Molecular Modelling, Dynamics and Protein Structure Prediction
Force fields, molecular mechanics, minimisation, MD ensembles, SCOP and CATH, homology modelling, threading, ab initio and AlphaFold (pLDDT, PAE) for CSIR NET.
- 2 min
On this page · 13 sections
- Prerequisites
- The Hook: A Fifty-Year Problem Solved in a Competition
- Part 1: Molecular Mechanics and Force Fields
- Part 2: Molecular Dynamics (MD)
- Part 3: Protein Structure Classification
- Part 4: Protein Structure Prediction
- Memory Tricks
- Don’t Confuse
- Common Mistakes
- Practice
- Quick Revision Table
- Key Takeaways
- What’s Next
Prerequisites
- Protein Structure, Conformation and Stability — Ramachandran plot, secondary structure, motifs, domains, folds and Anfinsen’s experiment.
- Chemical Bonds, Stabilising Interactions and Biophysical Chemistry — van der Waals, electrostatic and hydrophobic interactions that force fields try to reproduce.
- Sequence Alignment, BLAST, FASTA and Multiple Alignment — template search and alignment are the first steps of homology modelling.
The Hook: A Fifty-Year Problem Solved in a Competition
How CSIR NET asks it. New unit, no past CSIR papers. Expect Part B on force-field terms, minimisation vs dynamics, ensembles, SCOP/CATH levels, which prediction method suits which situation, and what pLDDT means. Expect Part C to ask for Lennard-Jones minima, electrostatic energies in different dielectrics, RMSD calculations, MD step counts, or interpretation of a model’s confidence plots.
Part 1: Molecular Mechanics and Force Fields
Molecular mechanics (MM) treats atoms as balls and bonds as springs. Electrons are not modelled explicitly (the Born–Oppenheimer separation lets us treat energy as a function of nuclear positions only). It is fast enough for proteins of thousands of atoms, but cannot break or form bonds: reactions need quantum mechanics or hybrid QM/MM methods. The 2013 Nobel Prize in Chemistry (Karplus, Levitt and Warshel) recognised multiscale models of this kind.
A force field = a functional form for the potential energy + a set of parameters for each atom type:
$$U = \underbrace{\sum_{\text{bonds}} k_b (b - b_0)^2}_{\text{stretching}} + \underbrace{\sum_{\text{angles}} k_\theta(\theta - \theta_0)^2}_{\text{bending}} + \underbrace{\sum_{\text{dihedrals}} \tfrac{V_n}{2}\,[1 + \cos(n\phi - \gamma)]}_{\text{torsion}} + \underbrace{\sum_{\text{impropers}}(\ldots)}_{\text{planarity, chirality}} + \underbrace{\sum_{iThe first four are bonded terms; LJ and Coulomb are non-bonded (between atoms in different molecules or separated by three or more bonds; 1–2 and 1–3 pairs are excluded, and 1–4 pairs are often scaled down).
Common force fields
| Force field | Typical use | Note |
|---|---|---|
| AMBER | Proteins, nucleic acids | All-atom; GAFF extends it to small ligands |
| CHARMM | Proteins, lipids, nucleic acids | All-atom |
| GROMOS | Biomolecules | United-atom: non-polar H merged into their carbons |
| OPLS-AA | Proteins, organic liquids | Fitted to liquid properties |
| MM2/MM3, MMFF94 | Small organic and drug-like molecules | Include cross terms |
| MARTINI | Membranes, large assemblies | Coarse-grained: about four heavy atoms per bead |
Solvent is either explicit (water models such as TIP3P, SPC/E, TIP4P) or implicit (a continuum: Poisson–Boltzmann or generalised Born). Explicit water is more realistic and far more expensive.
Energy minimisation
Minimisation moves atoms downhill on the energy surface until the forces are close to zero.
| Algorithm | Uses | Behaviour |
|---|---|---|
| Steepest descent | First derivative (gradient) | Robust far from the minimum (removes bad clashes); slow to converge |
| Conjugate gradient | Gradient plus history of previous steps | Faster convergence near the minimum |
| Newton–Raphson | Second derivatives (Hessian) | Fewest steps but expensive per step; L-BFGS approximates it cheaply |
Minimisation finds the nearest local minimum, not the global one, and it gives a structure at 0 K with no motion. To explore conformations, use conformational search: systematic search (small molecules), Monte Carlo, simulated annealing (heat, then cool slowly), genetic algorithms, or molecular dynamics.
