ChroCompare

ChroCompare predicts whether two compounds can be adequately separated on a given chromatographic method, and helps identify which of several available methods is most likely to succeed. For each candidate column it predicts the retention factor of both compounds and, from a calibrated distribution of model errors, a probability that the pair will be separated within a required region and by a required margin. This supports choosing the most promising chromatographic method for a separation, as well as assigning likely elution order and peak identity.

This repository contains the code used to train the retention-factor models, calibrate the separation-probability predictions, and reproduce the pairwise test-set evaluation described in the accompanying manuscript (DRAFT18_paper.pdf).

Software requirements

• Python 3.9+ (developed and tested with CPython; no OS-specific dependencies)
• RDKit (descriptor calculation)
• pandas, numpy, scipy
• scikit-learn (RandomForestRegressor)
• cloudpickle
• tqdm

No requirements.txt / dependency pinning is included yet — see the note in CONTRIBUTING.md about the project's current maturity level.

Repository contents

File	Purpose
generate_fingerprints.py	Calculates normalised physicochemical descriptor fingerprints for a set of SMILES strings, using the descriptors and normalisation constants in physico_descs.csv.
generate_calibration.py	Trains a random forest retention-factor model per chromatographic column (as listed in columns.csv) and builds the calibration PDF/CDF used to turn a pair of predictions into a separation probability. Saves models to ./Models/ and calibration data to ./CalibrationData/.
ChroCompare.py	Core prediction function (compare): given two fingerprints, returns the predicted retention factor of each on every column and the probability of separation, plus the best column overall.
testing.py	Runs compare() for every pair of compounds in a held-out test set and writes a detailed per-pair results file (results.csv), including comparisons against two "naive" (non-probabilistic) baseline strategies.
run_single.py	Runs the full pipeline for a single pair of SMILES strings, without needing a pre-built dataset file. See "Usage" below.
analysis.py, percentage_plot.py, AUC_curve.py	Analyse and plot the results produced by testing.py (calibration curves, ROC/AUC, method-recommendation accuracy).
columns.csv	List of chromatographic columns/methods used, with their run length (used to normalise retention time to retention factor).
physico_descs.csv	List of RDKit physicochemical descriptors used as model input, with the mean/standard deviation used to normalise each one.

Usage

1. Train models and calibration data (one-off, requires a labelled dataset with a Fingerprint column, as produced by generate_fingerprints.py, and one retention-time column per chromatographic method):

   python generate_calibration.py

   This populates ./Models/ and ./CalibrationData/, which are required by everything below.

2. Predict for a single pair of compounds:

   python run_single.py "<SMILES_1>" "<SMILES_2>"

3. Evaluate on a full test set (a CSV with SMILES and Fingerprint columns, plus one retention-time column per method, as TEST_SET.csv):

   python testing.py

   followed by analysis.py, percentage_plot.py, and AUC_curve.py to analyse and plot the resulting results.csv.

Citing

If you use this code, please cite the accompanying manuscript (will attach).

License

Apache License 2.0 — see LICENSE.md.