The completed XAI benchmark includes an offline scientific validation suite covering source bias, adversarial separability, permutation controls, dummy and nearest-neighbour baselines, matched subsets, calibration, decoy sensitivity, stability, and explanation sanity. See docs/XAI_VALIDATION.md. Current status is benchmark-supported with cautions; it is not independently externally or experimentally validated.
Independent-model evaluation against audited BindingDB-deposited PubChem BioAssays is documented in docs/PUBCHEM_EXTERNAL_VALIDATION.md. The model remains locked, ChEMBL-deposited assays—including AID 162060—are excluded, and sparse inactive outcomes are reported without metric inflation claims.
DockForgeAI is a reproducible, resource-aware molecular docking, virtual-screening, interaction-analysis, and hit-prioritization workflow with an interactive Streamlit dashboard and portable scientific reports.
The project turns prepared protein and ligand structures into validated redocking poses, geometric interaction summaries, small-library screening rankings, RDKit descriptor profiles, and consensus prioritization. It exists to keep those steps auditable and practical on an Apple Silicon laptop with 8 GB RAM.
DockForgeAI does not turn docking scores into experimental claims. Vina scores are approximate ranking values, RDKit descriptors are not full ADMET predictions, and every predicted hit requires experimental validation.
- Protein retrieval, inspection, cleaning, and receptor preparation
- Ligand chemistry correction, conformer handling, and PDBQT preparation
- AutoDock Vina redocking, corrected multi-pose SDF export, and crystallographic RMSD validation
- Protein–ligand interaction comparison and catalytic Asp25 analysis
- Resumable, low-memory virtual screening and consensus hit prioritization
- Interactive py3Dmol pose viewer and RDKit 2D molecule rendering
- Offline PDF/HTML reports, manifest, download bundles, Docker packaging, and CI
Interactive overview of shortlisted candidates, consensus ranking, docking predictions, and molecular-dynamics status.
End-to-end computational workflow from molecular input through prediction, docking, consensus ranking, and molecular-dynamics validation.
Candidate-specific molecular structure, descriptors, RMSD, hydrogen bonds, contacts, and scientific interpretation.
Interactive comparison of shortlisted compounds using ligand RMSD, hydrogen-bond persistence, contacts, and distance metrics.
Browser-based visualization of representative and final protein-ligand structures.
Input structures → protein/ligand preparation → docking box → Vina redocking
→ RMSD validation → interaction analysis → demonstration screening
→ descriptors/alerts → consensus prioritization → PDF/HTML reports
The src/ tree contains testable domain modules, scripts/ contains command-line entry points, app/ provides Streamlit, config/settings.yaml centralizes paths and parameters, and data/processed/ retains validated outputs. See architecture.
The HIV-1 protease reference case retains chains A/B and the catalytic Asp25 dyad. The corrected redocking result is:
- Top-ranked Vina score: −11.575 kcal/mol
- Crystallographic heavy-atom RMSD: 0.518 Å
- Rank 1 passes the 2 Å criterion
- Interpretation: Validated strongly
The three-compound virtual screen completed 3/3 docking runs; compound 3 ranked first at −2.848 kcal/mol. This validates software execution only. It is not a biological discovery and does not establish any compound as an inhibitor.
Add project-owned screenshots at docs/images/dashboard.png, docs/images/viewer.png, and docs/images/export-center.png. These placeholders intentionally avoid external assets.
conda env create -f environment.yml
conda activate dockforgeFor an existing environment, use conda env update -n dockforge -f environment.yml --prune. Detailed platform notes are in installation.
Launch the dashboard:
conda run -n dockforge streamlit run app/streamlit_app.pyGenerate final reports from existing outputs:
conda run -n dockforge python scripts/generate_final_report.py --overwriteOpen reports through the Reports & Exports page or under data/processed/final_reports/.
docker compose up --buildThen open http://localhost:8501. Docker uses a Linux micromamba image and persistent data, reports, and logs mounts. Apple Silicon package availability and validation status are documented in Docker notes.
