[ENH] clean implementation of `DeepAptamer`

[fkiraly]

After prioritization, we decided that DeepAptamer would be a nice next algorithm to "clean up" for easy use.

code paper pdf

[fkiraly]

1st thing to do:

• review code base
• try to attach a taxonomy to it based on Taxonomy and high-level API design of algorithms #1
• come up with high-level API before implementing, post here how the class/method etc structure would look like

[satvshr]

Blocker: DeepAptamer uses 4 shape properties which they got from a web server called DNAshape, there is an R repository by some authors of that paper called DNAshapeR, there is no python repository for the same. My question is: Is there any way to run R code in python?

[satvshr]

@KubiczekD @JaBirke In the paper (page 6) under "Pre-trained DeepAptamer could identify key nucleotides
essential for target binding" they mention VAE (which I think suggests where the aptamer can bind with the target protein) as a use-case. I remember Dennis talking about the use of DeepAptamer for this, but I would like you to know VAE is not a part of the architecture but the output from DeepAptamer is something VAE uses (There is no code on how the VAE is constructed either), just to avoid misunderstandings! Kindly correct me if I am wrong.

[satvshr]

Inconcistencies:

• The paper mentions using a 2D convolutional filter in page 4 under "The integrated DeepAptamer model demonstrated superior performance compared with the individual CNN or BiLSTM model" but the image mentions using a 1D convolutional model instead.
• The image linked above also includes an attention layer for some reason, but when you try searching for the word "attention" in the official paper, it only links to 2 citations in references.

[fkiraly]

DeepAptamer uses 4 shape properties which they got from a web server called DNAshape

What are these as objects? Type, size, etc

[fkiraly]

VAE (which I think suggests where the aptamer can bind with the target protein)

I would read this as "variational auto-encoder", this is a type of generative neural network

[satvshr]

I would read this as "variational auto-encoder", this is a type of generative neural network

I was not doubting what it is, I am confused whether it is a part of DeepAptamer, but I am fairly certain it is not.

DeepAptamer uses 4 shape properties which they got from a web server called DNAshape

What are these as objects? Type, size, etc

I was going to mention it later on but here is where it gets interesting: there are 4 properties but the shape matrix is supposed to be 162 flattened dimensions long, which implies that not all shapes properties have vectors of equal length (given 162 does not get divided by 4)

[satvshr]

DNAhapeR is an R/BioConductor package for ultra-fast, high-throughput predictions of DNA shape features. The package allows to predict, visualize and encode DNA shape features for statistical learning. DeepAptamer uses 4 shape properties namely: "Four pentameric DNA shape features, namely MGW, Roll, ProT, andHelT, were extracted from Monte Carlo simulation using sliding window" as quoted from the paper under the sub-topic "DNA shape features" on page 11.

[fkiraly]

@satvshr, is there a python drop-in for these features?

If not, we may want to consider implementing them, they do not look too difficult.

[fkiraly]

@satvshr, regarding R packages:

I would suggest you do not look at the R code, because it is restrictively licensed (GPL, this is copyleft). Instead, look only at how it is used from within python and try to derive specifications.

Then we estimate how difficult it is to reimplement from paper.

Since you said you do not actually understand R code and have never coded in R, I even suggest we leave it at this, as this state of affairs protects us from license contamination.
(R is fading out anyway, as it looks like)

[satvshr]

I would suggest you do not look at the R code, because it is restrictively licensed (GPL, this is copyleft). Instead, look only at how it is used from within python and try to derive specifications.

I just set up everything to start understanding the R repo, I will verify if it is used within python but as I said in the daily, it most likely is a part of the input.

Edit: Over here shape features are being extracted from line which is being derived from the input file. The class deepselex is called in main.py and if we match arguments data_dir is the variable being supplied which is defined here. Given it is an absolute path from the authors device:

1. We can't look at the input.
2. The shapes vector was a part of the input.
   There is a python repository with more forks than the official one, but I think they use a different method to obtain these shapes. This is the webserver the authors of DeepAptamer used to get their shapes (same authors of DNAshapeR).

[satvshr]

I observed a few things:

1. DNAshape and DNAshapeR are both owned by Rohs Lab.
2. I was scrolling through the Rohs lab website and I found that they have 4 projects concerning the shape of DNA:

• DeepDNA shape
• methyl-DNAshape
• DNAshapeR
• DNAshape
  Now the last 2 points above are the same thing and are being used by DeepAptamer, but the first point is a python package with a free-to-use license(?)

My question to @KubiczekD @JaBirke: Does "DeepDNA shape" or "methyl-DNAshape" provide the same functionality as DNAshape/DNAshapeR?

My question to @fkiraly: Why is the R package not free-to-use but the python package is?

[fkiraly]

My question to @fkiraly: Why is the R package not free-to-use but the python package is?

The R package is subject to the GPL-2 license which is copyleft - this makes it unusable for most commercial purposes https://en.wikipedia.org/wiki/Copyleft

The python package is BSD-3, which is a permissive license (with attribution requirement), same that sktime also has.