LOGO Analysis

This repository provides a Python script to perform Leave-One-Group-Out (LOGO) analysis on combinatorial catalyst/substrate MLR models.
It interfaces directly with the bidirectional stepwise MLR function in mlr_utils.py from the Sigman Lab’s python-modeling environment.

---

Environment

The scripts are compatible with the standard Sigman Lab python-modeling environment.
If you encounter dependency issues, a clean conda environment can be created from the included logo_env.yml:

conda env create -f logo_env.yml
conda activate logo

---

Usage

Run locally:

python logo.py sctipt_test.xlsx > logo.out &

Run on the University of Utah CHPC (Notchpeak) cluster with the included submission wrapper:

submit_cli.sh python logo.py script_test.xlsx

Numerically unstable values can cause issues: Make sure too high feature values are filtered out first (>10^10)

---

Parallelization

As the last step of the bidirectional stepwise regression of mlr_utils.py is not parallelized and takes the majority of the time in model training, LOGO analysis of big datasaets and big models can take very long. The script parallelizes the model training of the independent LOGO splits to reduce computation time.The script uses two levels of parallelism:

1. Outer parallelization → distributes LOGO groups (leave-one-group-out folds) across workers.

   • Controlled by the number of CPUs you request in SLURM (--cpus-per-task) or by available cores on a local machine (total_cores variable)
   • Each group runs independently and writes its own log file (<LeftOutGroup>.log).

2. Inner parallelization → inside each LOGO worker, the bidirectional stepwise MLR search is parallelized across candidate models.

   • This passes the inner_threads input to the n_processors variable inside mlr_utils.py.
   • By default, the script automatically uses at least 1/8 of cores per inner job and minimum 2.

Adjusting cores manually

In logo.py, the split between outer and inner threads is defined like this:

total_cores = multiprocessing.cpu_count()

# Inner: number of threads per LOGO worker (minimum 2 and 1/8 of the total cores)
inner_threads = max(2, total_cores // 8)

# Outer: number of LOGO workers
n_logo_workers = max(1, total_cores // inner_threads)

• Increase n_logo_workers if you want to run more groups in parallel.
• Decrease it if you want to give each LOGO worker more cores for the inner MLR search.
• On the cluster, total_cores will automatically respect your SLURM allocation (SLURM_CPUS_PER_TASK).

---

Output

For each left-out group, the script generates:

• A .log file with detailed training/testing results and model parameters
• A .png parity plot (predicted vs. measured)
• A row in the final LOGO_results.csv summary file

The final CSV is sorted by LOGO MAE for quick comparison.

---