Metropolis Monte Carlo. Make a random move; accept it if the energy falls; if it rises by ΔE, accept with probability \(e^{-\Delta E/k_BT}\). At 300 K (RT ≈ 2.49 kJ/mol), an uphill move of 5 kJ/mol is accepted about 13% of the time, and one of 10 kJ/mol about 2%. The procedure samples the Boltzmann distribution.
Part 2: Molecular Dynamics (MD)
MD solves Newton’s equations of motion for every atom:
$$\mathbf{F}_i = -\nabla_i U = m_i \mathbf{a}_i$$Positions and velocities are advanced in small time steps with an integrator such as Verlet, leapfrog or velocity Verlet:
$$\mathbf{r}(t+\Delta t) = 2\mathbf{r}(t) - \mathbf{r}(t-\Delta t) + \mathbf{a}(t)\,\Delta t^2$$- Time step: must be short compared with the fastest motion, X–H bond vibration (period of the order of 10 fs). So 1 fs, or 2 fs when bonds to hydrogen are constrained with SHAKE or LINCS.
- Initial velocities are drawn from the Maxwell–Boltzmann distribution at the target temperature; temperature is computed from the kinetic energy (equipartition: average kinetic energy per atom \(= \tfrac32 k_BT\)).
- Periodic boundary conditions: the box is surrounded by copies of itself to avoid surface artefacts. Short-range interactions use a cut-off (typically around 1 nm); long-range electrostatics use particle-mesh Ewald (PME).
Ensembles
| Ensemble | Held constant | Needs | Use |
|---|---|---|---|
| NVE (microcanonical) | Number of particles, volume, energy | Nothing extra | Checks energy conservation of the integrator |
| NVT (canonical) | N, V, temperature | Thermostat (Berendsen, Nosé–Hoover, velocity rescaling, Langevin) | Heating, equilibration |
| NPT (isothermal–isobaric) | N, pressure, temperature | Thermostat + barostat (Berendsen, Parrinello–Rahman) | Production runs that mimic laboratory conditions |
A typical simulation protocol
graph LR
A["Starting structure
(PDB or model);
add H, protonate"] --> B["Solvate in water box;
add ions to
neutralise"]
B --> C["Minimise
(remove clashes)"]
C --> D["Heat and equilibrate
NVT then NPT,
restraints released"]
D --> E["Production run
ns to µs"]
E --> F["Analysis:
RMSD, RMSF, Rg,
H-bonds, SASA, PCA"]Analysing a trajectory
| Quantity | What it shows |
|---|---|
| RMSD vs time (to the start or an average structure) | Overall drift; a plateau suggests equilibration |
| RMSF per residue | Flexibility: loops and termini high, core low (compare crystallographic B-factors) |
| Radius of gyration (Rg) | Compactness: folding, unfolding, swelling |
| H-bonds, salt bridges, SASA | Specific interactions and burial |
| PCA (essential dynamics) | Dominant collective motions (domain hinge, opening/closing) |
| Free energy (FEP, thermodynamic integration, MM-PBSA/GBSA) | Binding or mutation free-energy differences |
Timescale problem. Routine simulations reach nanoseconds to microseconds; many biological events (folding of larger proteins, slow conformational changes) take longer. Enhanced sampling methods help: replica exchange (copies at different temperatures swap), metadynamics (bias fills visited free-energy wells), umbrella sampling, steered MD, and coarse-grained models. Common packages: GROMACS, AMBER, CHARMM, NAMD, OpenMM.
