Regressions to find regulators of splicing

Graph power 4

Data preprocessing

The PSIs and TPMs are first evaluated elsewhere (repos quantif_exon_skipping and stringtie_quantif).

R/prefilter_quantifs.R takes PSI and TPMs and prefilters to keep only relevant cases. R/prep_graph_power4.R makes the train/test split, and saves the matrices ready to be used directly in main script.

Network computation

Main file: graph_power4_oncluster.R. Can be called from CLI with right arguments: use src/dsq_create.sh to create a job array. This file describes a workflow using the steps defined in functions_steps.R, and there are many helper functions in loss_functions.R (which could be a package).

Note: for permutations, if permutation == 0 , no permutation, any number bigger than 0 indicates that permutation should be performed (the value doesn't matter). The seed for the entire script is set by sum(permutations)` to avoid having independent runs identical, but there shouldn't be much else affected (the diagonal in CLIME for a big penalty is also randomly generated).

Runs

Everything run as a dSQ (dead simple queue) joblist. In src/dsq_create_*, scripts to create the job from a joblist in joblists/.

• First run joblist noperm_glassoQuicClime.txt to try all algorithms with a few penalties (used to generate heatmap Fig. 7B). In parallel, run noperm_noperm_Scio.txt with less memory, more time (note the names are wrong, in fact it's CLIME that is performed separately).

• Then run joblist nopermglasso_k.txt, focusing on a single algorithm but trying out different k in knn imputation (Fig. S5B).

• Then "perm.txt" uses single algorithm and k, and runs permutation tests at all penalties (Fig. S5C, Fig. 7C). Finally, "final.txt" reruns a single time at best penalty (and algo and k, and a single permutation), but saves the full intermediate object for in-depth examination.

Post-analysis

After being run, download all csv files, and use analysis_from_cluster.R, which sources analysis_helpers.R and performs the comparisons (Fig 7B,C, S5B,C). The adjacency matrix is loaded and analyzed in analysis_network.R to create the subnetworks (Fig 7D,E, S5D).

Some "manual" explorations are in graph_power4_postanalysis.R (used to e.g. compare the impact of individual changes, not included in the paper).

Older versions

"Forward" fit

In regression_functions.R, fit PSI = C * TPM

Simulations

In simulation_perm_test.R

Permutation tests

Adapted to run on Ruddle: ruddle_permutation_test_PSI_vs_TPM.R, ruddle_permutation_test_PSI_vs_TPM_sim.R, then results interpretation in simulation_perm_test.R and results_permutation_test.R.