Captured at (local ISO): 2026-05-18 05:17:05

Preface

When “water from desert air,” “custom proteins,” “gene editing,” AI, and billions in capital appear together at the AI AS 2025 summit, the link may seem unclear—yet that is the core appeal of AI for Science: distant breakthroughs become reachable through AI. Held in San Francisco in late October, AI AS 2025 focused on deep AI–science integration. Nobel laureates, leading scientists, and industry leaders used three cases to open the curtain on AI for Science.

What Is AI for Science?

AI for Science is not simply using AI to process experimental data. It embeds artificial intelligence across the full research pipeline—from hypothesis generation and experiment design to data production and validation—as a fifth paradigm alongside experiment, theory, and computation. It breaks the bottlenecks of slow trial-and-error, long cycles, and heavy reliance on experience, letting scientists explore domains that were previously out of reach.

A core consensus at the summit: AI is no longer just a “helper” in research; it is becoming a subject of discovery. The three Nobel-level cases below are vivid proof.

Nobel Case Studies

I. AI + Materials Science — “Growing” Water from Desert Air

2025 Nobel laureate in Chemistry Omar Yaghi’s work shows how AI can reshape materials R&D. His MOF (metal–organic framework) materials act like “molecular sponges”—a few grams can expose a surface area comparable to a football field, adsorbing moisture from air, capturing CO₂, or storing hydrogen.

Before AI, MOF development was slow: labs screened metals and organic linkers by trial and error for 2–3 years. Yaghi’s team built a “virtual lab”:

• AI read 250 papers and extracted MOF synthesis conditions at ~95% accuracy, cutting two weeks of work to hours;
• Seven AI agents handled literature search, synthesis planning, machine programming, and 3D printing;
• A “super material” that 100 research groups had failed to fully synthesize was achieved in 84 experiments over 2–3 weeks.

The technology is commercialized: a non-electric solar version produces 850 L/day of water; an electric version reaches 2000 L/day without grid, seawater, or heavy infrastructure—bringing “water freedom” to remote areas and island data centers. The next goal: AI-optimized CLF (advanced MOF) plants in 600 major cities to remove excess atmospheric CO₂ in 3.5 years.

Here, AI for Science means accelerated screening + active design: materials R&D shifts from luck to precision, turning ambitious goals into deployable tech.

II. AI + Biology — “Building” Proteins Like LEGO

Proteins are the basis of life; nature took 3.5 billion years to produce on the order of 10¹² variants. 2024 Chemistry Nobel laureate David Baker used AI to go beyond nature—designing proteins that never existed.

Traditional work could only predict structure; AI enabled design on demand:

• Neural networks trained on known structures can organize amino-acid sequences into plausible new proteins;
• Give AI a task such as “activate only in the tumor microenvironment and stay dormant in normal tissue” to reduce drug side effects;
• Generate thousands of candidates at once and explore families systematically.

Baker’s analogy: like an image model drawing “a cat on a table,” protein design is “build function to spec.” The RF3 model, with more metagenomic data, improves structure prediction and opens doors for precision medicine and new drugs.

The breakthrough: from reading nature to creating nature—scientists are no longer limited to natural proteins.

III. AI + Gene Editing — Feeding Machine Learning with CRISPR

CRISPR is the “genetic scissors” for precise DNA edits. 2020 Chemistry Nobel laureate Jennifer Doudna showed how CRISPR and AI empower each other.

A core bottleneck is data: biological data are sparse and messy, limiting ML. Doudna’s approach: use CRISPR to generate training data:

• Design millions of guide RNAs to knock out or edit genes in mice;
• Use automated microscopy at scale to map gene → function;
• High-quality data let AI predict protein function, parse gene networks, and model RNA structure.

Therapies moved faster: from the 2012 discovery to FDA approval for sickle-cell disease in 2023 took about 10 years—fast for biomedicine. Patient Victoria Gray recovered fully; the team is also working to cut costs.

Here, AI for Science solves the data gap: when existing data are insufficient, experiments create data for AI, and AI accelerates the next breakthrough.

The Essence of AI for Science: Amplifying Creativity, Not Replacing Scientists

From materials to biology, from desert water to gene therapy, these cases show one theme: AI does not replace scientists—it frees them from repetitive trial-and-error and data wrangling so they can focus on asking questions and setting goals.

As Microsoft Research AI for Science director Christopher Bishop said at the summit, AI-driven discovery is a fifth paradigm—combining empirical experiment, rigorous theory, efficient computation, and active exploration.

Challenges remain: training data may run out in 4–5 years, hardware hits physical limits, and science demands high data quality. AI AS 2025 showed scientists, industry, and capital aligning—when Nobel-level insight meets AI compute and funding, breakthroughs will only accelerate.

AI for Science will spread to climate models, new energy materials, quantum computing, and more—helping humanity solve grand challenges faster: climate, disease, and the nature of life itself. The revolution’s opening chapter was written in San Francisco.