GHS Chemical Hazard Classification (Interpretable ML)
Sareer Ahmad
Interpretable Machine Learning for Predicting GHS Chemical Hazard Classifications
A multi-label classification framework that predicts all nine GHS hazard pictograms from molecular structure alone, and explains every prediction in chemical terms.
Author: Sareer Ahmad — Federal Directorate of Education, Islamabad, Pakistan ORCID: 0009-0003-2580-091X Contact: sareerkh9194@gmail.com Target journal: Computational Toxicology (Elsevier) Repository: https://github.com/sareer555/ghs-hazard-classification Archived code (v1.0.2): 10.5281/zenodo.21903565 · all versions Archived data (v1.0.0): 10.5281/zenodo.21876611 · all versions
▶ Try it live
https://ghs-hazard-classification.streamlit.app — no installation, no registration. Enter a chemical name, CAS number, PubChem CID or SMILES string.
The hosted instance runs the same XGBoost model file that produced every result below. It sleeps after a period of inactivity, so the first load can take around 30 seconds to wake; subsequent pages are immediate.
Structures you enter there are sent to the hosting provider. If you are screening proprietary or unpublished structures, run the application locally instead (see Quick start) — nothing then leaves your machine.
Citing this work. Cite the DOIs above rather than the GitHub URL. The archives are permanent and version-locked; a repository URL can move or be deleted. Cite the version DOI, not the "all versions" one — the latter follows whichever release is newest, so it cannot identify what produced a particular result. The data archive (366 MB) holds the curated dataset, descriptor matrix and trained models that are too large for GitHub. Every published result was produced by code release v1.0.0. v1.0.2 is the release to cite: no pipeline step differs, so it reproduces every result identically, but it corrects the workflow figure's claim that SMOTE was applied and adds the applicability-domain warning for small molecules.
Note on what is and is not in this repository. The code, models needed to run the application, figures, tables and manuscript are all here (~27 MB). The large intermediate arrays and datasets — about 8 GB — are excluded by.gitignoreand are archived separately on Zenodo. Re-running the pipeline regenerates every one of them from PubChem. See Reproducing from scratch below.
Quick start
The environment is already built at
.venv. To use the trained model:To launch the web interface, then open http://localhost:8501:
If you rebuild the environment from scratch, read the Streamlit note under "Documented deviations" first. Two issues stop the web app from starting on a fresh install, and both are already fixed in this copy.
To re-run the whole analysis:
What this project does
GHS hazard classification governs how chemicals are labelled, stored and handled worldwide, but the experimental testing it depends on has been done for only a small fraction of chemicals in industrial use. In March 2019 improperly identified chemical waste discharged into the Sungai Kim Kim river at Pasir Gudang, Johor, affected more than 2,500 people, most of them schoolchildren.
This project asks whether the nine GHS pictograms can be predicted from structure alone, and whether those predictions can be made transparent enough for a regulator to act on.
The nine hazard classes
Column names follow the official United Nations GHS pictogram numbering (10th revised edition), which is also what PubChem, ECHA and Malaysia's DOSH use.
