The aim of this project is to translate the ConsensusOPLS method from its
original MATLAB version (available on the GitLab
https://gitlab.unige.ch/Julien.Boccard/consensusopls) to an R version.
For this purpose, a mind map of the method's functions was created:
graph TD
classDef Title stroke-width:0, color:blue, font-weight:bold, font-size: 19px;
0[Functions of ConsensusOPLS method]
subgraph Number1[Step 0 : Translate MATLAB function into R]
0 --> 1
1(Step0_1_matrix2saisir)
0 --> 2
2(RVConsensusOPLSPerm)
0 --> 11
11(VIP)
1 --- |include in matrix2saisir| 3
3(add code)
2 --> 4(RVConsensusOPLS)
2 --> 5(2 RV_modified)
4 --> 6
6[koplsScale]
4 --> 7
7[1 koplsKernel]
4 --> 8
8[koplsModel]
4 --> 5
4 --> 9
9[3 ConsensusOPLSCV]
4 --> 10
10[DQ2]
9 --- |see graph below| 9
end
class Number1 Title;
To make it easier to read, we have reproduced the previous graph using the ConsensusOPLSCV function only.
graph TB
classDef Title stroke-width:0, color:blue, font-weight:bold, font-size: 19px;
0[Other functions of <br> ConsensusOPLS method]
subgraph Number1[Step 0 : from the ConsensusOPLSCV]
0 --> 1
1(ConsensusOPLSCV)
1 --> 2
2(koplsDummy)
1 --> 3
3(koplsReDummy)
1 --> 4
4(koplsScaleApply)
1 --> 5
5(koplsCenterKT*)
1 --> 6
6(koplsPredict)
1 --> 7
7(koplsRescale)
1 --> 8
8(koplsMaxClassify)
1 --> 9
9(koplsBasicClassify)
1 --> 10
10(koplsSensSpec)
1 --> 11
11(koplsConfusionMatrix)
1 --> 12
12(koplsCrossValSet)
1 --> 14
14(koplsModel)
1 --> 15
15(koplsScale)
1 ---|use koplsModel <br> and koplsScale as <br> RVConsensusOPLS <br>
function| 1
12 --> 3
5 ---|KTeTe, KTeTr, KTrTr|5
6 --> 5
6 --> 7
14 --> 5
10 --> 2
11 --> 2
14 --> 15
end
class Number1 Title;
This structure has been reproduced for the git tree: one branch per function
and per code file. Next, the functions were tested on example datasets
(demo_data proposed by Julien Boccard in its Matlab version) and verified.
Finally, all the branches will be merged to finalize the method.
A second folder has also been created. It is named ConsensusOPLS. It
contains the publishable version of the R code, with the structure of a CRAN R
package. In this folder, the previous codes are grouped by functionality. They
have also been optimized. Gradually, only this folder is used for further
package development.
At the meeting on 09/10/2023, it turned out that the KOPLS package had already been translated into R. To avoid redundancy or errors, the source code of the kopls package was searched for. A version was found on the following GitHub: https://github.com/sdechaumet/ramopls/tree/master/inst/package (version of 05/08/2020). Codes previously translated from Matlab to R were compared with those in the GitHub repository. References to the authors of the KOPLS package have been added at the start of each function's code.
The main branch of this project is shared with members of the Swiss
Institute of Bioinformatics (SIB) on the following GitHub:
https://github.com/sib-swiss/consensusOPLS. The other development branches are
only available on the present GitLab repository.
The development of the package can be summarized in the graph below:
flowchart TD
A(Translate Matlab -> R) -->|Script-to-script conversion <br>
With all that this implies: <br> format of objects to be adapted, <br>
calling/using functions, etc.| B
B(Redesign step)
B --> |From scripts to package| C
C{Generalizing and optimizing}
C --> D(Validation)
D --> E[Demo data]
D --> F[Real data]
Indeed, the first major step was to convert the Matlab scripts into R scripts. Switching from one language to another inevitably implies changes in code structure, object format, function calling or use, etc., and this is where we came in.
Next, we need to redesign the codes to move towards a package format. As mentioned above, this means grouping certain functions together for better comprehension, and unit testing the lowest levels of functions to ensure a certain quality of results. It also requires, among other things, building a helper, checking and partitioning input formats.
Once this first level of code validation had been achieved, we were able to work on adding functionality. For example, we adapted numerous mathematical formulas and restructured the code to reflect these changes. We modified the kernel type: in Matlab, the polynomial kernel could only be of order 1 (linear kernel). Now, the user can use a polynomial kernel of order greater than 1 (non-linear) or Gaussian. Similarly, we've gone back to predictive components: in Matlab, the method allowed only one component. Now, the user can vary this number to determine the optimal model according to the type of data being analyzed. In both the previous and current versions, the number of orthogonal components can also be greater than 1.
Code optimization, which took place in parallel with the generalization of the method, concerned several points. Firstly, function by function. The codes were simplified, as were the output formats, to make them more suitable for R. Indeed, during translation, the objects were reproduced with their Matlab structure, which is not the most suitable and optimal in terms of computation time for R. Next, permutations were parallelized. This led to a restructuring of the various code levels, including the KOPLS functions, to facilitate execution.
Next, the validation stage was very important. Unit tests were carried out, as was the integration test that guided the previous stages. The aim was to ensure that what we were proposing on R would deliver exactly the same results as on Matlab. These tests do not take into account the new functionalities added. To achieve this integration, we used the demo dataset proposed by Julien Boccard in Matlab. This was followed by a validation phase on real data. To do this, different sets were tested (two sets of internal project data, not shown; one set on public data, from ProMetIS (https://github.com/IFB-ElixirFr/ProMetIS/tree/master)). In this way, we ensured that on these 3 real data sets, matlab and R results were strictly identical when reproducing similar code between the two software packages.
Finally, we've added everything that will enable a non-experienced user to use the method: R function help, execution progress displays, user-friendly results.
The next step is to introduce variable selection to this ConsensusOPLS-DA method.
Step 1: before method utilisation.
Step 2: into the kernel construction.
(Coming soon)