AlphaFold Tutorial for Beginners: Free Course (16 Lectures)

In 2024 the Nobel Prize in Chemistry went partly to Demis Hassabis and John Jumper of Google DeepMind for AlphaFold. For fifty years, working out the 3D shape of a single protein could take a PhD student years in the lab. AlphaFold does a good first attempt in minutes.

Most AlphaFold tutorials skip straight to the tool and assume you already know structural biology. My students kept asking for something that starts from zero, in Urdu, and ends with analysis they could put in a thesis. So I recorded a full lecture series and wrote this guide to go with it. Treat it like a short book: one chapter at a time, with the lectures embedded where they belong.

Lecture 01 · 9:48

Your body contains tens of thousands of different proteins, and almost everything that happens in a cell is done by one of them. Enzymes speed up chemical reactions. Haemoglobin carries oxygen. Antibodies recognise viruses. Insulin tells cells to take up sugar. Most medicines work by sticking to a protein and changing what it does.

The building blocks: amino acids

A protein is a chain of amino acids. Every amino acid has the same backbone (an amino group, a central carbon and a carboxyl group) and a different side chain. Twenty standard side chains are used by every living thing. Their chemistry decides where each one ends up once the protein folds:

Amino acids link head to tail through peptide bonds. A chain can be a few dozen amino acids long or several thousand. Each link in the chain is called a residue, a word you’ll see in every AlphaFold output.

From gene to protein

The order of amino acids is written in DNA. A gene is copied into messenger RNA, and the ribosome reads that RNA three letters (one codon) at a time, adding one amino acid per codon. This is the central dogma of molecular biology, and it sets up the whole course:

AlphaFold works on the middle arrow: from protein chain to 3D shape. Chapters 8 to 10 deal with the last arrow, from shape to function.

Reading a protein sequence

Bioinformatics writes sequences with the one-letter codes, so the start of the human haemoglobin alpha chain is MVLSPADKTNVKAAW…. Sequences are stored in FASTA files: one header line starting with >, then the letters. Almost every tool in this course, AlphaFold included, takes FASTA as input.

>sp|P69905|HBA_HUMAN Hemoglobin subunit alpha
MVLSPADKTNVKAAWGKVGAHAGEYGAEALERMFLSFPTTKTYFPHFDLSHGSAQVKGHGKK
VADALTNAVAHVDDMPNALSALSDLHAHKLRVDPVNFKLLSHCLLVTLAAHLPAEFTPAVHA
SLDKFLASVSTVLTSKYR

Every known protein sequence lives in UniProt, with an accession ID such as P69905 for the protein above. That ID is how you’ll look a protein up in the AlphaFold Database later.

Next: Chapter 2 · What are the four levels of protein structure? →

Lecture 02 · 12:48

A newly made chain doesn’t stay a loose string. In milliseconds to seconds it folds into a specific 3D shape, and that shape is what lets it do its job. Biologists describe the shape in four levels. AlphaFold’s output only makes sense once these are clear.

Why proteins fold

The main driving force is the hydrophobic effect. Water-hating side chains pack together in the middle, like oil drops in water, and the charged and polar ones stay on the surface. Hydrogen bonds then lock helices and sheets in place. Some proteins get help from “chaperone” proteins that stop them clumping while they fold.

When folding goes wrong, the results can be serious. Misfolded proteins that stick together are linked to Alzheimer’s and Parkinson’s disease, and one wrong amino acid in haemoglobin causes sickle cell disease. Shape is not a detail; it is the protein.

Measuring a structure in the lab

All experimental structures go into the Protein Data Bank (PDB), which holds over 200,000 of them. That sounds like a lot until you compare it with the hundreds of millions of known sequences. Most proteins had no structure at all.

Could we just predict the shape from the sequence? In 1969 Cyrus Levinthal pointed out that a protein can’t find its fold by trying every possibility, because the number of possible shapes is astronomically large, yet proteins fold reliably within seconds. So the information must be in the sequence. Reading it out became known as the protein folding problem, and it stayed unsolved for half a century.

Next: Chapter 3 · What is AlphaFold, and how do you use the AlphaFold Database? →

Lecture 03 · 22:12

Since 1994, structure prediction groups have competed every two years in CASP, a blind test where you predict structures that have been solved but not yet published. Progress was slow for decades. At CASP14 in 2020, AlphaFold 2 scored a median above 90 out of 100 on the main accuracy scale, close to experimental quality for most targets. The organisers said the problem was, in large part, solved.

DeepMind and EMBL-EBI then ran AlphaFold on nearly every known protein and published the results free as the AlphaFold Protein Structure Database. It now holds more than 200 million predicted structures. In the lecture I use it live. The steps are:

1. Go to alphafold.ebi.ac.uk and search by protein name, gene name or UniProt accession.
2. Pick the right organism. “Insulin” alone returns dozens of species.
3. Look at the 3D model. Its colours show confidence (explained below).
4. Check the PAE plot under the model.
5. Download the PDB or mmCIF file to open in a viewer.

