Can AI Advance Science? DeepMind's VP of Science Weighs In

Episode Summary

Executive Summary: The episode argues that AI is becoming essential infrastructure for science, not just a productivity boost. Using AlphaFold as the central case, Pushmeet Kohli and Vijay Pandey explain how DeepMind moved from a classical protein-folding approach to an end-to-end model that transformed structural biology, enabled open access at global scale, and is now influencing drug discovery, genomics, materials, weather, math, and systems biology.

Main Topics: AlphaFold’s origin and technical evolution (Priority: 5/5): The conversation traces AlphaFold from DeepMind’s 2017 protein-structure project to AlphaFold 1’s success at CASP, then explains why the team discarded that approach and rebuilt AlphaFold 2 end-to-end for better training and performance. Why structure prediction matters for biology and drug discovery (Priority: 5/5): The speakers emphasize that protein structure is a foundational biological signal that helps infer function, understand disease, and design therapeutics more rationally. Open sourcing scientific AI and global access (Priority: 4/5): They discuss DeepMind’s decision to release AlphaFold and its database freely, the scientific and ethical reasoning behind that choice, and the tradeoffs involved in keeping some later models closed. AI as a new engine for scientific discovery (Priority: 5/5): Beyond AlphaFold, they highlight AI breakthroughs in weather forecasting, geometry, materials, quantum chemistry, and pure mathematics as evidence that AI can uncover novel science rather than only automate existing workflows. Changing economics of research and startups (Priority: 4/5): Vijay Pandey argues AI may dramatically reduce the cost and time of research, enabling smaller teams, faster startup iteration, and eventually more efficient clinical trials and personalized medicine. Data, evaluation, and the limits of current science AI (Priority: 4/5): Both speakers stress that scientific AI depends on high-quality data, strong benchmarks, and careful due diligence; systems biology is highlighted as a major frontier still constrained by data availability and evaluation difficulty. The cultural shift toward machine-assisted reasoning (Priority: 3/5): They frame AI as a tool that will move from novelty to necessity, with society eventually viewing certain scientific and analytical tasks as inappropriate for humans to do unaided.

Key Arguments: Protein structure prediction is a foundational problem: solving it unlocks understanding of function, disease mechanisms, drug design, and broader biological insight. AlphaFold 1 proved the concept by outperforming the state of the art, but its two-stage distance-to-structure pipeline limited end-to-end learning; AlphaFold 2 was necessary to fully solve the task. CASP-style blind evaluation was crucial because machine learning can easily overfit or “cheat”; rigorous benchmarking separated genuine progress from apparent progress. Open sourcing AlphaFold maximized scientific impact because its uses were too broad and unpredictable to be contained within a closed system. AI is not just efficient automation; in several domains it is becoming a prerequisite for making sense of data volumes that exceed human cognitive capacity. Scientific AI should be judged by whether it helps move from experimentation by trial-and-error toward predictive, model-based science. The biggest future gains may come in clinical trials, systems biology, and human-specific biological modeling, where current experimentation is slow, expensive, and limited. Data is the differentiator in biology and healthcare: many valuable datasets are inaccessible, dark, or never measured, so progress depends on both AI methods and better data collection. Open source accelerates science because others can immediately validate, build on, and extend results; it creates a cumulative “skyscraper” effect. Some outputs should remain closed or partially released when commercial, safety, or downstream-development incentives require it, as seen with AlphaMissense.

Data Points: AlphaFold development start: 2017 - DeepMind began the protein structure prediction effort around this year. CASP entry for AlphaFold 1: 2018 - AlphaFold 1 was entered in the Critical Assessment of Structure Prediction competition. AlphaFold database users: 1.6–1.7 million - Reported users of the AlphaFold database worldwide. Countries accessing AlphaFold database: 190 - Global reach of the AlphaFold database. Known protein structures at project start: ~150,000 - Approximate number of curated structures available when DeepMind started. Known proteins in AlphaFold database: ~250 million - The database contains predicted structures for nearly all known proteins. AlphaFold scientific interest figure: 1.7 million - Number of people interested in protein structure prediction, cited as a positive sign. AlphaFold 1 metric: State of the art - AlphaFold 1 outperformed prior methods in the 2018 competition. Accuracy threshold crossed: 80 GDT - A major milestone in AlphaFold 2 development. Higher accuracy target: 90 GDT and beyond - Further progress pursued after crossing 80 GDT. Model coverage expansion: Beyond proteins to DNA, RNA, small ligands, and biomolecules - Next-generation AlphaFold direction described by DeepMind. AlphaMissense prediction scope: 71 million human missense variants - Predictions were released rather than the full model for this use case. Material discovery result: 400,000 novel stable compounds - A materials model expanded the number of stable compounds known by more than an order of magnitude. Historical structure-solving time: Up to 5–10 years or a PhD - Illustrates how labor-intensive experimental protein structure solving used to be.

Pivotal Quotes: "AI is not sort of nice to have. It's basically almost a necessity for us to make sense and reason about any problem that we are now looking at." — Pushmeet Kohli: Used to frame AI as essential for modern scientific inquiry. "What we have entered is basically an age where a single human mind cannot comprehend the data that we are gathering about the universe." — Pushmeet Kohli: Explaining why AI becomes necessary across disciplines overwhelmed by data. "To some degree, I think what AlphaFold did is it took the structural biology of proteins and made it a database lookup." — Vijay Pandey: Describing the transformation from experimental bottleneck to scalable computational access.

Implications: AI is shifting science from artisanal experimentation toward predictive, scalable discovery. Expect faster labs, cheaper research, more open scientific infrastructure, and major breakthroughs in biology, medicine, and materials—if data, benchmarks, and safety are handled well.