make ppt of alphafold in depth details for seminar presentation and use standard sources
Reading File
Searching the Web
AlphaFold protein structure prediction deep learning mechanism 2024
Searching PubMed
AlphaFold protein structure prediction
Reading File
Searching Images
AlphaFold protein structure prediction neural network architecture
Reading Web Page
https://en.wikipedia.org/wiki/AlphaFold
Running Command
ls -la /home/daytona/workspace/ 2>/dev/null || echo "workspace empty"
Running Command
mkdir -p /home/daytona/workspace/alphafold-ppt
Writing File
~/alphafold-ppt/build.js
const pptxgen = require("pptxgenjs");
const pres = new pptxgen();
pres.layout = "LAYOUT_WIDE"; // 13.3" x 7.5"
pres.author = "Seminar Presentation";
pres.title = "AlphaFold: Revolutionizing Protein Structure Prediction";
// ─── COLOUR PALETTE ───
const C = {
deepBlue: "0A1628", // dominant dark BG
midBlue: "0D2137", // card BG
accentTeal: "00C8C8", // primary accent
accentGold: "FFB347", // secondary accent / Nobel
accentGreen: "38E07B", // success / accuracy
white: "FFFFFF",
lightGrey: "C8D8E8",
mutedBlue: "4A7BA8",
red: "FF5C5C",
};
// ─── HELPER: section header bar ───
function addHeaderBar(slide, text, subtitle) {
// Full-width gradient bar via rectangle
slide.addShape(pres.ShapeType.rect, {
x: 0, y: 0, w: 13.3, h: 1.35,
fill: { type: "solid", color: C.midBlue },
line: { width: 0 },
});
// Accent left stripe
slide.addShape(pres.ShapeType.rect, {
x: 0, y: 0, w: 0.18, h: 1.35,
fill: { color: C.accentTeal },
line: { width: 0 },
});
slide.addText(text, {
x: 0.38, y: 0.12, w: 12.5, h: 0.6,
fontSize: 26, bold: true, color: C.white, fontFace: "Calibri",
margin: 0,
});
if (subtitle) {
slide.addText(subtitle, {
x: 0.38, y: 0.75, w: 12.5, h: 0.45,
fontSize: 13, color: C.accentTeal, fontFace: "Calibri",
margin: 0,
});
}
}
// ─── HELPER: info card ───
function addCard(slide, x, y, w, h, title, body, accent) {
const ac = accent || C.accentTeal;
slide.addShape(pres.ShapeType.rect, {
x, y, w, h,
fill: { color: C.midBlue },
line: { color: ac, width: 1.5 },
shadow: { type: "outer", color: "000000", opacity: 0.3, blur: 6, offset: 3, angle: 45 },
});
// top accent strip
slide.addShape(pres.ShapeType.rect, {
x, y, w, h: 0.07,
fill: { color: ac },
line: { width: 0 },
});
if (title) {
slide.addText(title, {
x: x + 0.15, y: y + 0.12, w: w - 0.3, h: 0.4,
fontSize: 13, bold: true, color: ac, fontFace: "Calibri",
margin: 0,
});
}
if (body) {
slide.addText(body, {
x: x + 0.15, y: y + (title ? 0.56 : 0.2), w: w - 0.3, h: h - (title ? 0.65 : 0.3),
fontSize: 11.5, color: C.lightGrey, fontFace: "Calibri",
valign: "top", margin: 0,
});
}
}
// ─── HELPER: bullet block ───
function addBullets(slide, items, x, y, w, h, fontSize) {
const fs = fontSize || 13;
const arr = items.map((t, i) => ({
text: t,
options: { bullet: { type: "bullet", characterCode: "25B6" }, color: C.lightGrey, fontSize: fs, breakLine: i < items.length - 1 },
}));
slide.addText(arr, { x, y, w, h, fontFace: "Calibri", valign: "top" });
}
// ─── HELPER: numbered list ───
function addNumbered(slide, items, x, y, w, h, fs) {
const fontSize = fs || 12.5;
const arr = items.map((t, i) => ({
text: `${i + 1}. ${t}`,
options: { color: C.lightGrey, fontSize, bold: false, breakLine: i < items.length - 1 },
}));
slide.addText(arr, { x, y, w, h, fontFace: "Calibri", valign: "top" });
}
// ─────────────────────────────────────────────
// SLIDE 1 — TITLE SLIDE
// ─────────────────────────────────────────────
{
const s = pres.addSlide();
s.background = { color: C.deepBlue };
// DNA helix-like decorative circles
for (let i = 0; i < 18; i++) {
s.addShape(pres.ShapeType.ellipse, {
x: 10.5 + Math.sin(i * 0.7) * 1.2,
y: 0.1 + i * 0.4,
w: 0.28, h: 0.28,
fill: { color: i % 2 === 0 ? C.accentTeal : C.mutedBlue },
line: { width: 0 },
});
}
// Main title block
s.addShape(pres.ShapeType.rect, {
x: 0.6, y: 1.3, w: 9.4, h: 2.6,
fill: { color: C.midBlue },
line: { color: C.accentTeal, width: 2 },
});
s.addShape(pres.ShapeType.rect, {
x: 0.6, y: 1.3, w: 0.22, h: 2.6,
fill: { color: C.accentTeal },
line: { width: 0 },
});
s.addText("AlphaFold", {
x: 1.05, y: 1.45, w: 8.7, h: 0.85,
fontSize: 52, bold: true, color: C.accentTeal, fontFace: "Calibri",
margin: 0, charSpacing: 3,
});
s.addText("Revolutionizing Protein Structure Prediction with AI", {
x: 1.05, y: 2.3, w: 8.7, h: 0.65,
fontSize: 20, color: C.white, fontFace: "Calibri",
margin: 0,
});
s.addText("From the Protein Folding Problem to the Nobel Prize in Chemistry 2024", {
x: 1.05, y: 2.95, w: 8.7, h: 0.5,
fontSize: 13, color: C.accentGold, fontFace: "Calibri", italic: true,
margin: 0,
});
// Badges row
const badges = [
{ label: "DeepMind / Alphabet", color: C.mutedBlue },
{ label: "Nobel Prize 2024", color: C.accentGold },
{ label: "214M+ Structures", color: C.accentGreen },
{ label: "Open Source", color: C.accentTeal },
];
badges.forEach((b, i) => {
s.addShape(pres.ShapeType.roundRect, {
x: 0.6 + i * 3.0, y: 4.25, w: 2.6, h: 0.48,
fill: { color: C.deepBlue },
line: { color: b.color, width: 1.5 },
rectRadius: 0.1,
});
s.addText(b.label, {
x: 0.6 + i * 3.0, y: 4.25, w: 2.6, h: 0.48,
fontSize: 11, bold: true, color: b.color, fontFace: "Calibri",
align: "center", valign: "middle",
});
});
// Footer
s.addText("Seminar Presentation • May 2026 • Sources: Nature, DeepMind, Nobel Committee, PubMed", {
x: 0, y: 7.0, w: 13.3, h: 0.4,
fontSize: 9.5, color: C.mutedBlue, align: "center", fontFace: "Calibri",
});
}
// ─────────────────────────────────────────────
// SLIDE 2 — AGENDA
// ─────────────────────────────────────────────
{
const s = pres.addSlide();
s.background = { color: C.deepBlue };
addHeaderBar(s, "Presentation Outline", "A structured deep-dive into AlphaFold");
const topics = [
["01", "The Protein Folding Problem", "50 years of an unsolved biological challenge"],
["02", "History & CASP Benchmarks", "AlphaFold 1 (2018) → AF2 (2020) → AF3 (2024)"],
["03", "AlphaFold 2 Architecture", "Evoformer, MSA, Pairwise Representations"],
["04", "Training & Data Pipeline", "PDB, UniRef, BFD, model confidence (pLDDT)"],
["05", "AlphaFold 3", "Pairformer, diffusion model, DNA/RNA/ligands"],
["06", "Applications & Impact", "Drug discovery, disease research, structural biology"],
["07", "AlphaFold DB", "214M+ structures, global access, limitations"],
["08", "Nobel Prize 2024", "Hassabis, Jumper & Baker"],
["09", "Limitations & Criticisms", "Dynamics, disorder, hallucinations"],
["10", "Future Directions", "AlphaFold + experiments, protein design"],
];
topics.forEach(([num, title, sub], i) => {
const col = i < 5 ? 0 : 1;
const row = i % 5;
const x = 0.45 + col * 6.4;
const y = 1.5 + row * 1.15;
s.addShape(pres.ShapeType.rect, {
x, y, w: 6.0, h: 0.95,
fill: { color: C.midBlue },
line: { color: i % 2 === 0 ? C.accentTeal : C.mutedBlue, width: 1 },
});
s.addText(num, {
x: x + 0.1, y: y + 0.05, w: 0.6, h: 0.85,
fontSize: 22, bold: true, color: i % 2 === 0 ? C.accentTeal : C.mutedBlue,
fontFace: "Calibri", align: "center", valign: "middle", margin: 0,
});
s.addText(title, {
x: x + 0.75, y: y + 0.05, w: 5.1, h: 0.4,
fontSize: 12.5, bold: true, color: C.white, fontFace: "Calibri", margin: 0,
});
s.addText(sub, {
x: x + 0.75, y: y + 0.48, w: 5.1, h: 0.35,
fontSize: 10, color: C.lightGrey, fontFace: "Calibri", margin: 0,
});
});
}
// ─────────────────────────────────────────────
// SLIDE 3 — THE PROTEIN FOLDING PROBLEM
// ─────────────────────────────────────────────
{
const s = pres.addSlide();
s.background = { color: C.deepBlue };
addHeaderBar(s, "The Protein Folding Problem", "A 50-year grand challenge in molecular biology");
// Left column — explanatory text
addCard(s, 0.35, 1.55, 5.8, 2.3, "What Is the Protein Folding Problem?",
"Proteins are chains of amino acids that fold into precise 3D shapes to perform biological functions. " +
"The sequence of amino acids (primary structure) ultimately determines the final 3D conformation — " +
"but predicting that 3D shape computationally from the sequence alone remained unsolved for ~50 years.", C.accentTeal);
addCard(s, 0.35, 4.0, 5.8, 2.5, "Why Is It Hard? (Levinthal's Paradox)",
"In 1969, Cyrus Levinthal showed that if a 100-residue protein sampled all possible conformations at " +
"random, it would take longer than the age of the universe — yet proteins fold in microseconds. " +
"This implies folding follows guided pathways, not random search.", C.red);
// Right column — key facts boxes
const facts = [
["~20,000", "Human protein-coding genes"],
["~100,000", "Distinct human protein isoforms"],
["~200,000", "Structures in PDB (experimental)"],
["~50 years", "Duration of the folding problem"],
["Anfinsen's Dogma (1972)", "Sequence determines structure (Nobel Prize)"],
];
facts.forEach(([val, lbl], i) => {
const y = 1.55 + i * 1.2;
s.addShape(pres.ShapeType.rect, {
x: 6.55, y, w: 6.35, h: 0.95,
fill: { color: C.midBlue },
line: { color: i < 3 ? C.accentGreen : C.accentGold, width: 1.2 },
});
s.addText(val, {
x: 6.65, y: y + 0.06, w: 2.4, h: 0.8,
fontSize: 18, bold: true, color: i < 3 ? C.accentGreen : C.accentGold,
fontFace: "Calibri", valign: "middle", margin: 0,
});
s.addText(lbl, {
x: 9.1, y: y + 0.06, w: 3.6, h: 0.8,
fontSize: 11.5, color: C.lightGrey,
fontFace: "Calibri", valign: "middle", margin: 0,
});
});
}
// ─────────────────────────────────────────────
// SLIDE 4 — HISTORY & CASP
// ─────────────────────────────────────────────
{
const s = pres.addSlide();
s.background = { color: C.deepBlue };
addHeaderBar(s, "History & CASP Benchmarks", "The evolution of computational protein structure prediction");
// Timeline line
s.addShape(pres.ShapeType.line, {
x: 0.6, y: 3.5, w: 12.1, h: 0,
line: { color: C.mutedBlue, width: 2, dashType: "dash" },
});
const events = [
{ year: "1994", label: "CASP\nFounded", sub: "biennial blind\nprediction challenge", color: C.mutedBlue, x: 0.3 },
{ year: "2018", label: "AlphaFold 1\n(CASP13)", sub: "GDT ~68.5\nbest free-modeling", color: C.accentTeal, x: 2.8 },
{ year: "2020", label: "AlphaFold 2\n(CASP14)", sub: "GDT ~92.4\nnear-experimental", color: C.accentGreen, x: 5.3 },
{ year: "2021", label: "AF-Multimer\n& Open DB", sub: "98% UniProt\n~350K structures", color: C.accentTeal, x: 7.8 },
{ year: "2022", label: "AF DB\n214M+ structures", sub: "Full UniProt\npredictions", color: C.accentGold, x: 9.5 },
{ year: "2024", label: "AlphaFold 3\n& Nobel Prize", sub: "DNA/RNA/ligands\nHassabis & Jumper", color: C.accentGold, x: 11.3 },
];
events.forEach(ev => {
// Dot
s.addShape(pres.ShapeType.ellipse, {
x: ev.x + 0.55, y: 3.33, w: 0.32, h: 0.32,
fill: { color: ev.color },
line: { width: 0 },
});
// Year label above
s.addText(ev.year, {
x: ev.x, y: 2.85, w: 1.5, h: 0.38,
fontSize: 13, bold: true, color: ev.color, fontFace: "Calibri",
align: "center", margin: 0,
});
// Label below
s.addText(ev.label, {
x: ev.x, y: 3.75, w: 1.6, h: 0.55,
fontSize: 10.5, bold: true, color: C.white, fontFace: "Calibri",
align: "center", margin: 0,
});
s.addText(ev.sub, {
x: ev.x, y: 4.3, w: 1.6, h: 0.55,
fontSize: 9.5, color: C.lightGrey, fontFace: "Calibri",
align: "center", margin: 0,
});
});
// CASP GDT improvement bar chart (simple rectangles)
s.addText("CASP GDT Score Progression (Best Free-Modeling)", {
x: 0.35, y: 5.1, w: 12.6, h: 0.35,
fontSize: 12, bold: true, color: C.accentTeal, fontFace: "Calibri", margin: 0,
});
const bars = [
{ label: "CASP11\n2014", gdt: 40, color: C.mutedBlue },
{ label: "CASP12\n2016", gdt: 45, color: C.mutedBlue },
{ label: "CASP13\n2018 (AF1)", gdt: 68.5, color: C.accentTeal },
{ label: "CASP14\n2020 (AF2)", gdt: 92.4, color: C.accentGreen },
];
const maxH = 1.2;
bars.forEach((b, i) => {
const barH = (b.gdt / 100) * maxH;
const x = 1.5 + i * 2.8;
const yBase = 7.15;
s.addShape(pres.ShapeType.rect, {
x, y: yBase - barH, w: 1.8, h: barH,
fill: { color: b.color },
line: { width: 0 },
});
s.addText(`${b.gdt}`, {
x, y: yBase - barH - 0.3, w: 1.8, h: 0.28,
fontSize: 11, bold: true, color: b.color, fontFace: "Calibri", align: "center", margin: 0,
});
s.addText(b.label, {
x: x - 0.1, y: 7.15, w: 2.0, h: 0.3,
fontSize: 9, color: C.lightGrey, fontFace: "Calibri", align: "center", margin: 0,
});
});
s.addText("Source: CASP – Critical Assessment of Protein Structure Prediction; Jumper et al., Nature 2021", {
x: 0.3, y: 7.3, w: 12.7, h: 0.2,
fontSize: 8, color: C.mutedBlue, fontFace: "Calibri", margin: 0,
});
}
// ─────────────────────────────────────────────
// SLIDE 5 — ALPHAFOLD 2 ARCHITECTURE
// ─────────────────────────────────────────────
{
const s = pres.addSlide();
s.background = { color: C.deepBlue };
addHeaderBar(s, "AlphaFold 2 Architecture", "Evoformer + Structure Module — end-to-end differentiable deep learning");
// Main flow boxes
const pipeline = [
{ label: "Input:\nAmino Acid\nSequence", color: C.mutedBlue, x: 0.3 },
{ label: "MSA\nConstruction\n(Multiple Seq. Align.)", color: C.accentTeal, x: 2.45 },
{ label: "Evoformer\nBlocks (×48)\nPair+MSA Representation", color: C.accentGreen, x: 4.95 },
{ label: "Structure\nModule\n(IPA Layers)", color: C.accentTeal, x: 7.65 },
{ label: "Recycling\n(3 Iterations)\nRefinement", color: C.mutedBlue, x: 10.15 },
];
pipeline.forEach((p, i) => {
s.addShape(pres.ShapeType.rect, {
x: p.x, y: 1.5, w: 2.0, h: 1.4,
fill: { color: C.midBlue },
line: { color: p.color, width: 2 },
});
s.addText(p.label, {
x: p.x + 0.08, y: 1.5, w: 1.84, h: 1.4,
fontSize: 10.5, bold: true, color: p.color, fontFace: "Calibri",
align: "center", valign: "middle", margin: 0,
});
if (i < pipeline.length - 1) {
s.addShape(pres.ShapeType.line, {
x: p.x + 2.02, y: 2.2, w: 0.42, h: 0,
line: { color: C.accentTeal, width: 2 },
});
}
});
// Output box
s.addShape(pres.ShapeType.rect, {
x: 10.15, y: 3.1, w: 2.0, h: 0.75,
fill: { color: C.midBlue },
line: { color: C.accentGold, width: 2 },
});
s.addText("3D Structure\n+ pLDDT Score", {
x: 10.15, y: 3.1, w: 2.0, h: 0.75,
fontSize: 10, bold: true, color: C.accentGold, fontFace: "Calibri",
align: "center", valign: "middle", margin: 0,
});
// Deep-dive cards
const cards = [
{
title: "Multiple Sequence Alignment (MSA)",
body: "Searches databases (UniRef, BFD, MGnify) for evolutionarily related sequences. " +
"Co-evolutionary patterns reveal which residue pairs are spatially close. " +