How much should you trust it? pLDDT

Each residue gets a pLDDT score from 0 to 100: AlphaFold’s own estimate of how right it is about that part of the model. The database colours the structure by this score:

PAE (predicted aligned error) answers a different question: if I line the model up on residue A, how far off might residue B be? Low PAE between two domains means AlphaFold is sure how they sit relative to each other. High PAE means each domain may be fine on its own while the angle between them is a guess. On the plot, dark green blocks along the diagonal usually mark separate, well-defined domains.

Assignment 1: test your database skills

Lecture 04 · 0:48

1. Search the proteins given in the lecture on the AlphaFold Database.
2. Record the pLDDT for each, and say which regions you trust and which you don’t.
3. Download the structure files (PDB) and describe the secondary and tertiary folding you see.
4. Submit your answers in the format given in the lecture.

Spend the most time on step 2. Being able to say “this part I trust, this part I don’t” is what separates people who use AlphaFold from people who screenshot it.

Next: Chapter 4 · What is the difference between AlphaFold 2 and AlphaFold 3? →

Lecture 05 · 5:44

AlphaFold 2 was about single proteins. But real biology happens when a protein meets something: a strand of DNA, a drug molecule, a metal ion, another protein. AlphaFold 3, published in Nature in May 2024 by Google DeepMind and Isomorphic Labs, models those complexes in one go.

The multiple sequence alignment is the trick behind both. AlphaFold gathers related sequences from other species. If two positions always change together across evolution, they are probably touching in 3D. AlphaFold 2 relied on this signal heavily. AlphaFold 3 still uses it, but the diffusion module, similar to the method behind AI image generators, refines a noisy cloud of atoms into a structure.

For drug discovery this is a big deal: you can ask where a small molecule sits in a pocket, not just what the empty protein looks like. Even so, AlphaFold 2 is still the right tool for many jobs. The database is built on it, and ColabFold runs it quickly for free.

Next: Chapter 5 · How do you read a protein structure? →

Lecture 06 · 19:08

The first time you open a 3D protein, it looks like a plate of jalebi: ribbons twisting everywhere. The trick is to stop staring at the whole thing and pick out four elements one at a time.

• α-helices, drawn as spiral ribbons. A rigid coil with about 3.6 residues per turn.
• β-strands, drawn as flat arrows. Strands lying side by side form a β-sheet; the arrow points along the chain.
• Loops and turns, the thin tubes connecting them. Binding often happens here, and pLDDT often drops here too.
• Domains, compact units that fold on their own. Big proteins are often several domains on a string, each with its own job.

Free viewers

In any viewer, first colour by secondary structure to see helices and sheets. Then colour by B-factor: AlphaFold stores pLDDT in the B-factor column, so this brings back the confidence colours and shows you which parts to ignore.

Assignment 2: analyse a sample structure

Lecture 07 · 3:19

In this lecture I break down a real model live, then hand it over to you:

1. Open the structure of the protein given in the lecture in any viewer.
2. Find and count its helices, sheets and loops.
3. Identify its domains (UniProt or InterPro will help you name them).
4. Submit a short structural analysis with screenshots.

Next: Chapter 6 · How do you predict a protein structure with AlphaFold Server? →

Lecture 08 · 16:39

The database covers proteins already in UniProt. What if your protein is a new mutant, a designed sequence, or bound to DNA or a drug? Then you run the prediction yourself on AlphaFold Server, which runs AlphaFold 3 in your browser.

1. Sign in with a Google account. The server is free for non-commercial use, with a daily job limit.
2. Paste your amino acid sequence as a protein entity. Add more entities (protein, DNA, RNA, ligand, ion) for a complex.
3. Name the job and submit. Small jobs usually finish in a few minutes.
4. Open the result: five ranked models, a pLDDT-coloured view, the PAE plot, and pTM and ipTM scores.
5. Download the zip. It contains mmCIF files for any viewer.

pTM scores the overall fold. ipTM scores the contacts between chains. As a rough guide, ipTM above 0.8 is a confident complex, and below 0.6 the predicted binding is probably wrong.

MQIFVKTLTGKTITLEVEPSDTIENVKAKIQDKEGIPPDQQRLIFAGKQLEDGRTLSDYNIQKESTLHLVLRLRGG

The full workflow, A to Z

Lecture 09 · 23:07

This longer lecture does it the way you would for a thesis. I start with a sequence I know nothing about, find it, set up the job properly, use the advanced options (multiple copies of a chain, ligands, seeds), and interpret what comes back.

Next: Chapter 7 · How does AlphaFold work? The papers and the masterclass →

Running a tool doesn’t make you a researcher. Knowing how it works, and where it breaks, does. This chapter sends you to the original sources and then pulls everything together in the longest lecture of the series.

Assignment 3: the EMBL-EBI course and the Nature papers

Lecture 10 · 1:21

1. Complete the free EMBL-EBI AlphaFold online course.
2. Read the AlphaFold 3 paper: Abramson et al., “Accurate structure prediction of biomolecular interactions with AlphaFold 3”, Nature 2024.
3. Write down how the method works, where it fails, and what pLDDT and PAE mean in the authors’ own words.
4. Submit your summary as explained in the lecture.

