AlphaFold — Conference-Ready Knowledge Guide

1. The Problem It Solved: The Protein Folding Problem

• Proteins are linear chains of amino acids that fold into precise 3D shapes determining their function.
• Predicting this 3D shape from sequence alone had been unsolved for 50+ years (since Anfinsen's 1972 Nobel Prize-winning insight that sequence determines structure).
• The Critical Assessment of Protein Structure Prediction (CASP) competition benchmarks progress every 2 years. Before AlphaFold, top models achieved ~40–50 TM-score. AlphaFold2 hit ~92 median GDT score at CASP14 (2020) — a near-experimental-quality leap.
• As of October 2024, the Protein Data Bank (PDB) holds only ~226,000 experimentally determined structures — covering less than 0.1% of all known protein sequences.

2. The Three Generations

AlphaFold 1 (AF1) — CASP13, 2018

• Integrated deep learning with evolutionary covariance analysis (co-evolution of paired residues as a structural signal).
• Predicted interresidue distances using convolutional neural networks.
• Came 1st at CASP13 but still with significant gaps to experimental accuracy.

AlphaFold 2 (AF2) — CASP14, 2020 → Published Nature 2021

• Revolutionary architecture: the Evoformer — a deep attention-based transformer that jointly reasons over:
  • A Multiple Sequence Alignment (MSA) representation (evolutionary context across homologs)
  • A pair representation (pairwise residue relationships in 3D space)
• A Structure Module then builds 3D coordinates using invariant point attention.
• Key innovation: end-to-end differentiable — the whole pipeline trains together.
• Achieved near-experimental accuracy (median TM-score ~0.92) — described by CASP assessors as "a solution to the protein folding problem."
• 2024 Nobel Prize in Chemistry awarded to John Jumper (DeepMind) and Demis Hassabis for AF2 (shared with David Baker for de novo protein design).

AlphaFold 3 (AF3) — Nature, May 2024

• Built by Google DeepMind + Isomorphic Labs.
• Core change: replaced the structure module with a diffusion model (similar to image diffusion like DALL-E).
• Extends beyond proteins to multi-molecular complexes:
  • Protein–protein
  • Protein–DNA/RNA
  • Protein–ligand (small molecules)
  • Post-translational modifications (glycosylation, phosphorylation)
• Also introduced a Pairformer (lighter replacement for Evoformer).
• Controversy: initial release was server-only, with restricted weights — criticized for limiting open science (RoseTTAFold All-Atom released open-source around the same time).
• Later, weights were partially released for non-commercial academic use.

3. The AlphaFold Database (AFDB)

• Launched 2021 jointly with EMBL-EBI.
• 2022: expanded to 200 million+ structures covering virtually the entire UniProt.
• 2024 update: 214 million protein sequences covered, enhanced PAE (Predicted Aligned Error) viewer, Google Cloud Public Dataset access, improved APIs. [PMID: 37933859]
• 2025 redesign: updated interface, isoform coverage, underlying MSA access, improved APIs. [PMID: 41273079]
• Structures include a pLDDT score (per-residue confidence, 0–100) and PAE (predicted error between residue pairs, useful for domain boundary identification).

4. Key Technical Concepts to Know

5. Applications

Drug Discovery

• Enables structure-based virtual screening without needing experimental crystallography.
• Used to predict binding pockets for target proteins with no known structure.
• Protein–ligand docking in AF3 accelerates early-stage drug design.
• Companies like Isomorphic Labs, Schrödinger, and Insilico Medicine now routinely use AF predictions.

Fundamental Biology

• Mapping entire proteomes (structuromes) — organisms whose proteins were structurally unknown are now fully modeled.
• Studying orphan proteins (no known homolog structure) — now tractable.
• Understanding disease mutations: predicting how SNPs/missense mutations destabilize protein structure (e.g., TREM2 R47H in Alzheimer's disease).

Synthetic Biology & Protein Design

• Inverse use: AlphaDesign — using AF as a fitness oracle to guide de novo protein design.
• RFdiffusion, ProteinMPNN combined with AF2 for protein engineering pipelines.
• Designing novel enzymes, binders, and vaccines.

Structural Bioinformatics

• Foldseek: fast protein structure search (Nature Biotechnology 2024) enabled by AF structures.
• Clustering the entire known protein universe by structure (Nature 2023).
• Function prediction from predicted structures — binding site prediction at proteome scale.

