Gene Druggability Finder

Gene Druggability Finder (GDF) is a leakage-aware, multi-evidence framework for predicting and benchmarking human gene druggability.

The pipeline:

1. Builds eight druggability target labels for 19,295 HGNC-approved protein-coding genes.
2. Generates 22 independent biological feature blocks.
3. Merges 2,285 feature columns into a gene-level feature matrix.
4. Creates 440 target-feature datasets.
5. Benchmarks Random Forest and XGBoost using stratified five-fold cross-validation.
6. Summarises model performance and feature importance.

Interpretation: GDF predicts database-constructed druggability labels. It is a target-prioritisation and evidence-integration framework, not definitive proof that a gene will become a clinically successful drug target.

Key Results

The reported internal benchmark covers 440 target-feature datasets:

Result	Value
Human protein-coding genes	19,295
Druggability targets	8
Feature blocks	22
Feature columns before model sanitisation	2,285
Feature-group configurations	55
Train-ready datasets	440
Best mean cross-validated AUROC	0.9809
Median AUROC	0.8488
Mean AUROC	0.8344 +/- 0.1005
Median AUPRC	0.7476
Median Matthews correlation coefficient	0.4411
XGBoost selected as best model	350 datasets (79.5%)
Random Forest selected as best model	90 datasets (20.5%)

The highest AUROC was achieved by Dataset413_GRP_NoStructure_T5, which predicted the Biologic/Modality Target with:

• AUROC: 0.9809
• AUPRC: 0.9817
• Matthews correlation coefficient: 0.8786
• Best model: XGBoost

Integrated feature sets performed better than most individual feature blocks:

Feature tier	Datasets	Best AUROC	Median AUROC	Median AUPRC	Median MCC
Cumulative	176	0.9809	0.9346	0.9045	0.6497
Thematic	88	0.9809	0.8289	0.7620	0.4367
Individual feature blocks	176	0.9421	0.7475	0.6377	0.3093

These are internal cross-validation results. External and prospective validation are still required.

Pipeline Design

GDF separates target-label construction from predictive feature generation:

LABEL SIDE
HGNC + ChEMBL + Open Targets + DGIdb + Pharos/TCRD
    -> eight gene-level druggability targets

FEATURE SIDE
22 independent biological and molecular evidence blocks
    -> merged feature matrix

ANALYSIS SIDE
feature matrix + target labels
    -> 440 datasets
    -> Random Forest and XGBoost benchmark
    -> performance and feature-importance summaries

Direct target-defining evidence from ChEMBL, Open Targets, DGIdb and Pharos/TCRD is reserved for the label side and excluded from model inputs.

Repository Workflow

Run commands from the repository root.

Step 1: Install Dependencies

Python 3.10 is recommended.

conda create -n druggability python=3.10 -y
conda activate druggability

pip install pandas numpy requests tqdm pyarrow openpyxl scipy \
  scikit-learn xgboost networkx h5py biopython matplotlib

Optional system tools:

• fpocket for Feature 10 pocket scores.
• SLURM for HPC array execution.
• curl or wget as download fallbacks used by some scripts.

Step 2: Download Target-Label Databases

python Dataset0-DownloadDatabases.py --all --opentargets-release 26.03

The downloader skips valid files that are already present. Use --overwrite only when a deliberate refresh is required.

Step 3: Build Druggability Labels

python Dataset1-GenerateData.py --save-raw-evidence
python Dataset1.1-GenerateDataCheckFeatures.py

Primary output:

Step0_Output/03_HumanGene_DruggabilityLabels.csv

Step 4: Download Core Feature Databases

Recommended first download:

python Feature0-DownloadDatabase.py --recommended

This downloads HGNC, UniProt, Ensembl, STRING and GTEx inputs.

Large optional downloads:

python Feature0-DownloadDatabase.py --recommended --include-alphafold
python Feature0-DownloadDatabase.py --sources interpro --include-interpro-huge --include-pfam
python Feature0-DownloadDatabase.py --full

The downloader skips existing valid files, resumes .part downloads, dynamically discovers the current AlphaFold human archive and uses current Pfam bulk filenames.

