AlphaFold’s Hidden Advantage

This week, something strange happened at Google DeepMind. Much of the original team behind AlphaFold, the protein-structure system that won a Nobel Prize, has now dispersed. Some researchers moved into Gemini projects, some joined Isomorphic Labs, and others left the company altogether.

This doesn’t mean that AlphaFold itself is going away. Its database still provides hundreds of millions of predicted protein structures, and DeepMind still describes it as the blueprint for its AI-for-science work.

The interesting part is what DeepMind appears to be building next. Rather than organizing a dedicated team around one scientific problem, it is increasingly developing broader systems that can generate hypotheses, debate ideas, use scientific tools, and help plan experiments.

That sounds like the natural next step. AlphaFold predicted the structure of a protein. A scientific agent could eventually decide which protein matters, what experiment to run, and what the result means.

There is one important difference, though. AlphaFold had an answer key.

What I mean by that is protein folding was an incredibly difficult problem, but it had been made unusually suitable for machine learning. That’s because scientists had spent decades depositing experimentally determined protein structures into public databases. The input was clear, an amino-acid sequence. The desired output was also clear, the protein’s three-dimensional structure.

The field even had CASP, a blind competition where researchers predicted structures that had been experimentally solved but not yet publicly released. Once the answers were revealed, everyone could see whose prediction was closest. AlphaFold did not have to persuade people that its output looked scientifically interesting. It could be measured against the real structure.

Drug discovery becomes much messier after that.

A model can predict a convincing structure and still point researchers toward the wrong target. It can design a molecule that binds tightly but also binds to something dangerous. It can propose a mechanism that works in a dish but disappears in an animal. A drug can survive all of that and still fail in patients because the disease biology was incomplete, the dose was impractical, or the benefit was too small.

There is no single answer at the end of that process. There are many answers, spread across years of experiments, clinical trials, and decisions.

This matters because the next generation of scientific AI is being asked to move into exactly these less tidy spaces. DeepMind’s co-scientist can generate and rank hypotheses using Gemini-based agents. AlphaEvolve can propose algorithms and improve them through repeated evaluation. But AlphaEvolve works especially well because computer programs can be run, scored, and rejected automatically. Biology rarely gives feedback that cheaply or cleanly.

So, the bottleneck may be shifting. Generating ideas is becoming easier. Generating plausible explanations is becoming much easier. The hard part is building experiments that can reliably tell us which ideas are wrong.

This is the lesson I would take from AlphaFold. Its success was real and enormous. But the hidden achievement was not only the model. The scientific community had already spent decades building the data, benchmarks, and experimental ground truth that allowed the model to prove itself.

DeepMind now wants AI systems that can participate in a much larger portion of science. Their success may depend less on how many hypotheses they can generate and more on whether biology can build an answer key fast enough to judge them.

Russian state researchers are building their own AI protein design platform because access to foreign tools is getting harder. The Federal Research Centre for Biotechnology at the Russian Academy of Sciences says it will take three years and run on Kurchatov Institute infrastructure. They also gave the reason plainly. Restricted access to foreign computational tools. Export controls aimed at chips are now shaping who gets to design proteins. The bigger question is whether a sovereign replacement arrives useful or already behind. [Link]

An open-source answer to Claude Science showed up about four weeks after Claude Science launched. The AICell Lab newsletter highlighted a model-agnostic research workbench positioned directly as the open alternative to Anthropic’s tool. The workbench itself may or may not matter. The speed of the response does. Biology-focused research software is now being copied, opened up, and repositioned within weeks, which is fast enough that neither side may have many real users yet. [Link]

A thirteen-year-old built a peanut-shell filter for pesticide residue and is now a national science finalist. Arika Kundu is one of ten finalists in the 2026 3M Young Scientist Challenge for LIGNEX, a biosorbent made from peanut-shell waste that removes residue from fresh produce. She says computational astrobiology is next. The timing is what makes the story stick. It landed the same week federal investigators widened a multi-state produce contamination inquiry. The agencies are still tracing the spill. A seventh grader is already building the filter. [Link]