For background, also skim the original paper: Jumper et al., “Highly accurate protein structure prediction with AlphaFold” (Nature, 2021). That is the work the 2024 Nobel Prize recognised. On a first read, don’t fight the methods section. Read the abstract, the figures and the limitations, then go back.

The complete AlphaFold masterclass

Lecture 11 · 1:33:38

This 93-minute session joins Chapters 3 to 6 and goes deeper into AlphaFold 3: how the architecture works, how to set up complex jobs, and how to interpret a difficult structure from start to finish. Watch it in two sittings with a viewer open, and repeat each step on your own protein.

Next: Chapter 8 · How do you find a protein’s function? GO, KEGG and COG →

Lecture 12 · 37:18

You have a sequence and a structure. Your supervisor’s next question will be: “So what does it do?” Three databases answer it from different angles.

Think of it this way: GO describes the job, KEGG shows the team the protein works in, and COG tells you its family tree. For known proteins, the UniProt entry already lists all three. For new sequences from your own genome or transcriptome, tools like eggNOG-mapper assign them in bulk.

Next: Chapter 9 · How do you predict protein function from structure? →

Lecture 13 · 27:32

Evolution changes sequences faster than shapes. Two proteins can share less than 20% of their sequence and still have the same fold and a similar job. So when a sequence search finds nothing, a structure search often still can.

• Structural search: upload your model to Foldseek (fast, searches the AlphaFold DB and PDB) or the DALI server. Hits with known functions give you a hypothesis.
• Pockets and active sites: look for clefts on the surface lined with conserved residues. Catalytic residues usually sit in high-pLDDT regions.
• Domains: match each domain to a family in InterPro.

Everything here is a hypothesis to test, not a result. A structural match tells you what a protein could do.

The whole pipeline in one picture

Lecture 14 · 3:32

This short lecture joins the course into one workflow you can reuse on any protein:

1. 1SequenceGet the FASTA from UniProt or your own data
2. 2StructureAlphaFold DB, AlphaFold Server or ColabFold
3. 3Check confidencepLDDT per residue, PAE between domains
4. 4CompareFoldseek or DALI against known structures
5. 5FunctionGO terms, KEGG pathways, COG groups

Need to run hundreds of proteins without the server’s daily limit? ColabFold runs AlphaFold 2 in Google Colab and is how many students get through large batches.

Next: Chapter 10 · Case study: what does AlphaFold show about p53? →

Lecture 15 · 13:55

p53 is a transcription factor that stops a damaged cell from dividing, or tells it to die. Its gene, TP53, is mutated in roughly half of all human cancers, which is why p53 is called the guardian of the genome. In the lecture we take it through every step of the course:

Next: Chapter 11 · What can’t AlphaFold do? →

Lecture 16 · 2:07

The last lecture looks ahead. After a new genome is assembled and its genes predicted, annotation normally compares sequences against databases. For many genes in less-studied organisms, that finds nothing. Predicting structures for those proteins and searching them with Foldseek can assign functions that sequence search misses. This is called structural annotation, and it is becoming a normal step in genome projects.

What AlphaFold can’t do

AlphaFold is a model, not an experiment. Keep these limits in mind before you write anything up:

• It predicts one static shape. Real proteins move, and many change shape when they bind something.
• Most single point mutations barely change the predicted structure, even when they destroy function. Don’t use AlphaFold alone to judge whether a mutation is harmful.
• Disordered regions come out as low-confidence spaghetti. That is information, not failure.
• Complexes with low ipTM, or ligands placed in poorly predicted pockets, can look convincing and still be wrong.
• For a key structure in a publication, experimental evidence (or an existing structure in the PDB) is still expected.

Glossary of AlphaFold and protein terms

Short definitions of every technical term used in this course

Amino acid

One of 20 building blocks that link together to form a protein chain.

Residue

A single amino acid within a protein chain. AlphaFold scores every residue.

FASTA

A plain-text format for sequences: a header line starting with “>”, then the letters.

UniProt

The main database of protein sequences and annotations, with IDs such as P04637.

PDB

The Protein Data Bank, the archive of experimentally determined 3D structures.

CASP

Critical Assessment of protein Structure Prediction, a blind test held every two years since 1994.

MSA

Multiple sequence alignment: related sequences lined up so evolutionary patterns can be read.

pLDDT

AlphaFold’s per-residue confidence score from 0 to 100. Above 90 is very high; below 50 is very low.

PAE

Predicted aligned error: AlphaFold’s expected position error between pairs of residues, used to judge domain arrangement.

pTM

Predicted TM-score: a 0–1 confidence for the overall fold of a model.

ipTM

Interface pTM: a 0–1 confidence for the contacts between chains in a complex. Above 0.8 is confident.

Domain

A compact part of a protein that folds on its own and often has its own function.

Intrinsically disordered region

A part of a protein with no single fixed shape. Usually shows very low pLDDT.

Foldseek

A fast tool that searches for proteins with similar 3D structures.

Gene Ontology (GO)

A standard vocabulary for protein function, biological process and cellular location.

All 16 lectures in one list

The full playlist, plus every lecture grouped by chapter

Frequently asked questions