StringScape Protein Modeller - Comparative Protein & Distance Matrix Visualizer

StringScape Protein Modeller

This app builds all-atom homology models of protein complexes (including DNA, RNA, small molecules, ions, and lipid membranes) by aligning AlphaFold models onto reference structures from the Protein Data Bank and computing membrane positions.

Tips

Using search terms can identify the wrong files from the Protein Data Bank if no AI is connected. I recommend using the specific PDB IDs for the structure you want modelled.
Click a residue, alignment letter, or matrix cell to highlight the equavalent ones in the others (these are synced). Shift+click in the 3D viewer to select more than one residue at once.
Open the Validation report to inspect where restraints are well supported and where the model does not yet satisfy them.

Examples:

Recent:

Select Species

Connect an AI

This app can work without AI, but the models tend to be better with AI. AI is used to help find the best reference structures to use in the Protein Data Bank and to summarise the available information on the protein.

Save and load sessions

If you connect a local folder, this web app can save sessions for you to reopen later

What do you want a model of?

Search terms, UniProt IDs, PDB IDs, or sequences

Searches made

Search	Returned PDB IDs	Status

Search results

PDB ID	Score	Species	Reference chain / protein	UniProt IDs	ID source	Match details	Action:

Chosen references

PDB ID	Score	Species	Reference chain / protein	UniProt IDs	ID source	Match details	Action	Group order	Conformation group	References in group

More options

Manually add Reference PDB IDs

Upload Structures

Manually set structures to align to references (AlphaFold IDs, UniProt IDs, or PDB IDs)

Upload structures

Sequence alignment

Choose the alignment method, match threshold and gap penalties.

Alignment methodMatch threshold (% identity)Gap open costGap extend cost

Reference structures

Use references directly

If this toggle is turned on, the app will use the reference structures directly, rather than aligning AlphaFold models onto the references.

Reference locking

Drafts lock aligned residues according to their source; all residues in the combined model are unlocked.

Backbone distance from unaligned residuesAligned residues in AlphaFold models sourced from the reference PDB will all be locked to the reference coordinates. Aligned residues in AlphaFold models not sourced from the reference PDB (manually supplied or from a STRING species different to the reference PDB species) that are farther than this many residues (the value above) from an unaligned region stay locked to the reference during draft builds.

Spring culling

Cull extra springs separately for each source. Backbone and base-pair springs are retained.

Remove % of the springs from the AlphaFold models with a residue at either end with pLDDT below 50Remove % of the springs from the AlphaFold models with residues at both ends with pLDDT of 50 or aboveRemove % of the intra-chain springs from the Reference modelsRemove % of the inter-chain/inter-molecule springs from the Reference models

Structure influence

Balance the reference and AlphaFold contributions to spring lengths and strengths.

Reference PDB100%

AlphaFold100%

Confidence & geometry

Tune confidence weighting and preserve local backbone shape.

PAE effect100%

pLDDT effect50%

Angular / dihedral restraints25%

Preserves local Cα bends and twists from the starting structures. AlphaFold restraints are weighted by pLDDT; zero turns these forces off.

Steric clash exclusion radius3.8 Å

Residue centres within this distance repel each other. Collision prevention is always enabled.

PAE and pLDDT can weaken AlphaFold spring targets and strengths. Zero ignores that confidence measure; higher values increase its effect. Unavailable scores leave the corresponding factor unchanged.

Elastic network

Set the spring behaviour and simulation step size.

Spring stiffness (k)2.0

Mobility / step size (μ)0.005

Max displacement per step (Å)Limits force-driven movement before clash and backbone projection.Spring cutoff threshold

Group simulation

The simulation relaxes each subunit together with its nearby neighbours in two passes. In pass 1, residues move flexibly. Steric forces are calculated at the selected step interval. Clash projection runs every step, recalculating distances and corrections from the current positions. Backbone connections remain constrained. In pass 2, residues move flexibly and steric clashes are checked and corrected every step. Locked reference residues remain fixed in both passes. The neighbour distance determines which nearby subunits join each core subunit’s group. The simultaneous residue budget limits how many unlocked residues can be active across groups at once; a single group may exceed the budget. A group stops when its core’s maximum residue speed falls below the stopping speed and it has no unresolved clashes, or when it reaches the maximum waiting time. Each group can run for up to half of the maximum waiting time in pass 1 and up to half in pass 2.