Part 3: Protein Structure Classification
Structure is more conserved than sequence: proteins with no detectable sequence similarity can share a fold. Classification databases organise domains hierarchically.
| Level | SCOP (Structural Classification of Proteins; manual curation) | CATH (semi-automatic) |
|---|---|---|
| Top | Class: all-α, all-β, α/β (mainly parallel β with alternating α, e.g. TIM barrel), α+β (separate α and mainly antiparallel β regions), multi-domain, membrane, small proteins | Class: 1 mainly α, 2 mainly β, 3 α–β, 4 few secondary structures |
| 2 | Fold: same major secondary structures in the same arrangement and topology | Architecture: overall arrangement of secondary structures, ignoring connections (e.g. sandwich, barrel, roll) |
| 3 | Superfamily: probable common ancestry, from structure and function, despite low sequence identity | Topology (fold): connectivity |
| 4 | Family: clear evolutionary relationship, usually from significant sequence identity | Homologous superfamily |
CATH = Class, Architecture, Topology, Homologous superfamily (further split into sequence families). Structural alignment programs used to compare folds include DALI, CE and TM-align. A SCOP successor (SCOP2) reorganised the hierarchy, but the four classical levels above are what exams test.
Part 4: Protein Structure Prediction
Levinthal’s paradox and the energy landscape
If each of 100 residues had just 3 backbone conformations, there would be \(3^{100} \approx 5\times10^{47}\) conformations; sampling them at 10¹³ per second would take about 10²⁷ years. Real proteins fold in milliseconds to seconds, so folding is not a random search: it follows a funnel-shaped energy landscape towards the native state. Prediction methods have to shortcut the search too.
Secondary structure and other 1D predictions
- Chou–Fasman: residue propensities for helix, sheet and turn from known structures. GOR: information theory using neighbouring residues. PSIPRED: neural networks on PSI-BLAST profiles. Using evolutionary profiles raised three-state accuracy (Q3) well above the first-generation methods.
- Transmembrane helices: hydropathy plots (Kyte–Doolittle; a hydrophobic stretch of about 20 residues spans the bilayer) and HMM predictors such as TMHMM.
- Signal peptides: SignalP. Intrinsic disorder: dedicated predictors; AlphaFold’s low-confidence regions often coincide with disorder.
Three classical routes to 3D
graph TD
Q["Query sequence"] --> B{"Template found by
BLAST / PSI-BLAST?"}
B -- "Yes, clear homologue" --> H["Homology (comparative)
modelling
MODELLER, SWISS-MODEL"]
B -- "No" --> T{"Fold recognised by
threading or
profile-profile?"}
T -- "Yes" --> TH["Threading /
fold recognition
Phyre2, I-TASSER"]
T -- "No" --> AB["Ab initio / de novo
Rosetta fragment
assembly"]
H --> V["Validate and refine"]
TH --> V
AB --> V
Q --> AF["Deep learning
AlphaFold2/3,
RoseTTAFold, ESMFold"]
AF --> VHomology (comparative) modelling
Based on two observations: similar sequences have similar structures, and the number of folds is far smaller than the number of sequences.
- Template search: BLAST/PSI-BLAST/HMM search of the PDB.
- Target–template alignment: the most important step; alignment errors cannot be fixed later.
- Backbone building: copy conserved-region coordinates. MODELLER builds the model by satisfaction of spatial restraints derived from the template; SWISS-MODEL is an automated server.
- Loop modelling: insertions and deletions (database fragments or ab initio).
- Side-chain modelling: rotamer libraries (for example SCWRL).
- Refinement (brief minimisation or MD) and validation.
Accuracy depends on sequence identity. Above roughly 50% identity, models can approach a medium-resolution experimental structure; in the 30–50% range they are usually useful with errors in loops and side chains; below about 30% (the twilight zone, roughly 20–35%), alignment errors become common and other methods are needed.
Threading (fold recognition)
When no template is detectable by sequence, thread the query through a library of known folds and score each fit with knowledge-based (statistical) potentials or 3D profiles that ask whether each residue suits its environment (buried or exposed, helix, strand or loop). Profile–profile and HMM–HMM comparison (for example HHpred) is closely related. Servers: Phyre2, GenTHREADER, I-TASSER (threading fragments reassembled by simulation). Works because folds are limited; fails for genuinely new folds.
Ab initio (de novo) prediction
Uses no template: physical or knowledge-based energy function + conformational search. Rosetta assembles short fragments (3 and 9 residues) taken from known structures with similar local sequence, using Monte Carlo moves. Historically feasible mainly for small proteins. Co-evolution analysis of large MSAs (correlated mutations between residue pairs suggest they are in contact) gave the contact maps that made de novo prediction far better, and was the bridge to deep learning.
