Veritasium: The Most Useful Thing AI Has Ever Done

What if the world's biggest challenges—from climate change to curing diseases and plastic waste—could be tackled by a single, invisible solution? This tiny solution lies in understanding proteins, the building blocks of life.

For decades, biologists struggled to determine protein structures, a problem likened to Fermat's last theorem in biology. Over 60 years, 150,000 protein structures were solved painstakingly. Then, AI changed the game by predicting 200 million structures in just a few years.

Proteins start as strings of amino acids, each with a central carbon atom bonded to an amine group, a carboxyl group, and one of 20 side chains. These fold into complex 3D shapes driven by molecular forces, defining their specific biological functions, like hemoglobin carrying oxygen.

Early protein structure determination used X-ray crystallography, requiring crystals and years of work. John Kendrew’s 12-year effort to solve myoglobin’s structure earned him the 1962 Nobel Prize. Yet, crystallization remains challenging and costly, motivating alternative approaches.

Scientists sought to predict protein folding from amino acid sequences, a cheaper and faster method. Linus Pauling predicted secondary structures like helices and sheets, but the full 3D folding remained elusive due to proteins’ evolutionary complexity and astronomical folding possibilities.

In 1994, the CASP competition challenged teams to predict protein structures from sequences. Early models like Rosetta pooled idle computer power but plateaued at low accuracy. Gamers playing Fold It helped solve HIV enzyme structures, showing human intuition complements computational power.

Demis Hassabis of DeepMind, inspired by Fold It, developed AlphaFold to use AI for protein folding. AlphaFold 1 used deep neural networks with evolutionary data but scored only 70 in CASP, below the 90 needed to solve structures reliably.

AlphaFold 2 introduced the EVO Former, a transformer-based architecture with two towers analyzing evolutionary and geometric data. It used attention mechanisms to refine amino acid relationships iteratively, producing highly accurate 3D protein models, surpassing previous methods.

AlphaFold 2’s breakthrough in CASP 14 was hailed as solving the protein folding problem, producing predictions nearly indistinguishable from experimental structures. This accelerated research by decades, aiding vaccine development, antibiotic resistance solutions, and disease understanding.

David Baker’s lab complements AlphaFold by designing new proteins using generative AI, like RF Diffusion, which creates synthetic proteins for applications such as human-compatible snake anti-venoms. This rapid design and testing cycle is dubbed 'Cowboy Biochemistry' for its speed and innovation.

The implications are vast: engineered proteins could revolutionize medicine, capture greenhouse gases, and break down plastics. AI-driven protein science exemplifies how technology can unlock fundamental biological mysteries and enable solutions to global challenges.

Beyond proteins, AI accelerates materials science, discovering millions of new crystals for future technologies like superconductors and batteries. These advances represent step-function changes in science, enabling discoveries and innovations previously thought impossible.

AlphaFold and related AI breakthroughs demonstrate how solving foundational problems unlocks entire new branches of knowledge. With continued AI progress, we can expect transformative impacts across science, health, and environment, heralding an exciting future for humanity.

While the future is promising, it also raises caution about AI’s role. Responsible development is crucial to harness AI’s power for good, ensuring it helps solve humanity’s biggest problems without unintended consequences. The protein folding revolution is just the beginning.