Predicting a protein structure with AlphaFold in Colab

A quick primer before the walkthrough:

In plain English: Predict an illustrative 3D protein structure hypothesis from an amino-acid sequence using AlphaFold in a Google Colab notebook — without requiring a local GPU cluster.

Key terms:

• AlphaFold — A deep-learning model developed by Google DeepMind designed to predict protein 3D structures from primary sequences.
• FASTA — A standard plain-text format for sequence data (typically a single header line followed by sequence characters).
• Google Colab — A cloud-hosted Jupyter notebook platform offering cloud GPU access.

What you'll need: A Google account and a representative protein sequence in FASTA format.

Comfort level: Useful if you have executed a Colab notebook and are familiar with basic protein sequence concepts. Not a beginner AI intro — consult introductory primers if this is your first AI notebook workflow.

New to this? You do not need to be a structural biologist to follow along. If you have never used an AI chat tool for lab work, read What is an LLM? A primer for life science teams first. A Colab notebook is just a web page with blocks of code you press play on — you are not installing anything on your computer.

Why AlphaFold in Colab is still the right starting point

For individual protein targets or small illustrative complexes, setting up a local AlphaFold installation with heavy database requirements may be unnecessary for initial exploration. The community-developed ColabFold notebook — which implements an optimized AlphaFold pipeline using MMseqs2 for fast multiple sequence alignment (MSA) generation — provides an accessible starting point. ColabFold documentation notes typical runtimes of 10–30 minutes for modest sequence lengths (as of 2024), depending on server load and cloud resource availability. For many bench scientists, it offers a fast path from FASTA sequence to preliminary PDB structural coordinates.

This tutorial provides an illustrative end-to-end walkthrough: sequence formatting, parameter selection, score interpretation, and structural evaluation. All steps are illustrative examples subject to your institution's specific SOPs and expert scientific judgment.

What you will need

• A Google account. (Note: Google Colab provides variable cloud GPU compute allocations under its terms of service. For proprietary or sensitive data, review Google's current privacy terms and your organizational vendor agreements.)
• The target amino acid sequence in single-line text format.
• Context regarding the target: expected monomer or multimer state, known domain boundaries, or predicted intrinsic disorder.

Step 1 — Open ColabFold

Locate the official ColabFold repository (e.g., "ColabFold AlphaFold2.ipynb" maintained by Mirdita et al.) or click the "Open in Colab" link on the repository page. Save a copy to your Google Drive so custom parameter selections persist across sessions.

Step 2 — Paste your sequence

The query_sequence input field accepts single or multi-chain entries using colon separators for illustrative targets:

• Monomer example: MKTIIALSYIFCLVFA...
• Homodimer example: MKTIIALSYIFCLVFA...:MKTIIALSYIFCLVFA...
• Heterodimer example: MKTIIAL...:MEVFPQ...

Signal peptides may be trimmed prior to input if analyzing mature protein forms, as un-cleaved N-terminal signal peptides frequently exhibit lower local confidence scores.

Step 3 — Set the important parameters

Adjust parameters based on your target requirements and internal guidelines:

• num_recycles: 3 is a common default starting point. Increasing to 6 may assist complex or hard-to-fold targets (e.g., larger structures or potential novel folds).
• num_relax: Setting to 1 applies an AMBER molecular mechanics relaxation step to top-ranked models. This step can help relieve minor steric clashes for downstream visualization, subject to your computational pipeline requirements.
• msa_mode: mmseqs2_uniref_env serves as the standard default search mode. single_sequence mode is typically reserved for testing targets without evolutionary MSA context.
• template_mode: none relies entirely on the neural network prediction; pdb100 allows structural templates from the PDB where available.

Step 4 — Run all cells

Select Runtime → Run all. Queue times for the public MMseqs2 alignment server vary based on concurrent network traffic.

Upon completion, ColabFold outputs five generated models, typically ordered by predicted metric scores (such as mean pLDDT for monomers or ipTM/pTM for multimeric complexes, depending on notebook settings).

Step 5 — Read the confidence outputs

Structure predictions should be evaluated against standard confidence metrics defined in the AlphaFold literature (Jumper et al., 2021; Mirdita et al., 2022):

• pLDDT (per-residue predicted Local Distance Difference Test): Measured on a 0–100 scale. Values > 90 generally indicate high local confidence; 70–90 denotes confident backbone modeling; 50–70 indicates lower confidence where local ribbon topology is speculative; < 50 often correlates with predicted flexible or disordered regions. Residues are color-coded in the embedded 3D viewer.
• pTM / ipTM (predicted TM-score): Scaled 0–1. For monomers, a pTM > 0.7 generally suggests high confidence in overall tertiary topology. For complexes, an interface pTM (ipTM) > 0.6 serves as a common heuristic for plausible inter-chain interactions, whereas ipTM < 0.4 indicates high interface uncertainty.