"Depth of MSA directly correlates with prediction accuracy.",
x: 0.3, y: 4.05, w: 4.1, accent: C.accentTeal,
},
{
title: "Evoformer Module",
body: "48 stacked blocks performing row- & column-wise attention over MSA and pairwise distance representation. " +
"Updates both representations simultaneously — amino acid relationships and spatial proximity signals " +
"inform each other iteratively. Key innovation over previous methods.",
x: 4.55, y: 4.05, w: 4.35, accent: C.accentGreen,
},
{
title: "Structure Module + IPA",
body: "Invariant Point Attention (IPA): attention mechanism that respects 3D geometry. " +
"Predicts backbone frames (rotation + translation) for each residue. " +
"End-to-end differentiable — FAPE loss (Frame Aligned Point Error) drives training.",
x: 9.05, y: 4.05, w: 4.0, accent: C.accentTeal,
},
];
cards.forEach(c => {
addCard(s, c.x, c.y, c.w, 2.95, c.title, c.body, c.accent);
});
s.addText("Reference: Jumper J et al. Highly accurate protein structure prediction with AlphaFold. Nature. 2021;596:583–589. PMID: 34265844", {
x: 0.3, y: 7.32, w: 12.7, h: 0.2,
fontSize: 8, color: C.mutedBlue, fontFace: "Calibri", margin: 0,
});
}
// ─────────────────────────────────────────────
// SLIDE 6 — TRAINING & CONFIDENCE
// ─────────────────────────────────────────────
{
const s = pres.addSlide();
s.background = { color: C.deepBlue };
addHeaderBar(s, "Training Pipeline & Confidence Metrics", "Data sources, loss functions, and pLDDT quality scores");
// Training data boxes
s.addText("Training Data Sources", {
x: 0.35, y: 1.55, w: 6.0, h: 0.38,
fontSize: 14, bold: true, color: C.accentTeal, fontFace: "Calibri", margin: 0,
});
const dataSrc = [
["PDB (Protein Data Bank)", "~170,000 experimental 3D structures used as ground truth"],
["UniRef90 / UniClust30", "Sequence databases for MSA construction"],
["BFD (Big Fantastic DB)", "2.2 billion sequence clusters from metagenomics"],
["MGnify", "Additional metagenomic protein sequences"],
["PDB70 Templates", "Structural templates for homology hints"],
];
dataSrc.forEach(([src, desc], i) => {
s.addShape(pres.ShapeType.rect, {
x: 0.35, y: 2.0 + i * 0.82, w: 6.0, h: 0.7,
fill: { color: C.midBlue },
line: { color: C.mutedBlue, width: 0.8 },
});
s.addText(src, {
x: 0.55, y: 2.05 + i * 0.82, w: 2.6, h: 0.58,
fontSize: 11.5, bold: true, color: C.accentGreen, fontFace: "Calibri", valign: "middle", margin: 0,
});
s.addText(desc, {
x: 3.2, y: 2.05 + i * 0.82, w: 3.0, h: 0.58,
fontSize: 10.5, color: C.lightGrey, fontFace: "Calibri", valign: "middle", margin: 0,
});
});
// pLDDT confidence scale (right side)
s.addText("pLDDT Confidence Score", {
x: 7.0, y: 1.55, w: 6.0, h: 0.38,
fontSize: 14, bold: true, color: C.accentTeal, fontFace: "Calibri", margin: 0,
});
s.addText("Per-residue Local Distance Difference Test (0–100)", {
x: 7.0, y: 1.95, w: 6.0, h: 0.3,
fontSize: 11, color: C.lightGrey, fontFace: "Calibri", margin: 0,
});
const plddt = [
{ range: "pLDDT > 90", label: "Very High Confidence", desc: "Suitable for binding site analysis, docking", color: C.accentGreen },
{ range: "70–90", label: "Confident", desc: "Reliable backbone — suitable for most analyses", color: "90EE90" },
{ range: "50–70", label: "Low Confidence", desc: "Use with caution; may be disordered", color: C.accentGold },
{ range: "< 50", label: "Very Low Confidence", desc: "'Spaghetti' regions — likely intrinsically disordered", color: C.red },
];
plddt.forEach((p, i) => {
const y = 2.4 + i * 1.15;
s.addShape(pres.ShapeType.rect, {
x: 7.0, y, w: 0.6, h: 0.95,
fill: { color: p.color },
line: { width: 0 },
});
s.addShape(pres.ShapeType.rect, {
x: 7.65, y, w: 5.35, h: 0.95,
fill: { color: C.midBlue },
line: { color: p.color, width: 1 },
});
s.addText(p.range, {
x: 7.7, y: y + 0.05, w: 5.2, h: 0.38,
fontSize: 12, bold: true, color: p.color, fontFace: "Calibri", margin: 0,
});
s.addText(`${p.label} — ${p.desc}`, {
x: 7.7, y: y + 0.45, w: 5.2, h: 0.4,
fontSize: 10, color: C.lightGrey, fontFace: "Calibri", margin: 0,
});
});
// Loss function note
addCard(s, 0.35, 6.25, 12.65, 0.85, null,
"Loss Function: FAPE (Frame Aligned Point Error) + torsion angle loss + distogram loss + LDDT head. " +
"Training used ~128 TPUv3 cores for several weeks. Self-distillation on PDB-70 templates improved generalization.",
C.mutedBlue);
s.addText("Source: Jumper et al., Nature 2021 | AlphaFold DB 2024, Varadi M et al., Nucleic Acids Res 2024 (PMC10767828)", {
x: 0.3, y: 7.32, w: 12.7, h: 0.2,
fontSize: 8, color: C.mutedBlue, fontFace: "Calibri", margin: 0,
});
}
// ─────────────────────────────────────────────
// SLIDE 7 — ALPHAFOLD 3
// ─────────────────────────────────────────────
{
const s = pres.addSlide();
s.background = { color: C.deepBlue };
addHeaderBar(s, "AlphaFold 3 (2024)", "Biomolecular interaction prediction — beyond single-chain proteins");
// Banner
s.addShape(pres.ShapeType.rect, {
x: 0.35, y: 1.5, w: 12.6, h: 0.6,
fill: { color: C.midBlue },
line: { color: C.accentGold, width: 1.5 },
});
s.addText(
"Published: Abramson J et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. " +
"Nature. 2024;630:493–500. | Co-developed: Google DeepMind & Isomorphic Labs",
{
x: 0.55, y: 1.5, w: 12.2, h: 0.6,
fontSize: 10.5, color: C.accentGold, fontFace: "Calibri", italic: true, valign: "middle", margin: 0,
}
);
// Key changes
const changes = [
{
title: "Pairformer (replaces Evoformer)",
body: "Simplified transformer architecture — processes pairwise representations. " +
"Drops MSA stack in favour of a single sequence representation with a learned template embedding.",
accent: C.accentTeal,
},
{
title: "Diffusion Module",
body: "Replaces the Structure Module with a generative diffusion model. " +
"Starts from a cloud of atoms and iteratively denoises guided by Pairformer output. " +
"Allows full-atom (all-atom) 3D prediction including ligand atoms.",
accent: C.accentGreen,
},
{
title: "Multi-Molecular Scope",
body: "Can model: proteins, nucleic acids (DNA + RNA), small-molecule ligands, ions, " +
"post-translational modifications (glycosylation, phosphorylation). " +
"≥50% improvement in accuracy over prior methods for protein–nucleic acid interfaces.",
accent: C.accentGold,
},
{
title: "Drug Discovery Focus",
body: "Developed in collaboration with Isomorphic Labs. " +
"Enables rational drug design by predicting how candidate molecules bind target proteins. " +
"Free AlphaFold Server for non-commercial research launched 2024.",
accent: C.mutedBlue,
},
];
changes.forEach((c, i) => {
const col = i % 2;
const row = Math.floor(i / 2);
addCard(s, 0.35 + col * 6.45, 2.25 + row * 2.5, 6.15, 2.3, c.title, c.body, c.accent);
});
s.addText("Source: Abramson J et al., Nature 2024 (PMID: 38594878) | AlphaFold Server: alphafoldserver.com", {
x: 0.3, y: 7.32, w: 12.7, h: 0.2,
fontSize: 8, color: C.mutedBlue, fontFace: "Calibri", margin: 0,
});
}
// ─────────────────────────────────────────────
// SLIDE 8 — APPLICATIONS & IMPACT
// ─────────────────────────────────────────────
{
const s = pres.addSlide();
s.background = { color: C.deepBlue };
addHeaderBar(s, "Applications & Scientific Impact", "Drug discovery, disease research, and structural biology revolution");
const apps = [
{
icon: "💊", title: "Drug Discovery",
body: "Identifies binding pockets in disease-related proteins. " +
"Accelerated lead compound identification for tuberculosis, Chagas disease, antibiotic resistance targets. " +
"Isomorphic Labs integrates AF3 into active drug pipelines.",
accent: C.accentGreen, x: 0.35, y: 1.5,
},
{
icon: "🧬", title: "Rare Genetic Diseases",
body: "Variant Effect Prediction: AlphaMissense (2023) classifies 71M missense variants. " +
"~89% of human variants classified as benign or pathogenic. " +
"Enables faster rare disease diagnosis and gene therapy target identification.",
accent: C.accentTeal, x: 6.85, y: 1.5,
},
{
icon: "🌱", title: "Structural Biology Acceleration",
body: "Nuclear pore complex (120 proteins) modelled computationally in months vs. decades experimentally. " +
"AlphaFold structures integrated into cryo-EM pipelines to resolve low-resolution ambiguities. " +
"Used to study protein evolution and metagenomics.",
accent: C.accentGold, x: 0.35, y: 4.05,
},
{
icon: "🔬", title: "Antimicrobial Research",
body: "Structural models of SARS-CoV-2 proteins aided vaccine and antiviral design. " +
"Malaria vaccine research (RTS,S and next-gen) benefited from Plasmodium protein structures. " +
"Enabled proteome-wide target identification for neglected tropical diseases.",
accent: C.red, x: 6.85, y: 4.05,
},
];
apps.forEach(a => {
addCard(s, a.x, a.y, 6.2, 2.4, a.title, a.body, a.accent);
s.addText(a.icon, {
x: a.x + 5.4, y: a.y + 0.1, w: 0.65, h: 0.6,
fontSize: 22, fontFace: "Segoe UI Emoji", margin: 0,
});
});
s.addText("Source: AlphaMissense – Cheng J et al., Science 2023 | Nuclear pore – Fontana P et al., Science 2022 | DeepMind blog", {
x: 0.3, y: 7.32, w: 12.7, h: 0.2,
fontSize: 8, color: C.mutedBlue, fontFace: "Calibri", margin: 0,
});
}
// ─────────────────────────────────────────────
// SLIDE 9 — ALPHAFOLD DATABASE
// ─────────────────────────────────────────────
{
const s = pres.addSlide();
s.background = { color: C.deepBlue };
addHeaderBar(s, "AlphaFold Protein Structure Database", "Largest open repository of predicted 3D protein structures");
// Stats row
const stats = [
{ val: "214M+", lbl: "Predicted Structures", color: C.accentGreen },
{ val: "48", lbl: "Organism Proteomes (full)", color: C.accentTeal },
{ val: "100%", lbl: "UniProt coverage predicted", color: C.accentGold },
{ val: "Free", lbl: "Open Access (CC BY 4.0)", color: C.accentTeal },
];
stats.forEach((st, i) => {
s.addShape(pres.ShapeType.rect, {
x: 0.35 + i * 3.2, y: 1.5, w: 3.0, h: 1.45,
fill: { color: C.midBlue },
line: { color: st.color, width: 2 },
});
s.addText(st.val, {
x: 0.35 + i * 3.2, y: 1.6, w: 3.0, h: 0.7,
fontSize: 28, bold: true, color: st.color, fontFace: "Calibri",
align: "center", margin: 0,
});
s.addText(st.lbl, {
x: 0.35 + i * 3.2, y: 2.3, w: 3.0, h: 0.55,
fontSize: 11, color: C.lightGrey, fontFace: "Calibri",
align: "center", margin: 0,
});
});
// File formats & access
addCard(s, 0.35, 3.15, 6.0, 2.1, "Data Formats & Access",
"Structures available in PDB, mmCIF, and binaryCIF formats. " +
"Metadata stored as JSON. API access at alphafold.ebi.ac.uk/api. " +
"FTP bulk download for all predictions. Integrated with UniProt, PDBe, and NCBI. " +
"Foldseek Cluster: structure-similarity clustering for all 214M entries.",
C.accentTeal);
// Organisms covered
addCard(s, 6.55, 3.15, 6.4, 2.1, "Coverage Highlights",
"Homo sapiens — full proteome (~20,386 canonical proteins). " +
"Model organisms: E. coli, S. cerevisiae, A. thaliana, D. melanogaster, C. elegans, M. musculus. " +
"Global health focus: Mycobacterium tuberculosis, Plasmodium falciparum, Leishmania species. " +
"Dark proteome: ~35% of human proteins were previously uncharacterised.",
C.accentGold);
// Limitation strip
addCard(s, 0.35, 5.4, 12.6, 1.1, "Known Limitations of the Database",
"Predictions represent static equilibrium structures — no dynamics or conformational ensembles. " +
"Low-confidence regions (pLDDT < 50) should not be used for drug docking. " +
"Multimeric complexes from AF-Multimer may have lower accuracy than monomeric predictions. " +
"Results require experimental validation for high-stakes applications.",
C.red);
s.addText("Source: Varadi M et al. AlphaFold Protein Structure Database in 2024. Nucleic Acids Res. 2024;52:D368–D375. PMC10767828", {
x: 0.3, y: 7.32, w: 12.7, h: 0.2,
fontSize: 8, color: C.mutedBlue, fontFace: "Calibri", margin: 0,
});
}
// ─────────────────────────────────────────────
// SLIDE 10 — NOBEL PRIZE 2024
// ─────────────────────────────────────────────
{
const s = pres.addSlide();
s.background = { color: C.deepBlue };
// Gold top banner
s.addShape(pres.ShapeType.rect, {
x: 0, y: 0, w: 13.3, h: 1.7,
fill: { color: "1A1200" },
line: { width: 0 },
});
s.addShape(pres.ShapeType.rect, {
x: 0, y: 0, w: 0.22, h: 1.7,
fill: { color: C.accentGold },
line: { width: 0 },
});
s.addText("🏆 Nobel Prize in Chemistry 2024", {
x: 0.45, y: 0.12, w: 12.5, h: 0.75,
fontSize: 30, bold: true, color: C.accentGold, fontFace: "Calibri", margin: 0,
});
s.addText('"For protein structure prediction" — Royal Swedish Academy of Sciences, October 2024', {
x: 0.45, y: 0.9, w: 12.5, h: 0.5,
fontSize: 13, color: C.lightGrey, fontFace: "Calibri", italic: true, margin: 0,
});
// Three laureates
const laureates = [
{
name: "Demis Hassabis", role: "CEO, Google DeepMind", share: "½ Prize",
bio: "Founded DeepMind in 2010. Led the AlphaFold project from inception. " +
"Drove the vision of applying deep reinforcement learning and attention mechanisms " +
"to biology. Also received Breakthrough Prize in Life Sciences 2023.",
color: C.accentGold,
},
{
name: "John Jumper", role: "Lead Researcher, Google DeepMind", share: "½ Prize",
bio: "Principal architect of AlphaFold 2. Designed the Evoformer and Invariant Point Attention " +
"modules. Previously at D.E. Shaw Research. " +
"Led technical development that achieved CASP14 near-experimental accuracy.",
color: C.accentGold,
},
{
name: "David Baker", role: "University of Washington", share: "½ Prize (separate)",
bio: "Awarded the other half of the Nobel Prize for computational protein design. " +
"Created Rosetta software suite. Designed novel proteins from scratch. " +
"Work complements AlphaFold: prediction vs. design.",
color: C.accentTeal,
},
];
laureates.forEach((l, i) => {
addCard(s, 0.35 + i * 4.32, 1.85, 4.1, 4.1, l.name, l.bio, l.color);
s.addText(l.role, {
x: 0.5 + i * 4.32, y: 2.3, w: 3.8, h: 0.38,
fontSize: 10.5, color: C.lightGrey, fontFace: "Calibri", italic: true, margin: 0,
});
s.addShape(pres.ShapeType.roundRect, {
x: 0.5 + i * 4.32, y: 5.5, w: 1.5, h: 0.38,
fill: { color: l.color },
line: { width: 0 },
rectRadius: 0.08,
});
s.addText(l.share, {
x: 0.5 + i * 4.32, y: 5.5, w: 1.5, h: 0.38,
fontSize: 10, bold: true, color: C.deepBlue, fontFace: "Calibri",
align: "center", valign: "middle", margin: 0,
});
});
addCard(s, 0.35, 6.15, 12.6, 1.0, null,
"The Nobel Committee noted: AlphaFold 2 represents one of the most significant advances in biology in decades. " +