Step 5: Generate the 22 Feature Blocks

Run the feature scripts after their source databases are available:

python Feature1_DeMap.py
python Feature2_String.py
python Feature3_Pathway.py
python Feature4_AlphaFold.py
python Feature5_InterProPfam.py
python Feature6_UniProt.py
python Feature7_GTEx.py
python Feature8_Ensmbl.py
python Feature9_Genetics_Contraint.py
python Feature11-ProteinFeatures.py
python Feature12-ProteinSequence.py
python Feature13_GWAS.py --download
python Feature14_GO.py --download
python Feature15_BioGRID.py --download
python Feature16_HPA.py
python Feature17_CTD.py --download
python Feature18_MGI.py --download
python Feature19_gnomAD_full.py
python Feature20_CORUM.py
python Feature21_PhosphoSitePlus.py --download
python Feature22_Paralogues.py

Feature 10 is designed as a chunked structure/pocket workflow:

sbatch Feature10-FPocket.sh
python Feature10-FPocket-merge.py

For a small local Feature 10 test:

python Feature10-FPocket.py 1 --chunk-size 100
python Feature10-FPocket.py 1 --chunk-size 100 --run-fpocket

Step 6: Merge All Feature Blocks

python Analysis1-MergeData.py --verbose

Output:

Dataset/Features_All.csv

The merger:

• Locates the expected gene-level output from each feature block.
• Normalises gene symbols.
• Prefixes columns with Feature1_ through Feature22_.
• Excludes identifiers, raw text, paths, URLs and other non-model columns.
• Outer-merges all available feature blocks.

Step 7: Generate Feature Statistics

python Analysis2-ListFeaturesStatistics.py \
  --features Dataset/Features_All.csv \
  --out Feature_Statistics.xlsx

python Analysis2.1-ListFeaturesStatistics.py

The first script calculates gene coverage, missingness and numeric summary statistics for every feature. The second prints feature counts per block.

Do not set --out to the existing Supplementary Material 2.xlsx unless you intend to replace it. Analysis2-ListFeaturesStatistics.py creates a new workbook and would overwrite the compiled performance and feature-importance sheets.

Step 8: Create Train-Ready Datasets

python Analysis3-MakeDatasets.py

This crosses eight target definitions with 55 feature-group configurations:

• 22 individual groups: F01 to F22
• 22 cumulative groups: CUM01 to CUM22
• 11 thematic groups

Thematic groups:

Group	Included feature blocks
GRP_Structure	4, 10
GRP_Network	2, 15
GRP_Expression	7, 16
GRP_Constraint	9, 19
GRP_Pathway	3, 14
GRP_Functional	5, 6, 11, 12
GRP_Omics	1, 18
GRP_Annotation	8, 13, 20, 21, 22
GRP_Literature	17
GRP_NoStructure	All except 4 and 10
GRP_Full	All 22 blocks

Each generated dataset contains:

Datasets/Dataset001_F01_T1/
    X_train.csv
    y_train.csv
    meta.json
    feature_columns.txt
    missingness_summary.csv

Missing values remain in X_train.csv and are imputed inside training folds.

Step 9: Train Models

Train one dataset locally:

python Analysis4-TrainModels.py 1 --no-permutation

Train all 440 datasets on SLURM:

sbatch Analysis4-TrainModels.sh

The training script:

• Removes identifiers and suspicious leakage-related columns.
• Converts selected predictors to numeric values.
• Replaces infinite and unsafe values with missing values.
• Performs median imputation inside each cross-validation fold.
• Evaluates Random Forest and XGBoost.
• Uses stratified five-fold cross-validation by default.
• Selects the best model using mean cross-validated AUROC.
• Writes fold predictions, metrics, confusion matrices and feature importance.

Main training outputs inside each dataset's Training/ directory:

00_training_metadata.json
00_feature_cleaning_report.csv
01_feature_columns_used.csv
02_target_info.csv
03_cv_all_predictions.csv
04_cv_fold_metrics.csv
05_model_comparison_summary.csv
06_best_model_summary.csv
07_feature_importance_all_folds.csv
08_feature_importance_summary.csv
09_top_features_best_model.csv
10_confusion_matrix_per_fold.csv
11_confusion_matrix_aggregate.csv
12_permutation_test_results.csv
13_leakage_audit.csv

Step 10: Audit Dataset and Training Outputs

python Analysis4.0-CheckDatasets.py

Analysis4.0-CheckDatasets.py currently contains an HPC-specific absolute path:

/data/ascher02/uqmmune1/DrugableGeneFinder/Final

Update BASE_DIR in that script before running it elsewhere.

