01 · Briefing · what happened

An AI that read 24 million papers points to a new Parkinson's target

Biotech & Longevity 2 min 6 sources

New tools this week all attacked biotech's oldest problem - the space of possible molecules and targets is astronomically large, so the whole game is narrowing the search.

24.4M

papers read by the AI

more than any person could

21,008

human genes searched

across 5,850 diseases

$110M

GSK's deal for data

to train discovery models

52

molecule templates

designed, not screened

At a glance

An AI called XunZi read 24 million papers and pointed to a new Parkinson's drug target.
Blocking that target saved brain neurons and eased movement - in mice, not yet people.
Other teams designed bacteria-killing molecules from scratch, not by screening a library.
GSK agreed to pay up to $110 million for data to train drug-discovery models.
There are more possible drug-like molecules than atoms in the solar system.
You cannot test them all, so the whole game is narrowing the search cleverly.

Forces in play

Size of the search High

more drug-like molecules than atoms in the solar system

AI narrowing power Building

models rank candidates before any lab test

Real-world proof Easing

the headline results are still only in mice

In play XunZi — AI that generated a Parkinson's target from 24M papers GSK and Relation — $110M pact to build discovery data and models Self-driving labs — robots that run and pick their own experiments

How it unfolded

Jul 30 GSK signs $110M data deal with Relation
Aug 1 review charts the rise of self-driving labs
Aug 3 designed peptides kill drug-resistant bacteria in mice
Aug 4 XunZi flags a new Parkinson's target

Full briefing

An AI biologist points to a Parkinson’s target

This week a system called XunZi, described in Nature Biomedical Engineering, read its way to a drug target no person had flagged.[1] It was trained on 24.4 million research papers and 613.6 terabytes of data, spanning 21,008 human genes and 5,850 diseases.[1] For Parkinson’s, it flagged two overactive enzymes, CHK2 and IRAK4, as places a drug might act.[1] Blocking CHK2 saved dopamine-making neurons and eased movement problems - in mice.[1] That last word is the whole caveat: most things that work in mice never work in people.

The point isn’t the machine. It is what it was up against. No person can hold 24 million papers in their head.[1] The link that connects one gene to one disease is real, but scattered across more journals than anyone can read.

Designing molecules instead of finding them

Two other results this week hit the same wall from the other side. In Nature Chemical Biology, researchers designed - from scratch - short protein chains that punch holes in bacteria.[2] They did not screen a library; they used physics simulations to pick sequences, then built 52 reusable templates.[2] One killed drug-resistant bacteria, including Acinetobacter baumannii, without harming human cells, and worked in infected mice.[2]

A separate team tackled the parts of our own proteins that have no fixed shape.[3] These “disordered” regions hold 37% of known disease-linked mutations but resist the usual tools.[3] The team paired pattern-searching with AlphaFold, the protein-structure AI, to map 1,300 interactions that were dark before.[3]

The search engines, and who pays for them

Behind these results sits new machinery for searching faster. A review in Nature Reviews Chemistry traced the rise of “self-driving labs.”[4] An algorithm proposes an experiment, a robot runs it, and the software reads the result and picks the next one.[4] A decade ago these were narrow automation; now they are discovery platforms.[4]

The money is following. On July 30, GSK agreed to pay the AI biotech Relation up to $110 million.[5] Relation will generate data on how human cells react to genetic and chemical nudges, then train models on it.[5] The data is the point: a model can only narrow a search as well as the examples it learned from.

What this changes

None of this is a cure, and the honest limits stack up: mouse results, preclinical data, a target that still needs a drug and years of trials.[1][2] What changed is the approach. The field is shifting from testing candidates to predicting which ones are worth testing at all. The next wave of tools aims to reason about physical reality directly, not just from text.[6] The bet under all of it is the same: you cannot try everything, so learn to guess well.

02 · Lesson · why it matters

The ocean too big to test

There are more possible drugs than atoms in the solar system, so the real skill is choosing which few are even worth testing.

How it works

The space of possible molecules is astronomically large
Testing every one is impossible - not enough time, money, or matter
So you narrow first: rules and structure rule most out
A model trained on past results predicts which few to test
You test the shortlist, not the ocean

The twist

The hard part of discovery was never the test. It is choosing which few things, out of an ocean, are even worth testing.

Where you've seen this

Web search

the web is too big to read, so a ranker guesses the ten links worth showing

Hiring

you cannot interview everyone, so a filter narrows thousands to a shortlist

Chess engines

too many move sequences to check, so the program prunes to the promising branches

The catch

A clever narrowing can search the wrong corner: if the answer lies outside what the model learned, the smart search walks right past it.

Full lesson

The number that breaks intuition

A drug is a small arrangement of atoms. The count of possible drug-like arrangements is often put near a 1 followed by 60 zeros. That is more than all the atoms in the solar system.

Proteins are worse. A short chain of 100 links can be built 20 different ways at each link. That is more combinations than there are seconds in the life of the universe, many times over.

So run the fastest lab on Earth, testing one candidate a second, since the Big Bang. You would have checked a rounding error of the whole. The problem was never “can we test this molecule.” It was “which molecule, out of an ocean, do we even pick up.”

Search, not luck

Last week the danger was the opposite one. Run enough tests and pure chance hands you a false hit; a finding is only as good as the bar it had to clear.

Here the trap flips. You can never run enough tests, so everything rides on the few you choose. The AI that read 24 million papers this week did not get lucky. It narrowed. Every real advance in this field is a way to throw candidates out before a lab ever sees them.

How you narrow

You narrow in layers.

First, rules. Physics and chemistry forbid most arrangements outright. They would not hold together, would not dissolve, would not reach the target. Rules delete the impossible for free. This week’s designed bacteria-killers skipped the old screen entirely: simulations picked the sequences worth building.

Second, data. Feed a model everything that has worked and failed, and it learns to rank the untested. That is what the big pharma deal this week actually bought. Not molecules. Data on how cells react, to teach a model where to look.

Third, speed. Self-driving labs test, read the result, and pick the next test themselves. The narrowing compounds, run after run.

The catch

A model can only narrow toward what it has seen. Aim it at a corner of the space its data never covered, and it will confidently walk past the answer.

The shapeless proteins mapped this week are the case in point. A quarter of our proteins have no fixed shape, so tools trained on the shaped ones were blind to them for years. A clever search is only as wide as its training. Narrow too hard on yesterday’s data, and the search settles into a dead corner it can’t see past.

Who the search leaves out

Where you point the search is not a neutral fact. A model learns from the diseases that already have data, the genes already studied, the trials already run. So the well-funded corners get searched again and again, and the neglected disease with no data stays dark. Not because its cure is missing from the ocean. Because no one’s map points there.

We are all inside that map. The drugs that reach us are the ones the search was aimed at, and the aim follows the data and the money. Even the largest model has read a sliver of what could be known. “We searched it” always means “we searched where we knew to look.” That is worth holding onto the next time a scan of everything promises to have found the answer.

03 · Lab · your turn

Search the ocean

The space of possible molecules is astronomically large, so discovery is a search problem not a testing problem: you cannot try everything, so everything rides on narrowing to the few worth testing - and a model can only narrow toward what it has already seen.

04 · Hope · carry this

For the first time we can search the vast space of possible medicines with real guidance, so cures once beyond reach - even for long-neglected diseases - are becoming findable.