Pass 1 steric interval (steps)Pass 1 clash projection max repeats per simulation stepNeighbour distance (Å)Simultaneous residue budgetStopping speed threshold (Å/s)Maximum waiting time (s)

How group simulation works

Each subunit serves once as a core. Its group includes subunits with any residue within the neighbour distance when that group starts. Each model runs two passes, each using half the maximum waiting time per group. Pass 1 uses flexible residues. Steric forces are recalculated at the selected interval. Clash projection recalculates distances and corrections every step; the repeat limit controls how many projection iterations run per step. Backbone connections remain constrained. These settings apply at the next build. Pass 2 uses flexible residues and checks sterics every step. Running time excludes pauses. Colours reset for the second pass. Overlapping groups share moving subunits; each core stops when its maximum residue speed is below the stopping speed and its group has no unresolved clashes, or when its time limit is reached. The residue budget applies when the next build starts. Additional groups start only when their combined active, unlocked residue count fits the budget; overlapping residues count once. A single group may exceed it. Dynamic starts at 10,000 and measures simulation-worker time per step every two seconds, increasing the budget for fast steps and reducing it for slow steps; newer groups may be deferred when the budget falls. Fixed reference residues remain locked.

Automatically run simulation

When off, all draft models are prepared, aligned and combined without simulation. Start simulation from the 3D viewer or Build stats to improve each prepared draft in order, then realign, merge and simulate the combined model. Animation settings and timing are preserved with updated coordinates.

This app builds homology models of protein complexes by aligning AlphaFold models onto one or more reference structures from the Protein Data Bank. This involves multiple steps:

Load one or more reference structures from the Protein Data Bank and load the AlphaFold structures from the AlphaFold database. If no AlphaFold model is available, the app just uses the protein sequence without any structural information.
- Optional: If STRING files have been uploaded, the app uses reference IDs, gene names and protein names to find target candidates in the aliases file, verifies them by sequence alignment, and falls back to a top-500 k-mer sequence search when needed. Repeated chains and conformational references reuse the same protein-identity assignment. UniProt accessions must match the target sequence and taxonomy; unresolved or ambiguous matches pause the build for review.
Align each of the chain/polynucleotide sequences from the reference with each of the AlphaFold sequences. Semi-global alignment fits each reference chain/polynucleotide within an AlphaFold sequence, allowing for extra AlphaFold residues at either end. Alignment Settings also offers local Smith–Waterman and global Needleman–Wunsch as alignment methods.
Calculate distance matrices. The distance matrix captures the spatial relationships between residues/nucleotides/atoms within each chain/polynucleotide/molecule and, for the reference structure, between each chain/polynucleotide/molecule.
Build the morph distance matrix. The distance matrices of the aligned reference structures and the AlphaFold models are combined to create the morph distance matrix. The per-reference PDB and AlphaFold sliders determine the blend of shared intra-chain distances and each values influence on the simulation.
Align AlphaFold models onto the reference structure. The AF PDB TM alignment aligns AlphaFold structures to the reference using rigid-body Template Modelling alignment. The simulation then relaxes the springs between residues, nucleotides and atoms using an overdamped elastic model. AlphaFold spring contributions are weakened where PAE indicates uncertain relative placement or pLDDT indicates uncertain local structure; local Cα angular and pseudo-dihedral restraints help preserve backbone bends and twists. Build Settings controls these effects. Each reference is used independently to build a draft model, keeping aligned interior residues locked while successive groups of nearby subunits relax.

