NGEP Validator

Predicting parvalbumin expression from neuronal morphology using deep learning

Current Scope: Pvalb in Neocortex

This validator currently predicts parvalbumin (Pvalb) expression in mouse neocortical neurons using deep learning on 14 engineered morphological features. The model was trained on 7,500 neurons from NeuroMorpho.org and validated against the Allen Brain Atlas.

Coming soon: Multi-gene support (GAD1, GFAP, VIP, SST), region-specific models across cortex, hippocampus, and cerebellum, and cross-species validation for human and rat neurons.

Backend Status: ● Initializing...

Run Validation

The model will generate predictions for this many neurons fetched from NeuroMorpho.org. Each run uses a random sample to test generalization to unseen data.

Validation Results

Summary

Processing...

Available Models

Models are automatically loaded from your backend. Click the button below to refresh.
Loading models...

About NGEP

NGEP (Neuron Gene Expression Prediction) is a computational framework for predicting cell-type-specific gene expression from neuronal 3D morphology. It uses a feedforward neural network trained on open-access data from NeuroMorpho.org and the Allen Brain Atlas to infer parvalbumin mRNA levels, demonstrating that neuronal shape carries significant information about molecular state.

Clinical Significance: Parvalbumin is expressed almost exclusively in fast-spiking GABAergic interneurons critical for gamma oscillations (30-100 Hz), temporal precision of neural firing, and seizure prevention. Pvalb dysfunction is implicated in autism, schizophrenia, epilepsy, Alzheimer's disease, fragile X syndrome, and ADHD. Rather than expensive, destructive molecular assays, morphology-based expression prediction enables rapid, non-invasive characterization of neuronal type.

Model Performance

R² Score: 0.3956 ± 0.0270 (explains ~40% of variance)
Pearson Correlation: 0.6322 ± 0.0200
RMSE: 0.6691 ± 0.0155
MAE: 0.5073 ± 0.0128
Statistical Significance: p-value < 10⁻⁵¹ across all validation folds

The model explains approximately 40% of variance in Pvalb expression from morphology alone. Remaining variance is attributable to gene regulatory networks, epigenetic state, developmental history, and local circuit context. The robust performance across 10-fold stratified cross-validation (R² std = 0.027) confirms generalization capability, with the lowest fold still achieving R² = 0.34.

Morphological Features

Five base measurements are extracted from each neuron's SWC reconstruction and expanded to 14 engineered features:

• Soma radius — Cell body size; relates to metabolism
• Total dendritic length — Cumulative extent of all branches
• Bifurcation count — Number of branch points
• Terminal count — Number of leaf nodes (synaptic endpoints)
• Branch density — Compactness of branching pattern

Features Coming Soon

Multiple Genes

Support for additional genes beyond Pvalb including GAD1, GFAP, VIP, SST, and other markers for comprehensive cellular characterization and disease-relevant gene expression mapping.

Brain Regions

Region-specific models for cortex, hippocampus, cerebellum, and other brain structures to improve prediction accuracy and enable circuit-level analysis across the whole brain.

Species Support

Cross-species validation for mouse, human, rat, and other organisms with species-specific prediction models to enable translational research and patient-derived iPSC analysis.

Patient Stratification

Tools for identifying circuit dysfunction from morphological imaging to enable targeted therapy development and precision medicine applications in neuropsychiatric disorders.

Biomarker Integration

Morphological features as accessible, imaging-based biomarkers for drug screening, therapy validation, and organoid quality control without requiring genomic assays.

Mechanistic Insight

Feature importance analysis to understand which morphological attributes drive expression, linking neuronal structure to molecular function at scale.