Computing with DNA the Manipulation of DNA to Solve Mathematical Problems Is Redeﬁning What Is Meant by “Computation”

Computing with DNA The manipulation of DNA to solve mathematical problems is redeﬁning what is meant by “computation”

by Leonard M. Adleman

omputer. The word conjures tunately, I had been remarkably unsuc- ated exclusively with the physical sci- up images of keyboards and cessful in communicating my ideas to ences. Biology was now the study of in- C monitors. Terms like “ROM,” the AIDS research community. So, in an formation stored in DNA—strings of “RAM,” “gigabyte” and “megahertz” effort to become a more persuasive ad- four letters: A, T, G and C for the bases come to mind. We have grown accus- vocate, I decided to acquire a deeper adenine, thymine, guanine and cyto- tomed to the idea that computation understanding of the biology of HIV. sine—and of the transformations that takes place using electronic compo- Hence, the molecular biology lab. There, information undergoes in the cell. There nents on a silicon substrate. under the guidance of Nickolas Chelya- was mathematics here! But must it be this way? The comput- pov (now chief scientist in my own lab- Late one evening, while lying in bed er that you are using to read these oratory), I began to learn the methods reading Watson’s text, I came to a de- words bears little resemblance to a PC. of modern biology. scription of DNA polymerase. This is Perhaps our view of computation is too I was fascinated. With my own hands, the king of enzymes—the maker of life. limited. What if computers were ubiq- I was creating DNA that did not exist Under appropriate conditions, given a uitous and could be found in many in nature. And I was introducing it into strand of DNA, DNA polymerase pro- forms? Could a liquid computer exist in bacteria, where it acted as a blueprint for duces a second “Watson-Crick” com- which interacting molecules perform producing proteins that would change plementary strand, in which every C is computations? The answer is yes. This the very nature of the organism. replaced by a G, every G by a C, every is the story of the DNA computer. During this period of intense learn- A by a T and every T by an A. For ex- ing, I began reading the classic text The ample, given a molecule with the se- Rediscovering Biology Molecular Biology of the Gene, co-au- quence CATGTC, DNA polymerase thored by James D. Watson of Watson- will produce a new molecule with the y involvement in this story began Crick fame. My concept of biology was sequence GTACAG. The polymerase Min 1993, when I walked into a being dramatically transformed. Biolo- enables DNA to reproduce, which in molecular biology lab for the first time. gy was no longer the science of things turn allows cells to reproduce and ulti- Although I am a mathematician and that smelled funny in refrigerators (my mately allows you to reproduce. For a computer scientist, I had done a bit of view from undergraduate days in the strict reductionist, the replication of AIDS research, which I believed and still 1960s at the University of California at DNA by DNA polymerase is what life believe to be of importance [see “Bal- Berkeley). The field was undergoing a is all about. anced Immunity,” by John Rennie; Sci- revolution and was rapidly acquiring DNA polymerase is an amazing little entific American, May 1993]. Unfor- the depth and power previously associ- nanomachine, a single molecule that

54 Scientific American August 1998 Computing with DNA Copyright 1998 Scientific American, Inc. DNA MOLECULES—with their sequences of adenine, thymine, guanine and cytosine (represented by the letters A, T, G and C)—can be used to store information and to perform computations. The molecule shown here in color, GCAGTCGGACTGGGCTATGTCCGA, encodes the solution to the sample Hamiltonian Path Problem on the next page.

