Discoveries in shrna Design

Transcription

1 Discoveries in shrna Design Hong Zhou 1, Xiao Zeng 2, Joseph Manthey 1 1 Saint Joseph College, 1678 Asylum Avenue, West Hartford, CT 06117, USA 2 Superarray Bioscience Corporation, 7320 Executive Way, Frederick, MD 21704, USA Abstract -- To successfully apply RNAi, the very first step is to design functional sirna or shrna sequences targeting specific genes. Many characteristics of functional synthetic sirnas have been identified, but it remains questionable if these characteristics can be equally applied to the design of functional shrnas. Our three-phase algorithm was developed based on the characteristics of functional sirnas. When it was used to design shrnas by Superarray Bioscience Corporation, it generated over 70% functional shrnas, a very high success rate for both sirna and shrna design that has never been reported experimentally. This indicates that the designs of sirna and shrna share the same concerns. The analysis of the shrna dataset of size 444 generated by Superarray reveals that the high free energy states of the two terminals have large positive impact on the shrna efficacy. While it has been reported that it is the free energy states of the first 2 or 5 base pairs of the two terminals from sirna sequences that correlated well with sirna efficacy, our analysis discovers that it is the free energy states of the first 7 base pairs from the 5 antisense strand and the first 6 base pairs from the 5 sense strand that can affect shrna efficacy. Enforcing the high energy states of the two terminals can further improve the success rate to 83.1%. We also find several base preferences at positions 3, 5, 10 and 15. However, positional base preferences found by different research groups conflict with each other. Therefore it is possible that all the positional base preferences identified so far might be an artifact of the biased datasets or analytical methods used. Keywords: shrna, sirna, RNAi 1. INTRODUCTION RNA interference (RNAi) is a pathway in which a short/small double stranded RNA induces the degradation of its sequence specific mrna, leading to the effect of gene silencing. Since its discovery, RNAi has become a powerful biological technique for gene function studies and drug discovery [1-3]. It also shows promise on some therapeutic applications [4-6]. The very first step in successfully applying RNAi is to design functional and sequence specific small interfering RNA (sirna) because randomly selected sirna is seldom functional [7]. Previous research has identified some empirical rules regarding the efficacy of sirna sequences. These rules include the GC content limitation (the optimal GC content is between 30% and 55%) [7-13], the thermo-stability preference (the 5 end of the antisense strand prefers lower stability for easy entry into the RISC complex) [11, 14, 15], and some base preferences in the sirna sequences [7, 16]. Recently, several studies employed computational methods to analyze the published functional sirna datasets whose sizes are relatively large and revealed more base preferences at particular positions [17-20]. The computational models and their discoveries seem to be promising. However, one challenge still remains: they are based on the existing datasets which are very biased and they have not been validated by biological experiments. Another challenge is that these existing datasets are usually for chemically synthesized sirna sequences. Currently there are two approaches to induce sirna sequences into cells. One is to transfect chemically synthesized sirna sequences into cells. Though this approach is more frequently used, its drawback is that it can not offer longterm gene suppression and some mammalian cell types are resistant to the transfection methods used to introduce synthetic sirnas into cells [21, 22]. The only highefficiency approach that enables long-term gene suppression is through the virus-mediated integration vector [22]. After integration, the vector produces small hairpin RNA (shrna) which is processed into active sirna. Another advantage of using the shrna approach is that once a functional vector is identified, it can be reproduced easily and inexpensively. Though many features of functional sirna have been identified, most of them are intended for chemically synthesized sirna. As suggested by Matveeva et al [19], some of these characteristics of functional sirna may not apply to shrna, though many of them may. In 2005, we developed a three-phase algorithm for computer-aided sirna design [23, 24]. This algorithm includes all the design considerations published by then and arranges all the necessary sirna selection rules in three groups of filters according to their impacts on the sirna efficacy and applys them to the design process in three steps. Each filter represents a specific design rule. Phase I filters eliminate all the sirna sequences containing the damaging elements for a functional sirna. Examples are those filters that prohibit the existence of internal palindrome or long GC stretch. All sirna candidates

