CANDID-CNS: AI Unlocks Stereochemistry and Beyond Rule of 5 to Predict CNS Penetration of Small Molecules

Neuroscience is the most difficult therapeutic area for pharmaceutical drug discovery. The blood-brain barrier (BBB) prevents ∼100% of large molecules and >98% of small molecules from penetrating the central nervous system (CNS). Most small molecule drugs designed for the CNS are Rule of 5 (Ro5) molecules. Thus, Beyond Rule of 5 (bRo5) molecules, which have significant potential to act on especially difficult drug targets and indications, represent an untapped chemical class for novel CNS therapeutics. However, bRo5 small molecules face three key challenges: (i) exclusion from most medicinal chemistry drug designs, (ii) difficulty penetrating the BBB, and (iii) inability of existing computational methods to predict the CNS penetration of bRo5 molecules. In addition, the stereochemistry of small molecules is a potential contributor to CNS penetration, but current computational methods do not adequately account for it. Here, we present CANDID-CNS, an artificial intelligence (AI) model with an attentive graph neural network (GNN) architecture, which overcomes the barriers in bRo5 and stereochemistry. Specifically, CANDID-CNS outperforms the industry standard at: (1) prioritizing bRo5 molecules for CNS penetration (CANDID-CNS 87% AUPRC compared to industry standard 56%), (2) distinguishing stereoisomers of differing CNS penetration (CANDID-CNS 68% AUROC compared to industry standard 50%), and (3) predicting the CNS penetration of a holdout set of our 1910 proprietary precision neuroscience molecules (CANDID-CNS 90% AUROC compared to industry standard 81%). CANDID-CNS represents a breakthrough in AI-driven drug discovery by unlocking the potential of bRo5 and stereochemistry to significantly expand the chemical space of small molecules that can be developed into therapeutics for debilitating diseases of the CNS.