“hops” onto a strand of DNA and slides and a mechanism called a finite control, My initial thinking was to make a along it, “reading” each base it passes which moved along the “input” tape DNA computer in the image of a Tur- and “writing” its complement onto a reading data while simultaneously moving machine, with the finite control re- new, growing DNA strand. While lying ing along the “output” tape reading and placed by an enzyme. Remarkably, es- there admiring this amazing enzyme, I writing other data. The finite control sentially the same idea had been sug- was struck by its similarity to something was programmable with very simple in- gested almost a decade earlier by Charles described in 1936 by Alan M. Turing, structions, and one could easily write a H. Bennet and Rolf Landauer of IBM the famous British mathematician. Tur- program that would read a string of A, [see “The Fundamental Physical Limits ing—and, independently, Kurt Gödel, T, C and G on the input tape and write of Computation”; Scientific Ameri- Alonzo Church and S. C. Kleene—had the Watson-Crick complementary string can, July 1985]. Unfortunately, while an begun a rigorous study of the notion of on the output tape. The similarities with enzyme (DNA polymerase) was known “computability.” This purely theoreti- DNA polymerase could hardly have that would make Watson-Crick comple- cal work preceded the advent of actual been more obvious. ments, it seemed unlikely that any exist- computers by about a decade and led to But there was one important piece of ed for other important roles, such as some of the major mathematical results information that made this similarity factoring numbers. of the 20th century [see “Unsolved truly striking: Turing’s toy computer This brings up an important fact Problems in Arithmetic,” by Howard had turned out to be universal—simple about biotechnologists: we are a com- DeLong; Scientific American, March as it was, it could be programmed to munity of thieves. We steal from the 1971; and “Randomness in Arithme- compute anything that was computable cell. We are a long way from being able tic,” by Gregory J. Chaitin; Scientific at all. (This notion is essentially the con- American, July 1988]. tent of the well-known “Church’s the- For his study, Turing had invented a sis.”) In other words, one could program “toy” computer, now referred to as a a Turing machine to produce Watson- Turing machine. This device was not Crick complementary strings, factor intended to be real but rather to be con- numbers, play chess and so on. This re- ceptual, suitable for mathematical in- alization caused me to sit up in bed and vestigation. For this purpose, it had to remark to my wife, Lori, “Jeez, these

be extremely simple—and Turing suc- things could compute.” I did not sleep ASHIMA ceeded brilliantly. One version of his the rest of the night, trying to ﬁgure out OMO NAR machine consisted of a pair of tapes a way to get DNA to solve problems. T

onsider a map of cities connected by certain nonstop Detroit Cflights (top right). For instance, in the example shown here, it is possible to travel directly from Boston to Detroit but not vice versa. The goal is to determine whether a path exists that will commence at the start city (Atlanta), finish at the end city (De- troit) and pass through each of the remaining cities exactly once. In DNA computation, each city is assigned a DNA sequence Chicago Boston (ACTTGCAG for Atlanta) that can be thought of as a first name (ACTT) followed by a last name (GCAG). DNA flight numbers can then be defined by concatenating the last name of the city of origin with the first name of the city of destination (bottom right). The complementary DNA city names are the Watson-Crick Atlanta complements of the DNA city names in which every C is replaced by a G, every G by a C, every A by a T, and every T by an A. (To sim- CITY DNA NAME COMPLEMENT plify the discussion here, details of the 3′ versus 5′ ends of the DNA molecules have been omitted.) For this particular problem, ATLANTA ACTTGCAG TGAACGTC only one Hamiltonian path exists, BOSTON TCGGACTG AGCCTGAC and it passes through Atlanta, CHICAGO GGCTATGT CCGATACA Boston, Chicago and Detroit in that DETROIT CCGAGCAA GGCTCGTT order. In the computation, this FLIGHT DNA FLIGHT NUMBER path is represented by GCAGTCG- ATLANTA - BOSTON GCAGTCGG END GACTGGGCTATGTCCGA, a DNA se- ATLANTA - DETROIT GCAGCCGA quence of length 24. Shown at the BOSTON - CHICAGO ACTGGGCT START left is the map with seven cities BOSTON - DETROIT ACTGCCGA and 14 nonstop flights used in the BOSTON - ATLANTA ACTGACTT CHICAGO - DETROIT ATGTCCGA actual experiment. —L.M.A. SLIM FILMS