"It solved a problem that biochemists had wrestled with since the 1970s and immediately democratised access to structural information.",
C.accentGold);
s.addText("Source: Nobel Prize Committee, Royal Swedish Academy of Sciences 2024 | NobelPrize.org", {
x: 0.3, y: 7.32, w: 12.7, h: 0.2,
fontSize: 8, color: C.mutedBlue, fontFace: "Calibri", margin: 0,
});
}
// ─────────────────────────────────────────────
// SLIDE 11 — LIMITATIONS & CRITICISMS
// ─────────────────────────────────────────────
{
const s = pres.addSlide();
s.background = { color: C.deepBlue };
addHeaderBar(s, "Limitations & Critical Perspectives", "What AlphaFold cannot do — and where caution is warranted");
const lims = [
{
title: "Static Structures Only",
body: "AlphaFold predicts a single static conformation. Proteins are dynamic — they undergo conformational changes, allosteric transitions, and induced-fit binding. " +
"Drug binding often requires sampling alternative states not captured by AF.",
color: C.red,
},
{
title: "Intrinsically Disordered Regions",
body: "Proteins with intrinsically disordered regions (IDRs) receive low pLDDT scores. " +
"IDRs are biologically functional (e.g., p53, FUS) but are misrepresented as structured in some outputs. " +
"Low confidence scores correctly flag this but are often ignored.",
color: C.accentGold,
},
{
title: "Hallucination Risk",
body: "AF3's diffusion model can occasionally 'hallucinate' physically impossible bond geometries or steric clashes — especially in regions with low evolutionary data. " +
"Post-processing validation (MolProbity, OpenStructure) is strongly recommended.",
color: C.red,
},
{
title: "No Co-Evolutionary Signal for Novel Proteins",
body: "Accuracy drops markedly for orphan proteins with few homologs in sequence databases (shallow MSA). " +
"De novo proteins and synthetic sequences may be poorly predicted. " +
"Experimental verification remains essential for novel fold regions.",
color: C.mutedBlue,
},
{
title: "Protein–Protein Complex Accuracy",
body: "AF-Multimer predicts protein complexes but accuracy (DockQ score) is lower than for monomers. " +
"Transient or weak interactions (Kd in µM range) are harder to capture. " +
"Requires experimental follow-up (pulldown, cross-linking MS).",
color: C.accentGold,
},
{
title: "Ethical & IP Concerns",
body: "AlphaFold 3 source code restrictions created debate in the open-science community. " +
"Dual-use risks: bioweapon-relevant protein prediction. " +
"Access inequality: high-compute prediction may remain unfeasible in low-resource settings.",
color: C.mutedBlue,
},
];
lims.forEach((l, i) => {
const col = i % 2;
const row = Math.floor(i / 2);
addCard(s, 0.35 + col * 6.45, 1.5 + row * 1.95, 6.15, 1.82, l.title, l.body, l.color);
});
s.addText("Source: Elofsson A, Curr Opin Struct Biol 2023 (PMID 37060758) | Chen L et al., Int J Mol Sci 2024 (PMID 39125995)", {
x: 0.3, y: 7.32, w: 12.7, h: 0.2,
fontSize: 8, color: C.mutedBlue, fontFace: "Calibri", margin: 0,
});
}
// ─────────────────────────────────────────────
// SLIDE 12 — FUTURE DIRECTIONS
// ─────────────────────────────────────────────
{
const s = pres.addSlide();
s.background = { color: C.deepBlue };
addHeaderBar(s, "Future Directions", "Where AlphaFold and structural AI are heading");
const futures = [
{
num: "01", title: "Protein Design (Inverse Folding)",
body: "RoseTTAFold and ProteinMPNN use AF-derived insights for de novo protein design. " +
"AlphaFold-based design already produced novel enzymes (RFdiffusion, Baker lab). " +
"Potential for therapeutic proteins, biosensors, and materials.",
color: C.accentGreen,
},
{
num: "02", title: "Dynamics & Conformational Ensembles",
body: "AlphaFlow and EigenFold explore predicting conformational ensembles instead of a single structure. " +
"Integrating molecular dynamics with AF predictions to simulate protein motion. " +
"Critical for GPCR agonist/antagonist modelling.",
color: C.accentTeal,
},
{
num: "03", title: "Multimodal Integration",
body: "Combining AF predictions with cryo-EM, NMR, and SAXS experimental data. " +
"Improved AlphaFold (Nov 2024, Linköping) integrates sparse experimental constraints. " +
"Hybrid experimental-computational pipelines becoming standard.",
color: C.accentGold,
},
{
num: "04", title: "Whole-Proteome Interactome Mapping",
body: "AF-Multimer applied to predict all protein–protein interactions in the human interactome. " +
"AlphaFold Metagenomic DB: >800K new protein families from environmental samples. " +
"Dark proteome illumination — unknown function proteins now structurally annotated.",
color: C.mutedBlue,
},
{
num: "05", title: "Therapeutic Protein Design",
body: "Design of novel antibodies, nanobodies, and enzyme therapeutics using AF as scaffold. " +
"Isomorphic Labs reported first AF3-guided drug candidates entering preclinical studies. " +
"Personalised neoantigen vaccine design leveraging AF structural insights.",
color: C.accentGreen,
},
{
num: "06", title: "Democratisation & Education",
body: "AlphaFold Server enables any researcher — globally — to predict protein structures free. " +
"Colab notebooks (ColabFold) allow local execution. " +
"Structural biology courses now incorporating AF predictions as first-line tools.",
color: C.accentTeal,
},
];
futures.forEach((f, i) => {
const col = i % 3;
const row = Math.floor(i / 3);
const x = 0.35 + col * 4.33;
const y = 1.5 + row * 2.8;
s.addShape(pres.ShapeType.rect, {
x, y, w: 4.1, h: 2.65,
fill: { color: C.midBlue },
line: { color: f.color, width: 1.5 },
});
s.addShape(pres.ShapeType.rect, {
x, y, w: 4.1, h: 0.07,
fill: { color: f.color },
line: { width: 0 },
});
s.addText(f.num, {
x: x + 0.1, y: y + 0.1, w: 0.55, h: 0.45,
fontSize: 18, bold: true, color: f.color, fontFace: "Calibri", margin: 0,
});
s.addText(f.title, {
x: x + 0.65, y: y + 0.1, w: 3.3, h: 0.45,
fontSize: 11.5, bold: true, color: C.white, fontFace: "Calibri", margin: 0,
});
s.addText(f.body, {
x: x + 0.12, y: y + 0.62, w: 3.86, h: 1.85,
fontSize: 10.5, color: C.lightGrey, fontFace: "Calibri", valign: "top", margin: 0,
});
});
s.addText("Source: Krokidis MG et al., Int J Mol Sci 2025 (PMID 40332289) | Wuyun Q et al., Molecules 2024 (PMID 38398585) | ScienceDaily Nov 2024", {
x: 0.3, y: 7.32, w: 12.7, h: 0.2,
fontSize: 8, color: C.mutedBlue, fontFace: "Calibri", margin: 0,
});
}
// ─────────────────────────────────────────────
// SLIDE 13 — COMPARISON: AF1 vs AF2 vs AF3
// ─────────────────────────────────────────────
{
const s = pres.addSlide();
s.background = { color: C.deepBlue };
addHeaderBar(s, "AlphaFold Version Comparison", "AF1 (2018) → AF2 (2020) → AF3 (2024)");
const headers = ["Feature", "AlphaFold 1 (2018)", "AlphaFold 2 (2020)", "AlphaFold 3 (2024)"];
const rows = [
["Architecture", "Distance-based CNN", "Evoformer + Structure Module", "Pairformer + Diffusion Module"],
["Input", "Sequence + MSA", "Sequence + MSA + Templates", "Sequence (+ optional templates)"],
["Output", "Distance distributions", "3D atomic coordinates", "Full-atom 3D (all molecules)"],
["Molecule Types", "Single-chain proteins", "Proteins + Multimer (2021 update)","Proteins, DNA, RNA, ligands, ions"],
["CASP Score", "GDT ~68.5 (CASP13)", "GDT ~92.4 (CASP14)", "Not CASP-evaluated (new scope)"],
["Confidence Metric", "None", "pLDDT (0–100)", "pLDDT + PAE + ligand confidence"],
["Open Source", "Yes (GitHub)", "Yes (GitHub + Colab)", "Partial (weights non-commercial)"],
["Database", "None", "AlphaFold DB launched 2021", "AlphaFold Server 2024"],
];
const colW = [2.2, 2.9, 3.5, 3.9];
const colX = [0.35, 2.58, 5.51, 9.04];
// Header row
headers.forEach((h, ci) => {
s.addShape(pres.ShapeType.rect, {
x: colX[ci], y: 1.5, w: colW[ci], h: 0.5,
fill: { color: ci === 0 ? C.mutedBlue : ci === 3 ? "0D2A1A" : C.midBlue },
line: { color: C.accentTeal, width: 0.8 },
});
s.addText(h, {
x: colX[ci] + 0.05, y: 1.5, w: colW[ci] - 0.1, h: 0.5,
fontSize: 11, bold: true,
color: ci === 0 ? C.lightGrey : ci === 1 ? C.mutedBlue : ci === 2 ? C.accentTeal : C.accentGold,
fontFace: "Calibri", align: "center", valign: "middle", margin: 0,
});
});
// Data rows
rows.forEach((row, ri) => {
const rowY = 2.05 + ri * 0.65;
row.forEach((cell, ci) => {
s.addShape(pres.ShapeType.rect, {
x: colX[ci], y: rowY, w: colW[ci], h: 0.62,
fill: { color: ri % 2 === 0 ? C.deepBlue : C.midBlue },
line: { color: "1A3050", width: 0.5 },
});
s.addText(cell, {
x: colX[ci] + 0.08, y: rowY + 0.04, w: colW[ci] - 0.16, h: 0.54,
fontSize: ci === 0 ? 10.5 : 9.8,
bold: ci === 0,
color: ci === 0 ? C.lightGrey : ci === 3 ? C.accentGold : C.lightGrey,
fontFace: "Calibri", valign: "middle", margin: 0,
});
});
});
s.addText("Sources: Jumper et al. Nature 2021 | Abramson et al. Nature 2024 | Senior AW et al. Nature 2020 | AlphaFold GitHub", {
x: 0.3, y: 7.32, w: 12.7, h: 0.2,
fontSize: 8, color: C.mutedBlue, fontFace: "Calibri", margin: 0,
});
}
// ─────────────────────────────────────────────
// SLIDE 14 — KEY REFERENCES
// ─────────────────────────────────────────────
{
const s = pres.addSlide();
s.background = { color: C.deepBlue };
addHeaderBar(s, "Key References & Sources", "Peer-reviewed publications and authoritative sources");
const refs = [
{
num: "1",
cite: "Jumper J, Evans R, Pritzel A et al. Highly accurate protein structure prediction with AlphaFold. " +
"Nature. 2021;596(7873):583–589. doi:10.1038/s41586-021-03819-2. PMID: 34265844.",
type: "Primary Paper – AlphaFold 2",
},
{
num: "2",
cite: "Abramson J, Adler J, Dunger J et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. " +
"Nature. 2024;630(8016):493–500. doi:10.1038/s41586-024-07487-w. PMID: 38594878.",
type: "Primary Paper – AlphaFold 3",
},
{
num: "3",
cite: "Senior AW, Evans R, Jumper J et al. Improved protein structure prediction using potentials from deep learning. " +
"Nature. 2020;577(7792):706–710. doi:10.1038/s41586-019-1923-7. PMID: 31942072.",
type: "AlphaFold 1",
},
{
num: "4",
cite: "Varadi M, Bertoni D, Magana P et al. AlphaFold Protein Structure Database in 2024: providing structure coverage for " +
"over 214 million protein sequences. Nucleic Acids Res. 2024;52(D1):D368–D375. PMID: 38015328. PMC10767828.",
type: "AlphaFold DB",
},
{
num: "5",
cite: "Krokidis MG, Koumadorakis DE, Lazaros K. AlphaFold3: An Overview of Applications and Performance Insights. " +
"Int J Mol Sci. 2025;26(8):3627. doi:10.3390/ijms26083627. PMID: 40332289.",
type: "Review 2025",
},
{
num: "6",
cite: "Chen L, Li Q, Nasif KFA. AI-Driven Deep Learning Techniques in Protein Structure Prediction. " +
"Int J Mol Sci. 2024;25(15):8426. doi:10.3390/ijms25158426. PMID: 39125995.",
type: "Review 2024",
},
{
num: "7",
cite: "Elofsson A. Progress at protein structure prediction, as seen in CASP15. " +
"Curr Opin Struct Biol. 2023;80:102594. doi:10.1016/j.sbi.2023.102594. PMID: 37060758.",
type: "CASP Review",
},
{
num: "8",
cite: "Nobel Prize Committee. The Nobel Prize in Chemistry 2024 – Scientific Background: Computational Protein Design and " +
"Protein Structure Prediction. Royal Swedish Academy of Sciences. October 2024. NobelPrize.org.",
type: "Nobel Committee",
},
];
refs.forEach((r, i) => {
const y = 1.5 + i * 0.72;
s.addShape(pres.ShapeType.rect, {
x: 0.35, y, w: 12.6, h: 0.65,
fill: { color: i % 2 === 0 ? C.deepBlue : C.midBlue },
line: { color: "1A3050", width: 0.5 },
});
s.addShape(pres.ShapeType.rect, {
x: 0.35, y, w: 0.42, h: 0.65,
fill: { color: i < 3 ? C.accentTeal : i < 6 ? C.mutedBlue : C.accentGold },
line: { width: 0 },
});
s.addText(r.num, {
x: 0.35, y, w: 0.42, h: 0.65,
fontSize: 11, bold: true, color: C.white, fontFace: "Calibri",
align: "center", valign: "middle", margin: 0,
});
s.addText(`[${r.type}] ${r.cite}`, {
x: 0.85, y: y + 0.05, w: 11.95, h: 0.55,
fontSize: 9.2, color: C.lightGrey, fontFace: "Calibri", valign: "middle", margin: 0,
});
});
}
// ─────────────────────────────────────────────
// SLIDE 15 — CLOSING / THANK YOU
// ─────────────────────────────────────────────
{
const s = pres.addSlide();
s.background = { color: C.deepBlue };
// Decorative circles
for (let i = 0; i < 22; i++) {
s.addShape(pres.ShapeType.ellipse, {
x: 0.3 + Math.cos(i * 0.55) * 0.9,
y: 0.2 + i * 0.32,
w: 0.22, h: 0.22,
fill: { color: i % 3 === 0 ? C.accentTeal : i % 3 === 1 ? C.mutedBlue : C.midBlue },
line: { width: 0 },
});
}
s.addShape(pres.ShapeType.rect, {
x: 1.3, y: 1.6, w: 10.7, h: 4.0,
fill: { color: C.midBlue },
line: { color: C.accentTeal, width: 2 },
});
s.addShape(pres.ShapeType.rect, {
x: 1.3, y: 1.6, w: 0.22, h: 4.0,
fill: { color: C.accentTeal },
line: { width: 0 },
});
s.addText("Thank You", {
x: 1.75, y: 1.75, w: 9.9, h: 0.9,
fontSize: 46, bold: true, color: C.accentTeal, fontFace: "Calibri", margin: 0,
});
s.addText("Key Takeaways", {
x: 1.75, y: 2.75, w: 9.9, h: 0.42,
fontSize: 15, bold: true, color: C.accentGold, fontFace: "Calibri", margin: 0,
});
const takeaways = [
"AlphaFold solved the 50-year protein folding problem using deep learning (Evoformer + IPA)",
"AlphaFold 2 achieved near-experimental accuracy at CASP14, revolutionising structural biology",
"AlphaFold DB provides 214M+ free structures — the largest structural biology resource ever",
"AlphaFold 3 extends prediction to protein–DNA/RNA/ligand complexes via diffusion modelling",
"2024 Nobel Prize in Chemistry awarded to Hassabis, Jumper, and Baker",
"Limitations remain: static structures only, low accuracy for disordered/novel proteins",
];
const arr = takeaways.map((t, i) => ({
text: `▸ ${t}`,
options: {
color: C.lightGrey, fontSize: 11.5, breakLine: i < takeaways.length - 1,
},
}));
s.addText(arr, {
x: 1.75, y: 3.22, w: 10.0, h: 2.2,
fontFace: "Calibri", valign: "top",
});
s.addText("Questions & Discussion", {
x: 1.75, y: 5.55, w: 10.0, h: 0.45,
fontSize: 16, bold: true, color: C.accentTeal, fontFace: "Calibri", margin: 0,
});
s.addText("AlphaFold DB: alphafold.ebi.ac.uk | AlphaFold Server: alphafoldserver.com | GitHub: github.com/google-deepmind/alphafold", {
x: 0, y: 7.0, w: 13.3, h: 0.38,
fontSize: 9.5, color: C.mutedBlue, align: "center", fontFace: "Calibri",
});
}
// ─── WRITE FILE ───
pres.writeFile({ fileName: "/home/daytona/workspace/alphafold-ppt/AlphaFold_Seminar.pptx" })
.then(() => console.log("✅ Presentation written successfully."))
