DNA computing approach to semantic knowledge ... - IOS Press

International Journal of Hybrid Intelligent Systems 2 (2005) 1–12 IOS Press

1

DNA computing approach to semantic knowledge representation Yusei Tsuboi∗, Zuwairie Ibrahim and Osamu Ono Institute of Applied DNA Computing, Graduate School of Science & Technology, Meiji University, 1-1-1 Higashimita, Tama-ku, Kawasaki-shi, Kanagawa, 214-8571, Japan Abstract. DNA computing has a lot of potential, in terms of ability to implement a relational database with circular molecules. In this work, a new DNA-based semantic model is proposed and described theoretically for implementing DNA based memories. This model, referred to as ‘semantic model based on molecular computing’ (SMC), has the structure of a graph formed by the set of all attribute-value pairs contained in the set of represented objects, plus a tag node for each object. Each path in the network, from an initial object-representing tag node to a terminal node represents the object named on the tag. Input of a set of input strands will result in the formation of object-representing dsDNAs via parallel self-assembly, from encoded ssDNAs representing (value, attribute)-pairs (nodes), as directed by ssDNA splinting strands representing relations (edges) in the network. The computational complexity of the implementation is estimated via simple simulation, which indicates the advantage of the approach over a sequential model. We believe that the semantic models are rather suitable for DNA-based memory, and that this proposal is the first such approach in the semantic networks area.

1. Introduction Our research group focuses on development in the area of semantic networks (semantic nets) [14], via a new computational paradigm. Human information processing often involves comparing concepts. There are various ways of assessing concept similarity, which vary depending on the adopted model of knowledge representation. In featural representations, concepts are represented by sets of features. In Quillian’s model of semantic memory, concepts are represented by the relationship name via links. Links are labeled by the name of the relationship and are assigned criteriality tags that attest to the importance of the link. In artificial computer implementations, criteriality tags are numerical values the represent the degree of association between concept pairs (i.e., how often the link is traversed), and the nature of the association. The association is positive if the existence of that link indicates some sort of similarity between the end nodes, and negative otherwise. For example, superordinate links (the term used for ‘is-a . . .’ relationships) have a positive association, while “is-not-a” links have a negative association. Just as there are at least two research communities that deal necessarily with questions of generalization in science, there are at least two bodies of knowledge concerned with representation of the known world as discovered and explained by science. On one hand, knowledge can be fundamentally procedural and causal; on the other, knowledge is fundamentally judgemental [19]. Naturally, the knowledge representation schemas are quite different; thus, the manners in which knowledge may be processed to generate new knowledge in the two models are also quite different. Semantic modeling provides a richer ∗

Corresponding author. Tel.: +81 44 934 7289; Fax: +81 44 934 7909; E-mail: [email protected].

