Leer en Español (Guía completa)
Reproducible virtual screening that closes the loop design → synthesizability filter → docking → interaction-quality scoring → ADMET, with an objective per-cavity scoring function, orthogonal confidence metric, and multi-objective Pareto optimization.
v1.2.1 — Standalone Zero-Server interactive HTML reports (3Dmol + Plotly), authentic fpocket cavity detection, and PyMOL transport tunnel visualizer bundles.
The recommended distribution route is Docker (or Linux/WSL). Because docking and interaction scoring involve floating-point operations across multiple C/Fortran libraries, running inside the container ensures exact reproducibility across any computer.
- Install and start Docker Desktop.
- Download
PoliScreen-Docker.batandPoliScreen.icofrom the latest release. - Double-click
PoliScreen-Docker.bat. It will pull the official pre-built image, create a desktop shortcut with the official icon, and openhttp://localhost:8501in your browser.
docker run --rm -it --init -p 127.0.0.1:8501:8501 \
-v "$PWD/proyectos:/data" ghcr.io/diegoanyg/poliscreen:latest
# then open http://localhost:8501For native conda installation for local development: docs/INSTALL.md. System requirements: docs/REQUIREMENTS.md.
PoliScreen features a modular web interface (available in both English and Spanish under Settings → Language):
-
Receptors — Download from the PDB (e.g.
8HTB) or upload a local.pdbfile. Inspect chains, keep cofactors, and extract the co-crystallized ligand as the control. (Tip: click Example (8HTB) to load the worked example immediately). -
Ligands — Upload your molecules (
.sdf,.mol2,.smi), build a series by reaction (with synthesizability verification), screen approved drugs from ChEMBL, or enumerate peptides. (Tip: click Example ligands (8HTB) to load 3 active Cruzain inhibitors and 3 decoys). - Run — Configure the search box around the cavity and launch docking (AutoDock Vina for small molecules, ADCP for peptides) and optional tunnel calculation (CAVER).
-
Results — Explore the ranking through:
- Interactive Pareto Frontier: Multi-objective landscape balancing binding affinity, interaction quality, and confidence score, featuring live 2D structure tooltips on hover.
- TOP Interaction Leaderboard: Geometric footprint polygons where vertices and edges visually represent specific protein contact residues and bond types.
- ADMET Profiler: Radar plots and endpoint estimations.
-
Transport Tunnels: Activation barrier (
$E_a$ ) and binding energy profile.
A detailed tutorial with the scientific rationale of each control is in docs/TUTORIAL.md.
Most screening panels rank strictly by affinity and contact count. That rewards promiscuity: a molecule touching many irrelevant residues can outrank one that anchors exactly where it should. PoliScreen is built on four core principles:
-
Objective per-cavity scoring, not similarity to the control. Interaction quality is the sum of each contact weighted by bond type (salt bridge > H-bond > pi-stacking > hydrophobic) and by residue role (catalytic, secondary, cavity, external). It is normalized against the fingerprint of the crystallographic ligand in its real pose — so a compound beats the control by making more and better productive contacts, not by copying it.
-
Confidence metric, orthogonal to the score.
confidence(0–1) is the geometric mean of binding-mode convergence across poses, affinity-interaction agreement, and (if enabled) Vina-neural-network consensus, attenuated when the control fails to reproduce its crystallographic pose. It quantifies how much to trust a result, not its magnitude. A high score with low confidence is a red flag. -
Synthesizability from the design stage. Analogues are filtered by real reaction feasibility (regioselectivity, OH classification, steric hindrance), so what gets docked is what a chemist can actually make.
-
Explicit reproducibility. Fixed seed, single thread per docking task, and a Methods export with every parameter and version. Inside one container image, two runs of the same configuration give the same result — and the image digest is what a paper cites.
poliscreen fingerprinthashes each stage so two machines can be compared.
The co-crystallised control defines the reference everything is measured against, so its chemistry is not guessed: bond orders come from the PDB's own chemical component dictionary, looked up from the ligand code. A single interaction engine (PLIP) feeds both the table and the diagrams: same author numbering, same bond types, no mismatch between what is measured and what is drawn.
- AutoDock Vina 1.2.5: Small molecule docking engine.
- AutoDock CrankPep (ADCP): Peptide docking (5–20 residues).
- PLIP: Macromolecular interaction profiler (hydrogen bonds, salt bridges, pi-stacking, hydrophobic contacts).
- fpocket: Cavity and binding pocket detection.
- CAVER 3.0.2: Transport tunnel detection (included in the container image).
- CaverDock: Transport trajectory and energy barrier profiling (via caver-translate).
- gnina: Optional deep learning neural-network re-scoring (GPU-accelerated).
- ADMET-AI: Molecular property and pharmacokinetic predictions.
- Rigid receptor: No side-chain flexibility during standard Vina docking.
- Non-covalent docking: Vina and compatible engines do not model covalent bond formation.
- ADMET predictions are estimates: For prioritization and filtering, not experimental values.
See CITATION.cff (GitHub renders it under Cite this repository). Cite also the underlying tools PoliScreen invokes; the full bibliography is provided in the app's How to cite panel and in the exported Methods file.
GNU GPL v3 or later (see LICENSE). The scientific tools PoliScreen invokes (Vina, ADCP, PLIP, RDKit, Open Babel, fpocket, gnina, CAVER) are independent programs keeping their own respective licenses; PoliScreen does not incorporate their source code.
Diego Cesar Anaya Guerrero