.catch(e => console.error("Error:", e));
Running Command
cd /home/daytona/workspace/alphafold-ppt && node build.js 2>&1
Running Command
ls -lh /home/daytona/workspace/alphafold-ppt/AlphaFold_Seminar.pptx
Your presentation is ready:
AlphaFold Seminar
Presentation · PPTX
What's Inside — 15 Slides
| # | Slide | Content |
|---|---|---|
| 1 | Title Slide | Bold dark theme with AlphaFold branding, Nobel badge, key stats |
| 2 | Outline | 10-topic structured agenda with numbered layout |
| 3 | The Protein Folding Problem | Levinthal's Paradox, Anfinsen's Dogma, scale of the problem |
| 4 | History & CASP | Visual timeline 1994→2024 + GDT score bar chart |
| 5 | AlphaFold 2 Architecture | Pipeline flow: MSA → Evoformer (×48) → Structure Module (IPA) → Recycling |
| 6 | Training & Confidence | PDB/UniRef/BFD data sources + colour-coded pLDDT scale |
| 7 | AlphaFold 3 | Pairformer, diffusion module, DNA/RNA/ligand scope, AF Server |
| 8 | Applications | Drug discovery, rare diseases, nuclear pore, antimicrobials |
| 9 | AlphaFold DB | 214M+ structures, coverage stats, limitations strip |
| 10 | Nobel Prize 2024 | Hassabis, Jumper, Baker profiles + Nobel Committee quote |
| 11 | Limitations | Dynamics, IDRs, hallucinations, IP concerns — 6 cards |
| 12 | Future Directions | Protein design, dynamics, multimodal, interactome, therapeutics |
| 13 | Version Comparison Table | AF1 vs AF2 vs AF3 across 8 features |
| 14 | References | 8 numbered citations (Nature, Nucleic Acids Res, Nobel Committee, PubMed) |
| 15 | Closing / Q&A | Key takeaways + resource links |
Design: Dark navy/teal/gold colour scheme with accent cards, timeline graphics, and a comparison table. All sources cited inline per slide and in a dedicated references slide (Jumper et al. Nature 2021, Abramson et al. Nature 2024, Varadi et al. NAR 2024, Nobel Committee 2024, and four PubMed-indexed reviews).
can you make the script for this seminar
Writing File
~/alphafold-ppt/AlphaFold_Seminar_Script.md
# AlphaFold Seminar — Presenter Script ### Full spoken script, slide by slide | Estimated delivery time: ~45–55 minutes --- ## SLIDE 1 — Title Slide **[Display as audience enters. Begin speaking once settled.]** Good [morning/afternoon/evening], everyone. Thank you for being here. Today we are going to talk about one of the most consequential scientific breakthroughs of the past two decades — AlphaFold. If you work in biology, medicine, pharmacology, or even computer science, this technology has already changed the landscape of your field — whether you know it yet or not. By the end of this talk, you will understand *what* AlphaFold does, *how* it works at a technical level, *why* it matters, and — importantly — what it *cannot* do. We will cover everything from the foundational biology problem it solves, to the deep learning architecture behind it, to its 2024 Nobel Prize, and where the field is heading next. Let's start at the beginning. --- ## SLIDE 2 — Presentation Outline **[Click to slide 2.]** Here is the structure of today's talk. We have ten core topics. We begin with the biological problem AlphaFold was built to solve — the Protein Folding Problem. We then trace the history of AlphaFold from its first appearance in 2018 through to AlphaFold 3 in 2024. After that, we go deep into the architecture — the Evoformer, the Structure Module, how training works, and how the model scores its own confidence. We then look at AlphaFold 3 specifically — which is a fundamentally different model from AlphaFold 2. From there we examine real-world applications, the public database, the Nobel Prize, the limitations, and finally the future. It is a lot to cover, so let's move quickly but thoroughly. --- ## SLIDE 3 — The Protein Folding Problem **[Click to slide 3.]** To understand AlphaFold, you first have to understand the problem it was built to solve. Proteins are the molecular machines of life. They catalyse biochemical reactions, carry oxygen in your blood, fight off pathogens, transmit signals between cells, and build the physical structures of your body. Every protein starts as a linear chain of amino acids — a string of chemical building blocks encoded by your DNA. That string spontaneously folds, in milliseconds, into a precise three-dimensional shape. And *that shape determines the protein's function*. A misfolded protein can cause disease — Alzheimer's, Parkinson's, cystic fibrosis, many cancers all involve proteins that are misfolded or misfunctioning. So the central question is: **given only the sequence of amino acids, can you predict the 3D structure?** This is called the Protein Folding Problem, and for about 50 years, it resisted solution. Now look at the numbers on the right. There are around 20,000 protein-coding genes in the human genome, producing potentially 100,000 distinct protein forms. But experimentally solving a structure — using X-ray crystallography, cryo-electron microscopy, or NMR — is slow, expensive, and often fails for membrane proteins or disordered regions. By 2020, the Protein Data Bank contained around 170,000 experimentally determined structures. Meanwhile, protein sequence databases held hundreds of millions of sequences. That gap — between known sequences and known structures — is what AlphaFold was built to close. The famous insight comes from Christian Anfinsen, who won the 1972 Nobel Prize in Chemistry for demonstrating that a protein's amino acid sequence alone contains all the information needed to determine its final 3D shape. That principle — Anfinsen's dogma — is the theoretical foundation AlphaFold rests on. And then there is Levinthal's Paradox. Cyrus Levinthal calculated in 1969 that if a 100-amino-acid protein tried every possible conformation randomly, it would take longer than the age of the universe to find the right fold. Yet proteins fold in microseconds. That means biology has found shortcuts — guided folding pathways. The challenge was figuring out what those shortcuts are, and encoding them in a computer program. --- ## SLIDE 4 — History & CASP Benchmarks **[Click to slide 4.]** So how do you benchmark progress on this problem? The answer is CASP — the Critical Assessment of Protein Structure Prediction — a biennial competition that has been running since 1994. The idea is elegant. Experimental groups determine new protein structures but keep them secret. Computational groups are given only the amino acid sequences and asked to predict the structures. The predictions are then compared against the experimental results and scored using a metric called GDT — Global Distance Test — where 100 is a perfect match. For decades, progress was incremental. By CASP12 in 2016, the best methods were achieving GDT scores in the mid-40s for the hardest targets — so-called free-modelling targets where no close template exists. These methods used co-evolutionary analysis and physics-based folding simulations, and they were good, but far from reliable. Then DeepMind entered. Their first system, AlphaFold 1, appeared at CASP13 in 2018. It jumped the GDT score to 68.5 for free-modelling targets — a significant leap. It won the competition but researchers noted it was still far from experimental accuracy. Promising, but not yet transformative. Two years later, at CASP14 in 2020, AlphaFold 2 appeared. And the scientific community was stunned. AlphaFold 2 achieved a median GDT score of 92.4 on the free-modelling targets. The next-best method scored in the 50s. Organisers called it a solution to the protein folding problem. John Moult, who co-founded CASP, said it was "a stunning advance" — the kind of result that comes once in a generation. You can see this in the bar chart here. The jump from CASP13 to CASP14 is not incremental — it is transformative. Between 2020 and 2024, DeepMind released the AlphaFold Protein Structure Database, first with 350,000 structures, then expanding to cover the entire known protein universe — over 214 million structures. And in May 2024, they released AlphaFold 3, which expanded beyond proteins entirely. --- ## SLIDE 5 — AlphaFold 2 Architecture **[Click to slide 5.]** Now let's get into the machine itself. How does AlphaFold 2 actually work? The architecture is shown in the pipeline across the top of this slide. At a high level, AlphaFold 2 is an end-to-end differentiable deep learning system — meaning every component is trained together, jointly optimising a single objective: predict the 3D structure as accurately as possible. Let me walk you through each stage. **Stage 1: Input.** You give the model an amino acid sequence — a string of letters, each representing one of the 20 standard amino acids. **Stage 2: Multiple Sequence Alignment (MSA).** The system then searches enormous protein sequence databases — UniRef90, BFD with 2.2 billion clusters, MGnify — to find evolutionarily related proteins from other species. These are aligned into a matrix called a Multiple Sequence Alignment, or MSA. Why does this matter? Because evolution is informative. If two amino acids in a protein have co-evolved — meaning when one mutates, the other tends to mutate to compensate — that is a strong signal they are spatially close in the 3D structure. The MSA encodes millions of years of evolutionary experiments. Deeper MSAs, meaning more related sequences found, generally mean more accurate predictions. **Stage 3: Evoformer.** This is the heart of AlphaFold 2. The Evoformer consists of 48 stacked transformer-like blocks, each of which jointly processes two representations: the MSA representation — capturing information across related sequences — and a pairwise representation — capturing relationships between every pair of amino acid positions. The key innovation here is that these two representations *update each other*. The MSA representation informs the pair representation, and vice versa, through what the authors call "triangle multiplicative updates" and row- and column-wise attention. This iterative communication allows the model to reason about both sequence conservation and spatial geometry simultaneously. **Stage 4: Structure Module.** The Evoformer outputs a refined pairwise representation and an updated single-sequence representation. The Structure Module takes these and predicts the actual 3D coordinates of every atom. The key component here is Invariant Point Attention — IPA. This is an attention mechanism designed to respect 3D geometry. It predicts a rigid-body frame — a rotation and translation — for each amino acid residue, and then the side-chain torsion angles within each residue. Because it operates in 3D space directly, the predictions are equivariant — they give the right answer regardless of how the input structure is oriented. **Stage 5: Recycling.** The entire pipeline is run three times. Each iteration takes the previous iteration's predicted structure as an additional input, allowing the model to progressively refine its prediction. This recycling mechanism was one of the critical contributions that pushed accuracy to near-experimental levels. **Output:** The final output is a full 3D atomic model of the protein, plus a per-residue confidence score — the pLDDT — which we will discuss on the next slide. --- ## SLIDE 6 — Training & Confidence Metrics **[Click to slide 6.]** Let's talk about how the model was trained and how it communicates its own uncertainty. **Training Data.** The model was trained on the Protein Data Bank — approximately 170,000 experimentally determined structures at the time. These are the ground-truth labels. The sequences paired with those structures were used, along with the giant databases on the left here — UniRef90 and UniClust30 for constructing MSAs, and BFD, which contains 2.2 billion clustered sequences largely derived from environmental metagenomics. This massive sequence diversity is critical because it provides the co-evolutionary signal the Evoformer learns to exploit. The training loss function is called FAPE — Frame Aligned Point Error. It measures the average distance between predicted and true atomic positions across all reference frames defined by backbone residues. This frame-aligned approach makes the loss invariant to global rotation and translation of the whole structure, which is mathematically important for training stability. Additional losses include torsion angle loss, distogram loss for predicted distance distributions, and a local distance difference test head. Training required approximately 128 Google TPUv3 chips running for several weeks — a substantial compute investment, though modest compared to large language models. **pLDDT Confidence Score.** Now the pLDDT — per-residue Local Distance Difference Test — is AlphaFold's self-assessment of its own prediction quality at each amino acid position. It ranges from 0 to 100. Look at the colour-coded scale on the right: - **Above 90**: Very high confidence. These regions are generally accurate enough for drug docking studies, binding site characterisation, and functional annotation. - **70 to 90**: Confident. The backbone is reliably predicted. Suitable for most downstream analyses. - **50 to 70**: Low confidence. Use with caution. These regions may be flexible or only conditionally ordered. - **Below 50**: Very low confidence. If you visualise these regions in PyMOL or ChimeraX, they often look like random coils — sometimes described as "spaghetti." These regions are likely intrinsically disordered and should *not* be used for structural analysis. The pLDDT score is stored in the B-factor field of the output PDB files — a clever repurposing of an existing file format field — and is colour-coded in all AlphaFold DB visualisations. Understanding pLDDT is not optional. Every researcher using AlphaFold predictions needs to check the confidence scores for their region of interest before drawing conclusions. --- ## SLIDE 7 — AlphaFold 3 **[Click to slide 7.]** In May 2024, Google DeepMind and its sister company Isomorphic Labs published AlphaFold 3 in *Nature*. It is not simply an updated AlphaFold 2 — it is a fundamentally different model with a fundamentally different scope. AlphaFold 2 was built for proteins. AlphaFold 3 was built for *biomolecular interactions*. Let me explain the key changes. **Pairformer replaces Evoformer.** AlphaFold 3 drops the full MSA processing stack and replaces the Evoformer with a simpler but powerful module called the Pairformer. The Pairformer processes a single sequence representation alongside pairwise representations, using learned template embeddings where available. This simplification makes the architecture more generalisable across molecule types — because you cannot build an MSA for a small molecule ligand. **Diffusion Module replaces the Structure Module.** This is the most dramatic architectural change. Instead of the deterministic Invariant Point Attention-based structure module, AlphaFold 3 uses a generative diffusion model. If you are familiar with image diffusion models like DALL-E or Stable Diffusion — the idea is analogous. The model starts with a random cloud of atoms in 3D space, and iteratively denoises them, guided by the Pairformer's output, until they converge on a physically plausible structure. This approach naturally handles the full atom detail of small molecules, nucleic acids, and post-translational modifications — all of which have very different atomic chemistry from protein backbone atoms. **Multi-Molecular Scope.** AlphaFold 3 can model proteins, DNA, RNA, small-molecule ligands, ions, and post-translational modifications like glycosylation and phosphorylation — all in one prediction. The authors reported at least 50% improvement in accuracy for protein–nucleic acid interfaces compared to existing specialist methods, and significant improvements for protein–ligand prediction. **Drug Discovery Integration.** This last point is not incidental. Isomorphic Labs was co-developer on AlphaFold 3, and their explicit goal is to use it for drug discovery. Understanding how a candidate drug molecule binds to a disease-relevant protein is one of the most important — and historically expensive — steps in pharmaceutical development. AlphaFold 3 directly addresses this. A free AlphaFold Server was launched simultaneously for non-commercial research, accessible at alphafoldserver.com. --- ## SLIDE 8 — Applications & Impact **[Click to slide 8.]** Let me now bring this out of the technical realm and into concrete scientific impact. What has AlphaFold actually been used for? **Drug Discovery.** AlphaFold is being used to identify and characterise binding pockets — the sites on a protein where small molecules can bind and modulate its function. Isomorphic Labs has reported active drug discovery programmes using AlphaFold-derived structural insight against targets including tuberculosis, Chagas disease, and antibiotic-resistant bacteria. The key advantage is speed: structure determination that once required months of crystallography experiments can now be initiated computationally in minutes. **Rare Genetic Diseases — AlphaMissense.** In 2023, DeepMind published AlphaMissense in *Science*. This is an AlphaFold-derived model that classifies missense mutations — single amino acid changes caused by point mutations — as likely benign, likely pathogenic, or uncertain. It classified approximately 71 million possible missense variants across the human proteome. Around 89% received a confident classification. For rare disease diagnosis, where identifying whether a patient's unique variant is disease-causing is critical, this tool is transformative. **Structural Biology Acceleration.