FBB2019 reproduction project

Belyaeva July, Bushmakin Ilya, Pushkarev Sergei

Intro

This project contains models and some test code for reproduction of some parts of article [1].

Methods

This repository contains:

1. Chainer CNN (oCNN) from original article [1].
2. PyTorch CNN (ptCNN) with the same architecture, but training code uses Cosine Annealing With Warm Restarts from article [2] to adjust learning rate and achieve lower loss and higher accuracy
3. Clustering code:
   1. Max pseudoclique clustering of ptCNN predictions (KMeans)
   2. Spectral clustering of DAFS scores
4. Visualisation and analysis code

Results

The horrifying code

1. The original code from the article has a little documentation. Instructions from README is not effective:
   1. dafs can not be installed from provided source without make clean due to broken links to BOOST
   2. dafscnn in *MakeNcRNAMatrix has to be replaced with dafs
   3. Provided bash script can not build dataset. It works only for 60 sequence datasets, and even so, it produces only first two files instead of 60.
   4. prepareData.py was written to replace original bash script and integrate functions of AssembleMatrix.py
   5. Original RNApairClassify.py can not be used for cross-validation with -v < 10 passed due to mysterious line val = 9 before the cycle. It is a patch to perform only the last part of cross-validation. This strange decision is also the reason of 'data vanishing', the 9/10 of the test dataset do not participate in training or validation of the CNN.
   6. We could not understand what is Unknown Family mode. All we know:

      In addition to the 10-fold cross-validation experiment, as a more difficult task and for the more practical purpose of finding new ncRNA families, the ncRNA families for generating the training data were chosen to be different from the ones for the test data. This experiment, called unknown-family validation, may evaluate the capacity of our one-dimensional CNN method for accurate clustering of ncRNA sequences of unknown families that do not exist in the training data.

   7. On some datasets DataLoader produces empty datasets for some reason.
2. NO DOCSTRINGS and almost no comments on the details of the script (metrics unmentioned, code for comparison not provided)

What was done?

1. New ConvNN (ptCNN) of the same architecture was created using PyTorch for deeper understanding of authors' ideas. Some new methods of training was used to reduce number of epochs before loss/acc convergence
2. DataSets and other data-related code were reimplemented with reliable numpy and sklearn functions. On load new code provides some stats of datasets, which can be useful to assess splitting quality and Model as well.
3. Clustering algorithms examined:
   1. DAFS score --> Spectral clustering
   2. ptCNN

References

1. Genta Aoki, Yasubumi Sakakibara, Convolutional neural networks for classification of alignments of non-coding RNA sequences, Bioinformatics, Volume 34, Issue 13, 01 July 2018, Pages i237–i244, https://doi.org/10.1093/bioinformatics/bty228
2. Loshchilov, Ilya, and Frank Hutter. "Sgdr: Stochastic gradient descent with warm restarts." arXiv preprint arXiv:1608.03983 (2016).