6. Limitations (Critical for Balanced Conference Discussion)

1. Disordered regions (IDRs): Intrinsically disordered proteins/regions score low pLDDT but AF gives a static structure — can be misleading.
2. Conformational flexibility: Predicts one conformation (energy minimum), not the full ensemble. Cannot capture conformational states or allosteric changes easily (though workarounds like MSA subsampling exist).
3. Novel folds: Proteins with few evolutionary homologs (small MSA) = reduced accuracy. Truly novel folds remain challenging.
4. Complexes (AF2): AF-Multimer has known failure modes for some homodimers (e.g., flavivirus NS4B) and large assemblies.
5. Dynamics: Static snapshot — not a substitute for MD simulations or NMR for understanding motion.
6. AF3 access: Initially restricted weights — ongoing debate about open vs. closed AI in science.
7. Template-based cases: Traditional homology modeling can still outperform AF3 for some targets with very close experimental templates (e.g., insect odorant receptors).

7. Competing & Complementary Tools

8. The Nobel Prize Context (2024)

• ½ to David Baker (University of Washington) — computational protein design
• ½ to Demis Hassabis & John Jumper (Google DeepMind) — AlphaFold

9. Quick Stats for Impact

• AF2 paper (Jumper et al., Nature 2021) — one of the most cited biology papers of the decade.
• AF3 paper (Abramson et al., Nature 2024) — protein–ligand docking accuracy comparable to physics-based tools.
• AFDB covers structures for organisms including humans, E. coli, A. thaliana, and essentially every major model organism.
• Structural biology output that took the entire PDB 50 years to accumulate was matched in scale within 18 months of AFDB launch.

10. Likely Conference Questions & Talking Points

• "Is the protein folding problem solved?" → Largely for single-chain static structures, yes. But folding dynamics, disordered proteins, and context-dependent conformations remain open.
• "What's the difference between AF2 and AF3?" → AF2 = Evoformer + proteins only; AF3 = diffusion model + multi-molecular complexes.
• "Can you trust the structures?" → Use pLDDT and PAE as guides. High-confidence regions are highly reliable; low-confidence regions need experimental validation.
• "How does it affect wet lab work?" → Doesn't replace experiments for dynamics/interactions, but massively reduces the need for X-ray crystallography/cryo-EM as a first step.

• Jumper et al., Nature 2021 — AF2 original paper
• Abramson et al., Nature 2024 — AF3 original paper (DOI: 10.1038/s41586-024-07487-w)
• Varadi et al., Nucleic Acids Res 2024 — AFDB 2024 update [PMID: 37933859]
• Krokidis et al., Int J Mol Sci 2025 — AF3 applications review [PMID: 40332289]

How to Simply Explain AlphaFold

Step 1 — Start with the Problem (30 seconds)

"Every protein in your body is a string of amino acids — like beads on a necklace. But that string folds into a very specific 3D shape, and that shape determines what the protein does. If we know the shape, we can understand disease, design drugs, and engineer biology. The problem? Figuring out that 3D shape experimentally takes months and costs tens of thousands of dollars. And we have millions of proteins we've never solved."

Step 2 — The Core Idea (1 minute)

"AlphaFold asks: can we predict the 3D shape just from the sequence of amino acids — using AI?"

"Think of it like autocomplete on your phone. Your phone has seen millions of sentences and learned the patterns. AlphaFold has seen millions of protein structures and learned the folding patterns. You give it a new sequence — it predicts the shape."

Step 3 — How It Actually Works (2–3 minutes, the core explanation)

🔷 Step A — Collect Evolutionary Clues

"AlphaFold doesn't just look at your protein. It searches databases for thousands of related proteins from other species — proteins that evolved from the same ancestor. This is called a Multiple Sequence Alignment (MSA)."

"The key insight: if two amino acids in a protein always change together across evolution — when one mutates, the other mutates too — they're probably physically touching in 3D space. Evolution is telling us about the structure."

🔷 Step B — Run It Through the AI (Evoformer)

"All this evolutionary information is fed into AlphaFold's neural network — called the Evoformer. It's a type of attention-based model (similar to what powers ChatGPT) that reasons about: which amino acids are close to each other? Which are far? What angles do they form?"

"Imagine you have 300 people standing in a room and you're trying to figure out who knows who. The Evoformer is continuously asking every person 'who are you most connected to?' and updating everyone's understanding of the whole network — simultaneously, many times over."

🔷 Step C — Build the 3D Structure

"Once AlphaFold has a confident map of which residues interact with which, it assembles the 3D coordinates — placing every atom in space. It even tells you how confident it is at each position using a score called pLDDT (0–100). High score = reliable. Low score = that region is probably floppy or disordered."

