Consensus Docking Pipeline
Quick Answer
What is this tool? Molecular docking predicts how a small-molecule ligand binds to a protein target. This tool runs a 3-stage consensus pipeline: ESMFold predicts the protein 3D structure from amino acid sequence, AutoDock Vina screens ligand poses, and GNINA CNN re-scores the top hits for better ranking. All runs stay local with no data egress.
How do I use it online? Enter the protein amino acid sequence and a list of ligand SMILES on the docking page, then start the consensus screening workflow. Results rank each ligand by binding affinity with the consensus score. No account required for basic runs.
Validated engines used: ESMFold (MIT) for structure, AutoDock Vina for pose generation, and GNINA with CNN scoring for affinity refinement.
What is this tool?
A 3-stage consensus molecular docking pipeline that predicts protein 3D structure from amino acid sequence using ESMFold, screens small-molecule ligands with AutoDock Vina, and re-scores the top hits with GNINA CNN deep learning for improved ranking accuracy.
Enter a protein amino acid sequence and a list of ligand SMILES to run the full consensus screening workflow.
Free trial: test with example ligands. No credit card required.
Last updated: July 2026 · Reviewed by VigyanLLM Research Team
The consensus molecular docking pipeline integrates ESMFold for protein structure prediction, AutoDock Vina for rapid pose generation, and GNINA with CNN scoring for binding affinity refinement — all running locally with no data egress.
| Component | Function |
|---|---|
| ESMFold | Protein structure prediction from amino acid sequence |
| AutoDock Vina | Rapid ligand-protein pose generation and scoring |
| GNINA (CNN) | Deep learning binding affinity refinement (ΔG/kcal/mol) |
| Consensus Scoring | Weighted average of Vina + GNINA + CNN scores |
Quick Answer: Can You Read a Docking Score Yet?
A docking “score” is a prediction, not a measurement. It answers one question: given this protein and this ligand, how energetically plausible is this pose? A single number means nothing until you know which engine produced it, how the kcal/mol value sits in context, and whether the pose matches anything already known about the target. Concretely, watch for three values in any result: the predicted affinity (kcal/mol), a confidence signal about whether the pose is plausible, and, when you have one, the distance to a known crystal pose. If only one of those looks good and the others are weak, treat the result as a maybe. The worked example below is designed to make each of those three signs familiar so you can recognize them in your own screens.
Imatinib vs ABL1: A Worked Docking Example
The system behind this tool is a consensus pipeline: ESMFold predicts a protein structure from sequence, AutoDock Vina screens many ligand poses, and GNINA re-scores the top hits with a deep-learning network. Here’s how that unfolds on the most famous example in the field — docking imatinib (Gleevec) against the BCR-ABL1 kinase:
Why use this pair as the tutorial? Because the answer is already known. Imatinib binds the ATP site shut, locking ABL1 in its inactive conformation — a biological fact verified over two decades of clinical use. When a docking pipeline places the ligand near that pocket and scores it near a literature-grade value, you are testing the program against ground truth instead of against an assumption. If that pose survives those checks, it is a reasonable benchmark for how the same tool will behave on a novel target.
| Step | What happens | What you see |
|---|---|---|
| ESMFold | Predicts an ABL1 3D structure from the sequence | 3D model, comparable to the known crystal |
| AutoDock Vina | Screens binding poses of imatinib | Top pose in the ATP-binding pocket, ~−9.8 kcal/mol |
| GNINA CNN | Re-scores poses with deep learning | Confidence score + overlap check |
Imatinib is not a random test case. It is the drug that turned BCR-ABL1 (the kinase driving chronic myeloid leukaemia) into a druggable target and transformed structure-based drug discovery. If a pipeline reproduces its known ATP-binding pocket at around −9.8 kcal/mol, it is doing the physics roughly right — the exact sanity check to run before you trust a tool with your own targets.
We run three independent methods because any single docking algorithm can miss the correct binding pose. When all three agree, we’re confident; when they disagree, the GNINA score is usually the better predictor. Consensus also protects against the subtler failure where two engines share a hidden assumption and produce correlated wrong answers.
How to Interpret Your Results
| Metric | Good | Bad | What it means |
|---|---|---|---|
| Binding affinity (Vina) | < −7 kcal/mol | > −5 kcal/mol | How tightly the ligand binds (more negative = tighter) |
| GNINA confidence score | > 0.7 | < 0.5 | CNN confidence that the pose is correct |
| RMSD vs crystal | < 2.0 Å | > 3.0 Å | How close the docked pose is to the known structure |
The three metrics are deliberately independent. A poster-child pose can score well on affinity while GNINA hesitates, or GNINA can be confident about a pose that sits somewhere unexpected relative to the crystal. Reading them together, rather than fixating on the kcal/mol value alone, is the single most transferable skill in structure-based screening — and the one that separates experienced docking users from people who chase the biggest negative number.
When Docking Goes Wrong
In our experience, the most common docking mistake is not preparing the binding target first. Removing water molecules and adding the correct hydrogens can shift binding affinities by 1–2 kcal/mol — enough to flip a promising hit into a reject. Beyond that, keep these five failure modes in mind:
| Symptom | Cause | Fix |
|---|---|---|
| ESMFold produces a garbage structure | Sequence has low-confidence regions | Check the pLDDT scores; if < 50, use AlphaFold DB |
| Vina gives very weak binding (> −5) | Ligand too large for the pocket / grid box missplaced | Try a smaller fragment; check the grid box coverage |
| GNINA score disagrees with Vina | Borderline pose | Trust GNINA for pose quality, Vina for affinity |
| Different PDB structures give different results | Conformational flexibility | Try 2–3 PDB structures and compare |
| Results look too good (< −12 kcal/mol) | Possible artifact (very small / highly charged ligand) | Check for close contacts; compare to known binders |
Finally, treat cross-tool agreement as evidence, not proof. A consensus call tells you the engines are internally consistent with their own assumptions; it cannot tell you whether those assumptions match your real biological context. The strongest validation trips always end in the wet lab — a predicted binder, or a predicted pose, remains a hypothesis until an experiment confirms it. What the consensus gives you is the confidence to prioritize which compounds and which poses are worth testing at all.
