A method for constructing an in-memory computing DNA convolutional neural network

Abstract

This invention discloses a method for constructing a storage-computing DNA convolutional neural network, relating to the field of molecular computing technology. The method includes the following steps: selecting specific DNA molecules to form a reaction system; writing and decaying the states of the reaction system to generate a DNA memristor; and defining the output states of the DNA memristor; constructing a computational module, including basic computational units and advanced computational units; integrating the DNA memristor and the computational module to construct a hierarchical architecture of the convolutional neural network; and connecting the various layers of the hierarchical architecture of the convolutional neural network through interfaces to form a storage-computing DNA convolutional neural network. This solves the problem in existing technologies that struggle to meet the advanced application requirements of complex intracellular disease diagnosis, which necessitate simultaneous long-term information storage, historical backtracking, and real-time decision-making.

Description

Technical Field

[0001] This invention relates to the field of molecular computing technology, and in particular to a method for constructing a storage-computing DNA convolutional neural network. Background Technology

[0002] DNA computing, as an emerging computing paradigm, offers the possibility of constructing ultra-low-energy, highly parallel convolutional neural networks due to its programmable sequence characteristics, predictable base pairing rules, and complex reaction kinetics. The main principle of DNA computing is to simulate the computational process using the biochemical reactions of biological macromolecules. Since up to 10¹⁸ DNA molecules can participate in the reaction simultaneously in a tiny test tube, DNA computing inherently possesses extremely high parallelism and storage density, while consuming very little energy.

[0003] Most existing DNA neural networks still rely on independent molecular computational units, failing to effectively utilize the high-density storage characteristics of DNA molecules to achieve integrated storage and computation. During computation, weight information often exists in a fixed concentration of molecular form, making it impossible to store and process information within the same unit as biological neurons. This separation not only limits computational efficiency but also makes it difficult to meet the advanced application requirements, such as the diagnosis of complex intracellular diseases, which require simultaneous long-term information storage, historical review, and real-time decision-making. Summary of the Invention

[0004] To address the above problems, this invention provides a method for constructing a DNA convolutional neural network that integrates in-memory computing.

[0005] The present invention provides a method for constructing a storage-computing DNA convolutional neural network, comprising the following steps:

[0006] S1: Select specific DNA molecules to form a reaction system, perform state writing and state decay on the reaction system to generate a DNA memristor, and define the output state of the DNA memristor;

[0007] S2: Build the computing modules, including basic computing units and advanced computing units;

[0008] S3: Integrate the DNA memristor and the computing module to build a convolutional neural network hierarchical architecture;

[0009] S4: Connect the layers of the convolutional neural network hierarchical architecture through interfaces to form a DNA convolutional neural network that integrates in-memory computing.

[0010] Furthermore, the specific DNA molecule includes fuel molecule F, auxiliary catalyst chain I, low-energy state molecule S0, and high-energy state molecule S1.

[0011] Furthermore, the state writing step is as follows:

[0012] The fuel molecule F serves as the input signal, and with the assistance of the catalyst in the auxiliary catalyst chain I, it undergoes a forward chain displacement reaction, releasing byproducts. Its reaction formula is expressed as:

[0013] ;

[0014] Wherein, fast indicates that the reaction rate is accelerated under the action of the catalyst;

[0015] The state decay steps are as follows:

[0016] When the fuel molecule F is absent as an input signal, the high-energy molecule S1 spontaneously undergoes a reverse reaction, reverting to the low-energy molecule S0, as shown in the following reaction equation:

[0017] ;

[0018] Here, "slow" indicates that the reverse reaction has a slow reaction rate.

[0019] Furthermore, the steps for defining the output state of the DNA memristor are as follows:

[0020] The output state S of the DNA memristor is defined as the ratio of the concentration difference between high-energy state molecules S1 and low-energy state molecules S0 to their sum, expressed as: .

[0021] Furthermore, the basic operation unit includes addition, subtraction, multiplication, and division operations; the advanced operation unit includes multiply-accumulate operations, comparison operations, and report operations.

