A Molecular Compass for Drug Discovery express

Original: "SEISMO: Increasing Sample Efficiency in Molecular Optimization with a Trajectory-Aware LLM Agent"
arXiv:2602.00663 · 2026-01-31 · CC BY 4.0 · 1 min
The program learns from every attempt and finds the right molecules several times faster.

📄 Showing the "Simple" version — "Advanced" is not ready yet. Add it to favorites to help prioritize it.

A chemist searching for a new drug is like a cook without a recipe: they have a list of desired properties but millions of possible ingredients. Each trial is costly and time-consuming, so you need to experiment wisely.

The new program SEISMO acts like a chef who tastes the dish after every step and immediately decides what to add next. It doesn't wait for a whole batch of attempts to accumulate; instead, it modifies the molecule after each measurement. At the same time, it remembers the entire journey: what has been tried, what worked, and what didn't.

In tests on 23 tasks, SEISMO achieved the desired properties 2-3 times faster than previous programs, often nearly reaching the maximum within 50 attempts.

If you explain to the program in words why a molecule didn't work — for example, "it dissolves too poorly in water"—it finds a solution even faster. This means that advice from an experienced chemist helps the program just as much as the measurements themselves.

Such an approach could accelerate the search for drugs and other useful molecules, saving time and expensive experiments.

The program SEISMO is named after the seismograph: just as the instrument detects the slightest ground vibrations, the algorithm detects the slightest improvements in molecular properties.

Scientists
Erwin SchrödingerHugh Everett III
Concepts
optimization ai drug discovery quantum computing resource optimization maximum likelihood drug design enzyme genetic algorithm personalized medicine active learning SLAM medicine atomistic simulation importance sampling agent protein language model Stochastic modeling automation sample size molecular dynamics acceleration symbolic regression