Molecular simulation

Why it matters

Experiments give snapshots; simulations and predictions fill in the gaps. Molecular dynamics (MD) shows how a protein moves, how a ligand slides into a pocket or how a membrane channel opens. Structure prediction with tools such as AlphaFold now gives a confident model for most protein sequences in minutes. Biochemists and biophysicists do not all need to write simulation code, but every one of them reads computational results, and many run them, so understanding what these methods do and where they fail is essential.

Molecular dynamics at a high level

MD treats atoms as balls and bonds as springs and moves them according to Newton's laws:

1. Start from a structure (from the PDB or a prediction), add hydrogens, put it in a box of explicit water molecules and ions at a realistic concentration.
2. Force field: a set of equations and parameters (such as AMBER, CHARMM or OPLS families) giving the potential energy from bond lengths, angles, torsions, van der Waals and electrostatic interactions.
3. Integrate: compute the force on every atom, move all atoms a tiny time step (typically 2 femtoseconds, 2 × 10⁻¹⁵ s), and repeat millions of times.
4. Control conditions: thermostats and barostats keep temperature and pressure realistic (for example 310 K, 1 bar).
5. Analyse the trajectory: RMSD (how far the structure drifts from the start), RMSF (which residues flex most), hydrogen bonds, distances, and binding free energies with special methods.

Scale check. One microsecond of simulation at a 2 fs time step is 500 million steps. On modern GPUs, a medium-size protein in water (around 100,000 atoms) might run roughly hundreds of nanoseconds per day, so microsecond runs take days to weeks. Many biological events (large conformational changes, folding of big proteins) take milliseconds or longer and need enhanced sampling methods or specialised hardware.

What MD is good at: local flexibility, water and ion networks, how mutations change motion, refining ligand poses, interpreting experimental data.

Limits: force fields are approximations (no bond breaking in classical MD; reactions need QM/MM methods); results depend on the starting structure and the sampling; a single short run can be misleading, so researchers run replicates.

Structure prediction: AlphaFold and friends

AlphaFold 2 (DeepMind, 2020–2021) used deep learning on sequences and known structures, plus information from evolutionarily related sequences (a multiple sequence alignment), to predict 3D structures with accuracy close to experiment for many proteins. Its developers shared the 2024 Nobel Prize in Chemistry with David Baker, whose lab pioneered computational protein design. The AlphaFold Protein Structure Database (with EMBL-EBI) holds over 200 million predicted structures. Newer tools (AlphaFold 3, RoseTTAFold and others) also model complexes with other proteins, nucleic acids and small molecules.

Confidence scores you must read:

• pLDDT (0–100) per residue: above 90 very high; 70–90 good backbone; below 50 low, often disordered.
• PAE (predicted aligned error): how confident the model is about the relative position of two parts, for example two domains. Low PAE between domains means their arrangement is reliable.

What prediction does not tell you: which conformation dominates in the cell, how the protein moves, the effect of most point mutations on stability, exact ligand positions, or post-translational modifications. Predictions are hypotheses, best combined with experiments.

How the pieces fit together

A modern project might go: predict a structure with AlphaFold → use it to solve an X-ray structure by molecular replacement or to fit a cryo-EM map → run MD to explore flexibility and why a mutation matters → test predictions with kinetics, binding or CD. Computation and experiment check each other.

Worked example: interpreting results

A paper reports a 500 ns MD simulation of a kinase showing a loop "opening" near the active site, and an AlphaFold model of the same protein where the loop has pLDDT of 45.

• The low pLDDT suggests the loop is flexible or disordered, consistent with motion.
• One 500 ns run showing opening once is weak evidence; three or more replicates, and an experimental check (for example, NMR or hydrogen–deuterium exchange) would make the claim convincing.
• Neither result alone tells you how often the loop is open in cells.

Common mistakes

• Treating a high-confidence prediction as proof of function or binding.
• Ignoring PAE when interpreting domain arrangements.
• Over-interpreting one short MD trajectory.
• Forgetting the environment: membrane proteins need a lipid bilayer in simulation; pH sets protonation states.

Recap

• MD integrates Newton's laws with a force field in femtosecond steps; it explores motion but is limited by force-field accuracy and timescale.
• AlphaFold predicts structures with per-residue (pLDDT) and pairwise (PAE) confidence.
• Both are hypothesis generators that work best alongside experiments.

Try it

This lesson has no activity. Check your understanding with the quiz on the Do tab.