How was protein structure determined before AI?

Historically, experimental techniques like X-ray crystallography were used, which involved creating protein crystals and analyzing how X-rays diffracted through them. This process was laborious, expensive, and could take years for a single protein structure.

What was the significance of the CASP competition?

CASP (Critical Assessment of protein Structure Prediction) was a competition that spurred the development of computational methods for predicting protein structures. It provided a standardized way to evaluate the accuracy of different algorithms.

How does AlphaFold work?

AlphaFold uses deep learning, specifically a modified Transformer architecture called Evoformer. It analyzes the amino acid sequence, evolutionary information (conserved mutations across species), and geometric constraints to predict the protein's 3D structure with high accuracy.

What made AlphaFold 2 so successful?

AlphaFold 2 combined maximum compute power, a large diverse dataset, and significantly improved AI algorithms, including the Evoformer architecture and a structure module. This allowed it to achieve accuracy comparable to experimental methods.

What are the real-world applications of AlphaFold?

AlphaFold has accelerated research in developing vaccines (like for malaria), understanding disease mechanisms (schizophrenia, cancer), breaking down antibiotic resistance, and studying endangered species. It has advanced biological research by decades.

Can AI design new proteins?

Yes, AI methods like David Baker's RF diffusion, inspired by generative AI art tools, can design entirely new proteins from scratch with specific desired functions, opening possibilities for new medicines and materials.

What is the broader impact of AI in science?

AI is creating transformative leaps across scientific fields, from materials science (DeepMind's GNome program) to drug discovery and environmental solutions. It's solving fundamental problems that historically blocked human progress, accelerating discovery at an unprecedented rate.

Key Moments

AlphaFold - The Most Useful Thing AI Has Ever Done

Veritasium

Education3 min read25 min video

Feb 10, 2025|10,468,905 views|319,216|12,788

veritasium science physics nobel prize proteins protein folding deepmind alphafold LLM neural network AI artificial intelligence

Save to Pod

Key Moments

TL;DR

AI's AlphaFold revolutionizes biology by predicting protein structures, accelerating scientific discovery and problem-solving.

Key Insights

Protein structure determination, a century-old challenge, was solved by AI's AlphaFold, predicting millions of structures rapidly.

AlphaFold's success stems from advanced deep learning, leveraging evolutionary data and attention mechanisms similar to LLMs.

The development of AlphaFold involved significant computational power and refined AI algorithms, moving beyond traditional experimental methods.

Beyond prediction, AI, like David Baker's RF diffusion, can now design entirely new proteins with specific functions.

These AI-driven advancements have immediate real-world applications, from medicine (vaccines, antivenoms) to environmental solutions (plastic breakdown).

AI is fundamentally changing the pace of scientific discovery, enabling breakthroughs previously unimaginable and accelerating progress across various fields.

THE CENTURY-OLD PROTEIN FOLDING CHALLENGE

For decades, determining the intricate 3D structure of proteins was a monumental scientific hurdle. Proteins, essential for virtually all biological processes, fold from simple amino acid chains into complex shapes that dictate their function. Experimental methods like X-ray crystallography were slow, expensive, and often yielded limited results, with resolving just one protein structure sometimes taking years or forming the basis of an entire PhD. This bottleneck profoundly limited biological research and drug development.

EARLY ATTEMPTS AND COMPUTATIONAL PREDICTION

Initial efforts to predict protein structure relied on understanding molecular dynamics and basic geometrical principles, leading to concepts like secondary structures (helices and sheets). However, the sheer number of possible configurations for even short amino acid chains made direct computational prediction infeasible. Competitions like CASP were established to encourage the development of computational models, but early algorithms struggled, achieving low accuracy scores and failing to match experimental data sufficiently.

THE EMERGENCE OF ALPHA FOLD

DeepMind's AlphaFold represented a paradigm shift. Its initial version, AlphaFold 1, utilized a deep neural network trained on known protein structures and, crucially, evolutionary clues. By analyzing how amino acid sequences co-evolve across different species, it predicted which amino acids were likely to be close in the final structure. This led to a 2D representation of distances and twists between amino acids, which was then used to fold the protein chain.

ALPHA FOLD 2 AND THE POWER OF TRANSFORMERS

AlphaFold 2 significantly improved accuracy through a more sophisticated architecture called an Evoformer. This model adapted the 'attention' mechanism, known from large language models like ChatGPT, to analyze sequential biological data. It processed evolutionary information and geometric constraints concurrently, allowing different parts of the AI to 'communicate' and refine predictions iteratively. This deep learning approach, combined with massive computational resources, dramatically boosted prediction accuracy.