to create de novo miraculous molecular were at hand. These tools were essen- weak forces such as hydrogen bonds. If machines such as DNA polymerase. tially the following: a molecule of DNA in solution meets a Fortunately, three or four billion years DNA molecule to which it is not com- of evolution have resulted in cells that 1. Watson-Crick pairing. As stated plementary (and has no long stretches are full of wondrous little machines. It earlier, every strand of DNA has its Wat- of complementarity), then the two mol- is these machines, stolen from the cell, son-Crick complement. As it happens, ecules will not anneal. that make modern biotechnology possi- if a molecule of DNA in solution meets 2. Polymerases. Polymerases copy in- ble. But a molecular machine that would its Watson-Crick complement, then the formation from one molecule into an- play chess has apparently never evolved. two strands will anneal—that is, twist other. For example, DNA polymerase So if I were to build a DNA computer around each other to form the famous will make a Watson-Crick complemen- that could do something interesting, I double helix. The strands are not cova- tary DNA strand from a DNA template. would have to do it with the tools that lently bound but are held together by In fact, DNA polymerase needs a “start

CHICAGO TO DETROIT

G BOSTON TO CHICAGO C ATLANTA TO BOSTON G A C C A WATSON-CRICK C G ANNEALING, T C G C T in which Cs pair G C A with Gs and As join G G A with Ts, will result in A G A DNA ﬂight-number C T A G T C strands (shown here: At- A T lanta to Boston, Boston T A G G to Chicago, and Chicago T C to Detroit) being held end- G C C to-end by strands encoding COMPLEMENT OF CHICAGO the complementary DNA city COMPLEMENT OF BOSTON names (shown here: Boston and Chicago).

56 Scientific American August 1998 Computing with DNA Copyright 1998 Scientific American, Inc. signal” to tell it where to begin making shorter strands moving more quickly that can be computed. To get your com- the complementary copy. This signal is than longer ones. Hence, this process puter to make Watson-Crick comple- provided by a primer—a (possibly short) separates DNA by length. With special ments or to play chess, you need only piece of DNA that is annealed to the chemicals and ultraviolet light, it is pos- start with the correct input information template by Watson-Crick complemen- sible to see bands in the gel where the and apply the right sequence of opera- tarity. Wherever such a primer-template DNA molecules of various lengths have tions—that is, run a program. DNA is a pair is found, DNA polymerase will become to rest. great way to store information. In fact, gin adding bases to the primer to create 6. DNA synthesis. It is now possible the cell has been using this method to a complementary copy of the template. to write a DNA sequence on a piece of store the “blueprint for life” for billions 3. Ligases. Ligases bind molecules to- paper, send it to a commercial synthesis of years. Further, enzymes such as poly- gether. For example, DNA ligase will facility and in a few days receive a test merases and ligases have been used to take two strands of DNA in proximity tube containing approximately 1018 operate on this information. Was there and covalently bond them into a single molecules of DNA, all (or at least most) enough to build a universal computer? strand. DNA ligase is used by the cell to of which have the described sequence. Because of the lessons of the 1930s, I repair breaks in DNA strands that oc- Currently sequences of length approxi- was sure that the answer was yes. cur, for instance, after skin cells are ex- mately 100 can be reliably handled in posed to ultraviolet light. this manner. For a sequence of length The Hamiltonian Path Problem 4. Nucleases. Nucleases cut nucleic 20, the cost is about $25. The molecules acids. For example, restriction endonu- are delivered dry in a small tube and ap- he next task was choosing a prob- cleases will “search” a strand of DNA pear as a small, white, amorphous lump. Tlem to solve. It should not appear for a predetermined sequence of bases to be contrived to fit the machine, and and, when found, will cut the molecule None of these appeared likely to help it should demonstrate the potential of into two pieces. EcoRI (from Escheri- play chess, but there was another im- this novel method of computation. The chia coli) is a restriction enzyme that will portant fact that the great logicians of problem I chose was the Hamiltonian cut DNA after the G in the sequence the 1930s taught us: computation is Path Problem. GAATTC—it will almost never cut a easy. To build a computer, only two William Rowan Hamilton was As- strand of DNA anywhere else. It has things are really necessary—a method tronomer Royal of Ireland in the mid- been suggested that restriction enzymes of storing information and a few simple 19th century. The problem that has evolved to protect bacteria from viruses operations for acting on that informa- come to bear his name is illustrated in (yes, even bacteria have viruses!). For tion. The Turing machine stores infor- the box on the opposite page. Let the example, E. coli has a means (methyla- mation as sequences of letters on tape arrows (directed edges) represent the tion) of protecting its own DNA from and manipulates that information with nonstop flights between the cities (ver- EcoRI, but an invading virus with the the simple instructions in the finite con- tices) in the map (graph). For example, deadly GAATTC sequence will be cut trol. An electronic computer stores in- you can fly nonstop from Boston to to pieces. My DNA computer did not formation as sequences of zeros and Chicago but not from Chicago to Bos- use restriction enzymes, but they have ones in memory and manipulates that ton. Your job (the Hamiltonian Path been used in subsequent experiments information with the operations avail- Problem) is to determine if a sequence by many other research groups. able on the processor chip. The remark- of connecting flights (a path) exists that 5. Gel electrophoresis. This is not able thing is that just about any method starts in Atlanta (the start vertex) and stolen from the cell. A solution of het- of storing information and any set of ends in Detroit (the end vertex), while erogeneous DNA molecules is placed in operations to act on that information passing through each of the remaining one end of a slab of gel, and a current is are good enough. cities (Boston and Chicago) exactly applied. The negatively charged DNA Good enough for what? For univer- once. Such a path is called a Hamilton- molecules move toward the anode, with sal computation—computing anything ian path. In the example shown on

