Leakage-controlled multi-laboratory RT revision release
This frozen release corresponds to the revised validation of the manuscript. It contains GAT, GCN, and FPNN checkpoints for six inner folds in each of three identity-grouped and three scaffold-aware outer repetitions; fitted stack models; preprocessing objects; the controlled FPNN one-hot checkpoints; configuration and analysis code; aggregate metrics; task-matched bootstrap outputs; hashed split assignments; and de-identified model predictions.
The previous top-level seven-checkpoint FPNN bundle has been superseded. The scientific analysis uses exactly six inner folds per outer repetition.
Validation scope
- Identity-grouped stack: R2 0.977 +/- 0.002, MAE 0.360 +/- 0.016 min, RMSE 0.578 +/- 0.057 min.
- Scaffold-aware stack: R2 0.928 +/- 0.004, MAE 0.670 +/- 0.013 min, RMSE 0.911 +/- 0.022 min.
- Scope: new molecular identity groups or scaffold components assigned to one of the 23 laboratory labels represented during training. No unseen-method claim is made.
Descriptor routing
GAT and GCN each use a five-descriptor branch. FPNN uses a Morgan fingerprint
and laboratory encoding but no molecular descriptors. The complete ExtraTrees
stack receives three base predictions plus the five descriptors. The public
ablation labelled base_predictions_only removes the five direct meta-learner
descriptor columns; it is not a descriptor-free retraining of GAT or GCN.
Reproduction
Create the recorded environment and run:
pip install -r requirements.txt
python reproduce.py --data /path/to/authorized/train.csv --output /new/empty/output --include-neural
The input must contain SMILES, Lab, and RT columns and must be obtained
through an authorized HighResNPS-derived data-access route. Training is the
long stage. Aggregate frozen outputs used for the paper are already included
under artifacts/summary.
Public/private data boundary
HighResNPS states that the service is for a closed user group, and recent
published uses describe license restrictions on public redistribution of its
metadata and spectra. This repository therefore does not redistribute SMILES,
experimental RT values, laboratory identifiers, or raw source rows. Each split
file instead uses a SHA-256 key computed from
record_index + SMILES + Lab + RT; an authorized user with the same modelling
table can reconstruct the key and verify every assignment. Public prediction
files contain that key and model outputs only, without source fields or targets.
Directory map
src/: model and feature code used by the frozen pipeline.revision/config/: frozen analysis configuration.revision/scripts/: split, training, summarization, figure/table source code.artifacts/{identity_grouped,scaffold_aware}/repeat_{1,2,3}/: six-fold checkpoints, fitted stacks, preprocessing, metrics, hashed splits, and de-identified predictions.artifacts/matched_lab_encoding/: architecture-matched one-hot FPNN checkpoints and aggregate comparison.artifacts/summary/: aggregate metrics, corrected paired bootstrap results, and figure/table source summaries.SHA256SUMS.txt: checksum manifest for the complete release.
No DOI is claimed. The repository commit and release tag identify the frozen public version.