A dance of hundreds of particles: some pairs spin slowly and gracefully, others twitch in a frantic rhythm. To understand the essence of the dance, you need to watch the slow movements: they are the main choreography of the system. In molecular science, such slow movements are called 'slow modes,' and they determine how a protein folds into a ball or how a drug binds to its target.

The new method SelfTICA learns to recognize this choreography without hints. It takes two snapshots of a molecular system separated by a small time interval (a lag) and tries to guess which parts in the snapshots are the same 'dance move' and which are just random passersby. This approach is called contrastive learning: the algorithm gets a reward for bringing similar states closer together and a penalty for confusing them with different ones. As a result, a map of the most important, slowest motions forms on its own inside the program.

The key trick is that learning and finding slow modes are separated. First, the method learns to extract useful features from the data, like a dancer first getting used to the rhythm of the music. Then, on the learned features, it applies a classic technique: time-lagged independent component analysis (TICA). This allows it to find the smoothest motions and even estimate their characteristic times.

This approach resembles the work of a choreographer who watches a rehearsal not from the beginning but from any point, and still guesses what move comes next. SelfTICA doesn't require long training trajectories and can work with data where rare transitions haven't happened yet. This means computer simulations of biological molecules will become faster and cheaper, and bring us closer to understanding how proteins, DNA, and drugs work.