User Manual

1. Overview

The L-GNM Server (Learned Graph Network Model) predicts per-residue protein flexibility from a single static structure. It combines:

• Elastic Network Models (ENM) — coarse-grained normal mode analysis using Cα atoms connected by harmonic springs
• Graph Neural Networks (GNN) — learned residue-specific spring constants instead of uniform coupling
• Quantum-inspired Neural PDE — efficient Hamiltonian dynamics with O(K) operator cost

2. Input Options

Fetch from RCSB

Enter a 4-character PDB ID (e.g., 6LU7, 1UBQ). The server downloads the structure from the RCSB Protein Data Bank.

Upload PDB File

Upload a .pdb or .cif file (max 50 MB). Drag-and-drop is supported. This includes AlphaFold predicted structures — pLDDT scores are auto-detected from the B-factor column.

Chain Selection

After loading, all available chains are listed. Select the chain you want to analyse. Only the selected chain is used for prediction — other chains are not included in the flexibility calculation.

3. Output Metrics

4. 3D Viewer

The interactive 3D viewer uses Mol* (Sehnal et al., Nucleic Acids Research, 2021). The structure is coloured by predicted flexibility using the B-factor column.

Mouse Controls

Toolbar

• Representations: Cartoon, Surface, Ball & Stick, Spacefill
• Reset: Return to default camera
• Auto Rotate: Continuous slow rotation for presentations
• Background: Toggle white/black
• Screenshot: Export high-resolution PNG

Sidebar

Use the collapsible sidebar (toggle arrow on right edge of viewer) to change colouring modes, palettes, highlight flexible/rigid residues, and download results. The sidebar can be re-opened anytime using the arrow tab.

5. Flexibility Profile Plot

This is the primary output — a line chart of predicted RMSF (Å) vs residue index.

How to read it

• Peaks correspond to flexible regions: loops, termini, disordered segments
• Valleys correspond to rigid regions: buried core, regular secondary structures
• Shaded red regions mark the top 20% most flexible residues (80th percentile threshold)
• Dots are coloured by normalised flexibility (blue = rigid, red = flexible)

What it means biologically

High RMSF residues are often at binding interfaces, allosteric sites, or regions that undergo conformational change. These regions may be drug targets or functionally important for protein-protein interactions.

6. Flexibility Map Plot

A bar chart showing normalised flexibility (0–1) per residue, coloured by the selected palette.

How to read it

• The sequence is annotated below the x-axis (every Nth residue)
• Red amino acid labels = flexible residues (Gly, Ser, etc.)
• Blue labels = rigid residues (Phe, Ile, Trp, etc.)
• Arrows mark the top 5 most flexible residues with their identity

7. Stiffness Heatmap

A 1D heatmap strip showing per-residue stiffness — the sum of spring constants connecting each residue to its neighbours within the cutoff distance.

How to read it

• Blue = low stiffness (soft, few/weak contacts) — typically surface residues, loops
• Red = high stiffness (rigid, many/strong contacts) — typically buried core residues
• Stiffness is inversely related to flexibility: stiff residues tend to have low RMSF

What it means

In the mENM model, spring constants are proportional to 1/r², so closely packed residues have stronger springs. In L-GNM, the GNN learns residue-specific spring constants from sequence and structural features.

8. Conformer Spread

Shows the average displacement of ENM-sampled conformers from the input structure at each residue position.

How conformers are generated

1. The Kirchhoff matrix Γ is computed from Cα contacts within cutoff distance d_c
2. Eigendecomposition gives normal modes (eigenvectors) and eigenvalues
3. Conformers are generated by displacing coordinates along the top-K slowest modes with random amplitudes, scaled by α (deformation scale)
4. The conformer spread at each residue is the RMS deviation across all sampled states

How to read it

• High spread = the residue samples many different positions → intrinsically flexible
• Low spread = the residue stays near its mean position → structurally constrained
• Pattern should roughly correlate with the RMSF profile

9. Normal Modes

Shows the top 5 slowest vibrational normal modes

What are normal modes?

Normal modes are the eigenvectors of the Kirchhoff matrix Γ. They describe collective motions — correlated movements of groups of residues. The slowest modes (lowest non-zero eigenvalues) capture the largest-amplitude functional motions.

How to read the plot

• Each line is one mode (Mode 1 = slowest, Mode 5 = 5th slowest)
• The y-axis shows the mode amplitude at each residue
• Sign changes (zero crossings) divide the protein into dynamical domains — residues on the same side of zero move together
• Mode 1 typically captures the dominant global motion (e.g., hinge bending, domain opening)
• Residues near zero crossings are hinge points — they don't move much but mediate the collective motion

Interpretation examples

• Single zero crossing in Mode 1 → the protein has two lobes that move relative to each other (hinge motion)
• Multiple zero crossings → more complex multi-domain motions
• Large amplitude at termini → floppy ends (expected for most proteins)
• Large amplitude in a loop region → functionally important conformational flexibility

10. Interpretation Guide

High flexibility (RMSF > mean + 1σ)

• Loops, turns, and termini
• Potential binding sites or allosteric regions
• Regions that undergo conformational change upon ligand binding
• Glycine (G) and serine (S) rich regions tend to be flexible
• Does not mean the structure is poorly determined — it reflects genuine dynamics

Low flexibility (RMSF < mean − 1σ)

• Protein core, hydrophobic packing
• Regular secondary structure (α-helices, β-sheets)
• Structurally critical regions (e.g., active site scaffold)
• Aromatic residues (W, F, Y) and branched aliphatics (I, V, L) tend to be rigid

Comparing methods

Validation

L-GNM predictions correlate with:

• Crystallographic B-factors (Pearson r ≈ 0.6–0.8)
• NMR order parameters S² (inverse correlation)
• MD simulation RMSF (Pearson r ≈ 0.7–0.9)
• ATLAS database experimental dynamics

11. FAQ

12. Citation

3D visualization: Sehnal D, Bittrich S, Deshpande M, et al. Mol* Viewer: modern web app for 3D visualization and analysis of large biomolecular structures. Nucleic Acids Research. 2021;49(W1):W431–W437.