1448-5869/05/$17.00 © 2005 – IOS Press, AKI and the authors. All rights reserved

2

Y. Tsuboi et al. / DNA computing approach to semantic knowledge representation

data structuring capability for database applications. In particular, research in this area has articulated a number of constructs that provide mechanisms for presenting structurally complex interrelations among data typically arising in commercial applications. Baum [4] first proposed the idea of using DNA annealing to perform parallel associative search in large databases encoded as sets of DNA strands. This idea is very appealing since it represents a natural way to execute a computational task in massively parallel fashion. Moreover, the required volume scales only linearly with the database size. Retrievals and deletions under stringent conditions occur reliably (98%) within very short times (100’s of milliseconds), regardless of the degree of stringency of the recall or the number of simultaneous queries in the input. Adleman’s [11] ground-breaking work demonstrated a way to use molecules for computational purposes. The extreme compactness of DNA as a data storage medium is nothing short of incredible. Since a mole contains 6.02 × 10 23 DNA base monomers, and the mean molecular weight of a monomer is approximately 350 grams/mole, then 1 gram of DNA contains 2.1 × 1021 DNA bases. Since each DNA base can encode 2 bits (there are 4 types of bases), it follows that 1 gram of DNA can store approximately 4.2 × 10 21 bits of information. In contrast, conventional storage technologies can store at most roughly 10 9 bits per gram, so that DNA has the potential of storing data on the order of 10 12 more compactly than conventional storage technologies [9]. In this paper, we propose a semantic model and molecular computation algorithm for implementing a DNA based semantic memory system using DNA reactions in a straightforward manner. It is demonstrated that a DNA-based molecular computing strategy may be applied in the area of semantic networks. The basis of the strategy is an Adleman-like scheme which exploits the random motion of DNA strands in the vessel to effect parallel computation. 2. Rotation of DNA The proposed model in this paper is represented by double-stranded DNAs (dsDNA) from theoretical point of view. In the vessel, after the strands have been mixed at defined conditions, circular strands will be formed via self-hybridization. Although most theoretical works on DNA-based computation do not consider computational feasibility, and very little experimental work has been reported [7,22], in this section two experimental works regarding database operations using rotation of DNA molecules are introduced. The mechanism of forming circular DNAs is made clear by using virtual DNA fragments. In addition, the essential tools of biochemical operations are outlined. 2.1. Related works A strategy for an inference engine based on circular DNA molecules derived from a plasmid is introduced, along with details regarding practical implementation [16]. A set of indirect rules may be interpreted as a rule tree. Indirect rules create the edges of the tree, each of which is implemented with an indirect rule molecule. At the tree nodes, a process of indirect rule self-hybridization is performed. Rules as edges of a graph have created the complete rule tress with a root at the top and leaves at the bottom. The resulting inferences can be ‘read’ following the experiments with higher precision and efficiency. Arita, et al. [12] suggests a method for encoding data, and reports experimental results for performing concatenation and rotation of DNA. This work uses fewer enzymes, and is more reliable. Furthermore, the concatenation and join procedure is much more reliable and efficient than a method previously proposed in [8,15]. This work’s principal significance is in the demonstration of the possibility of DNA computing in terms of a relational database with circular molecules. These works also contributed to the

Y. Tsuboi et al. / DNA computing approach to semantic knowledge representation A

B

C

3

D

5'

3' 3' AJ B

B

JC

C

JD

D

JA

5'

a. DNA fragments representing path AJ BJ CJ DJ A

A

B AJ B

D JA

D

B JC C JD

C

b. Circular molecule: path AJ BJ CJ DJ A

Fig. 1. Mechanism of circular DNA formation.

development of ‘memory machines’, which are machines employing a very large memory that implement simple operations. However, these works are not based on semantic networks. A semantic network would be an effective model to store a set of knowledge in circular molecules. It is thought that one method of constructing a memory with power near to that of man is to construct a semantic model based on molecular computing. In this light, we ask: ‘what type of the model is most suitable for implementing such a DNA-based architecture?’ We then present a semantic model derived from the concepts of basic semantic networks, with simple chemical protocols for implementation. 2.2. DNA Computing approach with principal bio-chemical operations In Adelman’s scheme [11], each edge and node of a Hamilton Path Problem instance graph is represented by a single-stranded DNA (ssDNA) of length 20 bases. The edge u → v from node u to nodev is constructed to be Watson-Crick complementary to the node sequences derived from the 3  10-mer of nodeu and the 5 10-mer of nodeu. For instance, Figure. 1(a) illustrates a set of DNA fragments representing a path from A → B → C → D → A, adding a last edge fragment D → A to form a closed circuit, which plays an important role in forming a circular DNA. We next outline the principal operations from DNA biotechnology involved in model implementation, such as hybridization and ligation. DNA computation also requires the tools of PCR (Polymerase chain reaction) and gel electrophoresis for analysis, as described below. Hybridization & Ligation: The set of DNA fragments is inserted into a test tube. Each ssDNA will anneal to a complementary sequence under defined reaction conditions in the tube. The DNA sequences are mixed in the presence of a DNA ligase. This enzyme will form a covalent bond between each pair of DNA molecules, as long as they are splinted by a third strand formed by adjacent complementary

