An AI designed to work through hypotheses, experiments, and corrections has created two methods capable of outperforming tools already used in bioinformatics.
The system is called The Little Scientist. Its principle is simple: instead of modifying code and keeping only what works, the agent must state a hypothesis. It then predicts the expected result, writes its program, tests it, and compares what actually happened with its prediction.
Source: Wikimedia Commons — Arek Socha, CC0
This difference matters when a trial fails. The system receives detailed information about the cases in which its program improves or regresses. It can therefore link a specific modification to its effects. In the next iteration, these results are used to construct a new hypothesis rather than starting over at random.
The researchers tested this method on two biology-related tasks. The first involves predicting the effect of a mutation on a protein's function. The system developed a strategy called Delta V, which combines several existing models and adjusts their results. On the ProteinGym benchmark used in the study, it ranked first according to the five official metrics evaluated.
The second test concerned DNA. Some proteins preferentially bind to particular short sequences. Identifying these motifs helps researchers understand how gene activity is controlled. Starting from scratch, the AI wrote an algorithm called DALE to search for these motifs in the data.
Across 132 proteins regulating gene activity, DALE achieved an average score of 0.842, compared with 0.803 for STREME, the reference tool. In this test, DALE also ran 11 times faster.
The system also provides a way out of a dead end. After several attempts without progress, a second agent suggests a new way to approach the problem, sometimes inspired by another field. In both experiments, this change in perspective contributed to improvements later developed by the main agent.
The code and the history of the trials are published to make it possible to reproduce the tests.