Modern small-molecule drug discovery typically starts with two things: a library of compounds to test, and an assay to identify compounds with the desired biological characteristics.
Perhaps you want to find molecules that inhibit a bad actor by binding to its natural ligand binding pocket; here, you would typically look for molecules that both:
- Bound to the target, and
- Caused inhibition of the kinase's activity, say, by looking for reduced phosphorylation of known targets.
However, if you don't have (a), you might instead find molecules that work through other means; perhaps by inhibiting an upstream factor which is required for that kinase's activity, or by preventing the kinase from dimerizing and getting to an activated state.
This is an example of a phenotypic screen: a screen for some biological consequence (a phenotype) which could be achieved through unknown mechanisms. This is of course not a bad thing, in general. Through a phenotypic screen you might identify an entirely new, and perhaps better, way of modulating a desired target. However, not knowing how your drug works can also be a recipe for disaster; that unknown mechanism might have completely unknown toxicities as well.
Fortunately, we can leverage external public data to help uncover the mechanism of action for unknown chemical matter. Treating a cell model with a chemical perturbation causes changes in the transcriptome landscape, both at the gene and gene transcript level, which together form a kind of "fingerprint" of cellular activity.
Perhaps these genes form recognizable pathways (e.g., interferon response), or are the result of changes to transcription factor activity which themselves cause a cascade of effects. We can take these fingerprints and overlay them on known pathway networks, or simply do an unbiased search among the hundreds of thousands of bulk RNAseq experiments found in public datasets or your own data to identify where these fingerprints have been seen before.
Often, this is enough: the top studies where these transcriptomic fingerprints were found contain obvious perturbations that can explain your results. If the findings are more nuanced, studies can be mined by hand or using AI: small/large language models purpose built to identify the patterns in these studies that provide strong hypotheses as to how the drug works.