How to Find a Cancer Drug Among Millions of Molecules ✎ express

Original: "Data-Efficient Active Learning Discovery of Transition Metal Photosensitizers for Type I Photodynamic Therapy"
arXiv:2603.19912 · 2026-03-20 · arXiv non-exclusive · independent analysis · 1 min
Scientists have taught a computer to pick molecules that kill cancer cells even without oxygen.

Tumors grow fast, and oxygen inside them often runs short. Conventional drugs that need oxygen are powerless there. But some molecules work differently: they pull electrons from neighboring molecules and trigger a chain reaction that destroys cancer cells. Such molecules are called photosensitizers: they are activated by light.

The catch is that suitable molecules number in the millions, and testing each one in the lab is impossible. So the chemists got clever. They built a database of 2.1 million metal complexes — molecules where, like ornaments on a Christmas tree, ligand "branches" surround a metal atom. Then they ran an algorithm that gradually learned to guess which molecule would be useful. It consulted the accurate but slow calculations only 300 times, and that was enough to find promising candidates.

The algorithm is like a seasoned gardener: rather than checking every seed, it can tell from the sprout which plant will yield the best crop.

Among the selected molecules, osmium complexes (based on a heavy, rare metal) were the most common. It turned out that success hinged on the "branches" around the metal being uneven: electron-donating on one side and electron-withdrawing on the other. This asymmetry helps the molecule trigger the desired reaction. These findings could be useful not only in medicine but also in solar cells, where such molecules convert light into electricity.

Osmium is one of the densest metals on Earth. A fist-sized chunk weighs as much as a hefty dumbbell.

Osmium is so dense that a liter of this metal weighs 22.6 kilograms — almost as much as two buckets of water.

In Andy Weir's novel The Martian, the protagonist uses solar panels to survive: those panels also rely on metal complexes similar to the ones the authors studied.

Scientists
Erwin SchrödingerHugh Everett IIIWilliam Lawrence Bragg
Concepts
active learning energy transfer radiotherapy ai planning