Stability Analysis of Non-Linear Classifiers using Gene Regulatory Neural Network for Biological AI
This work addresses the challenge of creating programmable biological AI systems by bridging gene networks and neural networks, though it appears incremental as it builds on existing analogies between GRNs and ANNs.
The authors tackled the problem of modeling biological gene regulatory networks as neural networks for reliable computing, developing a Gene Regulatory Neural Network (GRNN) and analyzing its stability for non-linear classification, with results showing that parameter adjustments can shift classification boundaries to enable programmable GRNNs.
The Gene Regulatory Network (GRN) of biological cells governs a number of key functionalities that enables them to adapt and survive through different environmental conditions. Close observation of the GRN shows that the structure and operational principles resembles an Artificial Neural Network (ANN), which can pave the way for the development of Biological Artificial Intelligence. In particular, a gene's transcription and translation process resembles a sigmoidal-like property based on transcription factor inputs. In this paper, we develop a mathematical model of gene-perceptron using a dual-layered transcription-translation chemical reaction model, enabling us to transform a GRN into a Gene Regulatory Neural Network (GRNN). We perform stability analysis for each gene-perceptron within the fully-connected GRNN sub network to determine temporal as well as stable concentration outputs that will result in reliable computing performance. We focus on a non-linear classifier application for the GRNN, where we analyzed generic multi-layer GRNNs as well as E.Coli GRNN that is derived from trans-omic experimental data. Our analysis found that varying the parameters of the chemical reactions can allow us shift the boundaries of the classification region, laying the platform for programmable GRNNs that suit diverse application requirements.