaplan What do you think? Can you do it? Contributors are welcome!

[n4ndoz]

Hi, Im working with prot structure prediction as my masters project. I implemented some of your operations as vectorized functions, and I include some physchem properties on input. Im even using your model, with slight modifications (adding dilated convolutions and aproximating to the architecture of Raptor-X-Contacts).

[hypnopump]

Hi there! Thanks for the interest and continuing the work on the project! I'm sharing my newest version here with you in a zip folder (it includes cool ideas, but i moved on to different projects before integrating them and producing good results).

It includes the following features:

• vectorized data processing functions
• handling protein arbitrary length
• predictions by X squared crops of Y*Y AAs each and averaging at prediction time
• deeper resnet
• physicochemical features of AAs

You can grab ideas and copy/paste from this codebase. I would like to ask you if, in return, you could integrate this code with the one in the repo and also share your dilated convolution model (don't overwrite the existing model code, just provide the option of using your model as well via loading a diferent model).

Also if you could share some results it would be very nice as well. Contact me via email so I can send you the newest version of the code: [email protected]

It would be awesome if you could do this! I'll add your name to the contributors / collaborators / authors list!

[n4ndoz]

Hi, I will reach you by email!

Thanks for sharing the code and ideas. I don't have the time now, at least until July, to add to the repo, since I'm still coding adn writing, but as of July I can update and integrate my models. Since I use a similar workflow than yours (except that I generally prepare the data before passing to a generator).