C C C G G T C G G C C A T A G G G A G A A C A T LIGASES connect C the splinted mole- G C T A G T A cules. Wherever the T A protein ﬁnds two G C C T strands of DNA in G proximity, it will C LIGASE covalently bond ASHIMA them into a single

strand. OMO NAR T

Computing with DNA Scientific American August 1998 57 Copyright 1998 Scientific American, Inc. page 56, it is easy to see that a unique Path Problem was shown to be “NP- from the set. Hamiltonian path exists, and it passes complete.” Without going into the the- c. For each vertex, check if through the cities in this order: Atlanta, ory of NP-completeness, suffice it to that path passes through Boston, Chicago, Detroit. If the start say that this finding convinced most that vertex. If not, remove city were changed to Detroit and the theoretical computer scientists that no that path from the set. end city to Atlanta, then clearly there efficient algorithm for the problem is 3. If the set is not empty, then would be no Hamiltonian path. possible at all (though proving this re- report that there is a Hamil- More generally, given a graph with mains the most important open prob- tonian path. If the set is directed edges and a specified start ver- lem in theoretical computer science, the empty, report that there tex and end vertex, one says there is a so-called NP = P? problem [see “Turing is no Hamiltonian path. Hamiltonian path if and only if there is a Machines,” by John E. Hopcroft; Scien- path that starts at the start vertex, ends tific American, May 1984]). This is This is not a perfect algorithm; never- at the end vertex and passes though each not to say that no algorithms exist for theless, if the generation of paths is ran- remaining vertex exactly once. The the Hamiltonian Path Problem, just no dom enough and the resulting set large Hamiltonian Path Problem is to decide efficient ones. For example, consider the enough, then there is a high probability for any given graph with specified start following algorithm: that it will give the correct answer. It is and end vertices whether a Hamiltoni- this algorithm that I implemented in the an path exists or not. Given a graph with n vertices, first DNA computation. The Hamiltonian Path Problem has 1. Generate a set of random been extensively studied by computer paths through the graph. Seven Days in a Lab scientists. No efficient (that is, fast) al- 2. For each path in the set: gorithm to solve it has ever emerged. In a. Check whether that path or my experiment, I sought a Ham- fact, it seems likely that even using the starts at the start vertex Filtonian Path Problem small enough best currently available algorithms and and ends with the end to be solved readily in the lab, yet large computers, there are some graphs of vertex. If not, remove enough to provide a clear “proof of fewer than 100 vertices for which de- that path from the set. principle” for DNA computation. I chose termining whether a Hamiltonian path b. Check if that path passes the seven-city, 14-flight map shown in exists would require hundreds of years. through exactly n vertices. the inset on page 56. A nonscientific In the early 1970s the Hamiltonian If not, remove that path study has shown that it takes about 54

