Screening the Bioactivity of the P450 Enzyme by Spiking Neural Networks

The current study investigated the potential usage of spiking neural networks to score P450 bioactivity, predicting the enzyme's ability to catalyze xenobiotics. Numerical experiments focused on enzyme P450 3A4 and 2C9 isoforms to evaluate optimal network configurations and molecular fingerprint length in forecasting compounds where P450 was active or not. Reacher structural representations of the molecules improved the SNN outcomes employing leaky-integrate-and-fire neuronal models. In addition, this neuron performed slightly better than the one that included synaptic decay. Overall, the accuracies obtained in a benchmark library of compounds were in line with other machine learning techniques encountered in the previous literature, which further supports the applicability of spiking neural networks for quantitative structure-activity analysis. The SNN hyperparameter evaluation found substantial stability across simulations. In the future, the application of selected SNN configurations in conjunction with neuromorphic hardware could be considered as a high-performance computing alternative for virtual chemoinformatics screening to significantly improve energy efficiency and achieve accelerated calculations.