2 must successfully pass all the phase I filter assessment. Phase II filters are used to rank eligible sirna sequences by a final score with the sum of gain and penalty points. There is a cutoff such that sirna candidates will be selected only if their final scores are no less than the cutoff point. Phase III filters represent those rules whose impact on the sirna functionality has yet to be clarified and therefore are considered optional. Most of the selective filters in Phase II are set to ensure the selection rule that the 5 end of antisense strand is less thermodynamically stable, especially when compared to the 5 sense end. This differential stability increases the probability that the antisense strand will be incorporated into the RISC complex since there are studies showing that the RISC complex prefers antisense strand to sense strand [15, 25]. This algorithm with its default settings was adopted by Superarray Bioscience Corporation and has been eventually used for their shrna design. Though some research has indicated that synthetic 29-mer shrna can be more potent inducers of RNAi than shorter shrna [26], the default settings of the algorithm are for 21-mers because sirna of 21 base pairs are more commonly used. Biological experiments conducted by Superarray have shown that 71.2% (316/444) of the shrnas designed by our algorithm are functional (i.e. can suppress at least 70% of the target gene expression), a high success rate that has never been reported experimentally with both shrna and chemically synthesized sirna. For example, the design algorithms employed recently by Root [22] and Moffat [21] for shrna generate less than 40% functional shrnas. Since our algorithm was originally developed based on discoveries and studies concerning sirna functionality, its high success rate in shrna design suggests that most design concerns for sirna sequences are applicable to shrna design. This article further extends our analysis on the available shrna dataset generated by Superarray Bioscience Corporation. Although this dataset is biased as it is specifically generated by our three-phase algorithm, the analysis reveals useful information that may help confirm the effectiveness of the rules used in the algorithm, modify the existing rules or rearrange them for better prediction and identify new rules. 2. RESULTS OF STATISTICAL ANALYSIS The two sets of biological experiments completely tested 444 shrnas targeting 125 genes. Of the 444 shrnas, 316 are found to be functional (71.2%). Considering the fact that variations exist in the experimentally measured suppression efficacy, we decide to remove some shrnas whose efficacy are in the range near 70% in hope of ensuring the validity of the dataset. The two ranges for removal selected are 60-75% and 55-80%. This means we exclude shrnas whose efficacy is between 60-75% and 55-80% respectively. With the 60-75% removal range, there are 351 shrnas for analysis in which 268 are considered functional (efficacy >= 75%). However, with the 55-80% range, there are only 289 shrna sequences for analysis in which 221 shrnas are considered functional (efficacy >=80%). By default, the three-phase algorithm sets a selection cutoff such that only shrnas which score at least 7 points will be selected for biological experiments. However, former experiments showed that there were a few genes for which our algorithm failed to design enough shrnas. Thus in the dataset there are some shrnas scoring less than 7 points. For this reason, the chi-square test for independence, is first applied to assess if the 7-pointscutoff is significant in distinguishing between functional and nonfunctional shrnas. Note that all further statistical results are obtained using the chi-square test unless otherwise noted. The analysis shows that the 7-pointscutoff is significant in making the distinction between the functional and nonfunctional shrnas. shrnas scoring at least 7 points have a higher probability of being functional. When the removal range is 60-75%, the p-value is With the removal range to be 55-80%, the statistical significance increases (p=0.0046). Several other removal ranges were tested using the chi-square test and it has been found that the removal range of 55-80% generates the most significant result. For example, if the removal range is further expanded to be 50-85%, the significance drops (p=0.037). Unless otherwise stated, the rest of our results were obtained with a removal range of 55-80%. To obtain 7 points for a shrna, it must pass multiple phase II filters. What filters/rules contribute to the statistical bias of the 7-points-cutoff? After analyzing all the design rules, we found that the shrnas scoring no less than 7 are perfectly correlated with the f-dga filter (R=1.0). The f-dga filter in phase II is defined such that any shrna whose free energy (ΔG) of the 5-mer at 5 -end of the antisense strand is no less than -3.2 would gain 1 point. The free energy of the 5-mer at 5 -end of the antisense 1 strand, ΔG as-5, shows some bias between functional and nonfunctional shrnas. Functional shrnas prefer the ΔG as-5 to be no less than -3.2 while disfavoring ΔG as-5 being less than -3.2 (p = ). This shows that the statistical bias behind the 7-points-cutoff is due to the ΔG as-5 values of shrnas, and suggests that having ΔG as-5 larger than -3.2 increases the probability obtaining functional shrnas as shown in Table 1. 1 Subscript as in ΔG as-5 represents the 5 -end of antisense strand while ss represents the 5 -end of sense strand, -5 means the first 5 nucleotides)