DNA POLYMERASE T A

T T

C G G A

C G

G A A G T C C C C T G G

A POLYMERASE CHAIN REACTION, or PCR, is used to replicate DNA molecules that begin with the start city (Atlanta) and terminate with the end city (Detroit). In this example, a primer—GGCT, G C representing the complement of the ﬁrst name of Detroit—is annealed to the right end of a DNA strand. The primer signals the polymerase to begin making a Watson-Crick complement of the strand. After the polymerase is done, the double helix is split into two strands so that Watson-Crick complements of each half can be made. This process is repeated to obtain a large number of copies of molecules that have the correct start and end cities. Gel electrophoresis is then used to isolate those molecules that have the right sequence length of 24.

58 Scientific American August 1998 Computing with DNA Copyright 1998 Scientific American, Inc. seconds on average to find the unique name. Atlanta’s complementary name problem contained just a handful of Hamiltonian path in this graph. (You becomes, for instance, TGAACGTC. cities, there was a virtual certainty that may begin now....) After working out these encodings, I at least one of the molecules formed To simplify the discussion here, con- had the complementary DNA city names would encode the Hamiltonian path. It sider the map on page 56, which con- and the DNA flight numbers synthesized. was amazing to think that the solution tains just four cities—Atlanta, Boston, (As it turned out, the DNA city names to a mathematical problem could be Chicago and Detroit—linked by six themselves were largely unnecessary.) I stored in a single molecule! flights. The problem is to determine the took a pinch (about 1014 molecules) of Notice also that all the paths were existence of a Hamiltonian path start- each of the different sequences and put created at once by the simultaneous in- ing in Atlanta and ending in Detroit. them into a common test tube. To be- teractions of literally hundreds of tril- I began by assigning a random DNA gin the computation, I simply added lions of molecules. This biochemical re- sequence to each city. In our example, water—plus ligase, salt and a few other action represents enormous parallel Atlanta becomes ACTTGCAG, Boston ingredients to approximate the condi- processing. TCGGACTG and so on. It was conve- tions inside a cell. Altogether only about nient to think of the first half of the one fiftieth of a teaspoon of solution DNA sequence as the first name of the was used. Within about one second, I city and the second half as the last held the answer to the Hamiltonian name. So Atlanta’s last name is GCAG, Path Problem in my hand. whereas Boston’s first name is TCGG. To see how, consider what transpires Next, I gave each nonstop flight a DNA in the tube. For example, the Atlanta-to- “flight number,” obtained by concate- Boston flight number (GCAGTCGG) nating the last name of the city of origin and the complementary name of Boston with the first name of the city of desti- (AGCCTGAC) might meet by chance. nation. In the example on page 56, the By design, the former sequence ends Atlanta-to-Boston flight number becomes GCAGTCGG. Recall that each strand of DNA has its Watson-Crick complement. Thus, each city has its complementary DNA

A C

T ASHIMA OMO NAR G A T

T with TCGG, and the latter starts with For the map on page 56, there is only T AGCC. Because these sequences are one Hamiltonian path, and it goes complementary, they will stick togeth- through Atlanta, Boston, Chicago and

G er. If the resulting complex now en- Detroit, in that order. Thus, the mole- counters the Boston-to-Chicago ﬂight cule encoding the solution will have the C G number (ACTGGGCT), it, too, will sequence GCAGTCGGACTGGGCT- C join the complex because the end of ATGTCCGA. G the former (TGAC) is complementary Unfortunately, although I held the so-