[0022] Furthermore, the construction steps of the addition reaction are as follows:

[0023] First addend chain A add The second addend chain B add It reacts with the first summing auxiliary chain G1 and the second summing auxiliary chain G2 respectively to generate the first summing intermediate and the second summing intermediate;

[0024] The first and second summation intermediates are respectively replaced by the first summation trigger chain T1 and the second summation trigger chain T2 to generate the same summation output chain SUM. The concentration of the summation output chain SUM is the sum of the addition reaction, expressed as follows: ;

[0025] The construction steps of the subtraction reaction are as follows:

[0026] Minuend chain A subGenerate the subtraction output chain SUB, and the subtraction chain B. sub An elimination chain H is generated, and the difference output chain SUB and the elimination chain H undergo an offset reaction and consume each other.

[0027] After the offsetting reaction reaches equilibrium, the concentration of the remaining subtractive output chain SUB is the difference of the subtraction reaction, expressed as: ;

[0028] The steps for constructing the multiplicative reaction are as follows:

[0029] First factor chain A mul The quadrature intermediate is generated by combining with the quadrature auxiliary chain G3, and the quadrature intermediate is combined with the second factor chain B. mul The product of the multiplication reaction is obtained by combining and replacing the multiplication trigger chain T3 with the multiplication output chain MUL. The concentration of the multiplication output chain MUL is the product of the multiplication reaction, expressed as: ;

[0030] The steps for constructing the division reaction are as follows:

[0031] Dividend chain A div Generate the quotient output chain DIV, and the divisor chain B. div An inhibitory complex is generated by reacting with the quotient auxiliary chain G4. The inhibitory complex reacts with the quotient output chain DIV and consumes each other.

[0032] After the inhibition reaction reaches equilibrium, the concentration of the remaining quotient output chain DIV is the quotient of the division reaction, expressed as: .

[0033] Furthermore, the construction steps for the multiply-accumulate operation are as follows:

[0034] The DNA memristors are arranged in a cross array structure, with each DNA memristor located at the intersection of the cross array structure.

[0035] A weight matrix W is constructed as the convolution kernel, and each weight element of the weight matrix W is stored at the corresponding crossover point position through the output state of the DNA memristor.

[0036] The input signal matrix X is transmitted along its own column direction and multiplied with each of the weight elements one by one to generate a product of multiple multiplication reactions;

[0037] The product of multiple multiplication reactions is propagated along the row direction of the weight matrix W, and the addition reaction is performed on the product of all multiplication reactions within the same convolution window belonging to the convolution kernel to generate a locally weighted sum Y, expressed as: ;

[0038] All the local weighted sums Y are arranged according to the spatial position of the signal matrix X to form the feature map of the convolution kernel;

[0039] The steps of the comparison operation are as follows:

[0040] For the first comparison chain A max Second comparison chain B max The subtraction reaction is performed in parallel to generate a difference signal, and the first comparison chain A is determined. max and the second comparison chain B max Size;

[0041] For the first comparison chain A max and the second comparison chain B max Perform reverse recovery and output the first comparison chain A. max and the second comparison chain B max Comparative chains with higher concentrations are represented as ;

[0042] The steps for the reporting operation are as follows:

[0043] The low-concentration noise signal in the output result of the calculation module is consumed by the threshold chain Th to generate a denoised signal;

[0044] By using the competing chain Wta, different categories of signal chains in the denoised signal react and annihilate each other, retaining only the winner chain with the highest initial concentration.

[0045] By amplifying chain F rep The concentration of the winner chain is amplified to generate a detection signal SIG.

[0046] Furthermore, the convolutional neural network hierarchical architecture includes an input layer, a convolutional layer, a pooling layer, and an output layer.

[0047] Furthermore, the construction steps of the input layer are as follows:

[0048] The pixels of the binarized image are converted into DNA concentration pulses, where data with a pixel value of 1 corresponds to a positive pulse signal, and data with a pixel value of 0 remains silent.

[0049] Each row of pixels in the binarized image is stored in one of the DNA memristors;

[0050] The steps for constructing the convolutional layer are as follows:

[0051] The concentration pulse of the DNA is preset with 16 sets of input combinations, and the weighted sum of each set of input combinations is pre-stored;

[0052] The comparison operation simulates the RelU function, removes negative values ​​from the local weighted sum Y, and outputs non-negative concentration feature values.

[0053] The steps for constructing the pooling layer are as follows:

[0054] The feature map of each convolutional kernel is divided into regions. Within each region, the maximum value of the non-negative concentration feature value is selected through the comparison operation, and a two-dimensional feature matrix is ​​output.

