Neuroplastic Self-Organizing Map (NPSOM): A Novel Approach for Analyzing Multiomics Data

Abstract

This paper introduces the Neuroplastic Self-Organizing Map (NPSOM), an unsupervised learning approach that enhances the traditional Self-Organizing Map (SOM) with neuroplasticity features. NPSOM incorporates mechanisms for synaptic decay, neurogenesis, and memory forgetting to improve the adaptability and performance of the SOM. By mimicking the brain’s ability to strengthen certain neurons’ synaptic connections and forget unnecessary information, NPSOM adapts to changing data distributions while retaining important information. The experimental results show that NPSOM outperforms MOSEGCN (best baseline) by 0.3% on the BRCA dataset while delivering similar performance on the ROSMAP dataset in terms of classification accuracy.

Methodology: Neuroplasticity in SOMs

NPSOM builds upon the traditional SOM by introducing three key biological-inspired mechanisms:

Synaptic Decay: Mimics the weakening of unused neural connections over time.
Neurogenesis: Allows the model to create new neurons to adapt to new data patterns.
Memory Forgetting: Enables the model to discard irrelevant or redundant information, focusing on the most salient features.

Figure 1: Importance Curve from NPSOM. This curve illustrates how the model identifies and prioritizes the most significant features across the learning process.

Experimental Results

We evaluated NPSOM on two major multiomics datasets: TCGA-BRCA (Breast Invasive Carcinoma) and ROSMAP (Alzheimer's detection).

Comparison with State-of-the-Art

Method	BRCA Acc.	ROSMAP Acc.
KNN	0.783	0.651
RF	0.768	0.754
LASSO	0.772	0.755
XGBoost	0.791	0.764
MOGONET	0.806	0.800
SEGCN	0.840	0.792
MOSEGCN	0.867	0.830
NPSOM+Proto Net (Ours)	0.8707 ± 0.012	0.8345 ± 0.016

Figure 2: BRCA Dataset Analysis. Final visualization of the NPSOM clustering and feature importance for the BRCA dataset.

Figure 3: ROSMAP Dataset Analysis. Final visualization of the NPSOM clustering and feature importance for the ROSMAP dataset.

Ablation Study

To understand the contribution of each neuroplasticity component, we conducted an ablation study.

Model	BRCA Acc (%)	ROSMAP Acc (%)
Full Model	87.07 ± 1.2	83.45 ± 1.6
Without Neurogenesis	68.74 ± 4.3	72.07 ± 2.7
Without Forgetting	82.25 ± 5.2	74.32 ± 5.7
Without Synaptic Decay	83.62 ± 3.1	72.09 ± 4.4

Conclusion

NPSOM demonstrates that incorporating neuroplasticity features into unsupervised learning models significantly enhances their ability to handle complex, high-dimensional multiomics data. By dynamically adjusting to data distributions, NPSOM provides a more robust and biologically plausible framework for feature selection and data analysis in bioinformatics.