Neural networks implemented with dsd circuits

Abstract

Neural networks can be implemented with DNA strand displacement (DSD) circuits. The neural networks are designed and trained in silico taking into account the behavior of DSD circuits. Oligonucleotides comprising DSD circuits are synthesized and combined to form a neural network. In an implementation, the neural network may be a binary neural network in which the output from each neuron is a binary value and the weight of each neuron either maintains the incoming binary value or flips the binary value. Inputs to the neural network are one more oligonucleotides such as synthetic oligonucleotides containing digital data or natural oligonucleotides such as mRNA. Outputs from the neural networks may be oligonucleotides that are read by directly sequencing or oligonucleotides that generate signals such as by release of fluorescent reporters.

1 . A method of creating a neural network with DNA strand displacement gates, the method comprising:
 receiving, at an electronic computing device, a specification of the neural network comprising at least one hidden layer, wherein the specification indicates at least one oligonucleotide sequence as an input to the neural network;   training neuron and connection weights of the neural network in silico with a loss function to create a trained neural network, wherein the loss function also maximizes a distance of a combination of inputs for a neuron in the at least one hidden layer from a decision boundary of the neuron;   compiling the trained neural network to a plurality of DSD circuits;   obtaining the plurality of DSD circuits; and   combining the plurality of DSD circuits in one or more physical locations.   
     
     
         2 . The method of  claim 1 , wherein the neural network is a binary neural network, each neuron in the at least one hidden layer has an activation function that specifies either a positive weight or a negative weight, and connections to neurons in the at least one hidden layer have either a positive weight or a negative weight. 
     
     
         3 . The method of  claim 2 , wherein the loss function is a margin ranking loss function; and 
 training neuron and connection weights of the neural network further comprises using stochastic gradient descent to minimize a derivative of the margin ranking loss function such that the absolute value of the weights of the inputs for the neuron is maximized.   
     
     
         4 . The method of  claim 1 , wherein compiling the trained neural network to a plurality of DSD circuits comprises compiling each input to the neuron as a pair of DSD gates in dual-rail configuration. 
     
     
         5 . The method of  claim 1 , wherein compiling the trained neural network further comprises comparing multiple different simulations of chemical reaction networks (CRNs) representing reactions of the neural network to identify one or more CRN designs that maximizes a distance of the combination of inputs for the neuron from the decision boundary of the neuron. 
     
     
         6 . The method of  claim 1 , wherein combining the plurality of DSD circuits in one or more physical locations comprises combining a same concentration of oligonucleotides for each DSD circuit in each neuron in the at least one hidden layer. 
     
     
         7 . The method of  claim 1 , wherein combining the plurality of DSD circuits in one or more physical locations comprises combining a first plurality of DSD circuits corresponding to a first neuron at a first physical location and combining a second plurality of DSD circuits corresponding to a second neuron at a second physical location. 
     
     
         8 . The method of  claim 1 , wherein combining the plurality of DSD circuits in one or more physical locations comprises anchoring a single-stranded oligonucleotide from a one of the plurality of DSD circuits to a substrate. 
     
     
         9 . The method of  claim 1 , wherein the input to the neural network comprises a mRNA molecule; and 
 further comprising classifying the mRNA molecule using the neural network.   
     
     
         10 . The method of  claim 1 , wherein the input to the neural network comprises an oligonucleotide storing digital data; and 
 further comprising classifying the oligonucleotide using the neural network.   
     
     
         11 . An analog binary-neuron implemented with DNA strand displacement circuits, the neuron comprising:
 a pair of DSD circuits for each input to the neuron, wherein a first one of the pair of DSD circuits responds to a positive incoming binary signal and a second one of the pair of DSD circuits responds to a negative incoming binary signal and wherein each pair of DSD circuits releases either positive displacement strands or negative displacement strands; and   a threshold component or annihilator component that consumes the least numerous of the positive displacement strands or the negative displacement strands.   
     
     
         12 . The neuron of  claim 11 , wherein the positive incoming binary signal or the negative incoming binary signal is an oligonucleotide that is an output signal of another neuron in a previous layer of a BNN. 
     
     
         13 . The neuron of  claim 11 , further comprising an output gate that releases a first output signal in response to the positive displacement strands being most numerous and releases a second output signal in response to the negative displacement strands being most numerous. 
     
     
         14 . The neuron of  claim 13 , wherein a negative weight is implemented for the neuron by an output signal having a different value than the most numerous displacement strand and a positive weight is implemented for the neuron by the output signal having a same value as the most numerous displacement strand. 
     
     
         15 . The neuron of  claim 11 , wherein the first one of the pair of DSD circuits responds to the positive incoming binary signal by releasing positive displacement strands when a weight of the input is positive and by releasing negative displacement strands when a weight of the input is negative, and 
 wherein the second one of the pair of DSD circuits responds to the negative incoming binary signal by releasing positive displacement strands when a weight of the input is negative and by releasing negative displacement strands when a weight of the input is positive.