Understanding Molecular Docking: From Protein to Binding Pose
Molecular docking predicts how a small molecule (ligand) binds to a target protein (receptor) to form a stable complex. The process involves two key components: pose sampling \u2014 exploring different orientations and conformations of the ligand within the protein binding site \u2014 and scoring \u2014 evaluating each pose using energy-based functions. Docking is fundamental to structure-based drug discovery, enabling virtual screening of millions of compounds to identify potential drug candidates before experimental testing. The quality of docking predictions depends on the accuracy of the protein structure, the flexibility of both protein and ligand, and the scoring function\'s ability to approximate true binding free energy.
Keep a firm handle on what a docking score is and is not. It is a ranking aid produced by a scoring function that approximates binding free energy — useful for ordering compounds against a target and for triage before experimental validation, but never a substitute for a measured affinity. Two molecules can dock at −9 and −10 kcal/mol and be indistinguishable in the lab, because the scoring function smooths over solvent, entropy, and protein flexibility. That realism is why the worked example below walks through reading a result honestly rather than over-trusting a single number.
Docking Scoring Functions: What the Numbers Mean
| Method | Score Type | Typical Range | Interpretation |
|---|---|---|---|
| Vina affinity | Empirical free energy | -4 to -14 kcal/mol | More negative = stronger predicted binding |
| GNINA CNN score | Deep learning confidence | 0 to 1 | Higher = more likely to be correct pose |
| SMINA custom | Weighted empirical | -6 to -16 kcal/mol | Adjacent to Vina but with additional terms |
| Consensus | Average of all methods | Combined | Reduces false positives from individual methods |
A Vina score of -7 kcal/mol or lower (more negative) generally indicates good binding, while scores above -5 kcal/mol suggest weak or non-specific binding. However, scoring functions are approximations \u2014 always validate top predictions experimentally using techniques like surface plasmon resonance (SPR) or isothermal titration calorimetry (ITC).
Preparing Structures for Docking
Well-prepared input structures are essential for reliable docking: (1) for proteins, ensure hydrogen atoms are added at the correct protonation state for pH 7.4, (2) remove water molecules and cofactors unless they are critical for binding, (3) assign correct bond orders and formal charges for ligand atoms, (4) generate all possible tautomers and protonation states for the ligand at physiological pH, and (5) define the binding site using known active site residues or co-crystallized ligand coordinates. The docking pipeline handles protein preparationhandles protein preparation automatically using ESMFold for structure prediction and AutoDockTools for PDBQT conversion.
What's the difference between this and running AutoDock Vina directly?
Running Vina directly gives you a pose and an affinity estimate from one physics-based engine. This pipeline adds two steps: ESMFold predicts a structure when you only have a sequence, and GNINA re-scores the top Vina poses with a deep-learning network. The extra scoring layer is what turns a single engine's guess into a more confident call.
Why use three methods instead of one?
Consensus. Any single docking algorithm can miss the correct binding pose, so the pipeline runs ESMFold, Vina, and GNINA together. When all three agree on a pose and score, the answer is far more trustworthy than any one engine alone; when they disagree, the GNINA confidence score is usually the better predictor.
What PDB format do you accept?
For proteins, standard PDB files and raw amino-acid sequences both work. For ligands, the tool accepts SDF or MOL2 format. If you only have a sequence, ESMFold builds the structure for you in stage one.
How accurate is ESMFold compared to AlphaFold?
ESMFold is faster than AlphaFold but generally less accurate on long, multi-domain proteins. For well-studied proteins, starting from an experimental PDB structure is the best choice; use ESMFold when you have only a sequence and speed matters.
Can I dock protein-protein interactions?
Not yet. This tool is protein-ligand docking only. If you need to study how two proteins interact, that is a separate workflow — see the protein-protein docking guidance for what to use instead.
How do I validate a docking result?
Compare your top pose to a co-crystallized structure (RMSD under 2.0 Å is a good sign), check that the affinity and GNINA score agree, and sanity-check against known binders for your target. Consensus across engines plus agreement with experimental structures is the strongest evidence you will get from computation alone.
Benchmarked Against Industry Standards
The consensus pipeline combines ESMFold (MIT), AutoDock Vina (Apache 2.0), and GNINA for drug-discovery-grade virtual screening.
ESMFold pLDDT Validation
pLDDT scores correlate with CASP14/15 competition results. Mean confidence ≥ 70% is the recommended threshold for reliable docking.
Vina Cross-Validation
AutoDock Vina scoring reproduced within 1.5 kcal/mol RMSE on PDBbind core set. Exhaustiveness=4 balances speed and accuracy.
GNINA CNN Refinement
CNN re-scoring improves hit rates by 20-35% over Vina alone on the DUD-E benchmark. Default CNN model trained on PDBbind v2020.
One Plan for Primer Design + Docking
All subscription plans include both primer design and molecular docking credits. From ₹99 (~$1.20 USD) (~$1.19 USD)/day to ₹49,999 (~$600 USD) (~$602 USD)/month.