It sounds like they trained a classification model using 39,000 molecules with known reactivity to MRSA. The molecules are vectorized text representations of the structures. Once trained, they can run arbitrary molecules through the model and see which ones are predicted to have antibiotic properties, or at least MRSA reactivity.

They likely fed in molecules from families of structures that seem likely to contain an antibiotic but are too numerous to manually test them all. They get a prediction of which ones are likely to have the properties they want, and then start the slow process of creating and testing the molecules in the lab.