Eight Druggability Targets

The target table contains eight operational definitions of druggability:

Target	Definition	Positive genes	Prevalence
T1 Clinical Target	Approved or clinically established target; the strictest primary definition.	1,052	5.45%
T2 Clinical Investigation Target	Phase 1-3 clinical-investigation evidence.	504	2.61%
T3 Small-Molecule Target	Small-molecule tractability evidence.	7,231	37.48%
T4 Chemical Tractability Target	Broader chemical tractability evidence.	7,058	36.58%
T5 Biologic/Modality Target	Biologic, antibody, targeted-degradation or other modality evidence.	9,699	50.27%
T6 Drug-Gene Interaction Target	DGIdb-supported drug-gene interaction evidence.	4,575	23.71%
T7 Potentially Druggable Family Target	Potentially druggable family/category annotation.	9,518	49.33%
T8 Broad Druggability Target	Positive for any preceding druggability definition.	14,791	76.66%

T1 is the primary conservative target. T8 is a broad evidence-union label and should not be interpreted as a strict clinical gold standard.

Label-Construction Databases

Database	Purpose	Access
HGNC	Defines the human gene universe and stable identifier mappings.	HGNC complete set
ChEMBL	Human targets, mechanisms, approved/phase 4 drugs and potent activity evidence.	ChEMBL releases
Open Targets	Target identity, tractability and clinical-indication evidence.	Open Targets platform downloads
DGIdb	Drug-gene interactions and druggable-category evidence.	DGIdb downloads
Pharos/TCRD	Target development levels: TCLIN, TCHEM, TBIO and TDARK.	Pharos GraphQL API

Feature Blocks and Database Links

The current feature-statistics workbook contains 2,285 feature columns:

Feature	Source	Columns	Access
1	DepMap functional genomics	90	DepMap downloads / file-index API
2	STRING v12.0 PPI network	48	STRING API
3	Reactome pathways	30	Reactome downloads
4	AlphaFold DB + RCSB PDB structures	33	AlphaFold DB / RCSB PDB
5	InterPro + Pfam domains	72	InterPro / Pfam FTP
6	UniProt protein annotation	91	UniProt REST API
7	GTEx tissue expression	108	GTEx portal / v8 median TPM
8	Ensembl gene structure	52	Ensembl human GTF
9	gnomAD + MobiDB + DisProt constraint/disorder	74	gnomAD / MobiDB / DisProt
10	fpocket binding-pocket features	186	fpocket
11	Protein physicochemical/sequence features	463	Derived locally from protein FASTA
12	Protein language-model embeddings + PCA	77	UniProt embedding archive
13	NHGRI-EBI GWAS Catalog	43	GWAS association download
14	Gene Ontology / QuickGO	82	GOA human GAF / go-basic.obo
15	BioGRID curated interactions	50	BioGRID latest organism release
16	Human Protein Atlas	252	HPA downloads
17	CTD chemical-gene interaction burden	72	CTD chemical-gene interactions
18	Mouse Genome Informatics phenotypes	56	MGI reports
19	Full gnomAD constraint	109	gnomAD downloads
20	CORUM protein complexes	122	CORUM
21	PhosphoSitePlus PTM sites	153	PhosphoSitePlus downloads
22	Ensembl BioMart/Compara paralogues	22	Ensembl BioMart
Total	All 22 blocks	2,285

