-
Categories
-
Pharmaceutical Intermediates
-
Active Pharmaceutical Ingredients
-
Food Additives
- Industrial Coatings
- Agrochemicals
- Dyes and Pigments
- Surfactant
- Flavors and Fragrances
- Chemical Reagents
- Catalyst and Auxiliary
- Natural Products
- Inorganic Chemistry
-
Organic Chemistry
-
Biochemical Engineering
- Analytical Chemistry
- Cosmetic Ingredient
-
Pharmaceutical Intermediates
Promotion
ECHEMI Mall
Wholesale
Weekly Price
Exhibition
News
-
Trade Service
This paper on DNA circuit scaling up molecular with DNA-based dna-based-based neural network, published in Nature, is led by Professor Qian Lulu of the California Institute of Technology, whose research focuses on molecular programming of synthetic nucleic acid systems, involving biochemical molecular loops, DNA neural networks and molecular robots.
molecular-level pattern recognition plays a crucial role in the basic functionof of biological organisms, and before that, researchers have reported on using neural networks based on the linear threshold loop and Hopfield loop DNA to deal with such problems, but they can handle them on a limited scale, with no more than four patterns, each of which consists of four different DNA molecules.
in this article, the author uses a loop design called "winner-take-all" that is more computationally capable and better able to escape the limits of the number and complexity of patterns to be identified. using the seesaw DNA developed by this team, the
, by cleverly designing DNA sequences to correspond to the signal values of each node in the neural network architecture at the concentration of a certain DNA molecule, the authors tested the DNA loop-to-MNIST handwriting digital data set based on the "winner-take-all" strategy on multiple test sets of small to large, pattern-type-to-simple to complex.
and they have also verified that when human-induced "pollution" is added, the molecular loop can still effectively classify according to similarity, demonstrating its tolerance for high complexity and noisy environments.
specifically, the network designed in this paper can be divided into five parts, each of which can be achieved by simple reactions between DNA: 1. Weight multiplication.
this step on the network refers to the binary (0 or 1) input xi multiplied by the weight wij, the corresponding reaction is the substance Xi catalytic Wij reaction to generate the intermediate product Pij.
the pij is generated only when Xij is present, and Pij's concentration is determined by Wij, i.e. the concentration of Wij becomes a weighting value.
2. Weighted sum.
in the network this refers to the weighted sum of i inputs to j intermediate nodes.
molecular reactions, SGj is able to react with Pij, all intermediate substances in the same neuron, to generate Sj.
3. Pairs are annihilated and the winners eat.
this is a very interesting step of design.
Winner-take-all, the winner-take calculation is a simple competitive neural network model in which only the weighting of binary inputs and the output of the largest neurons are turned on.
this step of the molecular reaction is called "pair annihilation", the different kinds of Sj pairs of the union of the anihilation, can also be said that each Sj will destroy his competitor Sj, until the last Sj - the highest concentration of the winner.
4. Signal recovery.
Sj is combined with the recovery material RGj to produce the output Yj.
5. Fluorescent report output.
the fluorology report group Repj is combined with Yj to detect that the intensity of the fluorescent signal corresponds to yj's output signal value.
the above reactions are the seesaw reactions and collaborative hybridization of DNA molecules, which are achieved through displacement and migration between DNA chains, and they simply use the target pattern as a weight for the selection of weighted molecules.
the authors first examined the feasibility of the winner-take-all strategy, experimented with two weighted and chain S1 and S2, and then validated the molecular loop's ability to recognize four-bit patterns using a complete network. in the article,
the authors also raise three questions for the loop: the number of target patterns that can be processed at the same time, whether patterns of contamination or damage can be identified, dimensional variations of the DNA loop in higher complexity, and the use of MNIST in the latter article to answer these questions one by one.
in the 1-9 digital classification problem, taking into account the limitations of individual pair annihilation methods for large-scale pattern recognition, the author adopts the grouping strategy, and recognizes a number by two sets of outputs, effectively saving the compute time and loop size.
this paper constructs a new and interesting biochemical molecular loop, which shows good classification effect in the face of the complexity and pollution-based pattern recognition problem.
the authors think that the winner-take-all strategy and the method of using target pattern as a weight are worthy of exploring the further expansion of DNA neural networks.
Source: Wang Chu Task Force.