3 Table 1 Higher ΔG as-5 increases the probability that shrnas are functional. ROC: Receiver Opearting Characteristic. P= shrna (ΔG as-5 >=-3.2) shrna (ΔG as-5 <-3.2) Number Functional rate 81.6% (155/190) 66.7% (66/99) ROC specificity 0.49 ROC sensitivity 0.70 Table 1 indicates that the enforcement of filter f-dga in the selection algorithm could improve the design success rate significantly at the expense of missing some functional shrnas. The results in Table 1 were obtained using a removal range of 55-80%. Even with no removal range, enforcing the filter f-dga significantly improves the shrna design success rate to 76.0% (222 out of 292). This suggests that the high energy state at the 5 -end of antisense strand has positive impact on shrna efficacy. The free energy of 5 -end of antisense is computed with the first 5 base pairs as suggested by [12, 27]. It is reasonable to ask if other base pair lengths could better represent the free energy of the 5 -end. The free energy of different base pair lengths (2 to 10) from the 5 -end of the antisense strand is computed and chi-square test for independence is conducted to evaluate if there is any significant correlation between the free energy values and the shrna efficacy. It is found that the free energy values of the first 6, 7 or 8 base pairs from the 5 -end of antisense strand have statistically significant correlation with the shrna efficacy (p values are , and respectively). Based on the p values, it looks like that ΔG as-7 is a better representation of the free energy than ΔG as-5. These results are summarized in Table 2. Motivated by these results, we tested to see if the free energy of the 5 -end of sense strand has any correlation with shrna efficacy. Taking base pairs of different lengths (2 to 10) from the 5 -end of sense strand, we find that ΔG ss-6 and ΔG ss-7 show significant correlation with shrna efficacy (p values are and respectively). The results shown in Table 3 indicate that shrna favors higher free energy at the 5 -end of the sense strand. Comparing the values of ROC specificity and sensitivity from Tables 2 and 3, it is not difficult to notice that the high energy state at the 5 -end of antisense strand can better differentiate functional and nonfunctional shrnas than the high energy state at the 5 -end of sense strand can. This confirms the strand bias discovered before [15, 25]. However, it turns out that it is more promising to use either ΔG as-7, ΔG ss-6 or both to predict the efficacy of shrnas. The results are shown in Table 4. Note the ROC sensitivity improves significantly compared to those given in Tables 2 and 3, though the ROC specificity decreases some. Our discovery differs from the recent findings on sirna made by Matveeva and Shabalina [19, 20]. Matveeva and Shabalina found that the ΔG of the first and last two base pairs in the antisense strand correlates well with sirna efficacy. The difference between our results and theirs might be due to the fact that different data sets were used or to the differences between chemically synthesized sirnas and shrnas. Without any removal range, all the shrnas whose ΔG as-7 >=-6.6 or ΔG ss-6 >=-7.0 or both have averaged suppression efficacy of 76.98%, while those shrnas with no high energy state at either end have averaged suppression efficacy of 65.83%. A two-sample t-test reveals that the difference between the two averaged efficacy is very significant (p= ). This confirms that the use of either ΔG as-7, ΔG ss-6 or both could significantly improve the selection of functional shrnas. Positional base preferences were also analyzed. With the removal ranges of 60-75% and 55-80%, some positions show preferences for certain bases as shown in Table 5. The positional base preferences motivated us to combine them in order to help predict shrna efficacy. We assigned 1 point gain to a favored base and -1 point penalty to a disfavored base at each of these positions. The final score is the sum of the gains and penalties. We then tried to determine whether or not the sum is helpful in improving the predication of the shrna efficacy. Unfortunately, the distribution of the sums seems to be random. This raises the possibility that the positional preferences at the 4 positions shown in Table 5 might be the result of a biased dataset. Nevertheless, it is also