Feature Script Catalog

Script	Purpose	Main output
Feature0-DownloadDatabase.py	Downloads core non-label feature databases.	feature_databases/
Feature1_DeMap.py	DepMap functional-genomics feature builder.	feature1.dmapp.database/
Feature1_DepMap.py	DepMap download-only helper.	Downloaded DepMap files
Feature2_String.py	STRING network topology and evidence-channel features.	feature2.string.database/
Feature3_Pathway.py	Reactome pathway membership and hierarchy.	feature3_pathway/
Feature4_AlphaFold.py	AlphaFold and PDB structure features.	feature4_structure/
Feature5_InterProPfam.py	InterPro/Pfam domain architecture.	feature5_interpro_pfam/
Feature6_UniProt.py	UniProt annotation and sequence features.	feature6_uniprot/
Feature7_GTEx.py	GTEx tissue-expression features.	feature7_gtex/
Feature8_Ensmbl.py	Ensembl genomic annotation and gene structure.	feature8_ensembl/
Feature9_Genetics_Contraint.py	Disorder and genetic constraint/intolerance.	feature9_disorder_constraint/
Feature10-FPocket.py	Chunked geometry and optional fpocket scoring.	feature10_pocket_geometry/processed/chunks/
Feature10-FPocket-merge.py	Merges Feature 10 chunks.	feature10_pocket_geometry/
Feature11-ProteinFeatures.py	Protein sequence composition and physicochemical features.	feature11_protein_sequence/
Feature12-ProteinSequence.py	Protein embedding and PCA features.	feature12_protein_embeddings/
Feature13_GWAS.py	GWAS association burden and pleiotropy.	feature13_gwas_catalog/
Feature14_GO.py	Gene Ontology features.	feature14_gene_ontology/
Feature15_BioGRID.py	Curated interaction-network features.	feature15_biogrid/
Feature16_HPA.py	HPA expression and localisation features.	feature16_hpa/
Feature17_CTD.py	Aggregate CTD chemical-gene burden.	feature17_ctd/
Feature18_MGI.py	Mouse orthology and phenotype features.	feature18_mgi/
Feature19_gnomAD_full.py	Expanded gnomAD constraint metrics.	feature19_gnomad_full/
Feature20_CORUM.py	Protein-complex membership features.	feature20_corum/
Feature21_PhosphoSitePlus.py	PTM and kinase-substrate features.	feature21_phosphositeplus/
Feature22_Paralogues.py	Human paralogue features.	feature22_paralogues/

Analysis and Modelling Scripts

Script	Purpose
Analysis1-MergeData.py	Merges all available feature blocks into Dataset/Features_All.csv.
Analysis2-ListFeaturesStatistics.py	Produces per-feature coverage, missingness and numeric statistics.
Analysis2.1-ListFeaturesStatistics.py	Prints feature counts per block.
Analysis3-MakeDatasets.py	Builds the 440 train-ready target-feature datasets.
Analysis4-TrainModels.py	Trains Random Forest and XGBoost for one dataset.
Analysis4-TrainModels.sh	Runs all 440 datasets as a SLURM array.
Analysis4.0-CheckDatasets.py	Audits dataset and training output completeness.

Target-Level Benchmark Results

Results below use the intended 440-dataset benchmark:

Target	Best AUROC	Median AUROC	Median AUPRC	Median MCC	Best feature group
T1 Clinical Target	0.9738	0.9142	0.5582	0.4521	CUM20
T2 Clinical Investigation Target	0.9302	0.8491	0.1966	0.1739	CUM22
T3 Small-Molecule Target	0.9394	0.8505	0.7849	0.5198	CUM18
T4 Chemical Tractability Target	0.9397	0.8488	0.7726	0.5147	CUM22
T5 Biologic/Modality Target	0.9809	0.8496	0.8666	0.5325	GRP_NoStructure
T6 Drug-Gene Interaction Target	0.8914	0.8197	0.6402	0.4277	CUM19
T7 Potentially Druggable Family Target	0.9142	0.8249	0.8312	0.4819	CUM20
T8 Broad Druggability Target	0.9557	0.8345	0.9418	0.4339	GRP_Full

The strict Clinical Target has low prevalence, so AUPRC and AUPRC enrichment are especially important alongside AUROC. The broad T8 target has high prevalence and should be interpreted as an evidence-union target.

Feature Importance

The supplied results contain:

• 516,672 model-derived feature-importance rows.
• 21,168 top-feature rows.
• Feature importance outputs for all trained datasets.

Highly ranked individual predictors include:

• PDB structure availability and PDB count.
• Reactome Ensembl/UniProt mapping indicators.
• STRING weighted degree, k-core and approximate closeness centrality.
• UniProt transmembrane, membrane and signal-peptide annotations.
• Gene Ontology author-statement/evidence-code counts.
• DepMap expression summaries.

The supplied manuscript reports the highest-ranked evidence blocks as STRING network features, DepMap functional genomics, Reactome pathways, InterPro/Pfam domains, UniProt annotations and AlphaFold/PDB structure features. This supports the central conclusion that predictive signal is distributed across multiple complementary evidence sources.

