search

leonjessen / tensorflow_rstudio_example

3 stars 0 forks source link

proposed revisions to post #2

Closed jjallaire closed 6 years ago

jjallaire commented 6 years ago

You can see the rendered version of the revisions here:

https://htmlpreview.github.io/?https://github.com/leonjessen/tensorflow_rstudio_example/blob/jjallaire-proposed-edits/README.html

I am quite certain I've butchered some descriptions of things since I have no knowledge of cancer immunotherapy! (my main goal was to illustrate what I think is good macro-organization for the post). Please cleanup as necessary.

The one thing I think the post still needs is some description of significance after we do the visualizations of the model's performance. Are these results good/promising? How do they compare with traditional methods? If they compare somewhat poorly why do we be believe that a properly tuned DL model will be successful?

One related question I have is whether convolutional methods would bear any improvement since they could about the significance of location invariant patterns?

leonjessen commented 6 years ago

It's a pleasure collaborating with a native English speaker. Even though I consider myself quite proficient in English, it is evident to me, that the flow of the text is greatly increased - Excellent!

The layout also looks very nice. I have a few corrections following your rewrite, but I can do these after PR merge?

The netMHCpan method I link to, which is state-of-the-art in predicting the MHCI-peptide interaction, relies on a quite complex architecture and takes more than just the peptide sequence into account and using a custom algorithm, called NNalign. Furthermore, it predicts the strength of the molecular interaction as a continuous variable link to paper - The SB, WB and NB classes are human-friendly assignments. So in summary the full blown method is much more complicated than just a 'simple' deep feed forward and 3 classes. What we are going to work on now, is to investigate CNNs - Perhaps we could expand the post, such that we first show the deep FFWD (~94% accuracy) and then show that we can do even better with a CNN - I can try to find some time to work on that? Also I am working on updating the PepTools - Which in its current form lacks documentation.

jjallaire commented 6 years ago

Yes, I think if we extended it by showing a CNN that does even slightly better that would be fantastic. We could then further qualify that the best answer will require lots more work (perhaps linking to that paper again), but at least this example would have demonstrated the first couple of steps on the path.

jjallaire commented 6 years ago

Let me know if you think you will have time to work on the CNN. It would certainly be much better with that, but if you didn't have time we could also explain that the next logical step would be trying CNNs, and that to get the kind of results your team is hoping for go even further.