** One of the most dramatic examples is the nuclear pore complex — a massive protein assembly of around 120 proteins that regulates what enters and exits the cell nucleus. Experimental structural determination of this complex had been a decades-long project. In 2022, a team at Harvard published a near-complete model by integrating AlphaFold predictions with sparse cryo-EM data. What might have taken another decade was accomplished in a fraction of that time. AlphaFold structures are now routinely used to interpret cryo-EM maps — where the electron density is known at moderate resolution but the atomic model needs to be fitted — dramatically accelerating structure determination. **Antimicrobial and Pandemic Research.** AlphaFold was used to model all major proteins of SARS-CoV-2 shortly after the pandemic began, aiding in vaccine target identification and antiviral design. For malaria, the structures of key *Plasmodium falciparum* proteins — many of which resisted experimental determination — are now accessible, opening new avenues for vaccine and drug development. --- ## SLIDE 9 — AlphaFold Protein Structure Database **[Click to slide 9.]** The AlphaFold Protein Structure Database — hosted by the European Bioinformatics Institute at EMBL-EBI — is possibly the most consequential scientific data resource created in the past decade. As of 2024, it contains over **214 million predicted protein structures** — covering virtually every protein in the UniProt sequence database. That is orders of magnitude more structural data than has been accumulated experimentally in the entire history of structural biology. Access is completely free under a Creative Commons Attribution licence. Structures are available via the web browser, programmatic API, and bulk FTP download. Files come in PDB, mmCIF, and binaryCIF formats. The database covers full proteomes for 48 organisms, including all major model organisms — *E. coli*, yeast, *Arabidopsis*, fruit fly, nematode, and mouse — as well as the complete human proteome with all 20,386 canonical proteins. There is a strong focus on global health: complete proteomes of *Mycobacterium tuberculosis*, *Plasmodium falciparum*, and *Leishmania* species are included. One of the most exciting aspects is what researchers call the **dark proteome** — proteins with no known structure and often no known function. Around 35% of human proteins were previously uncharacterised structurally. AlphaFold illuminates this dark proteome, providing structural hypotheses for proteins that were previously black boxes. **However, and I want to be clear about this**, the database has known limitations. Every prediction represents a single static, equilibrium conformation. There are no dynamics. There are no alternative conformations. Low-confidence regions should not be used for drug docking. Multimeric predictions from AF-Multimer carry additional uncertainty. And all results benefit from experimental validation before high-stakes decisions are made. The database is a starting point — a powerful one — not a replacement for experimental structural biology. --- ## SLIDE 10 — Nobel Prize in Chemistry 2024 **[Click to slide 10.]** In October 2024, the Royal Swedish Academy of Sciences awarded the Nobel Prize in Chemistry to three scientists. Demis Hassabis and John Jumper received one half of the prize for protein structure prediction. David Baker of the University of Washington received the other half for computational protein design — which is the complementary problem of *designing* new proteins with desired functions, rather than predicting the structures of existing ones. **Demis Hassabis** founded DeepMind in 2010 with the mission of using artificial intelligence to accelerate scientific discovery. He recognised that protein structure prediction was a problem ideally suited to deep learning — high-dimensional, data-rich, and with a clear objective function. He drove the strategic direction of the AlphaFold project and led the organisation that made it possible. **John Jumper** was the principal technical architect of AlphaFold 2. His specific innovations — the Evoformer with its triangle attention mechanisms, the Invariant Point Attention module, the recycling strategy, and the FAPE loss function — are what made the CASP14 result possible. Before DeepMind, he worked at D.E. Shaw Research on protein simulation. His combination of deep learning expertise and structural biology knowledge was precisely the interdisciplinary synthesis the problem required. **David Baker** at the University of Washington has spent three decades developing computational protein design tools, culminating in the Rosetta software suite and more recently diffusion-based design approaches. His work represents the other side of the structural biology coin: not predicting what nature made, but designing what nature has not yet made. The Nobel Committee's statement was striking. They said AlphaFold solved a problem that had confronted biochemists for half a century, and that it had immediately democratised access to structural information that was previously reserved for well-resourced experimental labs. Both Hassabis and Jumper had already received the Breakthrough Prize in Life Sciences and the Albert Lasker Award for Basic Medical Research in 2023 — so this Nobel Prize, while historic, was not a surprise to the scientific community. --- ## SLIDE 11 — Limitations & Critical Perspectives **[Click to slide 11.]** I want to spend real time on limitations, because any tool used uncritically is a tool used dangerously. **Static Structures Only.** This is probably the most important limitation. Proteins are not rigid objects — they breathe, flex, change shape upon binding partners, and populate multiple conformational states. Many drug targets, particularly GPCRs and ion channels, are interesting precisely *because* of their conformational dynamics. AlphaFold gives you one conformation — typically the lowest-energy ground state. It tells you nothing about the range of motion, the transition pathways, or the alternative states that a drug might need to target. **Intrinsically Disordered Regions.** Many proteins — particularly transcription factors, signalling proteins, and hub proteins in interaction networks — contain large intrinsically disordered regions, or IDRs. These regions do not adopt a fixed 3D structure; they are functionally flexible. AlphaFold correctly assigns low pLDDT scores to these regions, but the 3D coordinates it assigns are essentially meaningless. The problem is that some users ignore the pLDDT and treat the coordinates as real. They are not. **Hallucination Risk in AlphaFold 3.** Because AF3 uses a generative diffusion model, it is subject to a class of errors familiar from generative AI — hallucination. Specifically, it can generate structures with physically impossible bond lengths, clashing atoms, or geometrically implausible backbone conformations. Post-processing validation using tools like MolProbity or OpenStructure is strongly recommended for any result used in downstream analysis or publication. **Shallow MSA Problem.** For proteins with few evolutionary relatives — newly evolved proteins, synthetic sequences, orphan proteins — the MSA is shallow or empty. AlphaFold's accuracy drops significantly in these cases because the co-evolutionary signal it depends on is absent. This is a fundamental limitation of the evolutionary information approach. **Protein Complex Accuracy.** AF-Multimer's accuracy for predicting protein-protein interactions, while impressive, is lower than for monomers. Weak or transient interactions are particularly difficult. Experimentally, these require pulldown assays, cross-linking mass spectrometry, or other binding assays that AlphaFold cannot replace. **Ethical and Policy Dimensions.** Finally, it is worth acknowledging the societal dimensions. The partial restriction on AlphaFold 3 source code — where model weights were released for non-commercial use only — generated significant debate about open science norms in the AI era. There are also genuine dual-use concerns: the same structural knowledge that accelerates drug development could, in principle, be used for harmful purposes. And like many computational advances, there is a question of access equity — whether researchers in low-resource settings can realistically leverage these tools. These are not reasons to avoid AlphaFold. They are reasons to use it thoughtfully. --- ## SLIDE 12 — Future Directions **[Click to slide 12.]** Where is all of this heading? **Protein Design.** The most exciting frontier is inverting the prediction problem: instead of predicting structure from sequence, designing a sequence that will fold into a desired structure. Tools like RFdiffusion and ProteinMPNN from the Baker lab now use AlphaFold-derived representations to design novel proteins from scratch — enzymes, binders, biosensors — with properties not found in nature. The first computationally designed protein-based drugs are moving toward clinical testing. **Dynamics and Conformational Ensembles.** New approaches like AlphaFlow and EigenFold are extending the AlphaFold framework to predict not just a single structure but an *ensemble* of conformations — a distribution over the structural landscape a protein can occupy. Combined with molecular dynamics simulation, this will eventually give us a complete picture of protein motion, not just a snapshot. **Multimodal Integration.** A November 2024 paper from Linköping University demonstrated that AlphaFold can be improved by incorporating sparse experimental data — partial cryo-EM maps, chemical cross-links, SAXS profiles — as constraints during prediction. Hybrid experimental-computational pipelines are rapidly becoming standard practice, and the boundary between computational prediction and experimental determination is blurring. **Whole-Proteome Interactome Mapping.** Researchers are applying AF-Multimer at scale to systematically predict all protein-protein interactions in the human interactome — a network of potentially hundreds of thousands of interactions. This will provide a structural atlas of cellular biology that was previously unimaginable. **Therapeutic Applications.** Isomorphic Labs has reported that AlphaFold 3-guided drug candidates have entered preclinical development pipelines. The timeline from target identification to clinical trial entry may compress dramatically. Personalised cancer immunotherapy — designing neoantigen vaccines tailored to a patient's specific tumour mutations — is another area where structural prediction is enabling new therapeutic strategies. **Democratisation.** And perhaps most importantly for the long-term health of science: the AlphaFold Server is free, the database is open access, and ColabFold allows AlphaFold 2 to be run on Google Colab with no specialist hardware. A graduate student anywhere in the world can now do structural biology that ten years ago required a major research institution with crystallography facilities. That democratisation of access is itself a scientific revolution. --- ## SLIDE 13 — Version Comparison Table **[Click to slide 13.]** Let me give you a quick side-by-side summary of the three generations. AlphaFold 1 in 2018 was a convolutional neural network that predicted distance distributions between residue pairs — it did not directly predict 3D coordinates. It used sequence and MSA as input and achieved a GDT of about 68 at CASP13. Impressive for its time, but not yet accurate enough for most applications. AlphaFold 2 in 2020 was the breakthrough version. New architecture — Evoformer plus Structure Module with Invariant Point Attention. Direct prediction of 3D atomic coordinates. GDT of 92.4 at CASP14. Introduced the pLDDT confidence score. Released as fully open source with the AlphaFold DB launch in 2021. AlphaFold 3 in 2024 replaced the Evoformer with the simpler Pairformer and replaced the Structure Module with a diffusion model. Extended scope to DNA, RNA, ligands, and ions. Not evaluated at CASP — the competition format doesn't cover its new molecular scope. Released with partial restrictions on commercial use, though the AlphaFold Server is free for non-commercial research. Each version represents a genuine architectural reimagining, not just incremental tuning. --- ## SLIDE 14 — Key References **[Click to slide 14.]** Here are the primary sources underpinning everything I have discussed today. The foundational paper is Jumper et al. in *Nature* 2021 — the AlphaFold 2 paper, one of the most-cited biology papers of the decade. AlphaFold 3 is described in Abramson et al., *Nature* 2024 — already cited over 9,000 times as of late 2025, which is extraordinary for a paper less than two years old. The original AlphaFold 1 system is described in Senior et al., *Nature* 2020. The AlphaFold database is documented in Varadi et al., *Nucleic Acids Research* 2024, which I encourage you to read if you plan to use the database — it has important guidance on data formats, limitations, and access methods. The two review papers — Krokidis et al. 2025 in *International Journal of Molecular Sciences* and Chen et al. 2024 — provide excellent synthesis of the current state of the field and are good starting points for deeper reading. The Nobel Prize scientific background document from the Royal Swedish Academy of Sciences is freely available at NobelPrize.org and is remarkably readable — I recommend it to anyone who wants a formal, authoritative account of why this work deserved the Nobel. --- ## SLIDE 15 — Closing / Thank You **[Click to slide 15.]** Let me close with the key takeaways. AlphaFold solved the protein folding problem — a challenge that had resisted the efforts of biochemists and biophysicists for fifty years — using deep learning. Specifically, using co-evolutionary reasoning encoded in the Evoformer and geometry-aware attention in the Structure Module. AlphaFold 2 achieved near-experimental accuracy at CASP14 in 2020, triggering what can only be described as a paradigm shift in structural biology. The AlphaFold Protein Structure Database provides 214 million free, open-access predicted structures — the largest structural biology resource in history. AlphaFold 3 extended the framework to protein interactions with DNA, RNA, and small molecules, directly enabling drug discovery applications. In October 2024, Demis Hassabis, John Jumper, and David Baker were awarded the Nobel Prize in Chemistry — recognition from the highest scientific authority that this work has fundamentally changed what is possible. And yet — AlphaFold has real limitations. Static structures, disordered regions, shallow MSAs, hallucinations. These are not minor caveats. They are boundaries that define where the tool is trustworthy and where it needs experimental support. The future is the integration of prediction, design, dynamics, and experiment into a unified structural biology pipeline. We are very much at the beginning of that story. Thank you. I am happy to take questions. --- ## APPENDIX: Anticipated Q&A **Q: How long does an AlphaFold prediction take?** For a typical protein of 200–500 residues, the AlphaFold Server returns results in minutes to tens of minutes depending on server load. The most time-consuming step is actually the MSA construction — the neural network forward pass itself is fast. For very large proteins or protein complexes, predictions can take hours. **Q: Is AlphaFold used in clinical medicine today?** Not directly in clinical diagnostic pathways yet, but pharmaceutical companies are using AlphaFold-derived structural models to accelerate drug discovery pipelines. AlphaMissense is being evaluated for integration into variant interpretation workflows for rare disease diagnosis. Clinical application in the traditional sense — informing individual patient care — is still emerging. **Q: Can AlphaFold predict membrane proteins?** Yes, with caveats. AlphaFold can predict membrane protein structures, and the database includes them. However, membrane proteins are often predicted without the lipid bilayer context, which can affect conformational accuracy. The pLDDT scores for transmembrane helices are often high, but the relative orientation of those helices — which determines, for example, the gating state of an ion channel — may not be accurately captured. **Q: How does AlphaFold compare to Rosetta?** Rosetta is a physics-based and knowledge-based software suite developed over 30 years by David Baker's lab. For protein structure *prediction*, AlphaFold 2 and 3 substantially outperform Rosetta. For protein *design* — generating new sequences for a desired structure — Rosetta's ProteinMPNN and newer RFdiffusion tools are at the frontier. The two approaches are increasingly complementary: AlphaFold predicts, Rosetta and its derivatives design. **Q: What is the AlphaFold license?** AlphaFold 2 code and weights are released under an Apache 2.0 licence — fully open, including commercial use. AlphaFold 3's model weights are available for non-commercial research under a custom DeepMind licence. The AlphaFold Database is under Creative Commons Attribution 4.0 International. The AlphaFold Server is free for non-commercial use. **Q: What comes after AlphaFold?** ESMFold from Meta AI uses protein language models instead of MSA, making it faster but slightly less accurate. RoseTTAFold2 from the Baker lab is competitive with AlphaFold 2. The field is moving toward models that can jointly predict structure *and* dynamics, and toward generative models that design novel proteins from scratch rather than predicting existing ones. --- *Script prepared for AlphaFold Seminar — May 2026* *Sources: Jumper et al. Nature 2021 | Abramson et al. Nature 2024 | Varadi et al. NAR 2024 | Nobel Committee 2024 | Chen et al. IJMS 2024 | Krokidis et al. IJMS 2025*
AlphaFold Seminar Script
MD File · MD
Here is what the script covers:
Structure: One section per slide, clearly labelled with click cues so you always know where you are.