C to the beginning of the latter (ACTG). lution in my hand, I also held about 100 In this manner, complexes will grow in trillion molecules that encoded paths length, with DNA flight numbers splint- that were not Hamiltonian. These had PRIMER ed together by complementary DNA to be eliminated. To weed out molecules city names. The ligase in the mixture that did not both begin with the start will then permanently concatenate the city and terminate with the end city, I chains of DNA flight numbers. Hence, relied on the polymerase chain reaction the test tube contains molecules that (PCR). This important technique re- encode random paths through the quires many copies of two short pieces different cities (as required in the first of DNA as primers to signal the DNA step of the algorithm). polymerase to start its Watson-Crick Because I began with such a large replication. The primers used were the number of DNA molecules and the last name of the start city (GCAG for

T C

G T G C G G C A G T C A

A A

G PROBE MOLECULES A

G C

C T

C C T G C G G A

ASHIMA C OMO NAR T

PROBE MOLECULES are used to locate DNA strands encoding paths that pass through the intermediate cities (Boston and Chicago). Probe molecules containing the complementary DNA name of Boston (AGCCTGAC) are attached to an iron ball suspended in liquid. Because of Watson-Crick affinity, the probes capture DNA strands that contain Boston’s name (TCGGACTG). Strands missing Boston’s name are then right length (in the exam- discarded. The process is repeated with probe molecules encoding the complementary ple on page 56, a length of DNA name of Chicago. When all the computational steps are completed, the strands left will be those that encode the solution GCAGTCGGACTGGGCTATGTCCGA. 24). All other molecules were discarded. This completed step 2b of the algorithm. To check the remaining se- Atlanta) and the Watson-Crick comple- The result was that molecules with quences for whether their paths ment of the first name of the end city both the right start and end cities were passed through all the intermediary (GGCT for Detroit). These two prim- reproduced at an exponential rate. In cities, I took advantage of Watson- ers worked in concert: the first alerted contrast, molecules that encoded the Crick annealing in a procedure called DNA polymerase to copy complements right start city but an incorrect end city, affinity separation. This process uses of sequences that had the right start city, or vice versa, were duplicated in a much multiple copies of a DNA “probe” mol- and the second initiated the duplication slower, linear fashion. DNA sequences ecule that encodes the complementary of molecules that encoded the correct that had neither the right start nor end name of a particular city (for example, end city. were not duplicated at all. Thus, by tak- Boston). These probes are attached to PCR proceeds through thermocycling, ing a small amount of the mixture after microscopic iron balls, each approxi- repeatedly raising and lowering the tem- the PCR was completed, I obtained a mately one micron in diameter. perature of the mixture in the test tube. solution containing many copies of the I suspended the balls in the tube con- Warm conditions encourage the DNA molecules that had both the right start taining the remaining molecules under polymerase to begin duplicating; a hot- and end cities, but few if any molecules conditions that encouraged Watson- ter environment causes the resulting an- that did not meet this criterion. Thus, Crick pairing. Only those molecules that nealed strands to split from their dou- step 2a of the algorithm was complete. contained the desired city’s name (Bos- ble-helix structure, enabling subsequent Next, I used gel electrophoresis to ton) would anneal to the probes. Then I replication of the individual pieces. identify those molecules that had the placed a magnet against the wall of the