Step 4 — What Changed with AlphaFold 3 (30 seconds)

"The latest version, AlphaFold 3, uses a diffusion model — the same type of AI that generates images like DALL-E. Instead of just folding proteins, it can now model proteins bound to DNA, RNA, and small drug molecules all at once. This is a big deal for drug discovery."

Step 5 — The Punchline (30 seconds)

"In 2021, AlphaFold predicted accurate structures for over 200 million proteins in a matter of months — more than the entire 50-year history of experimental biology combined. It won the 2024 Nobel Prize in Chemistry. The protein folding problem, which stumped science for half a century, was essentially solved by a neural network."

One-Slide Visual Flow (for your slides)

Amino acid sequence
        ↓
Search for evolutionary relatives (MSA)
        ↓
Evoformer AI: "Which residues are close? Which interact?"
        ↓
Structure Module: Place every atom in 3D space
        ↓
Output: 3D structure + confidence score (pLDDT)

Audience-Specific Tweaks

The One-Liner to Remember

"AlphaFold reads a protein's evolutionary history like a blueprint, runs it through a transformer AI, and produces a 3D structure in minutes — for free — that used to take years and cost a fortune."

Full Conference Script — How AlphaFold Works

OPENING — Hook the Audience

"Let me start with a question.

How many of you have heard of the protein folding problem?

(pause for show of hands)

For those who haven't — this was one of the greatest unsolved problems in all of science. It sat unanswered for over 50 years. Thousands of researchers dedicated their entire careers to it. And then in 2020, a team of AI engineers at Google DeepMind — not biologists, not chemists — solved it.

Their tool is called AlphaFold. And today I want to walk you through exactly how it works, why it matters, and why the scientific world called it — and I quote — 'a solution to a 50-year-old grand challenge in biology.'"

SECTION 1 — Setting the Stage: What is a Protein?

"Before we talk about AlphaFold, we need to understand what it's actually predicting. So let me take 60 seconds to explain proteins.

Every living cell — in your body, in a bacterium, in a plant — runs on proteins. Proteins are the machines of life. They carry oxygen in your blood, fight infections, digest your food, send signals between your neurons, and copy your DNA.

Now, a protein starts as a simple chain. A long string of smaller molecules called amino acids — there are 20 different types — linked together like beads on a necklace. A typical protein might have 300 to 1,000 of these beads in a specific order.

But here is the critical thing: that string doesn't stay flat. It folds. It twists and coils and bends into a very precise 3-dimensional shape. And that shape is everything. The shape determines the function. Change the shape — you change or destroy the function. This is why mutations cause disease. A single wrong amino acid can cause the protein to misfold, and that misfolded protein can cause cancer, Alzheimer's, cystic fibrosis.

So if we can look at a protein's sequence and know its shape — we can understand diseases, find drug targets, design new medicines. That is the prize."

SECTION 2 — The Problem: Why Was This So Hard?

"So why did this take 50 years?

The challenge is this: we can easily read a protein's sequence using DNA sequencing — it's fast and cheap. But determining the 3D shape experimentally? That requires techniques like X-ray crystallography, cryo-electron microscopy, or NMR spectroscopy. These methods take months to years per protein. They require specialized equipment. They cost tens of thousands of dollars per structure. And some proteins simply refuse to cooperate with these methods.

As a result, by 2020 — after 50 years of effort — the global Protein Data Bank held about 170,000 experimentally solved protein structures. One hundred and seventy thousand. That sounds like a lot, until you realize there are an estimated 200 to 400 million unique protein sequences in nature.

We had solved less than 0.1 percent of the problem.

The gap between what we know at the sequence level and what we know at the structural level — that gap was enormous. And it was blocking biology, blocking drug discovery, blocking our understanding of life itself.

That gap is what AlphaFold was built to close."

SECTION 3 — The AlphaFold Approach: The Big Idea

"So how does AlphaFold approach this?

The core idea is elegant. And it starts with a profound insight from evolutionary biology.

Here it is: evolution is a record of what works.

When a protein is critical to life, it gets preserved across millions of years of evolution. You and a yeast cell share proteins that are 40% identical — because that protein is so important that evolution couldn't afford to change it much. But it does change a little. Mutations happen. Some mutations are tolerated — the protein still folds and works. Others are lethal — the protein misfolds, the organism dies.

Now here's the genius part. If you look across thousands of species and compare the same protein — you'll notice something. Certain pairs of amino acids always change together. When position 47 mutates from one amino acid to another, position 83 almost always mutates at the same time. Why? Because positions 47 and 83 are physically touching each other in the 3D structure. If you change one without changing the other, the protein breaks. Evolution has been quietly telling us about the 3D structure all along — we just needed a way to listen.

AlphaFold listens."

