Notes: AI is Revolutionizing Scientific Discovery

This post summarizes John Jumper’s talk at Y Combinator (July 2025) on how AI is revolutionizing scientific discovery, with a focus on AlphaFold and the future of AI for science.

John Jumper’s Journey

John Jumper’s path to leading AlphaFold was unconventional:

• Originally trained as a physicist - Thought he would study the laws of the universe
• Dropped out of physics PhD - The work didn’t grab him
• Worked at a computational biology company - Fell in love with using computers to understand biology
• Returned to grad school - Studied biophysics and chemistry, got interested in machine learning (which they called “statistical physics” back then because ML was “disreputable”)
• Joined Google DeepMind - To apply powerful AI technologies to advance science

His guiding principle: Build tools that enable scientists to make discoveries. With about 35,000 citations of AlphaFold, tens of thousands of scientists are using these tools for vaccines, drug development, and understanding how the body works.

The Protein Folding Problem

Why Proteins Matter

• Humans have about 20,000 different types of proteins
• Proteins do practically every function in your cell
• DNA is the instruction manual for building these tiny machines
• Biology has evolved an incredible mechanism to build nano-machines out of atoms

The Folding Process

1. DNA gives instructions to build a protein (a sequence of amino acids)
2. After assembly, the protein folds spontaneously into a 3D shape
3. Like an IKEA bookshelf that builds itself
4. This shape determines function

Why Structure Prediction is Hard

• Proteins are incredibly small (a few nanometers - smaller than light wavelength)
• Traditional methods require:
  • Convincing proteins to form crystals (can take over a year)
  • Using synchrotrons to shine X-rays
  • Enormous cleverness and many failed attempts
• About 200,000 protein structures are known (growing at ~12,000/year)
• But billions of protein sequences are being discovered (3,000x faster than structures)

How AlphaFold Works

The goal: Go from a sequence of letters (amino acids) to a 3D structure prediction.

The Three Components of ML Success

1. Data: 200,000 protein structures (everyone has the same data)
2. Compute: 128 TPU v3 cores for 2 weeks (within academic scope)
3. Research: This is where AlphaFold was differentiated

Key insight: We tell too many stories about data and compute, not enough about research.

Research Matters More Than You Think

A careful experiment showed that AlphaFold 2 trained on just 1% of the data was as accurate as AlphaFold 1 (the previous state-of-the-art). This means:

Research was worth a hundredfold of data.

It’s Not About One Big Idea

When people saw AlphaFold, they said “equivariance is the answer!” But ablation studies showed:

• AlphaFold 2 was ~30 GDT better than AlphaFold 1
• Equivariance only explained 2-3 points of improvement
• Many midscale ideas add up to a transformative system

The real breakthrough came from grinding 1% at a time until crossing the accuracy threshold that mattered to experimental biologists.

The Importance of Blind Assessment

Protein structure prediction has had blind assessment since 1994 (CASP competition):

• Every 2 years, researchers predict structures of ~100 proteins whose answers aren’t known
• AlphaFold 2 had about 1/3 the error of any other group
• This is far ahead of LLMs in terms of rigorous evaluation

Why this matters:

• We all overfit to benchmarks
• Real-world problems are almost always harder than training problems
• External benchmarks are critical to figuring out what actually works

Making AlphaFold Available

Two key decisions:

1. Open-sourced the code - Specialists could use it immediately
2. Released a database - Started with 300,000 predictions, grew to 200 million (essentially every protein from organisms with sequenced genomes)

The Social Proof Effect

After CASP, structure predictors were convinced. But general biologists weren’t sure (“Maybe CASP was easy”).

When the database came out:

• People clicked in and compared to their unpublished structures
• “How did DeepMind get access to my unpublished structure?”
• Word of mouth built trust
• Everyone had a protein they hadn’t solved, or knew someone who did

Users Do the Darndest Things

Unexpected Applications

Two days after code release: A researcher put two proteins together with something in between - essentially “prompt engineering for proteins.” This became the best protein interaction prediction in the world.

The molecular syringe example:

• Scientists at MIT wanted to use a protein for targeted drug delivery
• They ran an AlphaFold prediction (not even a great one)
• Immediately saw: “Those legs at the bottom must be how it recognizes cells”
• Re-engineered it with designed proteins
• Now can target specific cells in a mouse brain for drug delivery

The Real Impact

• Special issue of Science on nuclear pore complex: 3 of 4 papers made extensive use of AlphaFold
• Over 100 mentions of “AlphaFold” - and DeepMind had nothing to do with it
• Scientists discovered new components of fertilization
• The whole field of structural biology is now 5-10% faster

Science is About Hypotheses, Not Just Validation

A common criticism: “Great predictions, but now we need to solve all proteins the classic way to validate.”

They’re right about one thing: Science is about experiments.

They’re wrong about another: Science is about making hypotheses and testing them - not about the structure of a particular protein.

AlphaFold enables scientists to:

• Form hypotheses faster
• Test thousands of interactions computationally
• Focus experimental effort on the most promising directions

The Future of AI for Science

AI as an Amplifier

Structure prediction and AI for science should be thought of as:

• An amplifier for experimentalists
• Starting from scattered observations (natural data - our equivalent of “all the words on the internet”)
• Training general models that understand underlying rules
• Filling in the rest of the picture

The Big Question

The most exciting question in AI for science is: How general will it be?

Will we find:

• A couple of narrow places with transformative impact?
• Or very broad systems?

Jumper expects it will ultimately be the latter as we figure it out. We’ll find more scientific knowledge within general systems (including LLMs) and use them for important purposes.

Key Takeaways

1. Research amplifies data and compute - Ideas are a core component, not just scaling
2. Many midscale ideas beat one big idea
3. Blind assessment is critical for knowing what actually works
4. Make tools available - Users will find applications you never imagined
5. AI for science is an amplifier - It makes experimentalists faster, not obsolete
6. Start where you have data, then find what problems it can be applied to