AlphaFold Designs Nuclear Fusion Catalysts While We Still Debate Decade-Long Trial-and-Error

Last week, while I was organizing wet lab results in the lab, a junior colleague working on protein dynamics simulations ran over. He said that for a specific enzyme mutant he predicted with AlphaFold3, the correlation coefficient between the predicted binding energy and subsequent SPR experimental data reached 0.95. Excitedly, he asked me: "Senior, are we soon going to stop doing wet experiments?"

I just smiled and told him that his dataset missed a key detergent condition, so the correlation might be an artifact. But a few days later, seeing this quote from Hassabis in his Nobel Prize interview—"Drug discovery will no longer require ten years of long trial-and-error; new clean energy sources may break through within a few years"—I suddenly realized that my junior's question might exactly reflect the cognitive gap facing the entire field.

Hassabis compares AGI's impact to being 10 times that of the Industrial Revolution. This number sounds like inflation in a sci-fi novel. But as someone who uses AI tools daily to handle protein sequence-structure-function mapping, I want to seriously dissect the technical feasibility of this statement—and where it can truly land.

From Protein Folding to Protein Design: AGI is Already Changing Our Workflows

Let's look at a concrete example. Traditional computational chemistry methods (like Rosetta) designing a high-affinity protein ligand require thousands of molecular dynamics simulations first, then calculating binding energies via free energy perturbation. The entire workflow takes weeks per candidate molecule. Now, end-to-end generative models (like RFdiffusion combined with ProteinMPNN) can generate hundreds of backbone sequences in minutes, even designing coupled reaction active sites that don't exist in nature.

# A simplified comparison of dry-lab/wet-lab iterative workflows
# Traditional: Wet lab screening -> Structure resolution -> Computational optimization -> Resynthesis -> Retesting
# 15-year cycle, 5 loops, each requiring 3 months of wet lab + 1 month of computation
# AGI-assisted: AI generates candidates -> High-precision structure prediction -> Virtual screening -> Key site wet lab validation
# 3-year cycle, 10 loops, each requiring 2 weeks of AI + 2 weeks of wet lab

# The key point: AI hallucinations need wet labs for correction, but the sampling density for correction can be drastically reduced
# For AlphaFold2 sites with pLDDT > 90, mutation experiment success rates are 4x higher than random sites (Broom et al., 2023, Structure)

This acceleration isn't linear; it's networked. When you can screen 100,000 candidates overnight, it compresses the past drug discovery phase of "lead compound optimization" into an iterable feedback loop. What Hassabis says about "no longer needing ten years of trial and error" is, at least for early-stage discovery of protein-based drugs, no longer a prophecy but something happening right now.

But Can AGI Really Make "Resource Scarcity" a Thing of the Past?

Here we need to distinguish two concepts: solving problems quickly vs. changing the problem itself.

The reason the Industrial Revolution's impact was revolutionary is that it changed the basic paradigm of energy use—from biological energy (human/animal labor) to fossil fuels. This allowed humans to break through the biological ceiling of "eat how much, work how much." The capabilities currently demonstrated by AGI are essentially improvements in information processing efficiency; they do not change the physical limits of matter and energy.

Take a bioinformatics example: We can use AI to predict the hydrolytic activity of a certain enzyme on cellulose, or even design mutants with 1000x increased activity. But to actually produce industrializable cellulosic ethanol, you still need to solve: 1) Thermal stability of enzymes at 50°C, 2) Mass transfer limitations in large-scale fermentation, 3) Energy consumption for downstream product separation. AI can accelerate solving the first two problems, but the third is determined by thermodynamics—you cannot eliminate distillation energy consumption through algorithms.

[!info] Key Judgment

The most likely landing point for AGI's impact on science isn't "eliminating resource scarcity," but "eliminating cognitive scarcity." It will shift scientific research from "trial-and-error driven" to "simulation driven." But the law of entropy increase in the material world won't change because of AI—synthesizing every molecule consumes real atoms, and every kilowatt-hour requires real energy. AI can only help you find better paths; it cannot conjure up the path itself.

So What Does Hassabis Mean by a 10x Impact?

I believe this number likely comes from his estimation of the magnitude of improvement in scientific research efficiency. The Industrial Revolution improved human production efficiency by 10x (measured by GDP, per capita output grew about 15x from 1800 to 2000). If AGI can also improve the efficiency of "discovering new knowledge" by 10x, its impact pattern on society will be different: not replacing manual labor, but replacing the search and optimization parts of intellectual labor.

Think about it: How many of our current research papers are just "running the same thing again in a different system under conditions others have already done"? How many are "using different algorithms but getting the same conclusion"? What AGI can truly do is quickly find the most worthwhile directions to validate within a massive hypothesis space—it becomes the "navigator" of scientific research, rather than just the "map."

Stage	Traditional Method	AGI-Assisted Method	Estimated Speedup
Literature Review	Manual reading + notes	Semantic search + Knowledge graphs + Hypothesis generation	20x
Experiment Design	Experience + Orthogonal testing	Bayesian optimization + Active learning	10x
Data Analysis	Manual fitting + Statistical tests	Automated pattern discovery + Uncertainty quantification	5x
Paper Writing	Word-by-word drafting + Revision	Structured generation + Auto-polishing	3x

Stacking these accelerations indeed might shorten the cycle for a protein structure from X-ray/Cryo-EM's 3 years to 3 months. But please note, these accelerations haven't changed a fundamental fact: Data from the physical world requires physical world experiments to generate. No matter how accurate AI-predicted structures are, they cannot replace electron density maps from crystal diffraction—they just make it clearer which direction to cut the crystals.

An Open Question

Hassabis says AGI is imminent, but I'm curious: When AGI can autonomously design room-temperature superconductors or custom biocatalysts, will our existing research ethics and patent systems keep up? If an AI generates 10,000 brand-new molecules in 24 hours, and 100 of them show unprecedented activity—who gets credit for the "discovery"? The training data, the engineers who wrote it, or the computing center running it?

Are we ready to accept that the subject of scientific discovery shifts from "humans" to a hybrid of "humans + AI"? I think the impact of this question may far outweigh the disappearance of resource scarcity.

Original Link: https://www.ithome.com/0/977/160.htm