The Predicted Aligned Error (PAE) matrix visualizes domain relative position confidence. Well-defined dark square blocks along the diagonal reflect rigid domains. Off-diagonal intensity indicates relative position confidence between domain pairs: light off-diagonal regions suggest that while individual domains may be predicted with high confidence, their relative orientation remains uncertain.

Step 6 — Download the files

The notebook compiles output files into an archive within your connected environment:

• *_relaxed_rank_001_*.pdb — The top-ranked model coordinate file (post-AMBER relaxation if selected).
• *_scores_rank_001_*.json — Numerical pLDDT and PAE matrix records.
• *_coverage.png — Sequence alignment depth across residue positions. Low coverage regions frequently correspond to lower confidence predictions, though exception cases occur based on template availability.

Step 7 — Look at the model in PyMOL or ChimeraX

Open the generated PDB coordinate file in molecular visualization software (such as PyMOL or UCSF ChimeraX) and color residues by B-factor, where ColabFold records the pLDDT confidence values. High confidence regions (typically blue/cyan) interspersed with flexible loops (red) are standard for many structural targets. Widespread low confidence scores across the full sequence may indicate insufficient MSA coverage or intrinsic protein disorder.

Things this workflow will not tell you

• Biological function or active state. AlphaFold models represent static structural hypotheses; they do not inherently determine enzymatic activity, biological function, or ligand-bound/free conformational transitions.
• Conformational ensembles. Standard runs return static structural snapshots. Sampling alternative conformational states requires specialized protocols (such as MSA subsampling methods) or specialized dynamic modeling tools.
• Post-translational modifications (PTMs). Standard AlphaFold2 models do not include non-protein chemical modifications, phosphorylation, or glycosylation.
• Cofactors, nucleic acids, and small-molecule ligands. The basic ColabFold AlphaFold2 pipeline processes amino acid chains only. Assessing ligand interactions requires specialized tools (such as AlphaFold3 or molecular docking suites).
• Clinical or diagnostic applicability. Generated structural models are intended strictly for research hypotheses and are not validated for clinical, diagnostic, dosing, or therapeutic decision-making.

An illustrative structural evaluation checklist

Before using predicted structural coordinates as working hypotheses for experimental design, consider evaluating the following heuristic indicators alongside expert structural biology review:

• Mean pLDDT across target functional domains exceeds 70.
• MSA depth across target regions meets your protocol's sequence coverage requirements (e.g., > 30 aligned sequences).
• PAE off-diagonal scores show low relative error for critical domain-domain interfaces.
• For multimeric complexes, ipTM exceeds 0.6 with interface residues exhibiting high local pLDDT scores (> 70) on participating chains.
• Geometry checks in visualization software (e.g., find_clashes in PyMOL) reveal no significant steric overlaps in the relaxed model.

Note: Satisfying these computational metrics provides a candidate structural hypothesis but does not replace formal experimental validation (e.g., X-ray crystallography, cryo-EM, or NMR).

If something goes wrong

• The notebook stops with a red error in the first cell. Usually the runtime disconnected or Colab did not give you a GPU. Choose Runtime → Disconnect and delete runtime, then Runtime → Change runtime type → GPU, and run all cells again.
• It has been sitting on the MSA step for a long time. The public alignment server queues jobs. Leave the tab open; if it fails outright, try again later or switch msa_mode to a different setting to test the pipeline.
• Everything comes back blue-and-red mush (low pLDDT across the whole chain). Two common causes: the sequence is genuinely disordered, or the alignment found very few related sequences. Check the *_coverage.png file — a thin, pale plot means shallow coverage and a weak prediction.
• The model looks fine but the two chains are in an odd arrangement. Look at the PAE plot, not the ribbon. Confident domains with an uncertain relative orientation is a very common (and honest) result.
• You are not sure whether the result is good enough. Walk the evaluation checklist below with a structural biologist before designing experiments around the model.

Plain-English glossary

• Sequence — the string of letters (one per amino acid) that describes a protein.
• MSA (multiple sequence alignment) — a stack of related protein sequences from other organisms; AlphaFold uses the evolutionary patterns in it as evidence.
• pLDDT — the model's confidence in each residue's local position, 0–100. Higher is better.
• PAE — a square plot showing how sure the model is about where one part of the protein sits relative to another.
• Relaxation — a quick physics clean-up that removes atoms that overlap slightly.
• PDB file — the standard text file holding 3D atom coordinates.