4

Y. Tsuboi et al. / DNA computing approach to semantic knowledge representation

O

A1

Am A2

V11

V12

Vnm

Fig. 2. A basic semantic net. (Ai, Vsi) (Ai, Vsi)

(Ai+1, Vt, i+1)

(Ai,Vti)

b. OR

a. AND

Fig. 3. AND/OR graph.

single-stranded overhangs (hybridization). Thus, the sequences are ligated to form a circular DNA duplex which represents an A → B → C → D → A path, as shown in Fig. 1(b). PCR: PCR is a technique which is used to amplify the number of copies of a specific region of DNA, in order to produce enough DNA to be adequately tested. This technique can be used to identify, with very high-probability, disease-causing viruses and/or bacteria, a deceased person, or a criminal suspect. Gel electrophoresis: Gel electrophoresis refers to the technique in which molecules are forced across a span of gel by application of an electrical potential gradient, as shown Fig. 8(a). Activated electrodes at either end of the gel provide the driving force. A molecule’s shape determines how rapidly the electric field drives it through the gelatinous medium.

3. Structure of basic semantic model In this section, we describe how to design a semantic model based on DNA computing. Semantic networks would mimic the intellectual ability of people if they were based on associative memories. The basic structure of a semantic network is a two-dimensional graph, similar to a network. It is relatively easy for humans to deal with a semantic network, because it represents an object (or concept) constructed from knowledge based on human memories. The semantic network consists of a set of ternary relations formed by: an Object, O; an Attribute, A; and an Attribute Value, V. In general, this list representation is denoted as: {< O, Ai , Vji > |i = 1, 2, ..., m; j = 1, 2, ..., n}

Y. Tsuboi et al. / DNA computing approach to semantic knowledge representation

(Color , Red)

(Apple)

5

(Size, Small)

(Shape, Circle)

(Color , Green)

(Size, Medium)

Fig. 4. Simple object model of an apple; The three determined attributes are shape, color, size.


 
 


  


  
  
  
  
  
 


    !

  ! 

 !  !

  !   ! 

  ! 

 !

  ! 

 !

  ! 

 !

  ! 

 !

  ! 

  !     

  

  !    ! 

   


 

Fig. 5. Semantic Model Based on Molecular Computing (SMC), which collectively models a set of objects, given a total number of attributes, m.

A basic semantic network may be described as a graph with nodes, edges, and labels representing the set of relations, as shown in Fig. 2. O is reasoned out by the relation between A i and Vji . Because the semantic network is simply standardized with vertexes and connecting edges, it is a suitable system to support the search for multiple objects in parallel, and for use as a knowledge-based system. In general, semantic network size increases with the number of attributes or attribute values. The other hand, it is imperative to transform complicated graphs into simpler ones. An AND/OR graph enables the reduction of graph size, and facilitates easy understanding. Thus, instead of using the standard existing semantic network structure described above, in the next section we define a new model, developed to make the most of DNA computing.

4. Newly-defined semantic model based on DNA computing Firstly, a tag with a name of an object is assigned to an initial vertex in the graph. After we determine the number and the kinds of the object’s attributes, each (attribute, attribute value) pair is assigned to an additional vertex. Secondly,the relationship between vertexes and edges is represented using a new defined AND/OR graph. As shown in Fig. 3(a), a directed edge in the terminal direction is inserted

6

Y. Tsuboi et al. / DNA computing approach to semantic knowledge representation

between the vertexes in series except for the following case. If there are two vertices which have same attributes but different attribute values, each of the directed edges is connected in parallel, as shown in Fig. 3(b). Each edge denotes only the connection between the vertexes in the directed graph. Finally, labels are attached to the vertexes, such as ‘(Tag)’ and ‘(Attribute, Attribute Value)’. Each node denotes either the name of the object or both an attribute and an attribute value. In short, a single path from an initial node to a terminal node corresponds to one object named on the tag. We newly define this graph as the knowledge representation model. This model represents an object, as described by the combinations between the nodes connected by the edges. For example, Fig. 4 illustrates this object representation in the context of an apple (named via the tag). An overall graph is then formed by the union of a set of such basic objects, each of which is described in similar, simple fashion. Figure 5 shows an example of such a network. We name such a graph a semantic model based on molecular computing (SMC). An SMC contains all of the attributes common to every object as well as each of the attribute values. Attribute layers consist of attribute values, lined up. If an object has no value for a certain attribute, that attribute value is assigned ‘no value’.