Adding membranes: The app estimates whether a protein complex resides in a lipid membrane and displays its predicted membrane position, curvature, and placement confidence.

Combine drafts when multiple references are loaded. Completed drafts are structurally TM-aligned and equivalent subunits are matched. By default, all references belong to one group: their sequences are combined, residue centres and spring values are averaged using the reference influence sliders, and unmatched subunits are retained. The combined model then undergoes group simulations with active subunits free of reference coordinate locks. When a connected AI identifies distinct conformational states, it names and orders separate reference groups. Only drafts within the same group are merged. The final combined model animates between these group models, morphing equivalent subunits and fading subunits that are absent from a state. The Conformation group selector beside Build process models displays an individual state. The generated animation has 100 frames between each group keyframe, with all variables synced except conformation.
Convert the simulation result into a mmCIF format and display in a Mol* viewer. The mmCIF file can then be downloaded.

Reference PDB: 6YKM

0 Residues

AlphaFold: P09348 + P0AF06

Colour

0 Residues

Sequence Alignment & Equivalent Chains

Calculating matches...

Reference PDB Chain	Length (PDB)	Equivalent AlphaFold Chain	Prefered name	Length (AF)	Sequence Identity	Sequence Similarity	TM-score	RMSD	Average TM-aligned Cα pair distance	Status	Action

Matrix palette:

Reference PDB Distance Matrix

AlphaFold Distance Matrix

Morph Distance Matrix

This combines the Reference PDB distance matrix and the AlphaFold distance matrix into a combined matrix used to set the spring lengths in the physics simulation.

Intra-chain: D_morph = (Σ w_ref,i · D_ref,i + w_AF · D_AF) / (Σ w_ref,i + w_AF) for available distances. If only one origin has a distance, its value is used. Inter-chain distances can come from each reference PDB and from a multi-chain AlphaFold structure.

Membrane size 1.00×

Residue size 1.0× Backbone size 2.0× Opacity of non-selected residues 100%

Extension of subunit patterns

Extension of DNA/RNA

Style:

Colour proteins by:

Select:

Waiting for AF PDB TM alignment.

Pulsing residues are not in the Reference PDB(s)

Chain: chains with the same protein ID share a colour family, with a different shade for each chain.

Ctrl + Drag: Move Chain Shift + Drag: Move Residue Alt + Drag: Rotate Symmetry Group

About this protein complex

AI overview

Connect an AI to generate an overview of this complex.

UniProt info

UniProt information will appear when available.

Performance diagnostics

Start before reproducing a stall. The report stays on your computer.

Report data

AF PDB TM alignment aligns each AlphaFold chain to the reference PDB. The simulation relaxes each subunit and its current neighbours in groups, keeping other subunits fixed.

Build timeline

Total elapsed —

Stage timings appear when a search or build starts.

Matrix times count numerical calculation only, excluding rendering and waiting. Other stages show elapsed wall time; total includes time between stages.

Build overview

Alignment quality and live simulation activity

Active sim. residues: 0
TM-score · AF PDB TM alignment: --
RMSD · AF PDB TM alignment: --
Weighted spring-length error: 0.00 a.u.
Average residue speed: -- Å/s
Max residue speed: — Å/s
Simulation step rate: 0.0 steps/s
Average simulation step rate: — steps/s
Dynamic max sim residue count: 10,000
Groups reaching the waiting limit: 0
Average max core residue speed at waiting limit: — Å/s
Active springs: 0
Nodes: 0
Estimated page memory: Unavailable
JavaScript heap in use: Unavailable
Estimated heap headroom: Unavailable

Simulation passes

Model	Pass 1 steps	Pass 1 time	Pass 2 steps	Pass 2 time	Final backbone deviations (excluding locked pairs)	Final clashes
Simulation statistics appear when a build starts.