Supported inputs include PDB receptors, SDF/SMILES ligands, PDBQT docking inputs/poses, and CSV libraries or external ADMET tables with declared provenance. Generated outputs include PDB/PDBQT/SDF structures, CSV/JSON/Markdown analyses, PNG/SVG figures, PDF/HTML reports, and a JSON report manifest.
data/processed/
├── docking_results, docking_tables, docking_poses
├── interaction_reports
├── screening_tables, screening_reports
├── prioritization_reports
├── report_assets/{figures,molecules}
└── final_reports/{pdf,html,report_manifest.json}
conda run -n dockforge python -m compileall app src scripts tests
conda run -n dockforge pytest -qSeeds, Vina settings, directory paths, component weights, and catalytic residues live in config/settings.yaml. Reports retain project-relative source provenance and the current Git hash when available. Expensive docking is excluded from report and CI tests.
The workflow is intentionally serialized and chunked for an 8 GB Apple Silicon system. It has no GPU requirement. Larger libraries need deliberate batching and storage planning.
- Docking scores are approximate; they are not experimental binding free energies.
- Geometric interactions depend on protonation and distance/angle assumptions.
- Descriptor and alert calculations do not replace measured ADMET.
- The demo screen is too small for discovery claims.
- Molecular dynamics, free-energy calculations, biochemical assays, and cellular validation are outside the current scope.
Phase XAI-1 constructs a curated HIV-1 protease benchmark from ChEMBL target CHEMBL243 activities and DUD-E target HIVPR decoys. Actives require exact IC50, Ki, Kd, or EC50 measurements with pChEMBL ≥ 6.0 (or potency ≤1,000 nM). DUD-E structures carry inactive_or_decoy / dude_decoy labels and are computational decoys, not experimentally proven inactive compounds. Ambiguous measurements are excluded, structure duplicates and active/decoy overlaps are removed, and intact Bemis–Murcko scaffold groups define the 70/15/15 train/validation/test assignments to reduce structural leakage.
Build and validate using:
conda run -n dockforge python scripts/build_hivpr_xai_dataset.py --target-count 2000 --active-count 1000 --decoy-count 1000 --resume
conda run -n dockforge python scripts/validate_hivpr_xai_dataset.py --dataset data/processed/xai_dataset/hivpr_xai_2000_compounds.csvSee XAI dataset documentation for source provenance, labels, exclusions, and offline/resume behavior. Phase XAI-1 does not perform docking, feature engineering, machine learning, SHAP, active learning, or make model-performance claims.
Potential later work includes feature engineering and leakage-aware modelling, but only after this dataset is independently reviewed. No roadmap item should weaken the current reproducible baseline.
XAI-2–XAI-6 now provide leakage-safe descriptors/fingerprints, repeated-seed scaffold-safe CPU benchmarks, validation-only calibration, training-derived applicability-domain warnings, model-behaviour explanations, cached Streamlit pages, model cards, and PDF/HTML reports. Run the local pipeline without data retrieval:
conda run --no-capture-output -n dockforge python scripts/run_complete_xai_pipeline.py --seeds 2026 2027 2028 2029 2030 --bootstrap-iterations 500 --overwrite --offlineThe strong benchmark separation must be interpreted cautiously because ChEMBL actives and DUD-E decoys can expose source-specific shortcuts. It is not evidence of binding or therapeutic activity. See XAI limitations.
If you use DockForgeAI in your research, please cite:
Pratik Ramchandra Chaudhari.
DockForgeAI: An Open and Reproducible AI-Assisted Computational Drug Discovery Platform.
Version 1.0.1, 2026.
GitHub: https://github.com/Pratik-2002-ux/DockForgeAI
A permanent archival DOI may be added in a future release.
DockForgeAI is released under the MIT License.
See LICENSE for the complete license terms.
DockForgeAI is intended for research and educational use.
Machine-learning predictions, docking scores, consensus rankings, molecular-dynamics results, and other computational outputs should not be interpreted as experimental validation, clinical evidence, or proof of therapeutic efficacy.
Experimental validation is required before biological or therapeutic conclusions are made.
Pratik Ramchandra Chaudhari
Computational Biology • Bioinformatics • AI-Assisted Drug Discovery • Scientific Software Development
Read the developer guide, keep changes scoped, add tests, preserve the 0.518 Å reference result, and run the full compile/test commands before proposing changes.
Further guides: user guide · reporting · Docker