REVOLUTIONIZING BIOLOGICAL DISCOVERY

The impact of AlphaFold has been profound and immediate. Within a few years, it predicted the structures of over 200 million proteins, vastly exceeding the cumulative experimental work of decades. This enormous dataset has accelerated research worldwide, aiding in vaccine development, understanding antibiotic resistance, and deciphering disease mechanisms. The sheer speed and scale of these predictions represent a step-change in scientific progress, advancing biology by decades.

AI-DRIVEN PROTEIN DESIGN AND FUTURE IMPLICATIONS

Beyond prediction, AI is now enabling the design of entirely novel proteins. Techniques like RF diffusion, inspired by generative AI for art, can create proteins with specific desired functions from scratch. This opens up possibilities for creating synthetic antivenoms, enzymes to break down plastics or capture greenhouse gases, and new therapeutic proteins for diseases. These advancements signal a future where AI doesn't just solve existing problems but actively creates solutions.

BROADENING AI'S SCIENTIFIC HORIZONS

The success of AlphaFold and protein design AI demonstrates AI's potential to revolutionize numerous scientific fields. Projects like DeepMind's GNoME have already discovered millions of new crystals and stable materials crucial for future technologies. AI is fundamentally altering the landscape of scientific inquiry, accelerating discovery at an unprecedented rate, and empowering researchers to tackle complex global challenges in medicine, materials science, and environmental sustainability.

Mentioned in This Episode

●Software & Apps

●Companies

●Concepts

●People Referenced

Comparison of Protein Structure Determination Methods

Data extracted from this episode

Method	Time per structure	Cost per structure	Number of structures determined (human effort)	Number of structures determined (AI effort)
X-ray Crystallography (manual)	12 years (e.g., Kendrew)	$10,000s	~150,000 (over 6 decades)	N/A
Protein Sequence Analysis (theoretical)	~$100 (sequence only)	~$100	N/A	N/A
AlphaFold 2	Minutes/Hours (prediction)	Low (compute cost)	N/A	>200 Million (minutes/hours)

Common Questions

The protein folding problem refers to determining the three-dimensional structure of a protein based solely on its amino acid sequence. This structure dictates the protein's function, and historically, predicting it was incredibly challenging and time-consuming.

Topics

Computational Biology CASP Competition

Mentioned in this video

Software & Apps

GNoME

DeepMind's AI program used in materials science, which has discovered millions of new crystals, including hundreds of thousands of stable materials relevant for future technologies.

Rosetta

An early algorithm developed for protein structure prediction, notable for its use of distributed computing (Rosetta@home) and its evolution into the game Foldit.

Foldit

A video game developed based on the Rosetta protein folding algorithm, which allows human players to contribute to scientific research by solving protein folding puzzles.

RF diffusion

A generative AI technique developed by David Baker, inspired by AI art generators like DALL-E, used to design entirely new proteins with specific functions by learning to denoise protein structures.

Evoformer

A specialized version of the Transformer architecture developed by DeepMind for AlphaFold 2, featuring two towers (biology and geometry) that process evolutionary and spatial information iteratively.

Concepts

myoglobin

An oxygen-storing protein found in muscles, notably in diving mammals, which was the target for early protein structure determination efforts by John Kendrew.

CASP

A biennial international competition (Critical Assessment of protein Structure Prediction) designed to assess the accuracy of computational methods for predicting protein 3D structures from their amino acid sequences.

CASP14

The 14th CASP competition where AlphaFold 2 achieved unprecedented accuracy, producing predictions virtually indistinguishable from experimental structures and exceeding the gold standard score of 90.

Nobel Prize in Chemistry 2024

Awarded in part to Demis Hassabis and John Jumper for the development of AlphaFold 2, recognizing its breakthrough impact on protein structure prediction and scientific understanding.

People

John Kendrew

A British biochemist who took 12 years to determine the structure of the protein myoglobin using X-ray crystallography, contributing significantly to early protein structure determination.

John Moult

Professor at the University of Maryland who founded the CASP competition to encourage the development of computer models for predicting protein structures from amino acid sequences.

David Baker

A biologist from the University of Washington who created the Rosetta@home distributed computing project and the game Foldit. He also developed RF diffusion for designing novel proteins and was a co-recipient of the 2024 Nobel Prize in Chemistry.