Innovative Fusion in Neuron Classification: Shared Information from Human and Mouse Electrophysiological Data

The scientific community has long faced significant challenges in neuron classification. Accurate classification of neurons is crucial for understanding brain function in both healthy and diseased states. This study, led by Ofek Ophir, Orit Shefi, and Ofir Lindenbaum from Bar-Ilan University, published in the journal Neuroinformatics, proposes a novel machine learning framework that classifies neurons by jointly using electrophysiological data from humans and mice.

Research Background

Neurons are the basic units of the nervous system, and their classification has been a core issue in neuroscience since Ramon y Cajal published Histology of the Nervous System of Man and Vertebrates in 1995. Classifying neurons aids in consistent identification across different laboratories and experimental conditions, which is crucial for understanding brain function and its alterations in healthy and pathological states.

Research Origin

This study was conducted by the Faculty of Engineering and the Leslie & Susan Gonda Multidisciplinary Brain Research Center at Bar-Ilan University (Ramat-Gan, Israel). The article was accepted on June 10, 2024, and published in the journal Neuroinformatics, DOI link: https://doi.org/10.1007/s12021-024-09675-5.

Research Process

Data Sources

The research data primarily comes from the Allen Cell Types Database (ACTD), which contains biological data from single-cell recordings of adult mice and humans. The mouse data includes recordings from 1920 cells, while the human data includes recordings from 413 cells.

Data Preprocessing

Through the extraction of electrophysiological features, the research team analyzed 41 electrophysiological tabular features. Data records included four types of stimulation conditions: noise stimulus, ramp stimulus, rectangular stimulus, and short rectangular stimulus, each designed to elicit different types of action potential (AP) responses.

Research Tasks

The research was divided into two main tasks:

1. Classification of broad neuron types (excitatory and inhibitory) in humans and mice
2. Classification of neuron subtypes using electrophysiological data from mice

Classification Models

The study employed two neural network models: - Domain-Adaptive Neural Network (DANN) - Locally Sparse Interpretable Network (LSPIN)

Domain-Adaptive Classification with DANN Model

To address the scarcity of human data samples, the research team used the DANN model to embed shared information from mouse data into human data. By aligning the distributions of the two domains, cross-domain classification was achieved.

Multi-Label Classification with LSPIN Model

Given the richness of mouse samples, the research team used the LSPIN model to handle the classification of five subtypes. This approach overcame the overfitting problem associated with low sample size data and achieved interpretable classification by predicting the most informative features for each sample.

Research Results

Task One: Broad Neuron Type Classification in Humans and Mice

Using the DANN model, the study demonstrated that whole-cell current clamp recordings in the mouse brain are similar to those in the human brain. The model showed high accuracy in classifying neuron types in both humans and mice, achieving 95.0% accuracy for human samples and 97.4% accuracy for mouse samples.

Task Two: Mouse Neuron Subtype Classification

Using the LSPIN model, the study achieved a rare high accuracy of 91.6% in classifying five subtypes, outperforming traditional machine learning models such as Random Forest (RF), Support Vector Classifier (SVC), and XGBoost. Additionally, the model provided interpretability in feature selection for each subtype.

Research Significance

Scientific Value

This study utilized a cross-species data fusion approach to improve the accuracy of neuron classification, contributing to a deeper understanding of the electrophysiological characteristics of neurons. This method can be used in real-time clinical applications, providing a basis for early diagnosis and treatment planning.

Practical Value

The DANN model effectively addresses the domain transfer issue in neuron classification, enabling the model to generalize across different organisms while maintaining high accuracy. The LSPIN model reduces overfitting through feature selection, enhancing classification interpretability, which is crucial for clinical applications.

Research Highlights

• Cross-Domain Classification: By integrating data from humans and mice, the study addresses data scarcity and domain transfer issues.
• Model Interpretability: The LSPIN model not only provides highly accurate classification results but also reveals the importance of features, offering new insights for biological feature research.

Future Research Directions

Future research could consider: 1. Expanding the range of species classified to explore the evolutionary conservation of different neuron types. 2. Applying the DANN method in an unsupervised manner to further improve classification accuracy for human neuron data. 3. Evaluating the model’s generalization capability and verifying its applicability across different laboratories and experimental conditions. 4. Deeply analyzing the biological characteristics of each neuron type to increase the interpretability of model predictions.

Conclusion

This study provides new insights into neuron classification methods by demonstrating the importance of cross-species data fusion and feature selection through the DANN and LSPIN models. The results not only enhance scientific understanding but also offer reliable tools for practical applications.