poral position of the response window to a shorter or longer value, depending on the subject's ability to perform (for example, shortening it if subjects were making few errors). The response window procedure obliged subjects to respond at speeds that did not permit high levels of accuracy and, consequently, error rates were substantial. (Mean latencies of highly motivated subjects under instructions to respond rapidly are typically between 550 and 650 ms.) The production of relatively high error rates allowed the priming effect-that is, the effect of the prime's con- gruence versus incongruence with the target's mean- ing-to be observed in subjects' error rates rather than in their response latencies to targets. With this response window procedure, priming took the form of lower error rates for congruent priming than for incongruent priming, reflecting some combination of facilitation by congruent primes and interference by incongruent primes. Priming was therefore measur- able by observing the difference between these error rates. Even though the procedure was designed to constrain response latencies to approximately the range of values that define the response window, response accuracy was analyzed for all trials except for a small percentage with latencies greater than 1500 ms, a value substantially longer than the time elapsed at the end of the window interval. The prima- ry measure used with these data was signal detection theory's d' measure of sensitivity of the target word's response to the prime word's meaning [see the fur- ther explanation in the legend to Fig. 1, and D. M. Green and J. A. Swets, Signal Detection Theory and Psychophysics (Wiley, New York, 1967)]. Results for this sensitivity measure were similar to those obtained with various alternative measures, such as the in- crease in error rate for incongruent relative to congru- ent primes or measured information transmitted from prime stimulus to target response. Prime words were presented for 17, 33, or 50 ms at a centered display location to which subjects were instructed to attend. In the visually masked (subliminal) prime condition, the prime was both preceded and followed, at the same screen location, by strings of consonants that served as forward mask (premask) and backward mask (postmask). (An example of a mask stimulus is the letter string GKQHYTPDGFQBYLG.) The pre- mask, prime, postmask, and target stimuli were pre- sented as black letters on a gray background. The premask lasted 100 ms and the postmask 17 ms. (Other pilot studies had shown that masking effec- tiveness was unaltered by increases in postmask du- ration beyond 17 ms, the minimum value obtainable with the 60-Hz computer display used in this re- search.) Subjects viewed the computer video display through a device that presented images from left and right halves of the display screen separately to left and right eyes. Although this dichoptic presenta- tion was not needed for the present procedures (which presented identical stimuli to both eyes at all times), its use has been found to increase mildly the effectiveness of visual masking. The combination of premask and postmask made the prime words dif- ficult or impossible to see for almost all subjects. By contrast, in supraliminal conditions the masking consonant strings were replaced by blanks (that is, the screen background color), as a consequence of which the prime words were easily legible de- spite their short (50 ms) duration. 12. Because preliminary findings revealed that direct measure performance was depressed by the re- quirement to respond rapidly, the response window procedure was not used during blocks of trials that obtained direct measures. Different discriminations on visually masked stimuli were requested in differ- ent experiments to allow opportunities to demon- strate that some types of information might pene- trate visual masking more readily than others. The basic properties of the results shown in Fig. 1 did not vary with the different direct measures, adding to confidence in generality of conclusions. 13. The conclusion that unconscious cognition is indi- cated by the presence of statistically significant inter- cept effects in the regression analyses of Fig. 1 rests on a methodological analysis by.A. G. Greenwald, M. R. Klinger, and E. S. Schuh [J. Exp. Psychol. Gen. 124, 22 (1995)] that extends the logic of an analysis introduced by P. M. Merikle and E. M. Reingold [J. Exp. Psychol Learn. Mem. Cognit. 