, is one of the most fundamental yet challenging decisions in drug development){kind=link}
Determining whether a drug should activate or inhibit a biological target, the direction of effect (DOE), is one of the most fundamental yet challenging decisions in drug development. A recent study introduces a computational framework that predicts DOE using gene and protein embeddings alongside genetic association data, providing a scalable way to refine therapeutic strategies across thousands of genes and diseases.
The researchers developed three complementary prediction models:
- Gene-level DOE-specific druggability model, which identifies druggable targets likely to respond to activation or inhibition, achieving a strong AUROC of 0.95.
- Gene-level DOE model, which predicts whether a target should be activated or inhibited independent of druggability, with an AUROC of 0.85.
- Gene-disease-specific DOE model, which captures how DOE may vary across disease contexts, reaching an AUROC of 0.59.
Notably, these models’ predictions correlated with clinical trial success rates, indicating potential for de-risking early-stage target discovery. The study also uncovered systematic biological differences between activator and inhibitor targets while activators remain concentrated in narrow classes like GPCRs.
The authors suggest that integrating these DOE predictions with existing target–disease association frameworks (e.g., Open Targets, Mantis-ML, GPS) could streamline therapeutic hypothesis generation, particularly for underexplored activation mechanisms. They note, however, that prospective validation is needed, as the current models are trained on historical drug data and rely on inferred genetic evidence.
Despite limitations this framework represents a significant advance in predictive pharmacology. By inferring DOE before clinical experimentation, researchers may better prioritize targets, avoid costly dead ends, and accelerate the path toward effective therapeutics.