AlphaFold, rational drug design, and what a cure really costs — and why the frontier runs on checked specialist tools, not one big brain.
August 2026 · Case 4 of 6
As you read — hold this question
If the cure already exists somewhere in a space of 1060 molecules, is the hard part inventing it — or noticing it?
50 yrs
the protein-folding problem — solved by AlphaFold in minutes
200M
protein structures predicted — nearly every one known to science
1060
drug-like molecules in chemical space — the cure already exists in it
The examinable core
The textbook version
A protein is a shape — and AlphaFold reads it
Proteins begin as a chain and fold into a precise 3-D shape; the shape decides the job. For about fifty years, predicting that fold was one of the hardest problems in science — a single structure could take much of a career. AlphaFold (DeepMind, Nature 2021; AlphaFold 3, 2024) learned to predict it in minutes, for ~200 million proteins — nearly every one known. It won the 2024 Nobel Prize in Chemistry.
This flips drug discovery from brute force to rational design: instead of throwing a million molecules at a target and seeing what sticks, you know the target’s exact shape and design the molecule that fits it.
Carry this
AlphaFold did not invent the folded shape — nature had folded proteins that way for billions of years. It noticed the pattern. Discovery, not invention.
What a cure costs
The rest of the story
The funnel, and the number that reframes it
~10,000
compounds screened at the top of the funnel
~250 → 5
reach preclinical work; ~5 reach human trials
1
becomes a medicine — ~9 in 10 that reach people still fail
The one survivor costs on the order of a billion dollars (Wouters et al., JAMA 2020); counting the failures behind it, closer to $2.6–3 billion (DiMasi et al., 2016). The number that reframes it all: the space of drug-like molecules is about 1060 (Bohacek, 1996) — more than the stars. Somewhere in it sits the cure; no lab could search that by hand, but a model can go looking.
The honest caveat — AlphaFold is a specialist, not an oracle
AlphaFold is a purpose-built biology model, trained on molecular structures, not the open internet. It makes predictions and mistakes, and the laboratory still has to check its work. The frontier of science runs on careful specialised tools, each checked by humans — not on one all-knowing chatbot.
The lesson
The other side
It finds the cure that was already there
These tools are pointed at real diseases now: malaria (WHO World Malaria Report 2024; the Pfs48/45 vaccine antigen), drug-resistant infections, Parkinson’s, Alzheimer’s and Huntington’s (mis-folding diseases), neglected diseases like Chagas and leishmaniasis (DNDi), and cancer.
The other side
The model does not invent the cure. It finds the one molecule, in a space of 1060, that was there all along — then a human checks it.
Quick recall — without looking back
Test yourself on this case
Question 1 of 3
Why does a protein’s shape matter, and what did AlphaFold do?
A protein folds into a 3-D shape that decides its job. AlphaFold predicted the fold for ~200 million proteins in minutes — a 50-year problem — winning the 2024 Nobel Prize in Chemistry.
Question 2 of 3
Walk the drug-discovery funnel and its cost.
~10,000 compounds → ~250 preclinical → ~5 to human trials → ~1 approved; ~9 in 10 that reach people fail. The survivor costs ~$1B, or ~$2.6–3B counting failures.
Question 3 of 3
Why is calling AlphaFold ‘an all-knowing AI’ wrong?
It is a specialist model trained on molecular structures, it makes mistakes, and the lab must verify its predictions. The frontier runs on checked specialist tools, not a general chatbot.