These values describe model contribution, not causal biological importance. Correlated predictors, annotation density and tree-model splitting behavior can affect importance rankings.

Data Leakage Controls

Direct target-defining evidence must never be used as model input:

• ChEMBL target/activity evidence.
• Open Targets tractability and known-drug evidence.
• DGIdb direct drug-gene evidence.
• Pharos/TCRD target-development labels.
• Approved-drug counts.
• Clinical-target labels.

The modelling pipeline also removes identifiers, text, accessions, paths, URLs, raw fields and columns matching suspicious leakage terms.

CTD is retained only as aggregate chemical-gene perturbation burden features. Chemical names and direct drug-label fields are excluded.

Supplementary Workbook

Supplementary Material 2.xlsx currently contains:

Sheet	Current contents
Feature Statistics	2,285 feature rows plus header.
Dataset Statistics	Dataset composition, target, feature-count and missingness records.
Model Performance	Best-model performance summaries for the benchmark datasets.
Feature Importance	Global and within-block feature-importance rankings.

Workbook Consistency Note

The current workbook contains 441 rows in Model Performance because Dataset ID 97 appears twice. The intended experimental design and reported benchmark contain 440 unique dataset IDs. Results in this README were calculated after deduplicating by Dataset ID.

Treat the workbook as a compiled results artifact and preserve a backup before running scripts that write Excel output.

Current Naming and Consistency Notes

Use the filenames that actually exist in this repository:

• Feature4_AlphaFold.py is internally described as the structure feature.
• Feature8_Ensmbl.py contains the filename spelling Ensmbl.
• Feature9_Genetics_Contraint.py contains the filename spelling Contraint.
• Feature11-ProteinFeatures.py implements protein physicochemical/sequence features.
• Feature12-ProteinSequence.py implements protein embeddings.

The workbook and supplied manuscript label Feature 12 as ESM-2 embeddings, while the current Feature 12 script documentation and download URL refer to the UniProt ProtT5 per-protein embedding archive. Confirm the intended embedding source before rerunning Feature 12 or reporting the method.

Expected Directory Structure

databases/
    HGNC/
    ChEMBL/
    OpenTargets/
    DGIdb/
    TCRD_Pharos/

feature_databases/
    AlphaFold/
    BioGRID/
    CORUM/
    CTD/
    Ensembl/
    Ensembl_Paralogues/
    GeneOntology/
    GTEx/
    GWASCatalog/
    HPA/
    InterPro_Pfam/
    MGI/
    PhosphoSitePlus/
    ProteinEmbeddings/
    STRING/
    UniProt/
    gnomAD/

Step0_Output/
feature1.dmapp.database/
feature2.string.database/
...
feature22_paralogues/

Dataset/
    Features_All.csv

Datasets/
    dataset_catalogue.csv
    Dataset001_F01_T1/
    ...
    Dataset440_GRP_Structure_T8/

Caching and Reproducibility

• Downloaders skip existing valid files unless explicitly forced.
• API-based feature scripts cache downloaded/raw responses where possible.
• Most feature scripts write summary text and JSON run metadata.
• Use the same HGNC release across all feature blocks.
• Archive database versions and download dates with final outputs.
• Keep raw downloads until processed outputs and mappings are validated.
• Impute missing values only inside model-training folds.
• Treat model performance as internal validation until external validation is completed.

Troubleshooting

Downloader reports no sources selected

Supply a mode:

python Feature0-DownloadDatabase.py --recommended

Existing files are skipped

This is expected:

[SKIP] Existing file: ...

Use --overwrite, --force-download or --force only when the relevant script documents the option and a refresh is required.

A feature source is unavailable

External APIs and FTP layouts can change. Keep caches, retry later or manually place the expected source file in the documented feature_databases/ subdirectory.

A full run is too large

Use each script's testing options first, for example:

python Feature4_AlphaFold.py --limit-genes 100
python Feature13_GWAS.py --max-rows 100000
python Feature18_MGI.py --max-pheno-rows 200000
python Feature22_Paralogues.py --limit-genes 1000
python Analysis3-MakeDatasets.py --dry-run

Data Availability

All source data used by the pipeline come from publicly available resources or resources available for academic use. No private or patient-identifiable data are used.

Repository: MuhammadMuneeb007/Gene_Druggability_Finder

Citation and Reuse

MIT LICENSE