AI methods: AlphaFold and successors
AlphaFold2 (Google DeepMind; CASP14 in 2020, published in 2021):
- Input: the query sequence, a multiple sequence alignment of its homologues, and optional templates.
- Evoformer: attention-based blocks that repeatedly update an MSA representation and a pair representation (residue–residue relationships, including geometric consistency through “triangle” updates).
- Structure module: places each residue as a rigid frame and builds atomic coordinates directly; the output is recycled through the network several times.
- Output: 3D coordinates plus two confidence measures.
| Output | Meaning | How to read it |
|---|---|---|
| pLDDT (per residue, 0–100) | Predicted local accuracy (local distance difference test) | > 90 very high; 70–90 confident (backbone generally right); 50–70 low; < 50 very low, often disordered regions |
| PAE (predicted aligned error, residue × residue, in Å) | Expected position error of residue x when the model is superposed on residue y | Low PAE within a block = a confident domain; high PAE between two domains = their relative placement is uncertain, even if each domain has high pLDDT |
Limitations to state in an answer: one static structure per sequence (no ensembles or dynamics); weaker for proteins with few homologues (shallow MSAs) and for designed or orphan sequences; AlphaFold2 does not model ligands, ions, cofactors or PTMs; effects of point mutations are poorly predicted; a confident prediction is still a model, not an experiment.
Successors and relatives: AlphaFold-Multimer (complexes); AlphaFold3 (2024) extends prediction to complexes of proteins with nucleic acids, small molecules, ions and modified residues, using a diffusion-based structure generator; RoseTTAFold (Baker laboratory; a “three-track” network on sequence, distance and coordinates); ESMFold (Meta; predicts from a single sequence using a protein language model, without an MSA, so it is fast). The AlphaFold Protein Structure Database (with EMBL-EBI) holds predicted models for over 200 million UniProt sequences.
Validating and comparing models
| Tool or metric | What it checks |
|---|---|
| Ramachandran plot (PROCHECK, MolProbity) | Backbone φ/ψ in favoured or allowed regions; outliers flag errors (Gly and Pro excepted) |
| MolProbity clashscore | Atomic overlaps |
| Verify3D, ProSA (Z-score), QMEAN, ERRAT | Whether each residue’s environment and overall statistics look native-like |
| RMSD (Å) | Average atomic deviation after superposition; sensitive to a few large errors and to length |
| GDT_TS (0–100) | Percentage of residues within distance cut-offs; used in CASP |
| TM-score (0–1) | Length-normalised similarity; above about 0.5 usually means the same fold |
| lDDT | Superposition-free local agreement (basis of pLDDT) |
Memory Tricks
Don’t Confuse
| This | vs | That |
|---|---|---|
| Molecular mechanics: fixed bonds, fast | vs | Quantum mechanics: electrons, bond breaking, slow |
| Energy minimisation: nearest local minimum, 0 K, no time | vs | Molecular dynamics: trajectory in time at temperature T |
| Monte Carlo: random moves, Boltzmann sampling, no time | vs | MD: deterministic integration, real time |
| All-atom (AMBER, CHARMM) | vs | United-atom (GROMOS) vs coarse-grained (MARTINI) |
| RMSD: deviation from a reference structure | vs | RMSF: fluctuation of each atom about its mean |
| SCOP superfamily: probable common ancestry | vs | SCOP fold: same topology, ancestry not implied |
| Homology modelling: needs a detectable template | vs | Threading: fold match without sequence similarity |
| Threading | vs | Ab initio: no template at all |
| pLDDT: local confidence per residue | vs | PAE: confidence in relative positions of residue pairs |
| TM-score: 0–1, length-normalised | vs | RMSD: Å, length-dependent |
Common Mistakes
Practice
All questions below are original practice questions written in CSIR NET style. They are not previous-year questions; this unit is new in the revised syllabus.
Level 1: Foundation
- Which ensemble keeps the number of particles, pressure and temperature constant?
- Expand CATH.
- What does a pLDDT above 90 indicate?
Answers
- NPT (isothermal–isobaric).
- Class, Architecture, Topology, Homologous superfamily.
- Very high local confidence: backbone and most side chains likely accurate.
