Brain-Inspired Biomolecular Networks for Reservoir Computing

Biomolecular structures assembled using the droplet interface bilayer (DIB) technique of lipid coated aqueous droplets in a hydrophobic medium allow construction of modular aqueous compartments separated by artificial lipid membranes. These membranes can be functionalized with active biomolecules to enable sensing and information processing, including but not limited to transmembrane proteins. When doped with alamethicin (Alm) or monazomycin (Mz), DIBs become volatile memristors that exhibit voltage-dependent short-term synaptic plasticity (STP). This property has led to their use as artificial synapses for neuromorphic hardware implementations of artificial intelligence algorithms, such as reservoir computing (RC).<br/>RC is an algorithm that uses specific device properties such as nonlinearity and fading STP to simplify classification of dynamic signals. We have previously demonstrated that individual Mz-doped DIB biomolecular synapses (BS) are suitable for use in RC classification and dynamic function learning, but little research has been done into forming BS networks for use in RC. An advantage of this approach is the ability to combine sensing and information processing into elements of the same network. Due to the customizability of individual compartment contents, sensing functions can be assigned to specific network locations by including thermoreceptors, mechanoreceptors, etc. into those droplets, but not others. In this work we present methods for rapid assembly and characterization of BS networks and schema for use in an RC framework, as well as hypothetical architectures for multimodal sensing.<br/>For assembly of BS networks, we develop a multi-welled “egg crate” style structure to hold droplets stationary to both maintain contact with neighboring droplets and electrodes in the bottom of each well. These electrodes are controlled and measured by multichannel NI-DAQs. Tested network structures will include linear networks, multilayer perceptrons of linear layers, and hexagonally packed clusters. To test the RC capability of each network, we use the UCI Epileptic Seizure Recognition Data Set to stimulate the reservoir. Additionally, using real world sensor data from a morphing airfoil in a wind tunnel, we input these signals as voltages into the droplet network and measure the conductance states of each node to classify and predict gusting behavior. For a visual task, we classify a custom set of 5x5 digits. We hypothesize that using a combination of Alm and Mz will yield higher accuracy than either alone in most cases.<br/>The above cases represent an electrical task, a mechanosensitive task, and a light sensitive task. In the future, these signals can be directly processed and classified by including the relevant transmembrane sensing channels used in organisms. Ongoing work seeks to stabilize these networks into devices that may be used in-device.