Estimated time: ~45–55 minutes at a normal speaking pace, leaving 10–15 minutes for Q&A in a standard 60–75 min seminar slot.
What each section does:
| Slide | Spoken focus |
|---|---|
| 1 – Title | Sets stakes — why this matters to everyone in the room |
| 2 – Outline | Brief roadmap |
| 3 – Folding Problem | Anfinsen's dogma, Levinthal's paradox, the gap between sequences and structures |
| 4 – History | Walks through CASP 1994→2024 with the GDT numbers explained verbally |
| 5 – AF2 Architecture | Full technical walkthrough: MSA, Evoformer (triangle attention), IPA, recycling |
| 6 – Training & pLDDT | FAPE loss, training data, pLDDT scale explained with clinical significance |
| 7 – AlphaFold 3 | Pairformer vs Evoformer, diffusion module explained via analogy, drug discovery angle |
| 8 – Applications | Nuclear pore complex, AlphaMissense, COVID, malaria — concrete examples |
| 9 – AFDB | 214M structures, dark proteome, limitations stated clearly |
| 10 – Nobel Prize | Profiles of all three laureates, Nobel Committee quote |
| 11 – Limitations | Six limitations explained with genuine depth — not glossed over |
| 12 – Future | Six future directions with specific tools named |
| 13 – Comparison | Verbal summary of the version comparison table |
| 14 – References | Brief note on each citation |
| 15 – Close | Key takeaways + invitation for questions |
Appendix: 6 anticipated audience questions with full answers (membrane proteins, clinical use, Rosetta comparison, licensing, ColabFold, etc.).
in word file
Reading File
Reading File
Writing File
~/alphafold-ppt/build_docx.js
const {
Document, Packer, Paragraph, TextRun, Table, TableRow, TableCell,
Header, Footer, AlignmentType, HeadingLevel, LevelFormat, BorderStyle,
WidthType, VerticalAlign, PageNumber, PageBreak, ShadingType,
UnderlineType
} = require('docx');
const fs = require('fs');
// ─── COLOUR HELPERS (hex without #) ───
const TEAL = "006B6B";
const NAVY = "0A3560";
const GOLD = "B8860B";
const GREY = "555555";
const LGREY = "777777";
const BLACK = "000000";
const WHITE = "FFFFFF";
const RED = "AA0000";
// ─── HELPER: plain body paragraph ───
function body(text, opts = {}) {
return new Paragraph({
children: [new TextRun({
text,
size: 24,
font: "Arial",
color: BLACK,
bold: opts.bold || false,
italics: opts.italic || false,
})],
spacing: { after: 120, before: opts.before || 0 },
alignment: opts.align || AlignmentType.LEFT,
});
}
// ─── HELPER: speaker cue paragraph (teal italic) ───
function cue(text) {
return new Paragraph({
children: [new TextRun({
text,
size: 22,
font: "Arial",
color: TEAL,
italics: true,
})],
spacing: { after: 80, before: 60 },
});
}
// ─── HELPER: bullet ───
function bullet(text, ref = "bullets", bold = false) {
return new Paragraph({
numbering: { reference: ref, level: 0 },
children: [new TextRun({ text, size: 24, font: "Arial", color: BLACK, bold })],
spacing: { after: 80 },
});
}
// ─── HELPER: Heading 1 (slide title) ───
function h1(text) {
return new Paragraph({
heading: HeadingLevel.HEADING_1,
children: [new TextRun({ text, size: 36, bold: true, font: "Arial", color: WHITE })],
shading: { type: ShadingType.SOLID, color: NAVY, fill: NAVY },
spacing: { before: 320, after: 160 },
});
}
// ─── HELPER: Heading 2 (sub-section) ───
function h2(text, color) {
const c = color || TEAL;
return new Paragraph({
heading: HeadingLevel.HEADING_2,
children: [new TextRun({ text, size: 28, bold: true, font: "Arial", color: c })],
spacing: { before: 200, after: 100 },
border: { bottom: { color: c, space: 1, style: BorderStyle.SINGLE, size: 6 } },
});
}
// ─── HELPER: Heading 3 (note label) ───
function h3(text, color) {
return new Paragraph({
children: [new TextRun({ text, size: 24, bold: true, font: "Arial", color: color || GOLD })],
spacing: { before: 160, after: 80 },
});
}
// ─── HELPER: page break ───
function pageBreak() {
return new Paragraph({
children: [new PageBreak()],
spacing: { after: 0 },
});
}
// ─── HELPER: horizontal rule (via border) ───
function rule() {
return new Paragraph({
children: [new TextRun({ text: "" })],
border: { bottom: { color: "CCCCCC", space: 1, style: BorderStyle.SINGLE, size: 4 } },
spacing: { before: 120, after: 120 },
});
}
// ─── HELPER: label + body inline ───
function labelBody(label, text) {
return new Paragraph({
children: [
new TextRun({ text: label + " ", size: 24, bold: true, font: "Arial", color: TEAL }),
new TextRun({ text, size: 24, font: "Arial", color: BLACK }),
],
spacing: { after: 100 },
});
}
// ─── HELPER: Q&A item ───
function qa(q, a) {
return [
new Paragraph({
children: [new TextRun({ text: q, size: 24, bold: true, font: "Arial", color: NAVY })],
spacing: { before: 160, after: 60 },
}),
new Paragraph({
children: [new TextRun({ text: a, size: 24, font: "Arial", color: BLACK })],
spacing: { after: 120 },
}),
];
}
// ─────────────────────────────────────────────────────────────────
// BUILD DOCUMENT
// ─────────────────────────────────────────────────────────────────
const doc = new Document({
numbering: {
config: [
{
reference: "bullets",
levels: [{
level: 0, format: LevelFormat.BULLET, text: "\u2022",
alignment: AlignmentType.LEFT,
style: { paragraph: { indent: { left: 720, hanging: 360 } } },
}],
},
{
reference: "dash",
levels: [{
level: 0, format: LevelFormat.BULLET, text: "\u2013",
alignment: AlignmentType.LEFT,
style: { paragraph: { indent: { left: 720, hanging: 360 } } },
}],
},
],
},
styles: {
default: {
document: { run: { font: "Arial", size: 24, color: BLACK } },
},
paragraphStyles: [
{
id: "Heading1", name: "Heading 1", basedOn: "Normal", next: "Normal", quickFormat: true,
run: { size: 36, bold: true, font: "Arial", color: WHITE },
paragraph: {
spacing: { before: 320, after: 160 },
outlineLevel: 0,
shading: { type: ShadingType.SOLID, color: NAVY, fill: NAVY },
},
},
{
id: "Heading2", name: "Heading 2", basedOn: "Normal", next: "Normal", quickFormat: true,
run: { size: 28, bold: true, font: "Arial", color: TEAL },
paragraph: { spacing: { before: 200, after: 100 }, outlineLevel: 1 },
},
{
id: "Heading3", name: "Heading 3", basedOn: "Normal", next: "Normal", quickFormat: true,
run: { size: 24, bold: true, font: "Arial", color: GOLD },
paragraph: { spacing: { before: 160, after: 80 }, outlineLevel: 2 },
},
],
},
sections: [{
properties: {
page: {
margin: { top: 1080, bottom: 1080, left: 1260, right: 1260 },
},
},
headers: {
default: new Header({
children: [
new Paragraph({
children: [
new TextRun({ text: "AlphaFold Seminar — Presenter Script", size: 18, font: "Arial", color: LGREY }),
],
alignment: AlignmentType.RIGHT,
border: { bottom: { color: TEAL, space: 1, style: BorderStyle.SINGLE, size: 4 } },
spacing: { after: 40 },
}),
],
}),
},
footers: {
default: new Footer({
children: [
new Paragraph({
children: [
new TextRun({ text: "May 2026 | Sources: Nature 2021/2024, Nucleic Acids Res 2024, Nobel Committee 2024 | Page ", size: 18, font: "Arial", color: LGREY }),
new TextRun({ children: [PageNumber.CURRENT], size: 18, font: "Arial", color: LGREY }),
],
alignment: AlignmentType.CENTER,
border: { top: { color: "CCCCCC", space: 1, style: BorderStyle.SINGLE, size: 4 } },
spacing: { before: 40 },
}),
],
}),
},
children: [
// ─── COVER PAGE ───────────────────────────────────────────
new Paragraph({
children: [new TextRun({ text: " ", size: 24 })],
spacing: { after: 480 },
}),
new Paragraph({
children: [new TextRun({ text: "AlphaFold", size: 72, bold: true, font: "Arial", color: NAVY })],
alignment: AlignmentType.CENTER,
spacing: { after: 100 },
}),
new Paragraph({
children: [new TextRun({ text: "Revolutionizing Protein Structure Prediction with AI", size: 32, font: "Arial", color: TEAL })],
alignment: AlignmentType.CENTER,
spacing: { after: 80 },
}),
new Paragraph({
children: [new TextRun({ text: "Seminar Presenter Script", size: 28, bold: true, font: "Arial", color: GOLD })],
alignment: AlignmentType.CENTER,
spacing: { after: 80 },
}),
rule(),
new Paragraph({
children: [new TextRun({ text: "15 Slides | ~45–55 Minutes Delivery | + Appendix Q&A", size: 24, font: "Arial", color: LGREY, italics: true })],
alignment: AlignmentType.CENTER,
spacing: { after: 80 },
}),
new Paragraph({
children: [new TextRun({ text: "Sources: Jumper et al. Nature 2021 | Abramson et al. Nature 2024 | Varadi et al. NAR 2024 | Nobel Committee 2024", size: 20, font: "Arial", color: LGREY, italics: true })],
alignment: AlignmentType.CENTER,
spacing: { after: 600 },
}),
new Paragraph({
children: [new TextRun({ text: "HOW TO USE THIS SCRIPT", size: 24, bold: true, font: "Arial", color: TEAL })],
alignment: AlignmentType.CENTER,
spacing: { after: 100 },
}),
new Paragraph({
children: [new TextRun({
text: "Each section below corresponds exactly to one slide in the AlphaFold_Seminar.pptx presentation. " +
"The slide number and title are shown at the top of each section. " +
"Stage directions appear in teal italics — e.g. [Click to next slide]. " +
"Sub-headings in gold mark key talking points. " +
"The Appendix at the end provides answers to six common audience questions.",
size: 24, font: "Arial", color: BLACK,
})],
spacing: { after: 480 },
alignment: AlignmentType.LEFT,
}),
pageBreak(),
// ─── SLIDE 1 ──────────────────────────────────────────────
h1("SLIDE 1 — Title Slide"),
cue("[Display as the audience enters. Begin speaking once the room has settled.]"),
body("Good [morning / afternoon / evening], everyone. Thank you for being here."),
body("Today we are going to talk about one of the most consequential scientific breakthroughs of the past two decades — AlphaFold."),
body("If you work in biology, medicine, pharmacology, or even computer science, this technology has already changed the landscape of your field — whether you know it yet or not."),
body("By the end of this talk, you will understand what AlphaFold does, how it works at a technical level, why it matters, and — importantly — what it cannot do. We will cover everything from the foundational biology problem it solves, to the deep learning architecture behind it, to its 2024 Nobel Prize, and where the field is heading next."),
body("Let's start at the beginning."),
pageBreak(),
// ─── SLIDE 2 ──────────────────────────────────────────────
h1("SLIDE 2 — Presentation Outline"),
cue("[Click to slide 2.]"),
body("Here is the structure of today's talk. We have ten core topics."),
body("We begin with the biological problem AlphaFold was built to solve — the Protein Folding Problem. We then trace the history of AlphaFold from its first appearance in 2018 through to AlphaFold 3 in 2024."),
body("After that, we go deep into the architecture — the Evoformer, the Structure Module, how training works, and how the model scores its own confidence."),
body("We then look at AlphaFold 3 specifically — which is a fundamentally different model from AlphaFold 2. From there we examine real-world applications, the public database, the Nobel Prize, the limitations, and finally the future."),
body("It is a lot to cover, so let's move quickly but thoroughly."),
pageBreak(),
// ─── SLIDE 3 ──────────────────────────────────────────────
h1("SLIDE 3 — The Protein Folding Problem"),
cue("[Click to slide 3.]"),
body("To understand AlphaFold, you first have to understand the problem it was built to solve."),
h2("What Are Proteins?"),
body("Proteins are the molecular machines of life. They catalyse biochemical reactions, carry oxygen in your blood, fight off pathogens, transmit signals between cells, and build the physical structures of your body. Every protein starts as a linear chain of amino acids — a string of chemical building blocks encoded by your DNA."),
body("That string spontaneously folds, in milliseconds, into a precise three-dimensional shape. And that shape determines the protein's function. A misfolded protein can cause disease — Alzheimer's, Parkinson's, cystic fibrosis, many cancers all involve proteins that are misfolded or misfunctioning."),
h2("The Core Question"),
body("So the central question is: given only the sequence of amino acids, can you predict the 3D structure?"),
body("This is called the Protein Folding Problem, and for about 50 years, it resisted solution."),
h2("The Scale of the Gap"),
body("There are around 20,000 protein-coding genes in the human genome, producing potentially 100,000 distinct protein forms. But experimentally solving a structure — using X-ray crystallography, cryo-electron microscopy, or NMR — is slow, expensive, and often fails for membrane proteins or disordered regions. By 2020, the Protein Data Bank contained around 170,000 experimentally determined structures. Meanwhile, protein sequence databases held hundreds of millions of sequences. That gap between known sequences and known structures is what AlphaFold was built to close."),
h2("Anfinsen's Dogma & Levinthal's Paradox"),
body("The famous insight comes from Christian Anfinsen, who won the 1972 Nobel Prize in Chemistry for demonstrating that a protein's amino acid sequence alone contains all the information needed to determine its final 3D shape. That principle — Anfinsen's dogma — is the theoretical foundation AlphaFold rests on."),
body("And then there is Levinthal's Paradox. Cyrus Levinthal calculated in 1969 that if a 100-amino-acid protein tried every possible conformation randomly, it would take longer than the age of the universe to find the right fold. Yet proteins fold in microseconds. That means biology has found shortcuts — guided folding pathways. The challenge was figuring out what those shortcuts are, and encoding them in a computer program."),
pageBreak(),
// ─── SLIDE 4 ──────────────────────────────────────────────
h1("SLIDE 4 — History & CASP Benchmarks"),
cue("[Click to slide 4.]"),
body("So how do you benchmark progress on this problem? The answer is CASP — the Critical Assessment of Protein Structure Prediction — a biennial competition that has been running since 1994."),
h2("How CASP Works"),
body("The idea is elegant. Experimental groups determine new protein structures but keep them secret. Computational groups are given only the amino acid sequences and asked to predict the structures. The predictions are then compared against the experimental results and scored using a metric called GDT — Global Distance Test — where 100 is a perfect match."),
h2("Decades of Incremental Progress"),
body("For decades, progress was incremental. By CASP12 in 2016, the best methods were achieving GDT scores in the mid-40s for the hardest targets — so-called free-modelling targets where no close template exists. These methods used co-evolutionary analysis and physics-based folding simulations, and they were good, but far from reliable."),
h2("AlphaFold 1 — CASP13 (2018)"),
body("Then DeepMind entered. Their first system, AlphaFold 1, appeared at CASP13 in 2018. It jumped the GDT score to 68.5 for free-modelling targets — a significant leap. It won the competition but researchers noted it was still far from experimental accuracy. Promising, but not yet transformative."),
h2("AlphaFold 2 — CASP14 (2020): The Breakthrough"),
body("Two years later, at CASP14 in 2020, AlphaFold 2 appeared. And the scientific community was stunned."),
body("AlphaFold 2 achieved a median GDT score of 92.4 on the free-modelling targets. The next-best method scored in the 50s. Organisers called it a solution to the protein folding problem. John Moult, who co-founded CASP, said it was 'a stunning advance' — the kind of result that comes once in a generation."),
body("You can see this in the bar chart on the slide. The jump from CASP13 to CASP14 is not incremental — it is transformative."),
h2("2021–2024: Database, Multimer, AlphaFold 3"),
body("Between 2020 and 2024, DeepMind released the AlphaFold Protein Structure Database, first with 350,000 structures, then expanding to cover the entire known protein universe — over 214 million structures. And in May 2024, they released AlphaFold 3, which expanded beyond proteins entirely."),
pageBreak(),
// ─── SLIDE 5 ──────────────────────────────────────────────
h1("SLIDE 5 — AlphaFold 2 Architecture"),
cue("[Click to slide 5.]"),
body("Now let's get into the machine itself. How does AlphaFold 2 actually work?"),
body("The architecture shown on this slide is an end-to-end differentiable deep learning system — meaning every component is trained together, jointly optimising a single objective: predict the 3D structure as accurately as possible."),
body("Let me walk you through each stage."),
h2("Stage 1: Input"),
body("You give the model an amino acid sequence — a string of letters, each representing one of the 20 standard amino acids."),
h2("Stage 2: Multiple Sequence Alignment (MSA)"),
body("The system then searches enormous protein sequence databases — UniRef90, BFD with 2.2 billion clusters, MGnify — to find evolutionarily related proteins from other species. These are aligned into a matrix called a Multiple Sequence Alignment, or MSA."),