4 possible that we haven t found the most effective method for using the positional preferences to improve shrna efficacy predication. Table 2 ΔG as-7 could be a better representation of free energy state than ΔG as-5. The first 7 base pairs shrna (ΔG as-7 >=-6.6) shrna (ΔG as-7 <-6.6) from 5 -end of antisense strand, p = Number Functional rate 82.9% (155/187) 64.7% (66/102) ROC specificity 0.53 ROC sensitivity 0.70 Table 3 Higher ΔG ss-6 increases the probability that shrna is functional. The first 6 base pairs from 5 -end of -sense strand, p = shrna (ΔG ss-6 >=-7.0) shrna (ΔG ss-6 <-7.0) Number Functional rate 83.9% (115/137) 69.7% (106/152) ROC specificity 0.68 ROC sensitivity 0.52 Table 4 Using either ΔG as-7, ΔG ss-6 or both to predict the efficacy of shrnas is more effective than using only one of them. p = ΔG as-7 >=-6.6 or ΔG ss-6 >=-7.0 ΔG as-7 < -6.6 and ΔG ss-6 <-7.0 Number Functional rate 83.1% (192/231) 50.0% (29/58) ROC specificity 0.43 ROC sensitivity 0.87

5 Table 5 Positional base preferences found in the shrna dataset. Positions and bases only refer to sense strand. position favored disfavore d P value (60-75%) P value of (55-80%) 3 C U U C A U C G Discussion and Conclusion The default setting of the three-phase algorithm is relatively stringent. Under this default setting, the algorithm cannot design shrna sequences for approximately 8% of genes from human Refseq database. This stringent setting is one of the reasons why the threephase algorithm has a high success rate in shrna design. It is possible that many functional shrna sequences are missed by this algorithm. However, when there could be enough shrna sequences designed for a given gene, this algorithm with the default stringent setting promises a good probability for functional shrnas. In shrna design, we believe that ensuring a few functional shrnas is more important than designing more than enough shrnas. Based on this belief, our future work will focus on how to minimize the number of false positive shrnas. It has been reported before that the local RNA target structure influences sirna efficacy [28, 29]. So including the local RNA target structure in our statistical analysis might shed more lights on the shrna efficacy predication. It has been confirmed by several studies that the free energy profile at the two terminals is the most critical factor relating to sirna efficacy [7, 15-20]. Our analysis on shrna dataset confirms that this belief applies to shrna design. However, there is difference. For chemically synthesized sirna, it was first found that the high free energy of first 5 base pairs at each terminal correlated well with sirna efficacy [12, 27], and later other researchers discovered that it might be the high free energy of the first 2 base pairs [19, 20]. Our results show that shrna efficacy can be predicted more effectively using the free energy profile of the first 6 base pairs at the 5 sense end and the first 7 base pairs at the 5 antisense end. Currently we are not clear about the reason behind the difference, though the difference might be due to the differences between chemically synthesized sirna and shrna. Reynolds [7] found several base preferences for chemically synthesized sirna including an A at position 3 and U at position 10. We find that positions 3 and 10 also have base preferences in shrna sequences, but position 3 favors C and position 10 favors A while disfavoring U. Matveeva [19] found several base preferences. Some of their results match and some conflict with that of Reynolds. So far the positional base preferences identified by different research groups are not consistent with each other [7, 16-20]. Hence there is some uncertainty on whether there really are positional base preferences for functional sirnas and shrnas. Perhaps positional base preferences are an artifact derived from the biased datasets or the methods of analysis employed. 4. REFERENCES 1. Elbashir, S. M., J. Harborth, W. Lendeckel, A. Yalcin, K. Weber, T. Tuschl. Duplexes of 21-nucleotide RNAs mediate RNA interference in cultured mammalian cells. Nature, 411: , Paddison, P. J., A. A. Caudy, G. J. Hannon. Stable suppression of gene expression by RNAi in mammalian cells. PNAS, 99: , Couzin, J. BREAKTHROUGH OF THE YEAR: Small RNAs Make Big Splash. Science, 298: , Surabhi RM, Gaynor RB: RNA interference directed against viral and cellular targets inhibits human immunodeficiency Virus Type 1 replication. J Virol, 76(24): , Xia H, Mao Q, Eliason SL, Harper SQ, Martins IH, Orr HT, Paulson HL, Yang L, Kotin RM, Davidson BL: RNAi suppresses polyglutamine-induced neurodegeneration in a model of spinocerebellar ataxia. Nat Med, 10(8): , Pai SI, Lin YY, Macaes B, Meneshian A, Hung C-F, Wu TC: Prospects of RNA interference therapy for cancer. Gene Ther, 13(6): , Reynolds, A., D. Leake, Q. Boese, S. Scaringe, W. S. Marshall, A. Khvorova. Rational sirna design for RNA interference. Nat Biotechnol, 22: , Elbashir, S. M., J. Harborth, K. Weber, T. Tuschl. Analysis of gene function in somatic mammalian cells using small interfering RNAs. Methods, 26: , Holen, T., M. Amarzguioui, M. T. Wiiger, E. Babaie, H. Prydz. Positional effects of short interfering RNAs targeting the human coagulation trigger Tissue Factor. Nucleic Acids Res, 30: , Hardin, C. C., T. Watson, M. Corregan, C. Bailey. Cation-dependent transition between the quadruplex and Watson-Crick hairpin forms of d(cgcg3gcg). BIOCHEMISTRY, 31: , 1992.