Level 2: Part B Style
Q1 (practice). For a Lennard-Jones pair with σ = 3.4 Å and well depth ε, the separation at the energy minimum and the energy there are:
- (a) 3.4 Å and 0
- (b) 3.82 Å and −ε
- (c) 3.82 Å and −ε/4
- (d) 6.8 Å and −ε
Answer and explanation
(b). Setting dU/dr = 0 gives \(r_{min} = 2^{1/6}\sigma = 1.1225 \times 3.4 = 3.82\) Å. Substituting: \(U = 4\varepsilon[(1/2)^2 - 1/2] = 4\varepsilon(1/4 - 1/2) = -\varepsilon\). At r = σ the energy is zero.
Q2 (practice). An all-atom MD run with bonds to hydrogen constrained uses a 2 fs time step. How many integration steps are needed for 100 ns?
- (a) 5 × 10⁵
- (b) 5 × 10⁷
- (c) 2 × 10⁸
- (d) 5 × 10⁹
Answer and explanation
(b). \(100\times10^{-9}\ \mathrm{s} / 2\times10^{-15}\ \mathrm{s} = 5\times10^{7}\) steps.
Q3 (practice). A protein’s best BLAST hit in the PDB is 14% identical over a short stretch with E = 4. A profile–profile search detects a likely TIM-barrel fold with high confidence. The most suitable classical modelling approach is:
- (a) homology modelling with the BLAST hit as template
- (b) threading / fold recognition
- (c) energy minimisation of an extended chain
- (d) NMR structure determination by MD
Answer and explanation
(b). The BLAST hit is not significant (E = 4) and identity is far below the twilight zone, so it is not a usable template. A fold is recognised without sequence similarity, which is exactly the threading scenario. (A deep-learning predictor such as AlphaFold would also be a natural modern choice.)
Q4 (practice). Two domains are placed in the same SCOP superfamily but in different families. This means:
- (a) they share a fold and probable common ancestry, but sequence identity is low
- (b) they have identical sequences
- (c) they have different folds
- (d) they share a fold by convergent evolution only
Answer and explanation
(a). Superfamily = probable common evolutionary origin inferred from structure and function; family = clear relationship usually visible in sequence. Different families within a superfamily share ancestry and fold but have low sequence identity.
Level 3: Part C Style
Q1 (practice). After superposing a homology model on the later-solved crystal structure, the Cα deviations (Å) for a four-residue segment are 0.5, 1.0, 1.5 and 2.0.
(i) Calculate the RMSD for this segment. (ii) If one more residue in a flexible loop had a deviation of 6.0 Å, what would the five-residue RMSD be? (iii) What does this say about RMSD as a quality measure?
Answer and worked solution
- \(\text{RMSD} = \sqrt{(0.25 + 1 + 2.25 + 4)/4} = \sqrt{7.5/4} = \sqrt{1.875} = 1.37\) Å.
- \(\sqrt{(7.5 + 36)/5} = \sqrt{8.7} = 2.95\) Å.
- One outlier more than doubled the RMSD. RMSD is dominated by the largest errors (it squares them), so a model with an accurate core and one misplaced loop can look poor. Measures that count the fraction of residues within cut-offs (GDT_TS) or are length-normalised (TM-score) are more robust; per-residue confidence (pLDDT) shows where the error is.
Q2 (practice). Using \(U = 1389\,q_1 q_2/(\varepsilon_r r)\) kJ/mol (r in Å, charges in units of e), calculate the Coulomb energy of a Lys⁺–Glu⁻ ion pair at 4 Å (i) in vacuum (ε_r = 1), (ii) in a protein interior (ε_r ≈ 4), and (iii) in bulk water (ε_r ≈ 80). (iv) At 300 K RT is about 2.5 kJ/mol. Comment on whether the surface salt bridge is stable.
Answer and worked solution
- Vacuum: \(1389 \times (-1)/(1 \times 4) \approx -347\) kJ/mol.
- Interior: \(-1389/16 \approx -87\) kJ/mol.
- Water: \(-1389/320 \approx -4.3\) kJ/mol.
- In water the interaction is under 2RT, so a surface salt bridge is only weakly stabilising and is frequently broken by thermal motion. Buried ion pairs are much stronger in principle, but burying charges costs a large desolvation penalty, so net stabilisation is not simply the interior Coulomb energy. This is why the dielectric treatment (explicit water, PME, or a continuum model) matters so much in force-field calculations.