17, 224 (1991)]. A concern in interpreting such intercept effects is the possibility that a spurious intercept may be produced when the predictor (in this instance, the direct mea- sure of prime perceptibility) is imperfectly measured. However, the regression analyses in Fig. 1 do not have the properties that can produce such spurious intercept effects. Such properties include both pos- itive regression slopes and average predictor scores substantially above zero. In contrast, the regression slopes that we obtained were approximately flat and predictor scores (that is, direct measures) were no- ticeably above zero only with prime duration of 50 ms. For a more detailed discussion of the possibility of spurious intercept effects, see (21) 14. The level of perceptibility of masked 50-ms primes can be read from the horizontal distribution of values in the lower panel of Fig. 1, A and C. Levels of direct measure performance corresponding to d' values <1.0 are commonly associated with self-reports of little or no perceptibility. Findings of SOA effects closely resembling those in Fig. 2B were obtained when the plotted variable was changed to magni- tude of intercept effect from regression analysis; that is, statistically significant intercept effects were found only for the 67-ms SOA. The intercept-effect alterna- tive analysis confirms that the pattern in Fig. 2B for subliminal priming as a function of SOA is indeed a pattern for unconscious priming. The plotted analy- sis in Fig. 2B, which includes all subjects who re- ceived masked priming, is properly comparable to the analysis in Fig. 2A for supraliminal priming (for which regression analysis is not an appropriate method). 15. The result shown in Fig. 3 is related to one previously reported by J. Cheesman and P. M. Merkle [Can. J. Psychol. 40, 343 (1986)]. They showed that supra- liminal priming was greater when there was a higher proportion of congruent priming trials in a block of trials, but subliminal priming showed no such effect. Their finding could be explained by the difference in two-trial sequential effects shown in Fig. 3. Combinations of three RNA or DNA strands produce triple helices that have been characterized in terms of base complemen- tarity and the relative orientations of sugar- phosphate backbones (1). Hydrogen bonds are formed between the available functional groups of each base pair in a Watson-Crick double helix and those of the third (Hoog- steen) strand located in its major groove. The sequence recognition that this allows 16. D. E. Broadbent, Perception and Communication (Pergamon, London, 1958); E. C. Cherry, J. Acoust. Soc. Am. 25, 975 (1953); A. M. Treisman, Q. J. Exp. Psychol. 12, 242 (1960). 17. U. Neisser, Cognitive Psychology (Appleton-Cen- tury-Crofts, New York, 1967). See also G. Sperling, Acta Psychol. 27, 285 (1967); D. A. Norman, Psy- chol. Rev. 75, 522 (1968). The language of infor- mation-processing stages used in this description has now surrendered paradigmatic center stage to neural network models. However, the empirical phenomena captured in this language remain sol- idly established. 18. The brevity of unconscious semantic activation mea- sured by the prime-target SOA should not be con- fused with the latency after masked prime presenta- tion at which semantic information is available. Se- mantic activation is presumed to occur after prelim- inary operations that may require a few hundred milliseconds for subliminal prime words (as well as for visible target words). 19. G. Sperling, Psychol. Monogr. 74, 11, (1960); T. A. Busey and G. R. Loftus, Psychol. Rev. 101, 446 (1994); V. di Lollo and P. Dixon, J. Exp. Psychol. Hum. Percept. Perform. 14, 671 (1988). 20. M. C. Potter, Mem. Cognit. 21, 156 (1993). 21. A. G. Greenwald and S. C. Draine, in 25th Carnegie Symposium on Cognition: Scientific Approaches to the Question of Consciousness, J. D. Cohen and J. W. Schooler, Eds. (Erlbaum, Mahwah, NJ, in press). 22. Cubic polynomial regression functions were used to capture possible nonlinear trends in the data. How- ever, it can be seen in the figure that intercept effects for these nonlinear functions were similar to those estimated by linear regression functions. 23. R. L. Abrams, S. C. Draine, A. G. Greenwald, paper presented at the 36th Annual Meeting of the Psy- chonomic Society, Los Angeles, CA, 10 to 12 No- vember, 1995. 24. Supported by grants from the National Institute of Mental Health (MH-41328) and National Science Foundation (SBR-9422242). 27 February 1996; accepted 25 July 1996 has been applied in the precise targeting of sequences in double-stranded DNA to ex- clude DNA binding proteins, including gene promoters (2), from their specific binding sites, and to direct single-site cleavage in chromosomal DNA (3). Triple helix forma- tion has also been presented as a mechanism for alignment of homologous sequences be- fore genetic recombination (4) and may take place in vivo (5). SCIENCE * VOL. 273 * 20 SEPTEMBER 1996 Parallel and Antiparallel (G-GC)2 Triple Helix Fragments in a Crystal Structure Dominique Vlieghe, Luc Van Meervelt, Alain Dautant, Bernard Gallois, Gilles Precigoux, Olga Kennard* Nucleic acid triplexes are formed by sequence-specific interactions between single- stranded polynucleotides and the double helix. These triplexes are implicated in genetic recombination in vivo and have application to areas that include genome analysis and antigene therapy. Despite the importance of the triple helix, only limited high-resolution structural information is available. The x-ray crystal structure of the oligonucleotide d(GGCCAATTGG) is described; it was designed to contain the d(G-GC)2 fragment and thus provide the basic repeat unit of a DNA triple helix. Parameters derived from this crystal structure have made it possible to construct models of both parallel and anti- parallel triple helices. 1 702