5. DNA computation for semantic memory In this section, a computational algorithm is described for a DNA-based semantic memory system, using SMC. Here, there are an input object and two reference objects. An input object is given by: {< O(I) , Ai , Vji > |i = 1, 2, ..., m; j = 1, 2, . . . , n}.

Reference objects are separately given by (p = q ): {< O(P ) , Ai , Vki > |i = 1, 2, ..., m; k = 1, 2, . . . , x}, {< O(Q) , Ai , Vli > |i = 1, 2, ..., m; l = 1, 2, . . . , y}.

Semantic information regarding O (P ) and O (Q) is stored in the semantic memory as knowledge bases. As a result of the algorithm, O (I) is classified into the two reference objects using DNA molecules. 5.1. DNA representation of SMC Each of the nodes and edges within an SMC may be represented by a DNA strand, as follows. Each node is mapped onto a unique ssDNA oligonucleotide, in a DNA library of strands. In the DNA library, rows indicate attributes, while columns indicate attribute values. Each DNA sequence is designed according to these relations to prevent mishybridization with other unmatching sequences. For example, if oligonucleotides are encoded with a strand length of 20 bases, the reference objects, O (P ) and O (Q) may be represented by the tag-node sequences shown in Table 1. Each edge is then constructed according to Adleman’s scheme. In addition, a fragment representing an edge from each terminal node to an initial node is added in order to form circular DNA. In this way, the model’s reference objects, O (P ) and O (Q) are represented by assembled dsDNAs. Figure 6 illustrates one of the paths shown for the apple model in Fig. 4, as represented by a dsDNA molecule (Apple) → (Shape, Circle) → (Color, Red) → (Size, Medium) → (Apple).

Y. Tsuboi et al. / DNA computing approach to semantic knowledge representation

7

Table 1 Object O(P ) and O(Q) sequences

(A pple)

Object

Sequence

O(P ) O(Q)

5’ATAGGCATTATCGGAGATCG3’ 5’TATCGGAAATTAAGTACCAG3’

(Shape, Circ le)

(Color, Red)

(Size, Midium )

5'

3' 3'

5' (A pple) J Shape,Circl e)

(Shape, Circl e) J C olor, Red)

(Col or , Red) J (Size, Mediu m)

(Size, Medium) J (A pple)

Fig. 6. One of the dsDNAs represented by the graph (Apple) in Fig. 4.

5.2. Experimental protocols Synthesis of a set of DNA fragments representing the reference objects, and an input object proceeds as follows. – Reference object The ssDNA representation of each edge, except a fragment representing an edge from each terminal node to an initial node, and tag in the network is synthesized as a knowledge based molecule. – Input object Attribute values are extracted from an input object separately, for a previously determined set of attributes, {Ai }. Using the attributes and attribute values, a ssDNA is synthesized as an input molecule to be combined with the DNA library. The following steps demonstrate the overall experimental protocol, using primitive chemical operations. Step 1: For reasonable operation, each of the knowledge based and the input molecules must first be amplified to a sufficiently high concentration. Knowledge based molecules are first inserted into a test tube, followed by addition of the input molecules. It is assumed that each ssDNA will spontaneously anneal to complementary sequences present in the mixture, under defined reaction conditions. The ssDNA sequences are then mixed in the presence of DNA ligase. As a result, all possible linear dsDNAs representing the paths are generated at random. Step 2: The ligation products are first divided into two equal aliquots in test tube P and Q as shown Fig. 7. PCR process is executed to each solution with different primers for increasing sufficiently. We could find the sequences for the primers from complement to the tag sequence and from complement to the terminal edge of SMC, not the final circularizing ssDNAs, such as (Color, Red) → (Size, Medium) strand in Fig. 6. Second, the final circularizing fragments amplified adequately in dsDNA representing reference objects O (P ) and O (Q) is separately added to each of the tubes. Finally, the linear dsDNAs come to circular dsDNAs in the presence of DNA ligase. Step 3: The resulting set of dsDNAs must then be analyzed to determine the specific set of represented objects produced by the reaction. The DNA mixtures are each subjected to gel electrophoresis to determine the presence of the dsDNAs representing the correct strands, which then appear as discrete bands on the gel. The length of each generated dsDNA, denoted as L S, is given by the simple relation: L S = L D(N A + 1),

