Noelia Ferruz, the chemist creating AI with ‘supernatural’ powers

In 1970, the British biologist John Maynard Smith came up with a simple analogy to describe one of the world’s most complex problems — protein evolution — in a paper called Natural Selection and the Concept of a Protein Space.

Maynard likened the evolutionary trajectory of proteins to a simple word transformation game. With a goal of evolving one word into another by means of single letter substitutions, Maynard offered a model of protein evolution. To transform ‘word’ to ‘gene’, it would look something like this: WORD → WORE → GORE → GONE → GENE. Word sequences containing words that made no sense would be discarded as not useful to the evolutionary process.

In this way, all proteins of interest would be connected to a greater or lesser extent in a continuous space. The English alphabet has 26 letters, and proteins have 20 basic units called amino acids. In 2010, the American engineer Frances Arnold realized that this simple game summarized an unfathomable space: the number of possible variants to create a single small protein of 100 units is 20 elevated to 100 — in other words, more than all the atoms in the universe.

Nature has explored only a tiny fraction of the entire universe of possible proteins over billions of years of evolution. There are countless proteins yet to be discovered, and “we can only dream of the immensity of their capabilities,” said Arnold, who won the 2018 Nobel Prize in Chemistry for her work on the directed evolution of enzymes.

Now, for the first time, artificial intelligence (AI) makes it possible to explore that entire universe. Its growing computational power has already described the three-dimensional shape of all the proteins that nature has invented — some 200 million — even predicting their interactions, an achievement announced just three months ago by a company linked to Google.

But how AI reached these results is a process inaccessible to humans. The machine gives the right answer but does not explain how it got there. Its creators can’t figure it out either, and other scientists are vetoed, because the technology company won’t reveal the codebase of its machine. AI is a black box.

This not only happens in questions of science, but also when AI is applied to transportation or finance: a self-driving car that jumps a traffic light and ends up killing a passerby, or a financial AI tool that denies a mortgage to a promising candidate because of his skin color.

Noelia Ferruz, 36, a chemist specializing in bioinformatics, has just been awarded a Starting Grant by the prestigious European Research Council (ERC) to solve this problem by creating a public, open and self-explanatory AI to be known as Athena.

“Today, artificial intelligence is already at the level of a PhD in chemistry,” explains Ferruz, who heads her own research group at the Centre for Genomic Regulation (CRG) in Barcelona. But thanks to its ability to devise and study compounds that nature has not invented, it will soon have a “supernatural” power, because it will exceed the known limits of nature.

Ferruz, the daughter of a housewife and a mechanic who never completed their studies, has just received €1.5 million to develop this system over five years. It will be an “intelligent agent,” a new type of AI capable of analyzing different kinds of information, such as the sequence of a protein, the three-dimensional image of the resulting molecule or the video of its different moving parts. Ferruz will be helped by a team of three bioinformaticians to train the AI while two molecular biologists will test the new molecules it designs in the laboratory. The researchers will inform the system if the experiments have worked, so that it will learn from its mistakes in a further step towards a more human-like intelligence.

The goal is to design new proteins that have a specific function, especially to save lives if they are used as drugs — for example, as more effective cancer antibodies. Ferruz wants to focus on enzymes, small proteins that speed up biochemical reactions. “Some 25% of current drugs have bonds called carbon-fluorine bonds, which are expensive and polluting to make, which contributes to the fact that some cost €60,000 per dose,” Ferruz explains. “Enzymes have the ability to make that bond as well, but they are not being used for this purpose yet.” Another possibility is to create an enzyme that binds to bisphenol A, a compound in plastic that interferes with human and animal fertility, to neutralize it.

The most difficult part of the project will not be to create this AI agent, but to understand it. Current language models like ChatGPT are made of layers of artificial neurons. “Each one does an operation and passes it on to the next one. In the end, we have up to a billion neurons in their latest versions. We see a constellation of neurons activating, but we don’t understand why; and the model doesn’t know how to explain it to us,” says Ferruz.

Ferruz’s goal is to understand how AI reaches its conclusions, which in the field is called explanatory AI. “It’s quite laborious, but we can open up layer by layer and see which neurons are activating in response to certain stimuli, which will show us what has been learned and help us understand how AI generates protein sequences better than a human,” she explains. The research has a striking parallel with the study of the brain, made up of 100,000 million neurons that establish 100 trillion connections, with the added paradox that, in this case, the intelligence was created by us.

Sitting in a glass office, a stone’s throw from the Mediterranean Sea, Ferruz says that she wanted to call this system Pandora. But a colleague reminded her of all the evils that came out of that particular box. “I asked ChatGPT and it suggested Athena. She’s the Greek goddess of wisdom, so it’s quite appropriate, although she’s also the goddess of war.”

The ERC is the most prestigious and demanding scientific body in the European Union. Each year, it selects emerging projects from young researchers to fund. Ferruz’s project is one out of a total of 494 projects that have been selected for funding.

Sign up for our weekly newsletter to get more English-language news coverage from EL PAÍS USA Edition