Column name Code Pictogram Meaning
GHS01_Explosive GHS01 Exploding bomb Explosive GHS02_Flammable GHS02 Flame Flammable GHS03_Oxidising GHS03 Flame over circle Oxidiser GHS04_CompressedGas GHS04 Gas cylinder Compressed gas GHS05_Corrosive GHS05 Corrosion Corrosive GHS06_AcuteToxicity GHS06 Skull and crossbones Acute toxicity GHS07_Irritant GHS07 Exclamation mark Irritant / harmful GHS08_HealthHazard GHS08 Health hazard Serious health hazard GHS09_Environmental GHS09 Environment Environmental hazardCorrection applied to the original study design
The research proposal named three of these columns
GHS07_HealthHazard, GHS08_Environmental and GHS09_Irritant, which rotates the descriptive suffixes relative to the UN scheme. The data were always bound to the numeric pictogram code and were therefore always correct — only the three labels were wrong. They have been renamed:Original proposal name Corrected name
GHS07_HealthHazard GHS07_Irritant GHS08_Environmental GHS08_HealthHazard GHS09_Irritant GHS09_EnvironmentalBecause the label matrix is a plain array whose column order never changed, no model was retrained and no prediction changed. The rename was applied by
src/migrate_column_names.py and the mapping is preserved in STEP2_ghs_label_schema.csv so that anyone holding an earlier copy of the outputs can translate between the two.Pipeline
Step Script What it does 1
src/step1_environment.py Environment setup and verification 2 src/step2_data_collection.py Harvest GHS annotations from PubChem 3 src/step3_data_cleaning.py Validate, deduplicate, reconcile labels 4 src/step4_descriptors.py Compute 1218 molecular descriptors 5 src/step5_scaffold_split.py Bemis-Murcko scaffold split (80/10/10) 6 src/step6_imbalance.py Class weights and metric selection. SMOTE is implemented but was not applied at this dataset size — see STEP6_smote_report.csv 7 src/step7_model_training.py Random Forest, XGBoost, SVM 8 src/step8_hyperparameter_tuning.py RandomizedSearchCV + refit on train 9 src/step9_evaluation.py Metrics, thresholds, bootstrap CIs, plots 10 src/step10_shap_analysis.py SHAP global and per-compound explanation 11 src/step11_malaysia_validation.py Malaysian sectors + Johor 2019 12 app.py, predict_ghs.py Web app and command-line tool 13 src/step13_publication.py Figures, tables, abstract, methodsSupporting analyses
Script What it does
run_full_pipeline.py Runs Steps 4–13 on the full 243,323-compound dataset src/controlled_size_experiment.py Does more training data help? Same test set, same hyperparameters, only size varies src/learning_curve.py Learning curve within a fixed dataset src/applicability_domain.py Where the model can be trusted: performance by molecule size src/build_supporting_information.py Compiles Tables S0–S5 and File S1 into one PDF src/redraw_figures_9_10.py Redraws Figures 9 and 10 from saved results, without retraining src/step4_full_from_colab.py Builds the full feature matrix from cached Colab descriptors src/migrate_column_names.py One-off rename of the three mislabelled GHS columns src/verify_interface.py 27 end-to-end checks on the app and CLI src/verify_deliverables.py Audits that every required output exists GHS_full_dataset_colab.ipynb Google Colab notebook for a full-dataset runDescriptors
1218 features per compound, reduced by a variance filter:
19 physicochemical — molecular weight, LogP, TPSA, H-bond donors and acceptors, ring counts, sp3 fraction, Labute ASA, Balaban J, Bertz CT
1024 Morgan / ECFP4 — circular substructure fingerprint, radius 2
167 MACCS keys — predefined substructure patterns
8 topological — Chi connectivity and Kappa shape indices
Why the scaffold split matters
Chemical datasets are full of near-duplicate molecules. A random split puts one member of a family in training and its twin in test, so the model appears far more accurate than it is. Bemis-Murcko scaffold splitting forces every compound sharing a chemical skeleton into the same split, so the reported scores reflect generalisation to genuinely new chemotypes.
Acyclic compounds have an empty Murcko scaffold. Rather than pooling them all into one enormous group — which would put most industrial solvents into a single split — each is treated as its own group.
How groups are allocated, and two ways of getting it wrong
Allocating scaffold groups to splits is harder than it looks, and two simpler schemes were tried and rejected. Both are documented in
src/step5_scaffold_split.py because both fail silently.Fill training first, then validation, then test. A single large scaffold group met after the training quota is full cannot fit in the 10% validation quota either, so it falls through to test. This produced an 80/3/17 split instead of 80/10/10.
Give each group to whichever split is furthest below its quota. The overall ratios come out exactly right, so this looks correct — but it starves the rare classes. Groups are processed largest first, so ring-bearing scaffolds fill the training quota early; after that the three splits have equal remaining capacity and single-compound groups are shared out roughly a third each. Every acyclic molecule is a single-compound group, and the rare classes are overwhelmingly small acyclic molecules — compressed gases are methane, nitrogen, hydrogen sulfide. The result:
Class Training share (should be 80%) GHS04 Compressed gas 30% GHS03 Oxidiser 39% GHS02 Flammable 54% GHS07 Irritant 81% ✓
The splitting algorithm was depriving the model of the examples it had fewest of, while every visible check passed.
