DECONbench : Benchmarking platform for deconvolution methods for tumor heterogeneity quantification

Organized by AArnaud - Current server time: May 31, 2020, 8:36 a.m. UTC

Current

Benchmark
Feb. 15, 2020, midnight UTC

End

Competition Ends
Feb. 15, 2021, midnight UTC

/!\ This benchmark will open soon, if you need an early access, please send an email to the organizers. /!\

Successful treatment of cancer is still a challenge and this is partly due to a wide heterogeneity of cancer composition across patient population. Unfortunately, accounting for such heterogeneity is very difficult. Clinical evaluation of tumor heterogeneity often requires the expertise of anatomical pathologists and radiologists.

This benchmark is dedicated to the quantification of intra-tumor heterogeneity using appropriate statistical methods on cancer omics data.

In particular, it focuses on estimating cell types and proportion in biological samples based on methylation and transcriptome data sets. The goal is to explore various statistical methods for source separation/deconvolution analysis (Non-negative Matrix Factorization, Surrogate Variable Analysis, Principal component Analysis, Latent Factor Models, …) using both RNA-seq and methylome data.

How to start ?

[1] Go on the challenge page, in the Participate tab, in the Files item : download the starting kit by clicking the Starting Kit button, and the public data sets by clicking the Public Data button for the 1st phase.

[2] On your local machine, unzip the just downloaded zip files stating_kit.zip and data_public.zip. Copy the data sets (D_met1_public.rds, D_rna1_public.rds, A_1_public.rds) into the unziped starting_kit directory. Then open R in the starting_kit directory, (e.g. open strating_kit.Rmd with RStudio).

The unziped strating-kit directory contains now:

  • A starting_kit.html corresponding to the vignette of the Benchmark (all useful information can be found here).
  • A submission_script.Rmd to modify and to use to submit your code.
  • The methylation and transcriptome D matrices, and the associated A matrix.

[3] In the R console launch the following command :

rmarkdown::render(input = "submission_script.Rmd")

How to submit your results ?

Now, let’s submit your code (a zip file generated by the sumbmission_script.Rmd file) in the Participate tab of the codalab challenge.

 

 

How is the scoring metric computed?

The discriminating metric will be computed on the A matrix: mean absolute error between the estimate and the groundtruth.

The matrix D of shape (N patients, M methylation sites) is provided. D = T A, with T the cell-type profiles (k cell types, M variables) and A the cell-type proportion per patients (N patients, k cell types).

Participants have to identify an estimate of A matrix.

During this benchmark, they have to submit a reproductible script (with their implemented solution) that compute A. This script will be applied on 10 simulated data sets to estimate 10 A matrices, and the mean of the MAE between those estimations and the simulated A matrices will be used for scoring.

 

 

By participating to this challenge, you accept to publicly share your submissions.

You can freely test your methods on this benchmark and compare yourself to the reference methods. When your development is finished or stable, please create a team with the following nomenclature : "Feature selection for DNAm" [met] + "Feature selection for RNA" [rna] / "Deconvolution method" [both]. The idea is to gather the similar approaches under the same team. Don't hesitate to send an email to the organisers if you have any questions or issues.

Benchmark

Start: Feb. 15, 2020, midnight

Description: Estimate the proportion matrix A from the DNAm and/or RNA-seq matrices D. The score "Time" is the average time in secondes (over 10 cases) to estimate one A matrix. The score "Data Type" is simply to indicate on which matrix D ("met" = 1, "rna" = 2, "both" = 3) the matrix A has been estimated.

Competition Ends

Feb. 15, 2021, midnight

You must be logged in to participate in competitions.

Sign In
# Username Score
1 codabench_4 0.0240
2 codabench_3 0.0370
3 codabench_1 0.0569