P ERSPECTIVES ic crystal quantum cas- 450 cade lasers are likely to CW Radiation cone ω BZ boundary InP be very desirable, for ex- Pulsed ample, in imaging appli- CW GaAs cations. More important- 300 Pulsed ly, substantial efforts are currently invested in pur- suing the realization of quantum cascade lasers mperature (K) 150 Te ω = ω at near-infrared telecom- 0 munication wavelengths. k Their property of being 0 y easily modulated at very 051020 50 100 k high speed would be very Wavelength (µm) x useful in high-throughput Quantum cascade successes. (Left) Maximum operating temperatures of the best existing quantum cascade lasers as data transmission. In this a function of their emission wavelength. (Right) Frequency ω of a representative waveguide mode (red line) as a func- respect, device integration tion of its wave vector kxy.The photonic crystal structure restricts the allowed wave vectors to the Brillouin zone (BZ), ω and vertical coupling of whose extension depends on the crystal periodicity. The portion of the (kxy) curve that, in the absence of the photon- the light would be as es- ic crystal, would be outside the BZ (dashed line) is “reflected” into the BZ, creating a photonic band. With this trick, a ω sential as for convention- guided mode of frequency 0 (yellow dot), which would otherwise have a large wave vector, can instead be formed in- ω al interband lasers. side the cone > c kxy (the “radiation cone”). This is necessary for obtaining radiation orthogonal to the x-y plane. Finally, photonic crys- tals could serve as waveguides in future fabrication technology. And, as we can see References cascade lasers realized in structures with from (1), their mainly planar emission is a 1. R. Colombelli et al., Science 302, 1374 (2003); published online 30 October 2003 (10.1126/ lower dimensionality, such as nanowhisk- perfect match for simple two-dimensional science.1090561). ers (11), which should provide better configurations. 2. J. Faist et al., Science 264, 553 (1994). threshold and temperature performances. But quantum cascade lasers offer even 3. F. Capasso et al., Phys. Today 55 No. 5, 34 (2002). One should also ask the reverse ques- more. As Colombelli et al. show (1), they 4. E. Yablonovitch, Phys. Rev. Lett. 58, 2059 (1987).

5. J. D. Joannopoulos, P. R. Villeneuve, S. H. Fan, Nature on June 17, 2008 tion: What can quantum cascade lasers can operate using surface-plasmon waves. 386, 143 (1997). give to photonic crystal science? They may Surface-plasmon photonic crystals may 6. M. Beck et al., Science 295, 301 (2002). provide the ideal laboratory system for de- enable the miniaturization of photonic cir- 7. R. Köhler et al., Nature 417, 156 (2002). 8. E. Cubukcu et al., Nature 423, 604 (2003). veloping new physics and device concepts. cuits, with great promise for subwave- 9. P. Russel, Science 299, 358 (2003). Because only electrons are involved, they length optics, data storage, and mi- 10. H. Benisty et al., in Photonic Crystals and Light do not suffer from surface recombination, croscopy (12). The availability of sources Localization,C.M.Soukoulis, Ed. (Kluwer, Dordrecht, emitting in surface-plasmon photonic Netherlands, 2001), pp. 117–128. and injection photonic crystal devices are 11. L. Samuelson, Mater. Today 6, 22 (2003). easily made. Furthermore, they can operate crystal modes will further fuel this area of 12. W. L. Barnes, A. Dereux, T. W. Ebbesen, Nature 424, at long wavelengths, greatly simplifying research. 824 (2003). www.sciencemag.org