[0055] The steps for constructing the output layer are as follows:

[0056] The two-dimensional feature matrix is ​​expanded into a one-dimensional feature vector;

[0057] The one-dimensional feature vector is multiplied and accumulated with the pre-trained weight matrix to generate probability scores for different target categories.

[0058] The reporting operation is performed on all categories of the target category to filter out the target category with the highest probability score, and then converted into a fluorescent signal output as the final recognition result.

[0059] In summary, the present invention has at least the following beneficial effects:

[0060] 1. This invention utilizes the memristor properties of DNA memristors to directly store input and weight information in the molecular concentration ratio (state S) of the reaction system, achieving a molecular-level "in-memory computing" solution through DNA strand substitution reactions. This design not only simulates the characteristics of biological synapses but also provides a reference for constructing in-memory computing DNA neural networks.

[0061] 2. The DNA memristor of this invention features a unique "state decay and reset" mechanism, and the auxiliary catalyst is not consumed during the reaction. This means that after completing a calculation, the system can be automatically reset or maintain its state via control input, supporting cascaded operations.

[0062] 3. This invention decomposes the network into independent general-purpose modules such as addition, subtraction, multiplication, division, multiplication-accumulation, comparison, and reporting. These modules, like building blocks, have standardized input and output interfaces. This means that researchers can flexibly add or remove layers (such as increasing the depth of convolutional layers) according to their needs without redesigning the global sequence, thus solving the problem of the difficulty in constructing deep molecular networks.

[0063] 4. This invention introduces a reporting module at the output end that includes three stages: "competition consumption," "lateral suppression," and "signal amplification." This mechanism effectively filters out thermodynamic noise and significantly amplifies the effective signal, enabling the system to achieve near-computer-level accuracy in MNIST handwriting recognition (96.56%). Attached Figure Description

[0064] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0065] Figure 1 This is a schematic diagram of the composition and computation process of a DNA convolutional neural network module that integrates in-memory computing;

[0066] Figure 2 This is a schematic diagram of the principle of a reusable DNA memristor;

[0067] Figure 3 This is a schematic diagram of the reaction principle of the addition reaction module;

[0068] Figure 4 This is a schematic diagram of the reaction principle of the subtraction reaction module;

[0069] Figure 5 This is a schematic diagram of the reaction principle of the multiplication reaction module;

[0070] Figure 6 This is a schematic diagram of the division reaction module.

[0071] Figure 7 This is a diagram of a multiply-accumulate module based on a DNA memristor cross-array.

[0072] Figure 8 This is a diagram of the overall architecture of a DNA convolutional neural network that integrates in-memory computing. Detailed Implementation

[0073] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0074] The following is in conjunction with the appendix Figures 1 to 8 The present invention will be described in further detail below.

[0075] This invention provides a method for constructing a storage-computing DNA convolutional neural network, comprising the following steps:

[0076] Step 1: Construct a reusable DNA memristor.

[0077] The reaction system consists of four key DNA molecules: fuel molecule F, auxiliary catalyst chain I, low-energy state molecule S0, and high-energy state molecule S1. Fuel molecule F represents the input signal; auxiliary catalyst chain I is the core molecular carrier that carries the historical state memory within the DNA memristor's own reaction system. It can retain the regulatory effect of historical inputs on the concentration ratio of low-energy state molecule S0 and high-energy state molecule S1 in the DNA memristor, so that the current output state of the DNA memristor is simultaneously influenced by the current input signal F and the combined effects of previous inputs; the relative concentrations of S0 and S1 determine the output state.

[0078] By utilizing the principle of DNA strand displacement reaction, and through two biochemical processes of state writing and state decay, the setting and resetting functions of DNA memristors are realized, enabling DNA memristors to possess the memristor characteristics of remembering historical inputs.

[0079] like Figure 2 As shown, state writing occurs when fuel molecule F is introduced into the reaction system, and a rapid forward chain displacement reaction takes place with the assistance of auxiliary catalyst chain I. The low-energy molecule S0 combines with fuel molecule F and rapidly transforms into the high-energy molecule S1, simultaneously releasing the byproduct F'. The reaction equation is as follows:

[0080] ;

[0081] Here, fast indicates that the reaction rate is accelerated by the catalyst.

