We summarise recently published work developing and validating machine-learning based protocols that permit functional magnetic resonance brain imaging (FMRI) to be used in drug discovery. To develop such protocols required strategies for identifying neuroimaging markers of drug efficacy that are robust to variability in the actions of effective compounds, and other factors , without requiring extensive new experimental data acquisition. Our approach is to learn signature brain activation modulations indicative of drug efficacy across databases of heterogeneous FMRI studies associated with the disease state of interest. Activity signatures are learnt through training classifiers to distinguish effective drugs from placebo. We validated the approach by developing a protocol for the identification of novel analgesics. This protocol successfully identified evidence for efficacy in 8 separate studies of established analgesic compounds. Further work has extended approaches to resting-state imaging. This work demonstrates a framework in which machine learning can facilitate translational applications of brain imaging. More broadly, this work shows how combining meta-analytic and machine learning approaches can broaden the inferences and applications possible with imaging technology.