3.1 MOA prediction

Drugs are assigned a class based on their MOA, the cellular pathway perturbed by the drug, as depicted in Figure 1. Given a set of drug profiles annotated with MOA classes, we can simulate reference and discovery drug sets in a leave-one-compound-out cross-validation (LOCOCV) scheme. At each fold of the cross-validation, we hold out a drug and predict its MOA class using a classifier trained on the remaining ‘reference’ drugs. The prediction is made as the nearest neighbor (1-NN) in cosine distance between drug profiles, d(p,p′)=1−cos θp,p′. This was proposed in (Ljosa et al., 2013) as an equitable way of comparing profiling algorithms. We settle for this lightweight approach as our focus here is on the discriminative power of the profiles.

MOA prediction is performed on an image via a phenotypic profile. The development of such a profile spans four ordered stages. Each stage may be accomplished by a variety of algorithms, the combination of which define a unique pipeline. Some stages may be omitted in certain pipelines, or subsumed to a common framework