Publication / Generative Models Validation via Manifold Recapitulation Analysis

Lumerical simulation setup and results for nanodisk and asymmetric nanostar geometries. A). STL files of (left) nanodisk and (right) asymmetric nanostars to show the geometry of the features imported into Lumerical for simulation. B) Lumerical simulation area setup: (Left) diagram/top-down view of the 75 nm nanostars on a substrate with a 500 nm pitch distance. Grid square is 100 nm × 100 nm; (Right) 3D diagram of Lumerical setup, where the red region represents silicon, the grey region represents silicon oxide, the yellow stars represents gold, and the orange box outlines the FDTD simulation area. The white box indicates the source location. C) Simulation results showing electric field magnitudes for 50 nm nanodisks and 75 nm nanostars at different pitch spacings. D) Simulation results demonstrating local field magnitude at the sharp tips of nanostars of several sizes compared to nanodisks of 50 nm.

Single-cell transcriptomics increasingly relies on nonlinear models to harness the dimensionality and growing volume of data. However, most model validation focuses on local manifold fidelity (e.g., Mean Squared Error and other data likelihood metrics), with little attention to the global manifold topology these models should ideally be learning. To address this limitation, we have implemented a robust scoring pipeline aimed at validating a model’s ability to reproduce the entire reference manifold. The Python library Cytobench demonstrates this approach, along with Jupyter Notebooks and an example dataset to help users get started with the workflow. Manifold recapitulation analysis can be used to develop and assess models intended to learn the full network of cellular dynamics, as well as to validate their performance on external datasets.

Read the full article here.