NFMap Architecture
Continuous Normalizing Flow Manifold Projections & Generative Topographic Mapping for Chemical Space.
The Three Pillars of NFMap
1. Normalizing Flows
Invertible, volume-preserving neural transformations (RealNVP / Glow) that map complex, high-dimensional Morgan fingerprint embeddings into smooth Gaussian latent density spaces without information collapse.
2. UMAP Topology
Riemannian manifold learning enforces local neighborhood preservation. Ensures activity cliffs, scaffold similarities, and pharmacophore geometries remain clustered naturally.
3. GTM Probabilistic Grid
Generative Topographic Mapping constructs continuous probabilistic reference landscapes, allowing instant client-side WASM projections and smooth property density overlays.
Deterministic Continuous Projection
Project any SMILES chemical structure onto precomputed continuous manifolds in real-time. Calculate exact coordinates, bioactivity predictions, and distance to active target centers.
Latent Space De Novo Sampling
Invert the Normalizing Flow to sample novel valid chemical structures from under-explored regions of the continuous manifold or traverse geodesics between active drug leads.