8

Y. Tsuboi et al. / DNA computing approach to semantic knowledge representation DNA products

Test tube P

Test tube Q

Fig. 7. After amplification with PCR, circular dsDNAs are formed by adding final circularizing DNAs.

where L D is the length of each synthesized DNA fragment, and N A is the number of attribute layers. From the tubes, the products from reaction P, Q are placed into lanes P and Q respectively. For instance, consider that the reference object is an apple as shown in Fig. 4, L D = 20, N A = 3. After gel electrophoresis is executed, the result of the analysis is shown in Fig. 8(b). We find out dsDNAs of 80 bp (base pair) exist in the lane.

6. Simple simulation approach The proposed model is a knowledge based memory, implemented via DNA molecules, which is in some sense similar to human memory due to its inherent massive parallelism. This performance is not realized in artificial, sequential models of computing. Although simulations will be interesting, the inherent advantages provided by the design will therefore be evident only when using real DNA molecules. It might be necessary to evaluate the advantages of the proposed DNA-based model, as compared with performance on a silicon-based computer. It is commonly regarded as difficult to evaluate the simulation of a chemical reaction on a silicon-based computer. DNA-based computers integrate software with hardware and calculate in parallel. A direct attempt to simulate the implemented reaction on a normal silicon-based computer will be compromised by the potential for a combinatorial explosion in the number of distinct path molecules. Some studies on artificial intelligence have been performed with regards to avoiding such increases in knowledge and computational complexity. For this reason, in order to demonstrate the advantage of the proposed model over a sequential model, we estimate the computational complexity required for solution, assuming that every ssDNA encounters all others in the test tube. It is possible to classify into a certain object by the combinations between input molecules

Y. Tsuboi et al. / DNA computing approach to semantic knowledge representation

9

DNA -

DNA length

Marker

Lane P

Lane Q

Long

80 bp

+

Short

a: A typical example of gel electrophoresis

b: Solution

Fig. 8. Analysis for solution with gel electrophoresis.

and knowledge based molecules. Therefore, it is reasonable to expect the number of combinations to increase with the number of objects and attributes. Figure 9 shows the relationship between the attribute layers and the number of combinations. The number of combinations is estimated separately for a simple, sequential architecture, and for a DNA-based architecture when 3, 100, and 1000 reference objects exist in the semantic memory in order to investigate the characteristic of exponential increase. The operation number required for a sequential architecture in which there are 3, 100 and 1000 objects is illustrated by blue, green and red lines, respectively. Each of these lines increases exponentially with the number of attribute layers. In contrast, a single light blue line indicates the operation number required for a DNA-based architecture for each of the case of 3, 100, and 1000 objects. This line increases logarithmically with the number of attribute layers. In this case, the number of combinations does not depend on the number of target objects, since the proposed application requires only DNA self-assembly, which proceeds for all objects in parallel. The transverse axis of Fig. 9 denotes the number of attributes. Even if this value increases, computational complexity logarithmically increases. It is clear that DNA computing approach is effective in spite of treating complicated representation. Although if reasonable computation is performed in vitro, a huge number of DNA molecules will be required for initial pools, our idea is concentrated to building a SMC proposed here. 7. Practical issues and applications for artificial intelligence It provides a necessary DNA computing-chemical process including experimental operation implementations which relied on previous works. However, the adopted protocol is very simple. Standard techniques from DNA biotechnology, such as annealing, ligation, and gel electrophoresis are required. Several experiments have been conducted to assess the performance of a DNA-based database realized by biochemical reactions. The result of these experiments indicates the interesting features of the method. In this work, each (attribute, attribute value) pair is represented by a random length-20 DNA oligonucleotide, encoded as a string on {A, T, C, G}. There are two issues to consider: the choice of