Time is active simulation running time; pauses and model preparation are excluded. Concurrent groups share each simulation step. Final backbone deviations count links at least 0.005 Å from their target, excluding links between two locked residues. Final clashes include locked pairs and use the residue clash radius.

Simulation step details

Concurrent cores share a worker step. Mobility is the effective value after the displacement limit; spring error is measured before integration. Time is worker group-step elapsed time.

Step	Draft model	Group (core)	Pass	Simulation step time (ms)	Clash and backbone projection repeats	Residue neighbour list rebuilt	Adaptive mobility	Spring error
Simulation steps appear when a build starts.

Spring-culling

Extra spring contributions removed before the sources are combined

Culling results appear when the spring network is built.

Fixing steric clashes and backbone issues

Pre-simulation projection, shown separately for each model

Projection statistics appear when a build starts.

Simulation kymograph plot

Plot

Displayed time rangeFull range

The graph starts when the simulation begins.

TM-score (Simulation) is recalculated for each graph sample from equivalent Cα positions in the current simulated model and the reference, in the fixed build coordinate frame. Higher (up to 1) means closer agreement. The TM-score for AF PDB TM alignment above is the fixed result from the initial chain alignments.

RMSD (Simulation) is the root-mean-square distance between equivalent Cα atoms in the current simulated model and the reference, in ångströms (Å). Lower means closer. The RMSD for AF PDB TM alignment above is fixed at the initial alignment result.

Weighted spring-length error is the square root of the mean stiffness-weighted spring potential, in arbitrary units (a.u.). Lower means the active spring network is more relaxed. It is neither an RMSD nor a physical energy.

Average residue speed is the mean distance each simulated residue moves per second between samples, in Å/s. Lower means the structure is changing more slowly.

Dynamic max sim residue count is the adaptive budget for concurrent active residues. It adjusts according to simulation-worker step time. Available in dynamic mode.

Simulation step rate is the number of completed simulation updates per second (steps/s), measured over a rolling two-second window. Higher means faster simulation throughput.

Adaptive mobility is the effective mobility used in the latest sampled simulation step, after the displacement limit.

Max residue speed of the core subunit(s) is the greatest residue speed across all currently active core subunits, measured over completed simulation steps in a short running-time window. Each core has its own stopping criteria.

Restraint satisfaction shows the mean agreement between active spring lengths and their target distances. A perfect match is 1.0; the score decreases as springs stretch or compress.

Overall model confidence combines distance-evidence support and spring satisfaction from the available restraints. Higher is better; it is a heuristic, not a calibrated probability.

Membrane placement confidence reports evidence supporting the inferred hydrophobic belt. It is unavailable when no membrane is detected.

Memory lines are browser estimates in MiB. Page memory includes workers where the browser supports measurement. Heap use and heap headroom can exclude workers, GPU and native memory; heap headroom is not available system RAM. Unsupported readings are omitted.

If normalised, each line is divided by its own maximum in the displayed time window, so 1.0 is that line’s current window maximum. The three validation scores are sampled intermittently and hold their last values between updates. Vertical lines mark group starts and changes to build controls.

Chain-merging

Reference fragments positioned along each full AlphaFold sequence

Reference structures may split one protein into separate chains or fragments. The app joins compatible fragments that align to the same AlphaFold protein so they can be modelled as one chain, while preserving the reference arrangement.

Build a model to see merged chains.

Click a row to select its chain in the 3D viewer; Shift-click to select multiple. Fragments on the same coloured row share one AlphaFold copy. Dashed lines connect merged fragments; grey fragments are not merged. Missing sections within a chain are shown at half opacity. Positions use AlphaFold sequence numbering.

Build a model to see reference fragments aligned to each AlphaFold sequence.

Compare two variables

Build a model to compare its variables.

X variable:

Y variable:

Hover over the plot to inspect values.

Mol* Viewer Model not built yet

Homology model Mol* Viewer

When the app has finished building the homology model, it will be displayed here.

