MDE Outperforms PCA, t-SNE in Tracking Cortical Network Development
New visual informatics framework tracks neuron activity from day 23 to 64 with high fidelity.
A new study published on arXiv (2502.20862) presents a visual informatics framework for analyzing how cortical neuronal networks evolve during development and in response to stimulation. The team compared three dimensionality-reduction techniques—Minimum-Distortion Embedding (MDE), Principal Component Analysis (PCA), and t-distributed Stochastic Neighbor Embedding (t-SNE)—on both simulated and in vitro human cortical cultures. MDE with a cosine metric consistently outperformed the others: it preserved the cosine-shape radius within each condition and the pairwise distances between condition centroids, enabling clear tracking of network maturation from Day in Vitro 23 (DIV23) to DIV64.
In stimulation experiments, MDE separated activity phases (weak vs. strong stimulation) more distinctly than PCA and retained transient changes in within-phase variability that PCA missed. The authors emphasize that metric choice is critical: cosine distance between population activity vectors yields embeddings that better reflect population activity pattern changes than Euclidean distance. This framework offers a quantitative tool for visualizing network development and stimulation-induced changes, with potential applications in neuropharmacology, regenerative medicine, and brain-computer interfaces.
- MDE with cosine metric tracks cortical network maturation from DIV23 to DIV64 in human cultures.
- MDE separates stimulation phases better than PCA and preserves transient variability changes.
- Choice of distance metric (cosine vs. Euclidean) is critical for faithful dimensionality reduction of neuronal data.
Why It Matters
Enables quantitative visualization of neural network development and drug responses, aiding neuroscience and therapy design.