10

Y. Tsuboi et al. / DNA computing approach to semantic knowledge representation

The number of combinations

With a DNA-based architecture, each of the case of 3,100, and 1,000 objects

The number of attribute layers Fig. 9. Simulation result.

appropriate strand length, and the sequence design of the DNA. As a given set of knowledge increases, it may not be possible to reliably represent each of the node and edges in the accompanying network with 20-mers. However, this issue can be resolved simply by adopting a longer fragment length. In practical sequences design, we have to consider an effective way to select sequences which avoid mismatched error hybridizations. Word design strategies for DNA-based computation have already been investigated. Substantial progress has been reported on this issue [1,23,6,10,14,17,21,22]. We expect that this issue, which is critical to overall process reliability, will be resolved satisfactorily in the near future. In the light of these issues, we will need to carefully design best sequences, prior to actual experiment. Some work in the artificial intelligence field has been focused on how to arrange and store a set of knowledge within memories. By semantic model composed of nodes and edges, not logical rule, a certain object is assigned by a tag node which links other nodes have information on itself. By picking the nodes up, it is possible to reach for the object at a stretch. If different objects have common attribute and attribute value nodes, they can share the node, which enables a reduction in the number of required DNA fragments. Moreover, SMC is represented by both attributes and attribute values with the concept of an AND/OR graph. As a result, the knowledge representation capacity of the proposed semantic model is very much greater than that of others [4,7,8,21,23]. Therefore, we believe that the proposed semantic model is very suitable for a DNA-based memory, and that this proposal is the first such approach in the semantic networks area. We have the idea of supporting emerging applications with the model, as shown in Fig. 10, which illustrates a generalized QA system, using DNA computing for the artificial intelligence. In a DNA-based semantic memory, DNA concatenation works in solution in an automatic and parallel fashion, as directed via encoded Watson-Crick complementary. A large scale semantic information set is stored in the vessel and reacts chemically, from which a correct answer is then extracted. Perhaps writing steps may be done with extreme rapidity. 8. Conclusion In this paper, a novel approach for a semantic model of DNA-based memory has been discussed. Following a description of the underlying semantic model, it has been shown how to store a knowledge

Y. Tsuboi et al. / DNA computing approach to semantic knowledge representation

Question

Analysis Answer

DNA-based semantic memo ry

Reaction

11

Fig. 10. Overall DNA computing-based QA system for artificial intelligence.

set by means of DNA circular molecules, along with a simple implementation of semantic retrieval. Our theory shows that a molecular approach is potentially valuable for building the most suitable model. It is possible to treat any objects if they are represented by the proposed model. This approach exhibits a number of advantages over traditional electronic machines. The simulation result suggests the proposed model be effective in parallel computation even if complicated models representing reference objects increase. The process of molecular self-assembly mimics the properties of an associative memory. In the future work a sophisticated model should be developed, with an improved knowledge-representation capacity. In addition, it is important to consider about a huge number of DNA molecules for initial pools. To achieve reliable performance, the optimization of some parameters, such as reaction-temperature, oligonucleotide concentrations, reaction time, etc. will be experimentally tested.

Acknowledgements The authors would like to thank Dr. J.A. Rose of the University of Tokyo for helpful comments and discussion that led to many improvements in this paper.