SECTION 4 — Step-by-Step: How AlphaFold Works

"Let me now walk you through the actual pipeline — step by step.

Step 1 — Input: Give AlphaFold a Sequence

"You start by giving AlphaFold a protein sequence. Just a string of letters — each letter representing one of the 20 amino acids. That's it. That's your only required input.

For example: MKTAYIAKQRQISFVKSHFSRQLEERLGLIEVQAPILSRVGDGTQDNLSGAEKAVQVKVKALPDAQFEVVHSLAKWKRQTLGQHDFSAGEGLYTHMKALRPDEDRLSPLHSVYVDQWDWERVMGDGERQFSTLKSTVEAIWAGIKATEAAVSEEFGLAPFLPDQIHFVHSQELLSRYPDLDAKGRERAIAKDLGAVFLVGIGGKLSDGHRHDVRAPDYDDWSTPSELGHAGLNGDILVWNPRLYANRFNKSGAFKTSQKGLQHGLPVSTDLDAWFKLK

That string of letters is all you give it. What comes out is a 3-dimensional molecular structure."

Step 2 — MSA: Search for Evolutionary Relatives

"AlphaFold's first job is to search massive protein databases — UniRef, UniClust, BFD — and find every protein from every species that is related to your query protein. This search produces what's called a Multiple Sequence Alignment, or MSA.

Think of the MSA as a table. Each row is a different organism's version of your protein. Each column is a position in the sequence. Now you can look down each column and ask: which positions vary together? Which positions are correlated?

This correlation pattern is the evolutionary fingerprint of the 3D structure. AlphaFold turns this into a numerical representation — essentially a two-dimensional map of all pairwise relationships between amino acids."

Step 3 — The Evoformer: The AI Brain

"Now comes the most sophisticated part — the neural network at the heart of AlphaFold 2, called the Evoformer.

The Evoformer is a type of transformer — the same fundamental architecture behind large language models like GPT. If you've used ChatGPT, you've interacted with a transformer. But instead of processing words and sentences, the Evoformer processes amino acids and their relationships.

It takes two inputs simultaneously:

• The MSA representation — rows of related sequences (the evolutionary context)
• The pair representation — a matrix capturing how every amino acid relates to every other amino acid

These two inputs talk to each other repeatedly — the Evoformer passes information back and forth between them, updating and refining its understanding in 48 consecutive layers. Each layer is asking: given everything I know, which amino acids are close in space? Which are far? What's the geometry?

By the end of the Evoformer, AlphaFold has built a very detailed internal model of the protein's spatial relationships — even before it's placed a single atom in 3D space.

A simple analogy: imagine you're trying to figure out the layout of a building, but you can't go inside. Instead, you interview thousands of people who've been in similar buildings. The Evoformer is the interview process — gathering clues, cross-referencing, refining — until you have a confident mental map of the interior."

Step 4 — Structure Module: Building the 3D Shape

"Once the Evoformer has its refined representation, it passes everything to the Structure Module.

The Structure Module places every atom in 3D space. It uses a technique called Invariant Point Attention — which means it builds the structure in a way that doesn't depend on arbitrary orientation. Rotate the protein, flip it, mirror it — the physics don't change, and AlphaFold's answer doesn't change either.

It builds the backbone first — the main chain of the protein — and then adds the side chains of each amino acid, placing every carbon, nitrogen, oxygen, and sulfur atom.

This step runs multiple times — the structure feeds back into the representation, the representation refines the structure — three times over, in what AlphaFold calls recycling. Each cycle, the model corrects itself. Like an artist sketching a portrait, then stepping back, then refining, then stepping back again.

The final output is a PDB file — a list of 3D coordinates for every atom in the protein — identical in format to what you'd get from X-ray crystallography."

Step 5 — Confidence Scores

"Critically, AlphaFold also tells you how much to trust each part of its prediction.

Every residue in the structure gets a pLDDT score — which stands for predicted Local Distance Difference Test — on a scale of 0 to 100.

• Above 90: Very high confidence. You can treat this like an experimental structure.
• 70 to 90: Good confidence. Generally reliable.
• 50 to 70: Low confidence. Use with caution — validate experimentally.
• Below 50: This region is likely intrinsically disordered — it probably doesn't have a fixed shape in real life.

This is visualized in color — blue for high confidence, red for low. When you look at an AlphaFold structure and see a floppy red tail, AlphaFold is telling you: 'I'm not certain here, and that might be biologically meaningful — this region may not have a fixed structure.'