[0082] State decay occurs when, in the absence of an input signal, a high-energy molecule S1 undergoes a slow, spontaneous reverse reaction, reverting to a low-energy molecule S0, resulting in the decay of the state value. The reaction equation is as follows:

[0083] ;

[0084] Here, "slow" indicates that the reverse reaction has a slow reaction rate.

[0085] The output state S of a DNA memristor is defined as the ratio of the concentration difference between high-energy state molecules S1 and low-energy state molecules S0 to their sum, i.e., the normalized concentration difference, expressed as: This design utilizes the fact that the concentration ratio of molecules in two states can change continuously, and then maps the concentration ratio to a continuous output state through a normalization formula, thereby simulating the storage of continuous states of a memristor, and the reading process does not disrupt the current molecular concentration balance.

[0086] Step 2: Construct a basic computational module based on DNA strand replacement.

[0087] like Figure 1As shown, the four arithmetic operations are implemented using a DNA strand substitution reaction network, serving as the atomic operations for neural network computation.

[0088] like Figure 3 As shown, the addition reaction module: input the first addend chain A add Second addend chain B add The two addend chains react with the first summation auxiliary chain G1 and the second summation auxiliary chain G2 respectively to generate the first summation intermediate and the second summation intermediate.

[0089] The first and second summation intermediates are respectively replaced by the first summation trigger chain T1 and the second summation trigger chain T2 to generate the same summation output chain SUM. Finally, the concentration of the summation output chain SUM is the sum of the concentrations of the two addend chains, expressed as... .

[0090] like Figure 4 As shown, the subtraction reaction module employs a competitive cancellation mechanism, with the minuend chain A... sub Generate the subtraction output chain SUB, and the subtraction chain B. sub An elimination chain H is generated. The difference output chain SUB undergoes an irreversible combination reaction with the elimination chain H and consumes each other.

[0091] After the reaction reaches equilibrium, the concentration of the remaining subtractive output chain SUB is the difference, expressed as: .

[0092] like Figure 5 As shown, the multiplicative reaction module employs a bimolecular cooperative mechanism, with the first factor chain A... mul Combined with the quadrature auxiliary chain G3, a quadrature intermediate is generated. The quadrature intermediate must be combined with the second factor chain B. mul Only when both inputs are combined can the quadrature trigger chain T3 be used to generate the quadrature output chain MUL; the absence of either input will result in no output. The concentration of the quadrature output chain MUL is proportional to the product of the two factor chains, expressed as: .

[0093] like Figure 6 As shown, the division reaction module employs the principle of competitive inhibition, with the dividend chain A... div The generation of the quotient output chain DIV is driven by the quotient chain B. div By combining with the quotient auxiliary chain G4, a suppression complex is generated. The suppression complex consumes the quotient output chain DIV and releases it back to the divisor chain B. div .

[0094] After the reaction reaches equilibrium, the concentration of the remaining quotient output chain DIV is compared with that of the dividend chain A. div Divisor chain B div The quotient is directly proportional to the quotient, expressed as: .

[0095] It should be noted that, as Figures 3-6 The numbers 1-12 shown represent functional domains of a DNA strand, which are abstract numbers representing DNA strand segments and used to characterize complementary recognition relationships and strand substitution reaction logic between DNA strands; number 1 * -12 * The Watson-Crick complementary domains correspond one-to-one with the DNA strand fragments numbered 1-12.

[0096] Step 3: Construct a high-level computation module based on DNA strand replacement.

[0097] To meet the complex processing requirements of neural networks, advanced computing modules are further constructed.

[0098] like Figure 7 As shown, the multiply-accumulate module arranges multiple DNA memristors into a cross-array structure, with the DNA memristors located at the array intersections. A weight matrix W is stored as the convolution kernel. Each weight element Wi of the weight matrix W corresponds to the output state of one DNA memristor in the cross-array.

[0099] The input signal matrix X is transmitted along its own column direction and multiplied with each weight element one by one to generate intermediate products.

[0100] The intermediate products are transmitted along the row direction of the weight matrix W. An additive module is used to aggregate all intermediate products within the same convolutional window belonging to the convolutional kernel, generating a locally weighted sum Y, represented as: , where i is the index variable. This implements weighted summation and supports sliding window operations for convolution calculations.