STRUCTURAL BIOLOGY Then, they synthesized their new and analyzed its unique folded structure by Learning to Speak nuclear magnetic resonance spectroscopy and x-ray crystallography. The final x-ray structure of the protein turned out to be an

the Language of almost exact copy of the structure modeled Downloaded from by the computer. The Kuhlman et al. work David T. Jones is a clear demonstration that structural bi- ologists have mastered the first stage of nderstanding the folding of proteins learner develops basic conversational learning the language of protein folding. can be likened to learning to speak a skills. Finally, after a great deal of practice, As with most languages, the language Unew language. Learning a language the learner begins to understand the actual of proteins is based on a relatively simple can be a long and difficult process. For syntax of the language, and the process is alphabet. In the case of most naturally oc- both young children and travelers in for- almost complete. Applying this analogy to curring proteins, an alphabet of just 20 let- eign countries, the first breakthrough in the report on page 1364 of this issue, we ters corresponding to 20 amino acids is learning a language is the ability to speak a can say that Kuhlman et al. (1) have spoken enough to construct an astonishing variety few important phrases, such as “I’m hun- a new sentence in the complex language of of protein structures that carry out a stag- gry!” or “Where is the railway station?” protein folding and have been understood gering array of biochemical processes. The and to be clearly understood. Eventually, perfectly. so-called protein-folding problem that has by uttering more and more of these phras- These investigators have successfully preoccupied structural biologists for more es and listening to the resulting replies, the created a new protein from scratch. Using than four decades can be most simply ex- an iterative computer program that opti- plained as the problem of discovering how mizes both sequence design and structure simple strings of 20 amino acids can en- The author is in the Department of Computer Science prediction, they have created a 93-residue code the complex three-dimensional (3D) and Department of Biochemistry and Molecular α β Biology, University College, London WC1E 6BT, UK. / protein that they call Top7, which has folded structures of proteins. Solving the

CREDIT: LEFT PANEL, S.AND J. BLASER FAIST, NEUCHÂTEL UNIVERSITY E-mail: [email protected] a unique, as yet unobserved, topology. protein-folding problem implies that we

www.sciencemag.org SCIENCE VOL 302 21 NOVEMBER 2003 1347 P ERSPECTIVES Kuhlman et al. (1): would have the ability to read any amino Novel protein fold acid sequence and deduce the correct na- tive folded structure for that protein. Even with recent progress (2), a complete solu- 93 tion to this problem lies somewhere in the future. Despite the difficulty in going di- rectly from sequence to structure, a number of researchers have been interested in the Harbury et al. (9):) inverse problem (3–5): deducing an amino Right-handed acid sequence that will, when synthesized, coiled coil self-assemble into a single desired 3D structure. Unlike the protein-folding prob- lem, which has only one desired solution (the native folded state of the protein), the Dahiyyy(at and Mayo (8)):

Protein chain length Zinc finger motif inverse case is likely to have many solu- 35 tions. There are many examples of pairs of proteins that fold in a strikingly similar way, but which have no evident similarity in their amino acid sequences (6). 28 Statistically at least, solving the inverse protein-folding problem with its many po- tential solutions should be easier than solv- ing the protein-folding problem, which has 1997 1998 2003 usually just a single correct solution. Year The earliest methods for designing pro- : The future is now. A timeline showing progress over the last 6 years in designing teins focused on the problem of building a new proteins using computer programs and then validating their synthesized versions experimen- hydrophobic core for a fixed backbone tally with techniques such as x-ray crystallography and NMR spectroscopy (1). [Images of structures structure. Because even single mutations were prepared using Swiss-PdbViewer.] usually result in conformational changes in on June 17, 2008 the protein backbone, these very restricted the target structure in atomic detail. steps together with essential experimental models were never likely to be successful. Despite this impressive result, the design of validation. We need to know whether the Probably the most useful idea to come out complete proteins has been restricted to methodology of Kuhlman and colleagues of these early methods was that of the side- short polypeptides with no complex fold- can be applied successfully to other design chain rotamer (7). Although most side ing topology. By designing a large complex problems, and we may in fact learn more chains have a lot of conformational flexi- folded protein from scratch and validating from failed attempts to apply their methods bility, they nonetheless have conformation- its structure in atomic detail, Kuhlman and than from the successes. The cycle of hy- al preferences, which can be expressed as a colleagues have taken the next leap forward pothesis, experimentation, and comparison relatively short list of likely conformations (see the figure). is the cornerstone of most scientific re- www.sciencemag.org for each of the 20 amino acids. Despite the It is tempting to lapse into futurology and search. Combining new protein design al- limitations of the fixed backbone assump- hold up the design of Top7 as the first step gorithms with structural genomics tech- tion, a number of groups used it to design toward a new world of disease treatments niques and high-throughput structure de- plausible protein sequences, which were at and manufacturing processes. However, termination should drive progress in under- least partially validated experimentally, and such a prospect may not be borne out by our standing how proteins fold and, perhaps, in one or two cases, fully validated. Most existing knowledge of the protein world. even how they evolve. Eventually, we will