There's also a second confidence metric called PAE — Predicted Aligned Error — which tells you the uncertainty in the relative position between any two residues. This is especially useful when you're looking at multi-domain proteins — it tells you whether two domains have a confident relative orientation, or whether they're flexibly connected."

SECTION 5 — AlphaFold 3: The Next Evolution

"In May 2024, Google DeepMind published AlphaFold 3. And it represents a fundamental architectural shift.

AlphaFold 2 was built for single proteins. AlphaFold 3 was built for molecular complexes — proteins interacting with other proteins, with DNA, with RNA, with small molecule drugs, with post-translational modifications like sugars and phosphate groups.

The technical change is significant: AlphaFold 3 replaces the Structure Module with a diffusion model — the same class of AI that powers image generators like DALL-E and Stable Diffusion.

Here's the intuition: a diffusion model starts with pure noise — random atomic positions — and then gradually denoises, step by step, guided by the model's learned understanding of molecular physics, until a coherent structure emerges.

Why is this powerful? Because it handles chemical diversity. Small molecules, nucleic acids, ions, modified residues — these are all fundamentally different chemistry. The diffusion approach can handle all of them in a unified framework, where the Evoformer-based approach struggled.

For drug discovery, this is transformative. You can now predict how a drug molecule docks inside a protein's binding pocket — without needing to crystallize the complex."

SECTION 6 — The Impact

"Let me put the impact in perspective.

AlphaFold DB — the public database of AlphaFold-predicted structures — launched in 2021 with 350,000 structures. By 2022, it had 200 million. Today it covers over 214 million protein sequences — essentially the entire known protein universe.

Fifty years of experimental structural biology produced 170,000 structures. AlphaFold produced 200 million in 18 months. For free. Accessible to anyone on the planet with an internet connection.

Researchers are now using AlphaFold to:

• Identify drug targets for neglected tropical diseases where no structural data existed before
• Understand how cancer mutations destabilize tumor suppressor proteins
• Design novel enzymes that don't exist in nature
• Predict antibody-antigen interactions for vaccine design
• Reannotate entire genomes of organisms we barely understand

In October 2024, the Nobel Committee recognized this by awarding the Nobel Prize in Chemistry — half to Demis Hassabis and John Jumper for AlphaFold, and half to David Baker for computational protein design. It was an explicit Nobel Prize for artificial intelligence applied to fundamental science."

SECTION 7 — Honest Limitations

"I want to be balanced here, because AlphaFold is not magic.

It predicts static structures — a single snapshot. Real proteins are dynamic. They flex, breathe, change shape when they bind a partner, transition between active and inactive forms. AlphaFold gives you the ground-state minimum energy structure. It doesn't give you the movie — just one frame.

Intrinsically disordered proteins — which make up perhaps 30% of the human proteome and are heavily involved in cancer and neurodegeneration — by definition don't have a fixed structure. AlphaFold's predictions for these regions are low confidence, and rightly so. But this means a huge swath of biology is still structurally inaccessible computationally.

Proteins with few evolutionary relatives — novel proteins, synthetic proteins, proteins from poorly sequenced organisms — give AlphaFold less evolutionary signal to work with, and accuracy drops.

And AlphaFold is not a replacement for experiment. It's a starting point. A hypothesis. The pLDDT score tells you how much to trust it, but critical conclusions still need experimental validation.

The tool is extraordinary. But it is a tool, not an oracle."

CLOSING — The One Takeaway

"So let me leave you with this.

For 50 years, the question was: can we read a protein's sequence and know its shape?

AlphaFold answered: yes. By listening to the language of evolution, processing it through a deep transformer AI, and assembling atoms in 3D space — it produces structures of near-experimental quality, in minutes, for any protein on Earth.

It didn't just accelerate structural biology. It democratized it. A researcher in a university with no crystallography lab, no cryo-EM, no million-dollar equipment budget — can now query the structure of any protein in the world in seconds, for free.

That is the significance of AlphaFold. Not just what it solved — but what it opened up.

Thank you."

Quick Reference — Script Timing

Status

cd /home/daytona/workspace/alphafold-conference

# Replace ELLIPSE with the correct shape name
sed -i 's/pres\.shapes\.ELLIPSE/pres.ShapeType ? pres.shapes.OVAL : "ellipse"/g' generate_presentation.js

# Or simply replace all ELLIPSE with RECTANGLE (simpler fix)
python3 -c "
with open('generate_presentation.js','r') as f:
    code = f.read()
code = code.replace('pres.shapes.ELLIPSE', 'pres.shapes.RECTANGLE')
with open('generate_presentation.js','w') as f:
    f.write(code)
"

node generate_presentation.js

What the Presentation Contains (18 slides, fully designed)