What is used now: a group-wise form of iterative stratification for multi-label data. Each group is scored against every split on the worst fractional overshoot the assignment would cause — measured on overall size and on each hazard class the group contains — and goes to the split with the lowest worst case. Groups holding rare-class compounds are placed first. Every class now lands within one percentage point of 80%, with exact 80/10/10 split sizes.
Step 5 now also checks for this: any class whose training share falls more than 15 points below the overall share is reported as starved. That check is a warning, not a failure — an earlier version treated it as fatal, which sent the run to a non-scaffold fallback split and silently destroyed the leakage guarantee. Trading the validity of the whole evaluation for a better class balance is never the right trade.
Documented deviations from the proposal
Every one of these is recorded in the step logs under
logs/ and reproduced in FINAL_PROJECT_SUMMARY_REPORT.pdf.Python installation (Step 1). No Python existed on the machine. The official python.org installer failed twice (WiX exit
0x3, no admin elevation for the chained MSIs) and conda was unavailable, so the proposal's conda fallback could not be used. A standalone CPython 3.11.15 was installed with uv, which needs no Windows installer.PubChem API change (Step 2). PubChem renamed
CanonicalSMILES to ConnectivitySMILES in 2025. The new property names are requested explicitly so that isomeric SMILES are obtained.CAS coverage (Step 2). CAS numbers come from synonym lists, which are slow to download. A 30-minute budget was applied, yielding CAS for 65% of compounds. CAS is metadata only and is never a model feature.
3b. Structure-download losses (Step 2). Seven batched property requests failed and were not retried compound-by-compound, costing 2,100 of the 245,807 annotated compounds (0.85%). The loss is random with respect to chemistry and does not bias any class.
Pictogram numbering (Step 2). Three column names in the original study design did not match the UN pictogram scheme. The data were bound to the numeric code and were unaffected; the labels have since been corrected. See the table above.
Modelling subset (Step 3f) — RESOLVED, no longer applies. Earlier versions modelled a 40,000-compound subset because of the 7.9 GB memory limit. A controlled experiment (
src/controlled_size_experiment.py) showed this was a real limitation, not a harmless one: mean AUC rises from 0.8187 to 0.8738 (+0.0551, four times the confidence interval) on going to the full 194,658-compound training set, and every class improves. All results now use the complete 243,323-compound dataset.Worth knowing: a learning curve computed within the subset appeared to plateau and suggested the opposite. That was an artefact — the subset had deliberately retained every rare-class positive, so those classes could not improve and the aggregate curve flattened prematurely. A learning curve on a non-representative subsample can actively mislead.
Imputation before variance filtering (Step 4). The proposal lists 4e before 4f. A column containing a NaN has undefined variance, so imputing first is required for the filter to work as intended.
Random Forest depth (Step 7).
max_depth=None for 200 trees × 9 classes would need several GB more than the machine has. Depth is capped and a minimum leaf size imposed.SMOTE versus class weights (Step 7). The proposal specifies both SMOTE-balanced input and
class_weight='balanced', which corrects the imbalance twice. Class weighting is used for the primary models and SMOTE is run as a separate ablation, so the effect is measured rather than assumed.SVM scale (Step 7). An RBF kernel costs O(n²). The SVM uses the top 100 features (the proposal's own fallback) and a subsampled training set. Its scores are therefore not strictly comparable with the other two algorithms.
Tuning budget (Step 8).
n_iter=30 with 5-fold CV was measured at well over a day. A timing probe sets the largest n_iter that fits a fixed wall-clock budget, with 3-fold CV.Refit after tuning (Step 8).
RandomizedSearchCV.best_estimator_ is fitted on whatever data the search saw — the validation set. Using it as the final model would waste 90% of the training data and leak validation information into the reported scores. The winning settings are taken from the search and a fresh model is fitted on the training set.SHAP sample (Step 10). SHAP values are computed on a 500-compound random sample of the test set, which is the proposal's own documented fallback.
