Neuroadaptive Intelligence: A Biologically-Inspired, Self-Repairing Neural Layer with Adaptive Learning for Molecular Graph Classification
Md. Azizul Hakim, Rashedul Arefin Ifty, Khaled Eabne Delowar, Samiul Azam Shuvo, Md Saimul Hoque Sawon, Md. Reduanul Haque Shakib · 2025
Inspired by the resilience and adaptability of biological neural systems, a novel neural network architecture is proposed to address persistent challenges in molecular graph classification. Drawing upon mechanisms such as calcium-dependent activity regulation, homeostatic control, and adaptive repair, the architecture is designed to maintain neuronal stability during training while mitigating common issues including saturation, neuron inactivity, and representational drift. The approach was evaluated on the benchmark AIDS Antiviral Screen dataset, where a classification accuracy of 99.8% was achieved using 10-fold cross-validation—surpassing existing state-of-the-art models by 0.25%. Enhanced performance was particularly evident in the classification of moderately active compounds, a category historically associated with high prediction uncertainty. Through biologically inspired modulation of internal states and targeted interventions upon signs of degradation, the proposed method enables more robust learning from sparse, imbalanced, and noisy molecular data. These findings suggest that the selective integration of neurobiological principles can lead to significant advancements in molecular machine learning and virtual screening for drug discovery.