[0101] All local weighted sums Y are arranged according to the spatial position of the signal matrix X to form the feature map of the convolution kernel.

[0102] Comparison module: Used to implement the ReLU activation function and pooling operations, for the first comparison chain A max Second comparison chain B max The difference signal is generated through parallel subtraction reactions, and the size of the two comparison chains is determined.

[0103] Then, the concentration of the larger value is recovered by using the difference signal in reverse, and it is expressed as: .

[0104] Reporting module: Used for classification decisions in the output layer, including the competitive consumption stage. It generates a denoised signal by using the low-concentration noise signal in the output result of the threshold chain Th consumption operation module.

[0105] In the lateral suppression stage, a competing chain Wta is introduced, which causes different types of signal chains in the denoised signal to react and annihilate each other, ensuring that only the winner chain with the highest initial concentration remains.

[0106] During the signal amplification phase, the winner chain and the amplification chain F rep The reaction continuously amplifies the concentration of the winner chain and generates the final detection signal SIG, achieving a high signal-to-noise ratio readout of the signal.

[0107] Step 4: Construct an in-memory computing DNA convolutional neural network architecture.

[0108] like Figure 8 As shown, the above modules are integrated into a complete convolutional neural network, which includes the following layers:

[0109] Input layer: Encodes the raw input image by converting the binarized image pixels (0 / 1) into DNA concentration pulses. The rule is that a positive pulse signal is generated for each pixel "1", while pixels "0" remain silent, with a pulse interval of 1 hour. The storage strategy uses a "one row, one pulse" approach, storing one row of pixel signals into one DNA memristor; a 28*28 image corresponds to 28 DNA memristors.

[0110] Convolutional Layer: Feature extraction utilizes a multiply-accumulate module array to perform convolution operations, setting a 2x2 convolution kernel with a stride of 2. For binary input in a 2x2 region, 16 possible input combinations are preset (2^4 = 16). The weighted sum of these combinations is calculated in parallel using a DNA memristor array, significantly reducing computational complexity. Finally, a comparison module is used to simulate the ReLU function (f(x) = max(0, x)) to remove negative values ​​from the convolution result.

[0111] Pooling layer: Perform 2*2 max pooling. Divide the 14*14 feature map into 7*7 regions, and use a comparison module within each region to select the maximum concentration value, thereby achieving feature dimensionality reduction.

[0112] Output layer: The fully connected layer utilizes the basic computation module to multiply and accumulate the pooled feature vectors with the pre-trained weight matrix, generating probability scores for each category (e.g., digits 0-9). Finally, it connects to the reporting module, which uses a winner-takes-all mechanism to identify the category with the highest score and outputs the corresponding fluorescence signal as the diagnostic result.

[0113] Example: Implementation process of handwritten digit recognition application scenario.

[0114] This embodiment demonstrates the specific steps for implementing MNIST handwritten digit recognition using the aforementioned DNA in-memory convolutional neural network.

[0115] Data preprocessing and binarization: Grayscale images from the MNIST dataset are selected as the original input, with an image size of 28*28 pixels. Then, binarization is performed, setting pixels with a grayscale value of 0 to logic "1" (representing validity) and setting the remaining grayscale values ​​to logic "0".

[0116] Input Layer: Establishes a mapping from binary values ​​to DNA signals. A logic "1" corresponds to generating a positive concentration pulse signal, while a logic "0" maintains a pulse-free state. The encoding interval between pulses is set to 1 hour to ensure discriminability. Image input employs a "one row, one pulse" storage strategy. Each row (28 pixels) of the image is converted into a sequence of temporal pulse signals, input and stored in a set of DNA memristors; one image occupies a total of 28 memristors. Convolutional kernel weights employ a "one element, one pulse" storage strategy. For a 2x2 convolutional kernel, its four weight values ​​are stored in four independent DNA memristor units.

[0117] Convolutional Layer: Convolution operations utilize a multiply-accumulate module based on a DNA memristor crossover array. A 2x2 convolution kernel is used with a stride of 2. For a 2x2 binary input region, all possible input combinations (totaling 2^32) are pre-computed. 4 =16 types). The neural network identifies the current input combination type and directly calls the corresponding DNA memristor array for parallel weighted summation. After convolution processing, the original 28*28 image is converted into a 14*14 feature map, where the value of each node represents the weighted sum of local features.

