AI for Science

Type: theme Slug: theme—AI-for-science Sources: highly-accurate-protein-structure-prediction-with---hassabis, highly-accurate-protein-structure-prediction-for-the-human-proteome—hassabis, protein-structure-predictions-to-atomic-accuracy-with-alphafold—hassabis, applying-and-improving-alphafold-at-casp14—hassabis, protein-complex-prediction-with-alphafold-multimer—hassabis, improved-protein-structure-prediction-using-potentials-from-deep-learning—hassabis, pushing-the-frontiers-of-density-functionals-by-solving-the-fractional-electron-problem—hassabis, advancing-mathematics-by-guiding-human-intuition-w—hassabis, nobel-prize-lecture-accelerating-scientific-discovery-with-ai—hassabis Last updated: 2026-05-13

---

Summary

AI for science is the newest and most forward-looking theme in the corpus, spanning protein structure prediction (6 papers), density functional theory (1), mathematical discovery (1), and the Nobel Lecture’s programme statement (1). The theme’s distinctive claim — articulated most clearly in the Nobel Lecture — is that deep learning can solve “impossible” scientific problems not by simulating physical processes but by learning statistical patterns from existing data. The learnable nature conjecture (lecture—nobel-prize-lecture-accelerating-scientific-discovery-with-ai) extends this into a general philosophical position about the nature of scientific knowledge.

Core content

Protein structure as proof of concept: The AlphaFold sequence (see theme—protein-folding) demonstrated that a pure pattern-recognition approach could solve a 50-year grand challenge. This established the template: large-scale experimental data → deep learning → scientific prediction at or above experimental accuracy.

Density functionals (2022): Pushing the frontiers of density functionals (paper—pushing-the-frontiers-of-density-functionals-by-solving-the-fractional-electron-problem) applied deep learning to solve the fractional electron problem — a known failure mode of conventional density functional theory in computational chemistry. This extends AI-for-science beyond structural biology into quantum mechanics.

Mathematical discovery (2021): Advancing mathematics by guiding human intuition (paper—advancing-mathematics-by-guiding-human-intuition-w—hassabis) used RL to guide mathematicians toward conjectures in knot theory and representation theory. Unlike AlphaFold (full automation), this is AI-as-collaborator — the system suggests patterns, the mathematician provides the proof.

The programme statement (2024): The Nobel Lecture (lecture—nobel-prize-lecture-accelerating-scientific-discovery-with-ai) frames AI for science as DeepMind’s primary mission going forward, with the learnable nature conjecture as its theoretical foundation.

Connections

• Theme: theme—protein-folding, theme—learnable-nature-conjecture
• Period: period—alphafold-era (bulk), period—post-alphafold (Nobel Lecture, density functionals)
• Projects: project—AlphaFold, project—AlphaFold2, project—AlphaFold-Multimer

Honest Gaps

• Only two scientific domains (protein structure, density functionals) have actual results — the “AI for science” programme is largely aspirational beyond these.
• The mathematics paper is a proof of concept, not a sustained programme.
• The learnable nature conjecture has no peer-reviewed support.
• No sources cover DeepMind’s internal prioritisation across scientific domains.