60 Scientific American August 1998 Computing with DNA Copyright 1998 Scientific American, Inc. test tube to attract and hold the metal cules left in the tube should be precisely that can compete with electronic com- balls to the side while I poured out the those encoding Hamiltonian paths. puters? That remains to be seen. Huge liquid phase containing molecules that Hence, if the tube contained any DNA financial and intellectual investments did not have the desired city’s name. at all, I could conclude that a Hamilto- over half a century have made electron- I then added new solvent and removed nian path existed in the graph. No ic computers the marvels of our age— the magnet in order to resuspend the DNA would indicate that no such path they will be hard to beat. balls. Raising the temperature of the existed. Fortunately, to make this deter- But it would be shortsighted to view mixture caused the molecules to break mination I could use an additional PCR this research only in such practical free from the probes and redissolve in step, followed by another gel-elec- terms. My experiment can be viewed as the liquid. Next, I reapplied the magnet trophoresis operation. To my delight, a manifestation of an emerging new to attract the balls again to the side of the final analysis revealed that the mol- area of science made possible by our the test tube, but this time without any ecules that remained did indeed encode rapidly developing ability to control the molecules attached. The liquid, which the desired Hamiltonian path. After molecular world. Evidence of this new C now contained the desired DNA strands seven days in the lab, the first DNA “molecular science” can be found in (in the example, encoding paths that computation was complete. many places. For example, Gerald F. went through Boston), could then be Joyce of Scripps Research Institute in poured into a new tube for further A New Field Emerges La Jolla, Calif., “breeds” trillions of screening. The process was repeated for RNA molecules, generation after gener- the remaining intermediary cities (Chi- hat about the future? It is clear ation, until “champion” molecules cago, in this case). This iterative proce- Wthat molecular computers have evolve that have the catalytic properties dure, which took me an entire day to many attractive properties. They pro- he seeks [see “Directed Molecular Evo- complete in the lab, was the most te- vide extremely dense information stor- lution,” by Gerald F. Joyce; Scientific dious part of the experiment. age. For example, one gram of DNA, American, December 1992]. Julius Re- At the conclusion of the affinity sepa- which when dry would occupy a volume bek, Jr., of the Massachusetts Institute rations, step 2c of the algorithm was of approximately one cubic centimeter, of Technology creates molecules that over, and I knew that the DNA mole- can store as much information as ap- can reproduce—informing us about how proximately one trillion CDs. They pro- life on the earth may have arisen [see T vide enormous parallelism. Even in the “Synthetic Self-Replicating Molecules,” tiny experiment carried out in one fifti- by Julius Rebek, Jr.; Scientific Ameri- eth of a teaspoon of solution, approxi- can, July 1994]. Stimulated by research 14 G mately 10 DNA flight numbers were on DNA computation, Erik Winfree of simultaneously concatenated in about the California Institute of Technology one second. It is not clear whether the synthesizes “intelligent” molecular com- A T fastest supercomputer available today plexes that can be “programmed” to could accomplish such a task so quickly. assemble themselves into predetermined G Molecular computers also have the structures of arbitrary complexity. There C C potential for extraordinary energy effi- are many other examples. It is the enor- ciency. In principle, one joule is sufficient mous potential of this new area that we for approximately 2 × 1019 ligation op- should focus on and nurture. erations. This is remarkable consider- For me, it is enough just to know that ing that the second law of thermody- computation with DNA is possible. In namics dictates a theoretical maximum the past half-century, biology and com- of 34 × 1019 (irreversible) operations per puter science have blossomed, and joule (at room temperature). Existing there can be little doubt that they will supercomputers are far less efficient, ex- be central to our scientific and econom- ecuting at most 109 operations per joule. ic progress in the new millennium. But Experimental and theoretical scien- biology and computer science—life and tists around the world are working to computation—are related. I am confi- exploit these properties. Will they suc- dent that at their interface great discov- ceed in creating molecular computers eries await those who seek them. SA

The Author Further Reading

LEONARD M. ADLEMAN received a Ph.D. in computer sci- Molecular Computation of Solutions to Combinatorial ence in 1976 from the University of California, Berkeley. In 1977 he Problems. Leonard M. Adleman in Science, Vol. 266, pages joined the faculty in the mathematics department at the Mas- 1021–1024; November 11, 1994. sachusetts Institute of Technology, where he specialized in algorith- On the Path to Computation with DNA. David K. Gifford in mic number theory and was one of the inventors of the RSA public- Science, Vol. 266, pages 993–994; November 11, 1994. key cryptosystem. (The “A” in RSA stands for “Adleman.”) Soon DNA Solution of Hard Computational Problems. Richard J. after joining the computer science faculty at the University of South- Lipton in Science, Vol. 268, pages 542–545; April 28, 1995. ern California, he was “implicated” in the emergence of computer Additional information on DNA computing can be found at http:// viruses. He is a member of the National Academy of Engineering. users.aol.com/ibrandt/dna_computer.html on the World Wide Web.