[0118] Pooling layer: A 2x2 max-pooling window with a stride of 2 is applied to the 14x14 feature map. The feature map is divided into 7x7 regions. Within each region, a comparison module uses a competitive reaction to select the signal with the highest concentration as the output, ultimately generating a 7x7 dimensionality-reduced feature matrix.

[0119] Output Layer: The pooled signal is input into the fully connected layer. The basic computation module performs multiplication and accumulation operations with the pre-trained weight matrix (obtained through offline supervised learning in MATLAB) to calculate the scores for the 10 categories representing the digits 0-9. The scores for the 10 categories are then input into the reporting module. Through thresholding to eliminate noise and competition chain lateral suppression, the highest-scoring category signal is preserved and specifically amplified by the amplification chain, ultimately outputting a unique classification result. Simulation tests show that this embodiment achieves a basic prediction accuracy of 85.15% on the MNIST test set. By optimizing the network structure (e.g., using four 3x3 convolutional kernels), the accuracy can be improved to 96.56%.

[0120] The above are merely preferred embodiments of the invention and are not intended to limit the invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the protection scope of the invention.

Claims

1. A method for constructing a storage-based DNA convolutional neural network, characterized in that, Includes the following steps: S1: Select specific DNA molecules to form a reaction system, perform state writing and state decay on the reaction system to generate a DNA memristor, and define the output state of the DNA memristor; S2: Build the computing modules, including basic computing units and advanced computing units; S3: Integrate the DNA memristor and the computing module to build a convolutional neural network hierarchical architecture; The convolutional neural network hierarchical architecture includes an input layer, a convolutional layer, a pooling layer, and an output layer; The steps for constructing the input layer are as follows: The pixels of the binarized image are converted into DNA concentration pulses, where data with a pixel value of 1 corresponds to a positive pulse signal, and data with a pixel value of 0 remains silent. Each row of pixels in the binarized image is stored in one of the DNA memristors; The steps for constructing the convolutional layer are as follows: The concentration pulse of the DNA is preset with 16 sets of input combinations, and the weighted sum of each set of input combinations is pre-stored; By simulating the ReLU function, negative values ​​in the pre-stored weighted sum of the input combination are removed, and non-negative concentration feature values ​​are output. The steps for constructing the pooling layer are as follows: The feature map of each convolutional kernel is divided into regions, and the maximum value of the non-negative concentration feature value is selected in each region to output a two-dimensional feature matrix. The steps for constructing the output layer are as follows: The two-dimensional feature matrix is ​​expanded into a one-dimensional feature vector; The one-dimensional feature vector is multiplied and accumulated with the pre-trained weight matrix to generate probability scores for different target categories. The target category with the highest probability score among all categories of the target category is selected and converted into a fluorescent signal output as the final recognition result; S4: Connect the layers of the convolutional neural network hierarchical architecture through interfaces to form a DNA convolutional neural network that integrates in-memory computing.

2. The method for constructing a storage-computing integrated DNA convolutional neural network according to claim 1, characterized in that, The specific DNA molecules include fuel molecule F, auxiliary catalyst chain I, low-energy molecule S0, and high-energy molecule S1.

3. The method for constructing a storage-computing integrated DNA convolutional neural network according to claim 2, characterized in that, The steps for writing the state are as follows: The fuel molecule F serves as the input signal, and with the assistance of the catalyst in the auxiliary catalyst chain I, it undergoes a forward chain displacement reaction, releasing byproducts. Its reaction formula is expressed as: ； Wherein, fast indicates that the reaction rate is accelerated under the action of the catalyst; The state decay steps are as follows: When the fuel molecule F is absent as an input signal, the high-energy molecule S1 spontaneously undergoes a reverse reaction, reverting to the low-energy molecule S0, as shown in the following reaction equation: ； Here, "slow" indicates that the reverse reaction has a slow reaction rate.

4. The method for constructing a storage-computing integrated DNA convolutional neural network according to claim 2, characterized in that, The steps for defining the output state of the DNA memristor are as follows: The output state S of the DNA memristor is defined as the ratio of the concentration difference between high-energy state molecules S1 and low-energy state molecules S0 to their sum, expressed as: .