Download Homology Model

Residue origin (Reference PDB = 1.00, AlphaFold = 2.00, Both = interpolated) is stored in _atom_site.b_iso_or_equiv.

Validation report

Sequence coverage—AlphaFold residues mapped

Restraint satisfaction—Mean spring agreement

Membrane placement confidence—When a membrane is detected

Overall confidence—Combined local score

Sequence coverage is the percentage of protein residues in the AlphaFold chains that align to at least one reference PDB residue. Each AlphaFold residue is counted once, even if it matches several references. Higher coverage means more of the model has a reference sequence match; it does not measure structural accuracy.

Equation

Coverage = 100 × (unique mapped AlphaFold protein residues) / (all AlphaFold protein residues)

Restraint satisfaction is the average agreement between each active protein-distance spring’s current length and its target length. Exact agreement scores 100%; the score falls exponentially as the difference grows, relative to the target distance. It describes fit to the active restraints, not the reliability of those restraints.

Equation

Pair score = exp(−|model distance − target distance| / max(2 Å, 0.2 × target distance))
Restraint satisfaction = 100 × mean(pair scores)

Membrane placement confidence measures support for the inferred membrane using hydrophobic density contrast, hydrophobic density inside the belt, effective hydrophobic support, and the detector score. It does not apply forces to the model. If no membrane is detected, it is unavailable.

Overall confidence averages the active residue-pair scores. Each pair combines support from available reference structures, AlphaFold pLDDT and PAE, agreement with its spring target. This is a heuristic consistency score, not a calibrated probability that the model is correct.

Equation

Pair confidence = evidence support × restraint satisfaction
Overall confidence = 100 × mean(pair confidences)

Sequence alignment by AlphaFold chain

Coverage and mean identity/similarity across matched reference chains

AlphaFold chain	Reference matches	Coverage	Average identity	Average similarity

Constraint confidence

How strongly the evidence supports each distance

LowHigh

Distribution

Spring compression / stretch

Current model distance relative to each spring target

CompressedTargetStretched

Distribution

Combined pair confidence

Evidence support × restraint satisfaction

LowHigh

Distribution

Grey cells have no active distance restraint or exceed the spring cutoff. Scroll to zoom a matrix, drag to pan, and click a cell for pair details. Coloured edge bars identify chains; dashed lines separate them. Evidence support combines available reference agreement (35%), AlphaFold pLDDT (15%), and pairwise PAE (50%), renormalising when a source is unavailable. Restraint satisfaction decreases with absolute distance error. Membrane placement confidence reports membrane-detection evidence independently of the spring simulation. These are heuristics, not calibrated probabilities.

Independent model-quality assessment

QMEANDisCo from SWISS-MODEL

QMEANDisCo is the default validation method used by the SWISS-MODEL homology modeller. It is a "single model method combining statistical potentials and agreement terms with a distance constraints (DisCo) score. DisCo evaluates consistencies of pairwise CA-CA distances from a model with constraints extracted from homologous structures. All scores are combined using a neural network trained to predict per-residue lDDT scores." (SWISS-MODEL website)

QMEANDisCo requires an email address for the SWISS-MODEL server to send you the result. Enter an email and then press Send to submit the current model to SWISS-MODEL. Opening this report does not submit it.

Not submitted.

QMEANDisCo GlobalUnavailable

Open QMEAN website

Select Species

Connect an AI

Save and load sessions

What do you want a model of?

Searches made

Search results

Chosen references

Add extra evidence

Build settings

Sequence alignment

Reference structures

Reference locking

Spring culling

Structure influence

Confidence & geometry

Elastic network

Group simulation

View homology model build process

Membrane statistics

Extension of subunit patterns

Extension of DNA/RNA

About this protein complex

Build timeline

Build overview

Simulation passes

Spring-culling

Fixing steric clashes and backbone issues

Chain-merging

Mol* Viewer Model not built yet

Download Homology Model

Sequence alignment by AlphaFold chain

Manually added evidence

Independent model-quality assessment