References [1] [2] [3] [4] [5] [6] [7] [8] [9]

A.G. Frutos, Q. Liu, A.J. Thiel, A.M.W. Sanner, A.E. Condon, L.M. Smith and R.M. Corn, Demonstration of a word design strategy for DNA computing on surfaces, Nucl. Acids Res. 25 (1997), 4748–4757. A. Marathe, A.E. Condon and R.M. Corn, On combinatorial DNA word design, J. Comp. Biol. 8 (2001), 201–219. D. Faulhammer, A.R. Cukras, R.J. Lipton and L.F. Landweber, Molecualr computation: RNA solution to chess problems, Proc. Natl. Acad. Sci. 97 (2000), 1385–1389. E.B. Baum, How to build an associative memory vastly larger than the brain, Science 268 (1995), 583–585. G. Kaplan, P. Cecchi and A. Ligchaber, DNA based molecular computation: template-template Interactions in PCR, Proc. 2nd Annual Meeting on DNA Based Computers, 1996, pp. 159–171. J.A. Rose, R. Deaton, D. Franceschetti, M. Garzon and S.E. Stevens, Jr., A statistical mechanical treatment of error in the annealing biostep of DNA computation, Proc. GECCO’99, 1999, pp. 1829–1834. J. Chen, R. Deaton and Y.-Z. Wang, A DNA-based memory with in vitro learning and associative recall, LNCS 2943, 2003, pp. 145–156. J.H. Reif, Parallel molecular computation: models and simulations, Proc. the 7th Annual Symaposium on Parallel Algorithms and Architectures, 1995, pp. 213–223. J.H. Reif, T.H. LaBean, M. Pirrung, V.S. Rana, B. Guo, C. Kingsford and G.S. Wickham, Experimental construction of very large scale DNA databases with associative search capability, Proc. a DIMACS Workshop: DNA Based Computers (DIMACS27), 2002, pp. 231–247.

12 [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] [21] [22] [23]

Y. Tsuboi et al. / DNA computing approach to semantic knowledge representation J. SantaLucia, H. Allawi and P. Seneviratne, Improved nearest-neighbor parameters for predicting DNA Duplex Stability, Biochemistry 35(11) (1996), 355–356. L.M. Adleman, Molecular computation of solutions to combinatorial problems, Science 266 (1994), 583–585. M. Arita, M. Hagiya and A. Suyama, Joining and rotating data with molecules, Proc. IEEE International Conference on Evolutionary Computation, 1997, pp. 243–248. M. Guarnieri, F. Fliss and C. Bancroft, Making DNA add, Science 273 (1996), 220–223. M.R. Quillian and M. Minsky, Semantic memory, semantic information processing, MIT Press, Cambridge, MA, 1968, pp. 216–270. P.W.K. Rothemund, A DNA and restriction enzyme implementation of turing machines, Proc. a DIMACS Workshop: DNA Based Computers (DIMACS 27), 1995, pp. 213–223. P. Wasiewicz, T. Janczak and J.J. Mulawka, The inference based on molecular computing, Cybernetics and Systems: An International Journal 31 (2000), 283–315. R. Deaton, M. Garzon, R.C. Murphy, J.A. Rose, D.R. Franceschetti and S.E. Stevens, Jr., Reliability and efficiency of a DNA-based computation, Phys. Review Letters 80 (1998), 417–420. R. Deaton, C.R. Murphy, M. Garzon, D.R. Franceschetti and S.E. Stevens, Jr., Good encodings for DNA-based solutions to combinatorial problems, Proc. DNA Based Computers DIMACS Series in Discrete Mathematics and Theoretical Computer, Science 44 (1999), 247–258. R.W. Blanning, Management applications of expert systems, Information and Management 7 (1984), 311–316. R. Penchovsky and J. Ackermann, DNA library design for molecular computation, Journal of Computational Biology 10(2) (2003), 215–229. R.S. Braich, C. Johnson, P.W.K. Rothemund, D. Hwang, N. Chevlyapov and L.M. Adleman, Solution of a satisfiability problem on a gel-based DNA computer, LNCS 2054 (2001), 27–42. R.S. Braich, N. Chevlypov, C. Johnson, P.W.K. Rothemund and L.M. Adleman, Solution of a 20-valiable 3-SAT problem on a DNA Computer, Science 296 (2002), 499–502. S. Roweis, E. Winfree, R. Burgoyne, N.N. Chelyapov, M.F. Goodman, P.W. Rothermund and L.M. Adleman, A sticker based model for DNA computations, Journal of Computational Biology 268 (1998), 615–629.