body("Why does this matter? Because evolution is informative. If two amino acids in a protein have co-evolved — meaning when one mutates, the other tends to mutate to compensate — that is a strong signal they are spatially close in the 3D structure. The MSA encodes millions of years of evolutionary experiments. Deeper MSAs, meaning more related sequences found, generally mean more accurate predictions."),
h2("Stage 3: Evoformer"),
body("This is the heart of AlphaFold 2. The Evoformer consists of 48 stacked transformer-like blocks, each of which jointly processes two representations: the MSA representation — capturing information across related sequences — and a pairwise representation — capturing relationships between every pair of amino acid positions."),
body("The key innovation here is that these two representations update each other. The MSA representation informs the pair representation, and vice versa, through what the authors call 'triangle multiplicative updates' and row- and column-wise attention. This iterative communication allows the model to reason about both sequence conservation and spatial geometry simultaneously."),
h2("Stage 4: Structure Module and Invariant Point Attention (IPA)"),
body("The Evoformer outputs a refined pairwise representation and an updated single-sequence representation. The Structure Module takes these and predicts the actual 3D coordinates of every atom."),
body("The key component here is Invariant Point Attention — IPA. This is an attention mechanism designed to respect 3D geometry. It predicts a rigid-body frame — a rotation and translation — for each amino acid residue, and then the side-chain torsion angles within each residue. Because it operates in 3D space directly, the predictions are equivariant — they give the right answer regardless of how the input structure is oriented."),
h2("Stage 5: Recycling"),
body("The entire pipeline is run three times. Each iteration takes the previous iteration's predicted structure as an additional input, allowing the model to progressively refine its prediction. This recycling mechanism was one of the critical contributions that pushed accuracy to near-experimental levels."),
h2("Output"),
body("The final output is a full 3D atomic model of the protein, plus a per-residue confidence score — the pLDDT — which we will discuss on the next slide."),
rule(),
new Paragraph({
children: [new TextRun({ text: "Reference: Jumper J et al. Highly accurate protein structure prediction with AlphaFold. Nature. 2021;596:583-589. PMID: 34265844", size: 18, font: "Arial", color: LGREY, italics: true })],
spacing: { after: 100 },
}),
pageBreak(),
// ─── SLIDE 6 ──────────────────────────────────────────────
h1("SLIDE 6 — Training Pipeline & Confidence Metrics"),
cue("[Click to slide 6.]"),
body("Let's talk about how the model was trained and how it communicates its own uncertainty."),
h2("Training Data Sources"),
body("The model was trained on the Protein Data Bank — approximately 170,000 experimentally determined structures at the time. These are the ground-truth labels. The sequences paired with those structures were used, along with the giant databases shown here — UniRef90 and UniClust30 for constructing MSAs, and BFD, which contains 2.2 billion clustered sequences largely derived from environmental metagenomics. This massive sequence diversity is critical because it provides the co-evolutionary signal the Evoformer learns to exploit."),
h2("Loss Function"),
body("The training loss function is called FAPE — Frame Aligned Point Error. It measures the average distance between predicted and true atomic positions across all reference frames defined by backbone residues. This frame-aligned approach makes the loss invariant to global rotation and translation of the whole structure, which is mathematically important for training stability. Additional losses include torsion angle loss, distogram loss for predicted distance distributions, and a local distance difference test head."),
body("Training required approximately 128 Google TPUv3 chips running for several weeks — a substantial compute investment, though modest compared to large language models."),
h2("pLDDT Confidence Score — A Critical Concept"),
body("The pLDDT — per-residue Local Distance Difference Test — is AlphaFold's self-assessment of its own prediction quality at each amino acid position. It ranges from 0 to 100."),
body("The colour-coded scale on the slide shows four confidence bands:"),
bullet("Above 90 — Very high confidence. These regions are generally accurate enough for drug docking studies, binding site characterisation, and functional annotation.", "bullets", true),
bullet("70 to 90 — Confident. The backbone is reliably predicted. Suitable for most downstream analyses."),
bullet("50 to 70 — Low confidence. Use with caution. These regions may be flexible or only conditionally ordered."),
bullet("Below 50 — Very low confidence. If you visualise these regions in PyMOL or ChimeraX, they often look like random coils — sometimes described as 'spaghetti.' These regions are likely intrinsically disordered and should not be used for structural analysis."),
body("The pLDDT score is stored in the B-factor field of the output PDB files — a clever repurposing of an existing file format field — and is colour-coded in all AlphaFold DB visualisations."),
body("Understanding pLDDT is not optional. Every researcher using AlphaFold predictions needs to check the confidence scores for their region of interest before drawing conclusions.", { bold: true }),
rule(),
new Paragraph({
children: [new TextRun({ text: "Source: Jumper et al., Nature 2021 | AlphaFold DB 2024, Varadi M et al., Nucleic Acids Res 2024 (PMC10767828)", size: 18, font: "Arial", color: LGREY, italics: true })],
spacing: { after: 100 },
}),
pageBreak(),
// ─── SLIDE 7 ──────────────────────────────────────────────
h1("SLIDE 7 — AlphaFold 3 (2024)"),
cue("[Click to slide 7.]"),
body("In May 2024, Google DeepMind and its sister company Isomorphic Labs published AlphaFold 3 in Nature. It is not simply an updated AlphaFold 2 — it is a fundamentally different model with a fundamentally different scope."),
body("AlphaFold 2 was built for proteins. AlphaFold 3 was built for biomolecular interactions."),
h2("Change 1: Pairformer Replaces Evoformer"),
body("AlphaFold 3 drops the full MSA processing stack and replaces the Evoformer with a simpler but powerful module called the Pairformer. The Pairformer processes a single sequence representation alongside pairwise representations, using learned template embeddings where available. This simplification makes the architecture more generalisable across molecule types — because you cannot build an MSA for a small molecule ligand."),
h2("Change 2: Diffusion Module Replaces Structure Module"),
body("This is the most dramatic architectural change. Instead of the deterministic Invariant Point Attention-based structure module, AlphaFold 3 uses a generative diffusion model."),
body("If you are familiar with image diffusion models like DALL-E or Stable Diffusion — the idea is analogous. The model starts with a random cloud of atoms in 3D space, and iteratively denoises them, guided by the Pairformer's output, until they converge on a physically plausible structure. This approach naturally handles the full atom detail of small molecules, nucleic acids, and post-translational modifications — all of which have very different atomic chemistry from protein backbone atoms."),
h2("Change 3: Multi-Molecular Scope"),
body("AlphaFold 3 can model proteins, DNA, RNA, small-molecule ligands, ions, and post-translational modifications like glycosylation and phosphorylation — all in one prediction. The authors reported at least 50% improvement in accuracy for protein-nucleic acid interfaces compared to existing specialist methods, and significant improvements for protein-ligand prediction."),
h2("Drug Discovery Integration"),
body("This last point is not incidental. Isomorphic Labs was co-developer on AlphaFold 3, and their explicit goal is to use it for drug discovery. Understanding how a candidate drug molecule binds to a disease-relevant protein is one of the most important — and historically expensive — steps in pharmaceutical development. AlphaFold 3 directly addresses this."),
body("A free AlphaFold Server was launched simultaneously for non-commercial research, accessible at alphafoldserver.com."),
rule(),
new Paragraph({
children: [new TextRun({ text: "Source: Abramson J et al., Nature 2024 (PMID: 38594878) | AlphaFold Server: alphafoldserver.com", size: 18, font: "Arial", color: LGREY, italics: true })],
spacing: { after: 100 },
}),
pageBreak(),
// ─── SLIDE 8 ──────────────────────────────────────────────
h1("SLIDE 8 — Applications & Scientific Impact"),
cue("[Click to slide 8.]"),
body("Let me now bring this out of the technical realm and into concrete scientific impact. What has AlphaFold actually been used for?"),
h2("Drug Discovery"),
body("AlphaFold is being used to identify and characterise binding pockets — the sites on a protein where small molecules can bind and modulate its function. Isomorphic Labs has reported active drug discovery programmes using AlphaFold-derived structural insight against targets including tuberculosis, Chagas disease, and antibiotic-resistant bacteria. The key advantage is speed: structure determination that once required months of crystallography experiments can now be initiated computationally in minutes."),
h2("Rare Genetic Diseases — AlphaMissense"),
body("In 2023, DeepMind published AlphaMissense in Science. This is an AlphaFold-derived model that classifies missense mutations — single amino acid changes caused by point mutations — as likely benign, likely pathogenic, or uncertain. It classified approximately 71 million possible missense variants across the human proteome. Around 89% received a confident classification. For rare disease diagnosis, where identifying whether a patient's unique variant is disease-causing is critical, this tool is transformative."),
h2("Structural Biology Acceleration"),
body("One of the most dramatic examples is the nuclear pore complex — a massive protein assembly of around 120 proteins that regulates what enters and exits the cell nucleus. Experimental structural determination of this complex had been a decades-long project. In 2022, a team at Harvard published a near-complete model by integrating AlphaFold predictions with sparse cryo-EM data. What might have taken another decade was accomplished in a fraction of that time."),
body("AlphaFold structures are now routinely used to interpret cryo-EM maps — where the electron density is known at moderate resolution but the atomic model needs to be fitted — dramatically accelerating structure determination."),
h2("Antimicrobial and Pandemic Research"),
body("AlphaFold was used to model all major proteins of SARS-CoV-2 shortly after the pandemic began, aiding in vaccine target identification and antiviral design. For malaria, the structures of key Plasmodium falciparum proteins — many of which resisted experimental determination — are now accessible, opening new avenues for vaccine and drug development."),
rule(),
new Paragraph({
children: [new TextRun({ text: "Source: AlphaMissense - Cheng J et al., Science 2023 | Nuclear pore - Fontana P et al., Science 2022 | DeepMind blog", size: 18, font: "Arial", color: LGREY, italics: true })],
spacing: { after: 100 },
}),
pageBreak(),
// ─── SLIDE 9 ──────────────────────────────────────────────
h1("SLIDE 9 — AlphaFold Protein Structure Database"),
cue("[Click to slide 9.]"),
body("The AlphaFold Protein Structure Database — hosted by the European Bioinformatics Institute at EMBL-EBI — is possibly the most consequential scientific data resource created in the past decade."),
body("As of 2024, it contains over 214 million predicted protein structures — covering virtually every protein in the UniProt sequence database. That is orders of magnitude more structural data than has been accumulated experimentally in the entire history of structural biology."),
h2("Access & Formats"),
body("Access is completely free under a Creative Commons Attribution licence. Structures are available via the web browser, programmatic API, and bulk FTP download. Files come in PDB, mmCIF, and binaryCIF formats. The database is integrated with UniProt, PDBe, and NCBI."),
h2("Coverage Highlights"),
body("The database covers full proteomes for 48 organisms, including all major model organisms — E. coli, yeast, Arabidopsis, fruit fly, nematode, and mouse — as well as the complete human proteome with all 20,386 canonical proteins. There is a strong focus on global health: complete proteomes of Mycobacterium tuberculosis, Plasmodium falciparum, and Leishmania species are included."),
h2("The Dark Proteome"),
body("One of the most exciting aspects is what researchers call the dark proteome — proteins with no known structure and often no known function. Around 35% of human proteins were previously uncharacterised structurally. AlphaFold illuminates this dark proteome, providing structural hypotheses for proteins that were previously black boxes."),
h2("Critical Caveat — Use With Care"),
body("However, and I want to be clear about this, the database has known limitations.", { bold: true }),
body("Every prediction represents a single static, equilibrium conformation. There are no dynamics. There are no alternative conformations. Low-confidence regions should not be used for drug docking. Multimeric predictions from AF-Multimer carry additional uncertainty. And all results benefit from experimental validation before high-stakes decisions are made. The database is a starting point — a powerful one — not a replacement for experimental structural biology."),
rule(),
new Paragraph({
children: [new TextRun({ text: "Source: Varadi M et al. AlphaFold Protein Structure Database in 2024. Nucleic Acids Res. 2024;52:D368-D375. PMC10767828", size: 18, font: "Arial", color: LGREY, italics: true })],
spacing: { after: 100 },
}),
pageBreak(),
// ─── SLIDE 10 ─────────────────────────────────────────────
h1("SLIDE 10 — Nobel Prize in Chemistry 2024"),
cue("[Click to slide 10.]"),
body("In October 2024, the Royal Swedish Academy of Sciences awarded the Nobel Prize in Chemistry to three scientists."),
body("Demis Hassabis and John Jumper received one half of the prize for protein structure prediction. David Baker of the University of Washington received the other half for computational protein design — which is the complementary problem of designing new proteins with desired functions, rather than predicting the structures of existing ones."),
h2("Demis Hassabis — CEO, Google DeepMind", TEAL),
body("Demis Hassabis founded DeepMind in 2010 with the mission of using artificial intelligence to accelerate scientific discovery. He recognised that protein structure prediction was a problem ideally suited to deep learning — high-dimensional, data-rich, and with a clear objective function. He drove the strategic direction of the AlphaFold project and led the organisation that made it possible."),
h2("John Jumper — Lead Researcher, Google DeepMind", TEAL),
body("John Jumper was the principal technical architect of AlphaFold 2. His specific innovations — the Evoformer with its triangle attention mechanisms, the Invariant Point Attention module, the recycling strategy, and the FAPE loss function — are what made the CASP14 result possible. Before DeepMind, he worked at D.E. Shaw Research on protein simulation. His combination of deep learning expertise and structural biology knowledge was precisely the interdisciplinary synthesis the problem required."),
h2("David Baker — University of Washington", TEAL),
body("David Baker at the University of Washington has spent three decades developing computational protein design tools, culminating in the Rosetta software suite and more recently diffusion-based design approaches. His work represents the other side of the structural biology coin: not predicting what nature made, but designing what nature has not yet made."),
h2("The Nobel Committee's Statement"),
body("The Nobel Committee's statement was striking. They said AlphaFold solved a problem that had confronted biochemists for half a century, and that it had immediately democratised access to structural information that was previously reserved for well-resourced experimental labs."),
body("Both Hassabis and Jumper had already received the Breakthrough Prize in Life Sciences and the Albert Lasker Award for Basic Medical Research in 2023 — so this Nobel Prize, while historic, was not a surprise to the scientific community."),
rule(),
new Paragraph({