notable of these was the complete redesign Unique protein folds are not a requirement truly be able to say that we are conversant Downloaded from of a 28-residue zinc finger motif (FSD-1) for generating new biochemical functions. in the language of proteins. by Dahiyat and Mayo (8). In this remark- For example, the naturally occurring “TIM able achievement, the designed protein not barrel” fold has been found in 21 distinct su- References only had the correct structure but folded perfamilies and is associated with more than 1. B. Kuhlman et al., Science 302, 1364 (2003). 2. J. Schonbrun, W. J. Wedemeyer, D. Baker. Curr. Opin. accurately without the benefit of a zinc- 60 different chemical reactions (10). Struct. Biol. 12, 348 (2002). binding metal center (the zinc-binding his- Clearly, nature uses protein folds economi- 3. J. W. Ponder, F. M. Richards, J. Mol. Biol. 193, 775 tidine residues at the core of FSD-1 were cally, and there is no reason to believe that (1987). replaced by phenylalanine residues). computational efforts to design enzymes 4. J. U. Bowie, R. Luthy, D. Eisenberg, Science 253, 164 (1991). The next breakthrough in protein design with new chemical activities need to consid- 5. K. Yue, A. Dill, Proc. Natl. Acad. Sci. U.S.A. 89, 4163 had to await formulation of algorithms (and er anything other than a naturally occurring (1992). perhaps also increased computational pow- scaffold structure as a starting point. Indeed, 6. C. A. Orengo, D. T. Jones, J. M. Thornton, Nature 372, 631 (1994). er) that enabled the main chain itself to attempts to reengineer the functions of natu- 7. J. Janin, S. Wodak, M. Levitt, B. Maigret, J. Mol. Biol. move in response to side-chain packing re- ral proteins—such as Bolon and Mayo’s re- 125, 357 (1978). quirements. Harbury and colleagues (9) design of an enzymatically inert bacterial 8. B. I. Dahiyat, S. L. Mayo, Science 278, 82 (1997). 9. P. B. Harbury, J. J. Plecs, B. Tidor, T. Alber, P. S. Kim, demonstrated the power of this more so- thioredoxin, rendering it catalytically active Science 282, 1462 (1998). phisticated approach by designing a set of (11, 12)—have been quite successful. 10. N. Nagano, C. A. Orengo, J. M. Thornton, J. Mol. Biol. right-handed, coiled-coil bundles (dimers, Perhaps the biggest impact of the new 321, 741 (2002). trimers, and tetramers) and experimentally work will be in protein structure prediction 11. D. N. Bolon, S. L. Mayo, Proc. Natl. Acad. Sci. U.S.A. 98, 14274 (2001). validating their tetrameric structure. The and modeling. Such advances will depend 12. D. N. Bolon, C. A. Voigt, S. L. Mayo. Curr. Opin. Chem. tetrameric coiled-coil structure matched on carrying out iterative protein design Biol. 6, 125 (2002).

1348 21 NOVEMBER 2003 VOL 302 SCIENCE www.sciencemag.org