The results show that reinforcement learning can do more than master board games. When trained to solve long-standing puzzles in protein science, the software excelled at creating useful molecules.