children: [new TextRun({ text: "Source: Nobel Prize Committee, Royal Swedish Academy of Sciences 2024 | NobelPrize.org", size: 18, font: "Arial", color: LGREY, italics: true })],
spacing: { after: 100 },
}),
pageBreak(),
// ─── SLIDE 11 ─────────────────────────────────────────────
h1("SLIDE 11 — Limitations & Critical Perspectives"),
cue("[Click to slide 11.]"),
body("I want to spend real time on limitations, because any tool used uncritically is a tool used dangerously."),
h2("1. Static Structures Only", RED),
body("This is probably the most important limitation. Proteins are not rigid objects — they breathe, flex, change shape upon binding partners, and populate multiple conformational states. Many drug targets, particularly GPCRs and ion channels, are interesting precisely because of their conformational dynamics. AlphaFold gives you one conformation — typically the lowest-energy ground state. It tells you nothing about the range of motion, the transition pathways, or the alternative states that a drug might need to target."),
h2("2. Intrinsically Disordered Regions", RED),
body("Many proteins — particularly transcription factors, signalling proteins, and hub proteins in interaction networks — contain large intrinsically disordered regions, or IDRs. These regions do not adopt a fixed 3D structure; they are functionally flexible. AlphaFold correctly assigns low pLDDT scores to these regions, but the 3D coordinates it assigns are essentially meaningless. The problem is that some users ignore the pLDDT and treat the coordinates as real. They are not."),
h2("3. Hallucination Risk in AlphaFold 3", RED),
body("Because AF3 uses a generative diffusion model, it is subject to a class of errors familiar from generative AI — hallucination. Specifically, it can generate structures with physically impossible bond lengths, clashing atoms, or geometrically implausible backbone conformations. Post-processing validation using tools like MolProbity or OpenStructure is strongly recommended for any result used in downstream analysis or publication."),
h2("4. Shallow MSA Problem", GREY),
body("For proteins with few evolutionary relatives — newly evolved proteins, synthetic sequences, orphan proteins — the MSA is shallow or empty. AlphaFold's accuracy drops significantly in these cases because the co-evolutionary signal it depends on is absent. This is a fundamental limitation of the evolutionary information approach."),
h2("5. Protein Complex Accuracy", GREY),
body("AF-Multimer's accuracy for predicting protein-protein interactions, while impressive, is lower than for monomers. Weak or transient interactions are particularly difficult. Experimentally, these require pulldown assays, cross-linking mass spectrometry, or other binding assays that AlphaFold cannot replace."),
h2("6. Ethical and Policy Dimensions", GREY),
body("Finally, it is worth acknowledging the societal dimensions. The partial restriction on AlphaFold 3 source code — where model weights were released for non-commercial use only — generated significant debate about open science norms in the AI era. There are also genuine dual-use concerns: the same structural knowledge that accelerates drug development could, in principle, be used for harmful purposes. And like many computational advances, there is a question of access equity — whether researchers in low-resource settings can realistically leverage these tools."),
body("These are not reasons to avoid AlphaFold. They are reasons to use it thoughtfully.", { bold: true }),
rule(),
new Paragraph({
children: [new TextRun({ text: "Source: Elofsson A, Curr Opin Struct Biol 2023 (PMID 37060758) | Chen L et al., Int J Mol Sci 2024 (PMID 39125995)", size: 18, font: "Arial", color: LGREY, italics: true })],
spacing: { after: 100 },
}),
pageBreak(),
// ─── SLIDE 12 ─────────────────────────────────────────────
h1("SLIDE 12 — Future Directions"),
cue("[Click to slide 12.]"),
body("Where is all of this heading?"),
h2("1. Protein Design (Inverse Folding)"),
body("The most exciting frontier is inverting the prediction problem: instead of predicting structure from sequence, designing a sequence that will fold into a desired structure. Tools like RFdiffusion and ProteinMPNN from the Baker lab now use AlphaFold-derived representations to design novel proteins from scratch — enzymes, binders, biosensors — with properties not found in nature. The first computationally designed protein-based drugs are moving toward clinical testing."),
h2("2. Dynamics and Conformational Ensembles"),
body("New approaches like AlphaFlow and EigenFold are extending the AlphaFold framework to predict not just a single structure but an ensemble of conformations — a distribution over the structural landscape a protein can occupy. Combined with molecular dynamics simulation, this will eventually give us a complete picture of protein motion, not just a snapshot. Critical for GPCR agonist/antagonist modelling."),
h2("3. Multimodal Integration"),
body("A November 2024 paper from Linkoping University demonstrated that AlphaFold can be improved by incorporating sparse experimental data — partial cryo-EM maps, chemical cross-links, SAXS profiles — as constraints during prediction. Hybrid experimental-computational pipelines are rapidly becoming standard practice, and the boundary between computational prediction and experimental determination is blurring."),
h2("4. Whole-Proteome Interactome Mapping"),
body("Researchers are applying AF-Multimer at scale to systematically predict all protein-protein interactions in the human interactome — a network of potentially hundreds of thousands of interactions. This will provide a structural atlas of cellular biology that was previously unimaginable."),
h2("5. Therapeutic Applications"),
body("Isomorphic Labs has reported that AlphaFold 3-guided drug candidates have entered preclinical development pipelines. The timeline from target identification to clinical trial entry may compress dramatically. Personalised cancer immunotherapy — designing neoantigen vaccines tailored to a patient's specific tumour mutations — is another area where structural prediction is enabling new therapeutic strategies."),
h2("6. Democratisation"),
body("And perhaps most importantly for the long-term health of science: the AlphaFold Server is free, the database is open access, and ColabFold allows AlphaFold 2 to be run on Google Colab with no specialist hardware. A graduate student anywhere in the world can now do structural biology that ten years ago required a major research institution with crystallography facilities. That democratisation of access is itself a scientific revolution."),
rule(),
new Paragraph({
children: [new TextRun({ text: "Source: Krokidis MG et al., Int J Mol Sci 2025 (PMID 40332289) | Wuyun Q et al., Molecules 2024 (PMID 38398585) | ScienceDaily Nov 2024", size: 18, font: "Arial", color: LGREY, italics: true })],
spacing: { after: 100 },
}),
pageBreak(),
// ─── SLIDE 13 ─────────────────────────────────────────────
h1("SLIDE 13 — Version Comparison Table"),
cue("[Click to slide 13.]"),
body("Let me give you a quick side-by-side summary of the three generations."),
h2("AlphaFold 1 (2018)"),
body("AlphaFold 1 was a convolutional neural network that predicted distance distributions between residue pairs — it did not directly predict 3D coordinates. It used sequence and MSA as input and achieved a GDT of about 68 at CASP13. Impressive for its time, but not yet accurate enough for most applications."),
h2("AlphaFold 2 (2020)"),
body("AlphaFold 2 was the breakthrough version. New architecture — Evoformer plus Structure Module with Invariant Point Attention. Direct prediction of 3D atomic coordinates. GDT of 92.4 at CASP14. Introduced the pLDDT confidence score. Released as fully open source with the AlphaFold DB launch in 2021."),
h2("AlphaFold 3 (2024)"),
body("AlphaFold 3 replaced the Evoformer with the simpler Pairformer and replaced the Structure Module with a diffusion model. Extended scope to DNA, RNA, ligands, and ions. Not evaluated at CASP — the competition format does not cover its new molecular scope. Released with partial restrictions on commercial use, though the AlphaFold Server is free for non-commercial research."),
body("Each version represents a genuine architectural reimagining, not just incremental tuning.", { bold: true }),
pageBreak(),
// ─── SLIDE 14 ─────────────────────────────────────────────
h1("SLIDE 14 — Key References"),
cue("[Click to slide 14.]"),
body("Here are the primary sources underpinning everything I have discussed today."),
h2("Primary Papers"),
bullet("Jumper J et al. Highly accurate protein structure prediction with AlphaFold. Nature. 2021;596(7873):583-589. PMID: 34265844. — The foundational AlphaFold 2 paper, one of the most-cited biology papers of the decade."),
bullet("Abramson J et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature. 2024;630(8016):493-500. PMID: 38594878. — Already cited over 9,000 times as of late 2025."),
bullet("Senior AW, Evans R, Jumper J et al. Improved protein structure prediction using potentials from deep learning. Nature. 2020;577:706-710. PMID: 31942072. — AlphaFold 1."),
h2("Database & Reviews"),
bullet("Varadi M et al. AlphaFold Protein Structure Database in 2024. Nucleic Acids Res. 2024;52(D1):D368-D375. PMC10767828. — Read this if you plan to use the database."),
bullet("Krokidis MG et al. AlphaFold3: An Overview of Applications and Performance Insights. Int J Mol Sci. 2025;26(8):3627. PMID: 40332289."),
bullet("Chen L et al. AI-Driven Deep Learning Techniques in Protein Structure Prediction. Int J Mol Sci. 2024;25(15):8426. PMID: 39125995."),
bullet("Elofsson A. Progress at protein structure prediction, as seen in CASP15. Curr Opin Struct Biol. 2023;80:102594. PMID: 37060758."),
h2("Authoritative Source"),
bullet("Nobel Prize Committee. The Nobel Prize in Chemistry 2024 - Scientific Background. Royal Swedish Academy of Sciences. October 2024. NobelPrize.org. — Freely available and remarkably readable; recommended for anyone wanting a formal account."),
pageBreak(),
// ─── SLIDE 15 ─────────────────────────────────────────────
h1("SLIDE 15 — Closing / Thank You"),
cue("[Click to slide 15.]"),
body("Let me close with the key takeaways."),
h2("Key Takeaways"),
bullet("AlphaFold solved the protein folding problem — a challenge that had resisted the efforts of biochemists and biophysicists for fifty years — using deep learning: co-evolutionary reasoning in the Evoformer and geometry-aware attention in the Structure Module."),
bullet("AlphaFold 2 achieved near-experimental accuracy at CASP14 in 2020, triggering a paradigm shift in structural biology."),
bullet("The AlphaFold Protein Structure Database provides 214 million free, open-access predicted structures — the largest structural biology resource in history."),
bullet("AlphaFold 3 extended the framework to protein interactions with DNA, RNA, and small molecules, directly enabling drug discovery applications."),
bullet("In October 2024, Demis Hassabis, John Jumper, and David Baker were awarded the Nobel Prize in Chemistry."),
bullet("AlphaFold has real limitations: static structures only, disordered regions, shallow MSAs, hallucinations. These define where the tool is trustworthy and where it needs experimental support."),
body(" "),
body("The future is the integration of prediction, design, dynamics, and experiment into a unified structural biology pipeline. We are very much at the beginning of that story."),
body("Thank you. I am happy to take questions.", { bold: true }),
pageBreak(),
// ─── APPENDIX ─────────────────────────────────────────────
new Paragraph({
children: [new TextRun({ text: "APPENDIX — Anticipated Q&A", size: 40, bold: true, font: "Arial", color: NAVY })],
spacing: { before: 100, after: 200 },
border: { bottom: { color: TEAL, space: 1, style: BorderStyle.SINGLE, size: 8 } },
}),
body("The following six questions are commonly raised after AlphaFold seminars. Answers are provided in full."),
body(" "),
...qa(
"Q1. How long does an AlphaFold prediction take?",
"For a typical protein of 200-500 residues, the AlphaFold Server returns results in minutes to tens of minutes depending on server load. The most time-consuming step is actually the MSA construction — the neural network forward pass itself is fast. For very large proteins or protein complexes, predictions can take hours."
),
body(" "),
...qa(
"Q2. Is AlphaFold used in clinical medicine today?",
"Not directly in clinical diagnostic pathways yet, but pharmaceutical companies are using AlphaFold-derived structural models to accelerate drug discovery pipelines. AlphaMissense is being evaluated for integration into variant interpretation workflows for rare disease diagnosis. Clinical application in the traditional sense — informing individual patient care — is still emerging."
),
body(" "),
...qa(
"Q3. Can AlphaFold predict membrane proteins?",
"Yes, with caveats. AlphaFold can predict membrane protein structures, and the database includes them. However, membrane proteins are often predicted without the lipid bilayer context, which can affect conformational accuracy. The pLDDT scores for transmembrane helices are often high, but the relative orientation of those helices — which determines, for example, the gating state of an ion channel — may not be accurately captured."
),
body(" "),
...qa(
"Q4. How does AlphaFold compare to Rosetta?",
"Rosetta is a physics-based and knowledge-based software suite developed over 30 years by David Baker's lab. For protein structure prediction, AlphaFold 2 and 3 substantially outperform Rosetta. For protein design — generating new sequences for a desired structure — Rosetta's ProteinMPNN and newer RFdiffusion tools are at the frontier. The two approaches are increasingly complementary: AlphaFold predicts, Rosetta and its derivatives design."
),
body(" "),
...qa(
"Q5. What is the AlphaFold licence?",
"AlphaFold 2 code and weights are released under an Apache 2.0 licence — fully open, including commercial use. AlphaFold 3's model weights are available for non-commercial research under a custom DeepMind licence. The AlphaFold Database is under Creative Commons Attribution 4.0 International. The AlphaFold Server is free for non-commercial use."
),
body(" "),
...qa(
"Q6. What comes after AlphaFold?",
"ESMFold from Meta AI uses protein language models instead of MSA, making it faster but slightly less accurate. RoseTTAFold2 from the Baker lab is competitive with AlphaFold 2. The field is moving toward models that can jointly predict structure and dynamics, and toward generative models that design novel proteins from scratch rather than predicting existing ones."
),
body(" "),
rule(),
new Paragraph({
children: [new TextRun({
text: "Script prepared for AlphaFold Seminar — May 2026 | " +
"Sources: Jumper et al. Nature 2021 | Abramson et al. Nature 2024 | " +
"Varadi et al. NAR 2024 | Nobel Committee 2024 | " +
"Chen et al. IJMS 2024 | Krokidis et al. IJMS 2025",
size: 18, font: "Arial", color: LGREY, italics: true,
})],
alignment: AlignmentType.CENTER,
spacing: { before: 120, after: 80 },
}),
], // end children
}], // end sections
}); // end Document
Packer.toBuffer(doc).then(buffer => {
fs.writeFileSync("/home/daytona/workspace/alphafold-ppt/AlphaFold_Seminar_Script.docx", buffer);
console.log("✅ Word document written successfully.");
}).catch(e => console.error("Error:", e));
Running Command
cd /home/daytona/workspace/alphafold-ppt && node build_docx.js 2>&1
Running Command
python /home/daytona/skills/docx/scripts/office/validate.py /home/daytona/workspace/alphafold-ppt/AlphaFold_Seminar_Script.docx 2>&1 && ls -lh /home/daytona/workspace/alphafold-ppt/AlphaFold_Seminar_Script.docx
Running Command
python3 -c "import zipfile; z = zipfile.ZipFile('/home/daytona/workspace/alphafold-ppt/AlphaFold_Seminar_Script.docx'); print('Valid ZIP. Contents:'); [print(' ', f) for f in z.namelist()]" 2>&1 && ls -lh /home/daytona/workspace/alphafold-ppt/AlphaFold_Seminar_Script.docx
The file is valid. Here it is:
AlphaFold Seminar Script
Word Document · DOCX
What's in the Word file:
- Cover page — title, subtitle, estimated timing, and a usage guide explaining how the document is structured
- Header — "AlphaFold Seminar — Presenter Script" on every page (top right)
- Footer — date, sources, and page number on every page
- 15 slide sections — each with a dark navy heading bar, teal sub-headings, speaker cue lines in teal italics (e.g. [Click to slide 3.]), gold talking-point labels, and inline source citations at the bottom
- Appendix — 6 fully answered Q&A items (timing, clinical use, membrane proteins, Rosetta comparison, licensing, what comes next)
- Formatting: Arial throughout, proper bullet lists (no unicode hacks), colour-coded for easy scanning during the talk
This is a shared conversation. Sign in to Orris to start your own chat.