The New Biology: Scientists Are Learning to Design Life

Artificial intelligence first astonished the world by learning to speak. Its more consequential achievement may come from somewhere quieter: laboratories where machines are learning the language of proteins, cells and disease, and beginning to design experiments of their own.

The artificial intelligence revolution arrived noisily, on laptops and phones, in classrooms and offices, producing essays, computer code and improbable pictures in seconds. Machines had learned to speak with us, and for a time that seemed extraordinary enough.

Yet in laboratories around the world, another revolution has been proceeding almost out of sight, among proteins folded into shapes too small to see, robotic arms moving samples between instruments, and computers searching through chemical possibilities that no human being could examine in a lifetime. Its language is written not in sentences but in amino acids, genes and molecular structures, and the machines are beginning to write in it.

Artificial intelligence can now help design molecules that have never existed, engineer proteins for particular purposes and search vast chemical landscapes for medicines. Other systems are beginning to formulate hypotheses, interpret experimental results and decide what experiment should be attempted next.

There is plenty of hype surrounding all of this, and biology has disappointed generations of people who thought they had finally mastered it. Yet something has changed now that AI assisted medicines are being tested inside human beings and some have reached Phase III clinical trials: predictions that once existed only inside computers have entered the body, where biology rather than software will decide whether they mean anything.

The Long Search

There is something almost archaic about the way medicines are still discovered.

Scientists can begin with thousands or millions of compounds, gradually eliminating them through laboratory experiments, toxicity studies and human trials. Most candidates fail. Sometimes they fail after years of work.

The problem is not a lack of sophistication. It is the extraordinary complexity of life.

A disease may involve networks of genes and proteins interacting inside cells, which are themselves responding to other cells and to the body around them. A molecule that performs beautifully in a laboratory dish can become useless inside a person.

Drug discovery is therefore partly a search through an immense landscape: finding the right biological mechanism, finding something capable of altering it, and then discovering whether changing that mechanism actually helps the patient without harming something else. Artificial intelligence does not abolish this uncertainty, but it can search relationships among enormous quantities of chemical and biological information and narrow millions of possibilities to a handful worth making. The laboratory remains the judge; the machine helps decide which suspects are worth bringing before it.

There is no guarantee that rentosertib will become a medicine; it may fail, and that is precisely why its journey matters. For years, AI drug discovery lived largely in presentations, research papers and computer simulations, where a machine could produce a molecule that appeared promising. Now the molecule has to survive an encounter with the far less accommodating reality of human biology.

Writing in the Language of Life

Proteins are among the tiny machines from which life is constructed.

They carry substances, transmit messages, accelerate chemical reactions and patrol the boundary between the body and the outside world. Much of what happens inside us depends upon the shapes into which they fold.

For decades, predicting those shapes was one of biology’s great problems. Artificial intelligence became remarkably good at it.

Then came the obvious, more unsettling question: if a machine can learn how a protein’s sequence determines its structure and behaviour, could the process be run backwards? Could we describe what we want a protein to do and ask a computer to help design it? That possibility marks a different kind of scientific threshold, the difference between learning to read the language of biology and beginning, however tentatively, to write in it.

Humans have always borrowed from nature, first finding medicines in plants and fungi and later learning to modify molecules and biological systems that evolution had already provided. Generative biology introduces another possibility: beginning not with what nature happens to have made, but with what we would like a molecule to do. We remain very far from ordering a cure from a computer, but the conceptual boundary has moved.

The Experiment That Never Sleeps

The larger transformation may begin when these systems are connected.

Imagine a laboratory late at night, after the researchers have gone home but the lights over the benches remain on and automated equipment continues moving samples between instruments. A computer has analysed the previous experiment and found the result disappointing, so the system changes the molecule; another machine decides which variation appears most promising, robotic equipment prepares it, an instrument tests it, and the result flows back into the computer. The hypothesis is adjusted and another experiment begins.

This is the idea behind the emerging self driving laboratory.

Something profound happens when that loop closes. One system can search the scientific literature while another analyses genomes, another predicts molecular behaviour and a specialised model designs a protein, with an agent coordinating the different parts until, in effect, AI begins using AI.

There remains, however, one participant that cannot be simulated away: nature itself. A protein must actually fold, a molecule must actually bind and a cell must live or die. The physical experiment returns an answer indifferent to whether the computer’s prediction was elegant or absurd, and the machine must absorb that answer and try again. This is what separates the emerging scientific system from the familiar chatbot, which largely works upon information that already exists; an experimental system can participate in creating information that has never existed before.

A Ghost Cell

Beyond individual proteins lies a more distant ambition: scientists want to build cells inside computers, not pictures of cells but computational models that behave sufficiently like living cells for researchers to experiment upon them, changing a gene, introducing a drug or altering a protein and then watching what happens.

In 2026 researchers reported a detailed three dimensional simulation of a minimal bacterial cell, modelling processes including DNA replication, metabolism and division.

It was an extraordinary achievement, but also a reminder of the distance still to travel, because a bacterium is among life’s simpler forms while a human cell belongs to another order of complexity.

Medicine has always had to deal in averages. Clinical trials tell doctors what happened to hundreds or thousands of people, while the physician standing beside one patient’s bed must decide what that evidence means for the particular person lying there. Yet patients are not averages, and two apparently similar cancers can respond very differently to the same drug. If sufficiently accurate virtual cells can eventually capture those differences, personalised medicine would cease to mean merely selecting from several existing treatments and could instead mean experimenting on a computational version of the disease before experimenting on the patient.

Biology Gets the Last Word

There are good reasons to be suspicious of this story. “AI designed” is becoming a marketing phrase, pharmaceutical companies have used computers for decades, and almost any modern discovery programme can now be presented as artificial intelligence if the definition is stretched far enough. Biological data is messy, correlations can masquerade as causes, models can be confidently wrong, and the human body remains magnificently inconvenient: a molecule can perform exactly as predicted and still fail as a medicine.

The real measure of this revolution will therefore not be the number of molecules announced by biotechnology companies, but what happens when those molecules encounter patients. The questions are stubbornly ordinary ones. Do these systems discover targets humans missed, produce better drugs, shorten development or reduce the extraordinary number of candidates that fail? Above all, do patients get better?

Rentosertib and GB 0895 cannot yet answer those questions, and Phase III trials may vindicate them, disappoint their developers or expose limitations that no computer model foresaw. Their importance for the moment is more modest but still consequential: they have crossed the boundary between computation and biology, carrying ideas shaped inside machines into human bodies where the claims made for this new science can finally be tested.

Behind them lies something potentially larger than either medicine, an emerging way of doing science in which machines can read accumulated knowledge, propose an idea, design an experiment, observe what happens and use the result to decide what should be tried next. For centuries that cycle has moved at something close to human speed, constrained not only by intelligence but by the hours in a working day, the time required to read, think, prepare an experiment, wait for a result and decide what the result means.

That rhythm may now begin to change. Some part of the scientific process can continue after the researchers have left the laboratory, as a proposed molecule is made and tested and the physical world supplies an answer that no model is permitted to argue with. The system can then alter its hypothesis and begin again, so that somewhere under the lights of a laboratory at night, while the scientists who built it are asleep, the next experiment is already under way.