Streamlit will not start on a fresh install (Step 12). Two separate faults, both fixed here, both of which will recur on any clean rebuild:
(a) First-run email prompt. Streamlit asks for an email address on the terminal before it starts its web server. Launched in the background the process looks alive but never opens a port, so the app appears to hang. Fixed by
~/.streamlit/credentials.toml containing an empty email, and by headless = true in .streamlit/config.toml.(b) Streamlit/Starlette incompatibility —
starlette is pinned. Streamlit 1.61.0 declares starlette<2,>=0.46.0, but that range is too permissive. Starlette 1.4.0 added a required keyword-only argument thread_minimum_size to GZipResponder.__init__, which Streamlit's own gzip middleware subclasses without passing. The server starts and reports itself healthy, then returns HTTP 500 for every request:An unconstrained
pip install streamlit picks the newest Starlette and reproduces the fault. The constraint starlette<1.0 is recorded in STEP1_environment_requirements.txt, and Step 1's smoke tests now check the GZipResponder signature so a broken combination is caught at setup rather than at first page load:.streamlit/config.toml also sets gatherUsageStats = false — this is a chemical safety tool, and no information about the chemicals a user submits should leave the machine.Results
Dataset: 243,323 unique compounds after cleaning; 40,000-compound label-stratified modelling subset; 40,000 × 817 feature matrix; scaffold split 32,000 / 4,000 / 4,000 with zero scaffold leakage.
Best model: XGBoost, mean AUC-ROC 0.890 across the nine classes.
Class Meaning AUC-ROC MCC
GHS04_CompressedGas Compressed gas 0.998 0.842 GHS03_Oxidising Oxidiser 0.997 0.828 GHS01_Explosive Explosive 0.973 0.534 GHS02_Flammable Flammable 0.956 0.698 GHS05_Corrosive Corrosive 0.868 0.521 GHS08_HealthHazard Serious health hazard 0.838 0.381 GHS09_Environmental Environmental hazard 0.831 0.347 GHS06_AcuteToxicity Acute toxicity 0.783 0.272 GHS07_Irritant Irritant / harmful 0.767 0.228Physically determined hazards (gas, oxidiser, explosive, flammable) are predicted almost perfectly from structure. Biologically mediated ones (acute toxicity, irritation) are hardest, because they depend on mechanism rather than bulk molecular properties.
Model selection. Three of the four models were statistically indistinguishable on AUC (bootstrap CI half-width 0.0139). RandomForest+SMOTE had the highest mean AUC by 0.0002, which is meaningless; XGBoost was selected on mean MCC (0.517 vs 0.500), the metric designated for the rare classes.
SMOTE versus class weighting — a question the proposal left open. SMOTE clearly beat plain class weighting for the Random Forest (MCC 0.500 vs 0.432), but neither beat gradient boosting.
Interpretability. The strongest single finding is that MACCS key 124,
[!#6;!#1]~[!#6;!#1] — two directly bonded heteroatoms — is the top predictor for both explosives (r = +0.96) and oxidisers (r = +0.97). That is the structural signature of nitro, nitrate, peroxide and azide groups, recovered by the model without being told.Malaysian validation (43 compounds): hazard recall 0.900 palm oil, 0.873 Johor 2019 incident, 0.871 petrochemicals, 0.667 rubber and semiconductor. On the Johor chemicals the framework recovered 10/10 flammable, 5/5 acute toxicity and 10/10 serious-health-hazard labels, but 0/3 corrosive — it missed hydrogen chloride. Inorganic acids sit outside the applicability domain of these descriptors, and the weakest sectors (rubber, semiconductor) are the most inorganic-heavy. This is a real boundary of the method, not a tuning problem.
Reproducibility
Random seed 42 is applied to Python's
random, NumPy, scikit-learn, XGBoost and every resampling procedure. The exact environment is pinned in STEP1_environment_requirements.txt. Steps 3 to 10 and 13 are deterministic: re-running them on the saved data reproduces every number exactly.One exception. Step 11 queries PubChem live for the structures and reference classifications of the Malaysian and Johor chemicals, so its results depend on PubChem's content and availability at the moment it runs. Between two runs the mean accuracy moved from 0.831 to 0.835 for this reason. To make Step 11 exactly reproducible, work from the saved snapshot
malaysia_validation/STEP11_malaysian_chemicals_raw_*.csv instead of re-querying.Disclaimer
These models are computational screening tools. They do not replace laboratory testing or regulatory assessment under Malaysia's Occupational Safety and Health (Classification, Labelling and Safety Data Sheet of Hazardous Chemicals) Regulations 2013.
Like this project
Posted Aug 28, 2026
XGBoost model predicting all 9 GHS hazard pictograms from molecular structure, with SHAP explainability. Deployed on Streamlit; now a paid report service.