5. The method for constructing a storage-computing integrated DNA convolutional neural network according to claim 1, characterized in that, The basic arithmetic units include addition, subtraction, multiplication, and division operations; the advanced arithmetic units include multiply-accumulate operations, comparison operations, and report operations.

6. The method for constructing a storage-computing integrated DNA convolutional neural network according to claim 5, characterized in that, The steps for constructing the addition reaction are as follows: First addend chain A add The second addend chain B add It reacts with the first summing auxiliary chain G1 and the second summing auxiliary chain G2 respectively to generate the first summing intermediate and the second summing intermediate; The first and second summation intermediates are respectively replaced by the first summation trigger chain T1 and the second summation trigger chain T2 to generate the same summation output chain SUM. The concentration of the summation output chain SUM is the sum of the addition reaction, expressed as follows: ; The construction steps of the subtractive reaction are as follows: Minuend chain A sub Generate the subtraction output chain SUB, and the subtraction chain B. sub An elimination chain H is generated, and the difference output chain SUB and the elimination chain H undergo an offset reaction and consume each other. After the offsetting reaction reaches equilibrium, the concentration of the remaining subtractive output chain SUB is the difference of the subtraction reaction, expressed as: ; The steps for constructing the multiplicative reaction are as follows: First factor chain A mul The quadrature intermediate is generated by combining with the quadrature auxiliary chain G3, and the quadrature intermediate is combined with the second factor chain B. mul The product of the multiplication reaction is obtained by combining and replacing the multiplication trigger chain T3 with the multiplication output chain MUL. The concentration of the multiplication output chain MUL is the product of the multiplication reaction, expressed as: ; The steps for constructing the division reaction are as follows: Dividend chain A div Generate the quotient output chain DIV, and the divisor chain B. div An inhibitory complex is generated by reacting with the quotient auxiliary chain G4. The inhibitory complex reacts with the quotient output chain DIV and consumes each other. After the inhibition reaction reaches equilibrium, the concentration of the remaining quotient output chain DIV is the quotient of the division reaction, expressed as: .

7. The method for constructing a storage-computing integrated DNA convolutional neural network according to claim 6, characterized in that, The construction steps for the multiply-accumulate operation are as follows: The DNA memristors are arranged in a cross array structure, with each DNA memristor located at the intersection of the cross array structure. A weight matrix W is constructed as the convolution kernel, and each weight element of the weight matrix W is stored at the corresponding crossover point position through the output state of the DNA memristor. The input signal matrix X is transmitted along its own column direction and multiplied with each of the weight elements one by one to generate a product of multiple multiplication reactions; The product of multiple multiplication reactions is propagated along the row direction of the weight matrix W, and the addition reaction is performed on the product of all multiplication reactions within the same convolution window belonging to the convolution kernel to generate a locally weighted sum Y, expressed as: ; All the local weighted sums Y are arranged according to the spatial position of the signal matrix X to form the feature map of the convolution kernel; The steps of the comparison operation are as follows: For the first comparison chain A max Second comparison chain B max The subtraction reaction is performed in parallel to generate a difference signal, and the first comparison chain A is determined. max and the second comparison chain B max Size; For the first comparison chain A max and the second comparison chain B max Perform reverse recovery and output the first comparison chain A. max and the second comparison chain B max The chain with the higher concentration of the two is denoted as ; The steps for the reporting operation are as follows: The low-concentration noise signal in the output result of the calculation module is consumed by the threshold chain Th to generate a denoised signal; By using the competing chain Wta, different categories of signal chains in the denoised signal react and annihilate each other, retaining only the winner chain with the highest initial concentration. By amplifying chain F rep The concentration of the winner chain is amplified to generate a detection signal SIG.

8. The method for constructing a storage-computing integrated DNA convolutional neural network according to claim 7, characterized in that, The method for selecting the target category with the highest probability score among all categories of the target category is implemented using the reporting operation; The method of selecting the maximum value among the non-negative concentration feature values ​​in each region is implemented using the comparison operation; The method of removing negative values ​​from the pre-stored weighted sum by simulating the ReLU function is implemented using the comparison operation. Specifically, the negative values ​​in the pre-stored weighted sum are the negative values ​​in the local weighted sum Y.

Images