1 Billion Compounds.

The screen ran on a Standard_ND96amsr_A100_v4 instance on Azure ML — 8 NVIDIA A100 80GB GPUs, 96 CPU cores, and 900GB of system RAM. The BBB-Nuke v0.12.0 pipeline was deployed as a streaming architecture: 8 GPU workers running batched GCN inference for pKa prediction, and 48 CPU workers handling standardization, physicochemical properties, CNS-MPO, and efflux RF classifiers in parallel.

Compounds were drawn from three sources: Enamine REAL (make-on-demand chemical space), PubChem (public bioactive compounds), and existing proprietary libraries. Each compound was standardized via RDKit, scored for 10 physicochemical properties, passed through a graph convolutional network for pKa prediction, evaluated by 7 efflux transporter RF classifiers, and assigned a final P(BBB) by the v5 gradient boosting classifier.

Post-screening analysis included Murcko scaffold decomposition to identify core chemical scaffolds, Morgan/ECFP4 fingerprint generation for structural similarity, and GPU-accelerated clustering via RAPIDS cuML (MiniBatch K-Means with k=1,000 clusters and HDBSCAN for fine-grained sub-structure).

The 3D visualization above uses UMAP (Uniform Manifold Approximation and Projection) to project the 2048-dimensional fingerprint space into 2D and 3D coordinates. A stratified subsample of 500,000 compounds — enriched for high-scoring hits and structurally diverse cluster representatives — is rendered via WebGL for interactive exploration.