The Rise of the AI Scientists

AI systems are increasingly used as powerful tools in many fields of science, and soon these systems may be able to design and perform their own experiments

(Pixabay, geralt)

From tool to designer

Machine learning and artificial intelligence-based systems are quickly developing as powerful research tools. Their capacity to handle ungodly amounts of data and scan for patterns help scientists better understand many research topics, from genomic regulation, over mass extinctions, to multi-factorial mental health conditions.

However, the AI systems remain fairly ‘dumb’ tools in these contexts. Sure, they can help scientists detect new patterns and generate novel hypotheses, but the machine learning/AI is not exactly asking its own questions or designing its own experiments.

Can it though? Can AI make the leap from ‘here are the patterns’ to ‘hey, this is weird’? Even more importantly, can it go from ‘hey, this is weird’ to ‘why is this weird?’. The question mark within the single quotes in the last sentence is important here.

Simulation time

When scientists are faced with a study object that is hard to actually study in person (say the Milky Way), they often develop simulations. After all, they can’t travel backwards or forward in time, or across the galaxy (yet?).

Aren’t we all in a simuation? (Unsplash, Markus Spiske)

Combing the relevant variables into a simulation can give them a closer — though virtual — look at what they’re interested in, whether it’s a galaxy or a molecule.

The virtual realm where these simulations dwell also happens to be the home of the AI and machine learning systems that are so adept at data handling and pattern detection.

Indeed, some recent work has shown the great promise of unleashing AI on scientific simulations.

With a deep neural architecture search, a group of (human) researchers was able to speed up scientific simulations. And not just by a little bit:

The method successfully accelerates simulations by up to 2 billion times in 10 scientific cases including astrophysics, climate science, biogeochemistry, high energy density physics, fusion energy, and seismology, using the same super-architecture, algorithm, and hyperparameters.

Now imagine an AI system using a scientific simulation to learn things. That could go rather quickly.

How about a neural network that goes beyond speeding up simulations, but actually derives certain principles from a simulation? Been there, done that… Researchers modeled a neural network after the physical human reasoning process, basically by forcing subnetworks to exchange information through connections with limited ‘bandwidth’. They fed the network data about the movement of the sun and Mars, as seen from Earth. And the network went ‘hey, it looks like both Earth and Mars are circling the sun’. Or, as the authors say it:

…the network finds the physically relevant parameters, exploits conservation laws to make predictions, and can help to gain conceptual insights, e.g., Copernicus’ conclusion that the solar system is heliocentric.

HAL’s watching… and learning (Wikimedia commons, Grafiker61)

Let’s zoom in, from the solar system to the miniature world of molecules.

Yep, here too AI is making discoveries.

One example is a deep learning model that was given a training dataset of over 2,000 molecules with a known antibacterial working.

Then, it was set loose in a database of over 107 million molecules without any further assumptions about the structure or function of the molecules. It was given the task to find new molecules with a potential antibiotic function.

The model identified several molecules, many of which resembled known antibiotics, but — and this is the key point — also some that did not look like any known antibiotic molecule. One of those, halicin (named after HAL) has already been tested in mice and proved effective against several pathogens.

Enter the physical realm

That’s all great. But it’s still pretty… virtual. It’s not as if the AI system is testing the molecule, or checking its findings.

Time for robots.

AI-driven robots. (Sounds dystopian, doesn’t it? Damn you, Hollywood.)

Meet Adam, the first of its kind (no, I’m not making this up). Adam is a lab robot infused with machine learning that predicts and tests gene functions in baker’s yeast.

A few years later, Adam was joined by Eve (nope, not making this up either). Much like the antibiotic prediction model discussed a little earlier, Eve screens for and selects compounds that might prove effective against one of several tropical diseases. Unlike the earlier antibiotic prediction model, Eve actually runs preliminary tests on promising candidates.

Drug/chemical compound discovery is exceptionally suited for these developments. There are huge databases (‘libraries’) of compounds waiting to be explored (enter machine learning/AI), and certain preliminary tests can be relatively easily standardized and automated (enter robotics).

Another field that shares these characteristics is materials science. A lot of materials and several standard tests.

You know what’s coming.

AI can discover new materials, and robotic systems can be automated to perform tests in material science. Only a matter of time before these are put together.

The intuitive leap

When an AI-drive robot watches an apple fall, is it able to develop a theory of gravity? Can a robot dream of a benzene ring (instead of electric sheep)?

These questions, I think, capture the intuition that some of the greatest scientific discoveries are the result of intuitive leaps, of flashes of insight, of breaking through onto a whole new plane of thought. Something — so we tell ourselves — robots and AI are not (yet?) capable of.

After all, intuition and creativity require something more than data and pattern detection, don’t they? Something ineffable, but still something (the feeling of awe might be a good candidate here. Let me know your candidate.).

On the other hand, figuring out heliocentrism by yourself — even in a simplified simulation — isn’t that bad. Took us a while. And coming up with new antibiotic molecules that have no know structural analogue among known ones has a bit of a ‘benzene story’ ring (pun very much intended) to it.

Let’s see what the automated AI scientists come up with next.