Synthetic Data Generation for Classifying Electrophysiological and Morpho-Electrophysiological Neurons from Mouse...

Area of Science:

• Neuroscience
• Computational Biology
• Machine Learning

Background:

• Accurate neuronal cell type classification is crucial for understanding brain organization.
• Existing multimodal neuron datasets are often scarce and imbalanced across subclasses.
• Developing effective data augmentation strategies is vital for improving classification performance.

Purpose of the Study:

• To benchmark synthetic data augmentation methods for classifying electrophysiology-defined neuronal classes (e-types).
• To evaluate the impact of augmentation on prediction accuracy using electrophysiology alone (E→e-type) and combined morphology/electrophysiology (M+E→e-type).
• To assess the biological realism of generated synthetic data.

Main Methods:

• Utilized the Allen Cell Types mouse visual cortex dataset with 17 e-type labels.
• Established real-data baselines using various classifier families.
• Applied Synthetic Minority Over-sampling Technique (SMOTE) and deep generative models (VAE, GAN, normalizing flows, DDPM) for data augmentation.
• Developed a fidelity framework to evaluate biological realism of synthetic data.

Main Results:

• Synthetic data augmentation yielded substantial generalization gains, particularly in high-dimensional feature spaces.
• Dimensionality reduction largely negated the benefits of augmentation.
• SMOTE provided the most robust and consistent improvements across tasks and augmentation levels.
• Most synthetic datasets maintained biological diversity, with minor deviations in rare subclasses.

Conclusions:

• Synthetic data augmentation is effective for improving neuronal subtype classification, especially with imbalanced datasets.
• SMOTE offers a reliable and consistent approach for enhancing classification performance.
• The developed fidelity framework is valuable for validating the biological realism of synthetic data in neuroscience research.