 permeability for complex small molecules that fall outside conventional medicinal chemistry rules.){kind=link}
A recently published study in the Journal of Chemical Information and Modeling describes an AI system capable of predicting blood–brain barrier (BBB) permeability for complex small molecules that fall outside conventional medicinal chemistry rules. Developed by the biotech firm 1910, the model (CANDID-CNS) addresses two long-standing challenges in neuroscience drug discovery, forecasting BBB penetration for Beyond Rule of 5 (bRo5) compounds and capturing stereochemical effects that influence transport into the central nervous system (CNS).
The blood–brain barrier excludes nearly all large molecules and the majority of small molecules, making accurate permeability predictions a central obstacle in CNS drug development. Most approved CNS therapeutics conform to Lipinski’s Rule of 5, leaving bRo5 compounds largely unexplored despite their therapeutic potential. Existing computational models struggle to evaluate these molecules and typically overlook stereochemical variation, which can substantially alter BBB transport.
The published work reports that CANDID-CNS, built using an attentive graph neural network architecture, improves on established benchmarks. According to the study, the model surpasses Pfizer’s CNS MPO scoring method in identifying BBB-permeant bRo5 structures and correctly distinguishing stereoisomers with different CNS penetration profiles. It also demonstrated strong performance when tested against 1910’s proprietary molecule library.
Notably, the authors link the model’s predictions to underlying physicochemical determinants. Correlations with quantum-mechanical hydration free energy suggest that the system learns features related to passive permeability rather than relying solely on statistical associations. This mechanistic signal, the authors argue, supports the model’s ability to generalize beyond its training set.
CANDID-CNS™ does not just classify molecules – it recovers the physicochemical principles that drive BBB transport. Its predictions correlate with quantum mechanical hydration free energy, indicating that the model implicitly learns the thermodynamic determinants of passive permeability. That mechanistic signal enables CANDID-CNS™ to generalize and identify brain penetrant bRo5 molecules and stereoisomers.
Jesse Collins, Ph.D., Senior AI Research Scientist at 1910 and lead author of the JCIM publication
The company reports that CANDID-CNS contributed to the discovery of a non-opioid covalent inhibitor being developed for chronic pain, a program supported in part by the NIH HEAL Initiative. While the therapeutic implications of bRo5 CNS agents remain early-stage, the study highlights a computational strategy for widening the chemical space considered in neuroscience drug discovery.