Wet TinyML: Chemical Neural Network Using Gene Regulation and Cell Plasticity

TL;DR

This study advances the GRNN concept by incorporating cell plasticity, to further exploit natural cell's adaptability, in order to diversify the GRNN search that can match larger spectrum as well as dynamic computing tasks.

Abstract

In our earlier work, we introduced the concept of Gene Regulatory Neural Network (GRNN), which utilizes natural neural network-like structures inherent in biological cells to perform computing tasks using chemical inputs. We define this form of chemical-based neural network as Wet TinyML. The GRNN structures are based on the gene regulatory network and have weights associated with each link based on the estimated interactions between the genes. The GRNNs can be used for conventional computing by employing an application-based search process similar to the Network Architecture Search. This study advances this concept by incorporating cell plasticity, to further exploit natural cell's adaptability, in order to diversify the GRNN search that can match larger spectrum as well as dynamic computing tasks. As an example application, we show that through the directed cell plasticity, we can extract the mathematical regression evolution enabling it to match to dynamic system applications. We also conduct energy analysis by comparing the chemical energy of the GRNN to its silicon counterpart, where this analysis includes both artificial neural network algorithms executed on von Neumann architecture as well as neuromorphic processors. The concept of Wet TinyML can pave the way for the new emergence of chemical-based, energy-efficient and miniature Biological AI.

Figures (9)

• Figure 1: Wet TinyML based on GRNNs extracted and searched in a bacterium to perform computing. Input chemicals trigger selective activation of relevant GRNN-subnetworks, rendering the GRNN a composite of many subnetworks. Gene products, diffusing into the cytoplasm forming cellular memory system that contributes to cell plasticity.
• Figure 2: In GRNN framework, gene-perceptron operate similarly to ANN perceptrons, processing inputs with weights influenced by multi-omic layer interactions (*), and time(t). Bacterial cells exhibit input-dependent plasticity by unique gene expression pathways varying with different inputs. Additionally, cells demonstrate temporal plasticity by altering GRNN subnetwork interaction weights over time.
• Figure 3: Illustration of the weight extraction process where a) elucidates the utilization of transcriptomic data considering gene $g_P$ as a target single-layer gene-preceptron and b) depicts the training process of minimizing the MSE between predicted and experiment expression levels.
• Figure 4: Power comparison between GRNN vs von Neumann and neuromorphic computing systems with respect to a) algorithmic complexity and b) structural complexity.
• Figure 5: Illustrations of a) the sparsity of gene expression and b) number of output node variations given the number of input nodes and the depth of the GRNN subnetwork.