Q3 (practice). An AlphaFold2 model of a 400-residue protein shows: residues 1–40, pLDDT below 40; residues 41–200 (domain 1) and 221–400 (domain 2), pLDDT above 90; residues 201–220, pLDDT about 55. The PAE plot shows low error within each domain but high error (about 25 Å) between domain 1 and domain 2. Which statements are correct?
(A) Residues 1–40 are probably intrinsically disordered or flexible, not necessarily mis-predicted. (B) The folds of the two domains are each likely to be accurate. (C) The relative orientation of the two domains in the model is reliable because both domains have pLDDT above 90. (D) A ligand-binding pocket in domain 1 should be checked with ligand-aware methods or experiment, because AlphaFold2 does not model bound ligands.
- (a) A, B and D only
- (b) A, B and C only
- (c) B and C only
- (d) A, B, C and D
Answer and worked solution
(a) A, B and D.
- (A) Correct: very low pLDDT often marks disorder.
- (B) Correct: high pLDDT and low intra-domain PAE.
- (C) Wrong: pLDDT is local. The high inter-domain PAE says the domains’ relative placement is uncertain, as expected with a flexible linker (residues 201–220, pLDDT about 55).
- (D) Correct: AlphaFold2 predicts the apo polypeptide; pocket shape may differ in the bound state.
Quick Revision Table
| Item | Fact |
|---|---|
| MM vs QM | MM: balls and springs, no bond breaking; QM/MM for reactions |
| Force field | Bonds + angles + torsions + impropers + LJ + Coulomb |
| LJ 12-6 | Zero at σ; minimum −ε at 2^(1/6)σ ≈ 1.12σ |
| Coulomb | 1389 q₁q₂/(ε r) kJ/mol·Å; screened by dielectric |
| Force fields | AMBER, CHARMM, OPLS-AA (all-atom); GROMOS (united-atom); MARTINI (coarse-grained) |
| Water models | TIP3P, SPC/E, TIP4P; implicit: PB, GB |
| Minimisers | Steepest descent (robust), conjugate gradient, Newton–Raphson / L-BFGS |
| Monte Carlo | Metropolis: accept uphill with e^(−ΔE/kT) |
| MD | F = −∇U; Verlet; 1–2 fs (SHAKE/LINCS); PBC; PME |
| Ensembles | NVE; NVT (thermostat); NPT (thermostat + barostat) |
| Analysis | RMSD, RMSF, Rg, PCA, H-bonds, SASA, MM-PBSA |
| SCOP | Class → Fold → Superfamily → Family |
| CATH | Class (1 α, 2 β, 3 α–β, 4 few SS) → Architecture → Topology → Homologous superfamily |
| Levinthal | 3^100 ≈ 5 × 10⁴⁷ conformations; folding funnel |
| Homology modelling | Template → alignment → backbone → loops → side chains → refine → validate; MODELLER, SWISS-MODEL |
| Identity guide | > 50% high accuracy; 30–50% useful; < 30% twilight zone |
| Threading | Fold library + knowledge-based potentials; Phyre2, I-TASSER |
| Ab initio | Rosetta fragment assembly; coevolution contacts |
| AlphaFold2 | MSA → Evoformer → structure module; pLDDT and PAE; CASP14 (2020) |
| pLDDT bands | > 90, 70–90, 50–70, < 50 |
| Successors | AlphaFold-Multimer, AlphaFold3 (ligands, nucleic acids), RoseTTAFold, ESMFold |
| Validation | Ramachandran, MolProbity, ProSA, QMEAN; RMSD, GDT_TS, TM-score (> 0.5 same fold) |
Key Takeaways
What’s Next
- Computer-Aided Drug Design: SBDD, LBDD, QSAR, Pharmacophores and ADMET — these force fields and structures put to work on ligands.
- Enzymes: Catalysis, Kinetics, Regulation and Isozymes — active sites and transition states that structures and simulations explain.
- Protein Structure, Conformation and Stability — revisit Ramachandran plots, folds and folding thermodynamics.
- Bioinformatic Resources and Databases — PDB records, resolution and the AlphaFold database.
- Chapter overview and Quick Revision.