TEMPL Pipeline

What is TEMPL Pipeline?

TEMPL Pipeline predicts protein-ligand binding poses by borrowing geometry from related protein-ligand complexes instead of starting from an unconstrained docking search. It is a template-based method from Jozef Fulop, Martin Sicho, and Wim Dehaen for rapid pose prediction when similar ligands and binding sites are already represented in structural data.

The method is strongest in interpolative settings: close analogs, conserved pockets, lead-optimization series, and targets with useful protein-ligand templates. It is less suitable for novel scaffolds, allosteric pockets with little template coverage, and cases where binding depends on large receptor rearrangements.

How to use TEMPL Pipeline online

A TEMPL Pipeline online run accepts a protein structure as PDB, ENT, or RCSB PDB ID and one organic small-molecule ligand as SMILES, PubChem, drawing, or SDF. ProteinIQ searches for related protein-ligand templates, generates constrained ligand poses, and returns ranked SDF files with shape, color, and combo Tanimoto scores.

Inputs

Input	Accepted values	Notes
Protein	.pdb or .ent file, Files library structure, or RCSB PDB ID	Protein atoms are required. Structures with nucleic acid atoms may need review before interpretation.
Ligand	SMILES text, PubChem lookup, structure drawing, or .sdf file	One covalent organic small molecule is expected. Disconnected salts, query atoms, metals, large peptides, and very complex carbohydrates are rejected or fail validation.
Job name	Optional text label	Stored with the run for later job history lookup.

For structures with missing atoms, alternate locations, or unusual residues, PDB Fixer can be used before running pose prediction.

Settings

Setting	Default	Description
Number of templates	100	Number of candidate protein-ligand templates to consider. Larger values broaden the search but can introduce less relevant templates.
Similarity threshold	Blank	Optional cutoff from 0.0 to 1.0. Leaving it blank keeps top-template selection by rank.
Number of conformers	200	Number of ligand conformers generated before alignment and scoring. More conformers can help flexible ligands but increases runtime.
Ranking metric	Combo Tanimoto	Metric used to select the ranked conformer. Combo Tanimoto averages shape and pharmacophore agreement; Shape Tanimoto and Color Tanimoto can be used when one signal should dominate.
Skip shape realignment	Off	Uses the generated conformer without the final shape realignment step. This is mainly useful for diagnostic comparisons.
Enable force-field optimization	Off	Applies force-field optimization after pose generation. It may improve local geometry but can also move a pose away from its template constraint.
Unconstrained conformer generation	Off	Disables MCS-constrained conformer generation. Use only when the shared scaffold constraint is undesirable.
Random seed	Blank	Optional non-negative integer for reproducible conformer generation and scoring behavior.

Results

Output	Description
Viewer	3D view of the submitted protein with ligand poses overlaid for inspection.
Data	Pose summary table with rank, metric, shape score, color score, combo score, template PDB ID, and file name.
Files	Downloadable SDF, JSON, and reference structure files.

The main result files are:

File	Contents
*_top3_poses.sdf	One top ligand pose for each scoring metric: shape, color, and combo.
*_all_poses.sdf	All ranked ligand poses generated during the run.
*_template.sdf	The selected template ligand used for MCS-guided pose generation.
*_pipeline_results.json	Template, MCS, target, and run metadata.
*_reference.pdb	Submitted protein structure included so the viewer can place ligand-only SDF poses in the target context.

File names may use the selected template PDB ID rather than the submitted target ID. That is expected for TEMPL outputs because the winning template determines the pose file prefix.

How TEMPL Pipeline works

TEMPL is a data-driven baseline for protein-ligand pose prediction. It searches a library of known protein-ligand complexes, identifies suitable templates, transfers ligand geometry into the target frame, and ranks generated poses by 3D similarity.

Template search

The pipeline uses protein embeddings to find structurally related protein-ligand templates. The released TEMPL dataset includes precomputed ESM-2 protein embeddings and processed ligand structures from PDBbind-scale complexes, which makes the search fast enough for interactive use.

Template quality matters more than template count. A close template from the same protein family with a related ligand can be more useful than many weakly related hits.

MCS-guided pose generation

After a template ligand is selected, TEMPL finds the maximal common substructure (MCS) between the submitted ligand and the template ligand. The shared atoms act as geometric anchors, and RDKit ETKDG conformer generation builds candidate 3D ligand conformations around that constraint.

This is why TEMPL is most useful for analog series. If the submitted ligand and template ligand share a clear scaffold, the MCS constraint gives the generated pose a strong geometric prior. If the ligand is a new chemotype with little shared substructure, the constraint becomes weaker or may fall back to less informative alignment.

Shape and pharmacophore scoring

Generated conformers are aligned to the template ligand and scored with RDKit shape alignment:

• Shape score: Agreement between the 3D molecular volumes.
• Color score: Pharmacophore-like feature agreement, such as donor, acceptor, aromatic, and hydrophobic feature overlap.
• Combo score: Average of the shape and color scores.

The top poses are not binding free-energy estimates. They are similarity-ranked hypotheses for where the ligand could sit if the selected template captures the relevant binding mode.

Interpreting results

Higher Tanimoto scores indicate closer agreement with the selected template. Scores near 1.0 suggest strong shape or feature overlap, while values closer to 0 indicate weak agreement. The useful threshold depends on ligand size, scaffold similarity, and whether the template ligand binds the same pocket.

Result field	Interpretation
Rank	Order in the result table for the reported top poses.
Metric	Scoring criterion used for that top pose: shape, color, or combo.
Shape score	3D volume overlap with the selected template ligand. Higher is better.
Color score	Pharmacophore feature overlap with the selected template ligand. Higher is better.
Combo score	Mean of shape and color scores. Useful as the default balanced ranking.
Template PDB	PDB ID of the protein-ligand complex that supplied the template geometry.
File	SDF file containing the corresponding ligand pose.

Visual inspection is essential. A high-scoring TEMPL pose should still be checked for steric clashes, impossible protonation assumptions, and whether key interactions are plausible in the submitted protein structure. The *_pipeline_results.json file is useful when documenting which template and MCS mapping drove a result.

When to use TEMPL Pipeline vs docking tools

TEMPL is a good first choice when the ligand resembles known binders and the protein has nearby structural templates. It can be faster and more stable than unconstrained docking for analog pose transfer because it starts from experimental geometry.

Method	Best fit	Main output
TEMPL Pipeline	Template-rich analog series and fast pose transfer	Similarity-ranked ligand pose SDFs
AutoDock Vina	General small-molecule docking with an interpretable empirical score	Binding poses with kcal/mol docking scores
GNINA	Docking plus CNN pose confidence and affinity-style rescoring	CNN-ranked poses and Vina-style scores
DiffDock-L	Blind docking or cases without a reliable known pocket template	Diffusion-generated poses with confidence scores
PandaDock	Physics-based protein-ligand docking with affinity-style ranking	Docked poses and binding-energy-oriented scores

For a lead series around a known chemotype, TEMPL can quickly test whether new analogs preserve a crystallographic binding mode. For chemically distant ligands or targets without useful templates, AutoDock Vina, GNINA, or DiffDock-L usually provide a better starting point.