Going Deeper

AlphaFold-Multimer: predicting protein complexes

AlphaFold 2 was originally designed for single chains. But most biologically interesting proteins don't work alone — they form complexes: homodimers (two identical chains), heterodimers (two different chains), or larger assemblies.

AlphaFold-Multimer (released in 2021, updated in 2022) extends AlphaFold 2 to handle complexes of up to 5 chains. Key changes include:

• Building a joint MSA that includes sequences for all chains simultaneously, allowing co-evolutionary signals between chains (inter-chain contacts) to be captured.
• Modified training on protein complexes from the PDB, so the model has seen inter-chain packing examples.
• Updated confidence metrics — ipTM (interface pTM) specifically measures confidence at the predicted interface.

ColabFold supports multimer predictions — simply enter multiple sequences separated by a colon (:) in the sequence field. The notebook handles the rest.

Practical uses of Multimer

• Predicting antibody–antigen complexes: Understanding how antibodies grip their targets for vaccine design and therapeutics.
• Enzyme–substrate interactions: Modelling how substrates bind to enzyme active sites in the context of the whole complex.
• Signalling complexes: How receptor kinases dimerize upon activation, or how scaffold proteins assemble multi-enzyme complexes.
• Structural genomics: Systematically predicting the structures of all protein interactions in an organism's protein–protein interaction network (interactome).

AlphaFold 3 in depth

We covered AlphaFold 3's architecture in Part 4. Here we focus on its expanded capabilities and when to use it over AlphaFold 2.

What AF3 adds

AF3 limitations vs. AF2

AF3 is not uniformly better than AF2 for all tasks:

• Single-chain protein accuracy: AF3 is roughly comparable to AF2 for single chains — neither dramatically outperforms the other on this task.
• Stochastic outputs: AF3's diffusion sampling means different runs can give somewhat different structures. Run multiple seeds and pick the highest-confidence result.
• Commercial use restrictions: The public AlphaFold Server has non-commercial restrictions. The AF3 code was released in late 2024 under a more permissive licence, enabling local installation.

Molecular dynamics: from snapshot to movie

As we discussed in Part 7, AlphaFold gives you a static structure. To study protein dynamics — how proteins flex, how binding events cause conformational changes, how inhibitors stabilise particular conformations — you need molecular dynamics (MD) simulation.

MD simulation numerically integrates Newton's equations of motion for every atom in the system (protein + water + ions), stepping forward in tiny time increments (typically 2 femtoseconds). Over millions of steps, you accumulate a trajectory showing how the protein moves.

AlphaFold + MD workflows

The combination is now a standard pipeline:

1. Get the AlphaFold structure (or experimental structure).
2. Set up an MD system: place the protein in a box of explicit water molecules, add ions to neutralise charge.
3. Energy minimise (relax close atomic contacts from the static structure).
4. Run MD simulation (nanoseconds to microseconds, depending on hardware).
5. Analyse trajectories: RMSD (how much does the structure drift?), RMSF (which residues fluctuate most?), principal component analysis (what are the dominant motions?).

Key MD software

Open-source alternatives to AlphaFold

AlphaFold is not the only game in town. A rich ecosystem of structure prediction tools has emerged, some offering advantages for specific use cases:

Protein design: the inverse problem

AlphaFold solves the direct problem: given a sequence, predict the structure. The inverse problem is equally revolutionary: given a desired structure or function, design a sequence that will fold into it.

This is the field of computational protein design, for which David Baker shared the 2024 Nobel Prize. Key tools:

• ProteinMPNN: A graph neural network that, given a 3-D backbone structure, designs amino acid sequences likely to fold into that structure. Works remarkably well.
• RFdiffusion: David Baker's diffusion model for generating entirely new protein backbones from scratch, then using ProteinMPNN to design sequences for them.
• AlphaFold as validation: After designing a sequence with RFdiffusion + ProteinMPNN, run AlphaFold on the designed sequence. If AlphaFold predicts the target structure with high confidence, the design is promising for experimental testing.

This loop — design → AlphaFold validation → experimental test — is now closing the gap between computational protein design and experimental reality at an unprecedented rate.

Where the field is heading

The structural biology field is moving rapidly. Active frontiers include:

• Ensemble prediction: Predicting the full range of conformations a protein adopts (not just the dominant one), using generative models like AF3's diffusion approach or dedicated tools like AlphaFlow.
• Function prediction: Going beyond structure to predict what a protein does — binding partners, catalytic activity, subcellular location — from sequence alone.
• Whole-proteome interaction networks: Using multimer prediction at scale to map all protein–protein interactions in a cell — an enormous computational undertaking now becoming feasible.
• RNA structure: Applying the same ideas to RNA — which also folds into complex 3-D structures. AlphaFold 3 handles some RNA, and dedicated RNA structure tools (like RhoFold and trRosettaRNA) are improving rapidly.
• Faster, lighter models: ESMFold showed that language models can predict structures without MSA. The next generation will be faster still — enabling real-time structure prediction in drug discovery pipelines.