6 11. Naito, Y., T. Yamada, K. Ui-Tei, S. Morishita, K. Saigo. sidirect: highly effective, target-specific sirna design software for mammalian RNA interference. Nucl. Acids Res., 32: W124-W129, Ui-Tei, K., Y. Naito, F. Takahashi, T. Haraguchi, H. Ohki-Hamazaki,. A. Juni, R. Ueda, K. Saigo. Guidelines for the selection of highly effective sirna sequences for mammalian and chick RNA interference. Nucleic Acids Res, 32: , Yuan, B., R. Latek, M. Hossbach, T. Tuschl, F. Lewitter. sirna Selection Server: an automated sirna oligonucleotide prediction server. Nucl. Acids Res., 32: W130-W134, Martinez, J., A. Patkaniowska, H. Urlaub, R. Luhrmann, T. Tuschl. Single-stranded antisense sirnas guide target RNA cleavage in RNAi. Cell, 110: , Khvorova, A., A. Reynolds, S. D. Jayasena. Functional sirnas and mirnas exhibit strand bias. Cell, 115: , Jagla, B., Aulner, N., Kelly, P.D., Song, D., Volchuk, A., Zatorski, A., Shum, D., Mayer, T., De Angelis, D.A., Ouerfelli, O., Rutishauser, U., Rothman, J.E. Sequence characteristics of functional sirnas. RNA, 11: , Ladunga, I. More complete gene silencing by fewer sirnas: transparent optimized design and biophysical signature. Nucleic Acids Res., 00:1-8, Vert, J., Foveau, F., Lajaunie, C., Vandenbrouck, Y. An accurate and interpretable model for sirna efficacy prediction. BMC Bioinformatics, 7: , Matveeva, O., Nechipurenko, Y., Rossi, L., Moore, B., Sætrom, P., Ogurstsov, A.Y., Atkins, J.F., Shabalina, S.A. Comparison of approaches for rational sirna design leading to a new efficient and transparent method. Nucleic Acids Res., 35:e63, Shabalina, S., Spiridonov, A.N., Ogurtsov, A.Y. Computational models with thermodynamic and composition features improve sirna design. BMC Bioinformatics, 7:65, Moffat, J., Grueneberg, D.A., Yang, X., Kim, S.Y., Kloepfer, A.M., Hinkle, G., Piqani, B., Eisenhaure, T.M., Luo, B., Grenier, J.K., Carpenter, A.E., Foo, S.Y., Stewart, S.A., Stockwell, B.R., Hacohen, N., Hahn, W.C., Lander, E.S., Sabatini, D.M., Root, D.E. A Lentiviral RNAi Library for Human and Mouse Genes Applied to an Arrayed Viral High-Content Screen. Cell, 124: , Root, D.E., Hacohen, N., Hahn, W.C., Lander, E.S., Sabatini, D.M. Genome-scale loss-of-function screening with a lentiviral RNAi library. Nature Methods, 3: , Zhou, H., Zeng, X., Wang, Y., Seyfarth, B.R. A threephase algorithm for computer aided sirna design. Informatica, 30: , Zhou, H., Wang, Y., Zeng, X. Fast and complete search of sirna off-target sequences. Proceedings of the international conference on bioinformatics & computational biology, pp , Las Vegas, USA, June 26-29, Martinez, J., A. Patkaniowska, H. Urlaub, R. Luhrmann, T. Tuschl. Single-stranded antisense sirnas guide target RNA cleavage in RNAi. Cell, 110: , Siolas, D., Lerner, C., Burchard, J., Ge, W., Linsley, P.S., Paddison, P.J., Hannon, G.J., Cleary, M.A. Synthetic shrnas as potent RNAi triggers. Nature Biotechnology, 23: , Levenkova, N., Q. Gu, J. J. Rux. Gene specific sirna selector. Bioinformatics, 20: , Overhoff, L., Alken, M., Far, R.K., Lemaitre, M., Llebleu, B., Sczakiel, G., Robbins, I. Local RNA target structure influences sirna efficacy: a systematic global analysis. J. Mol. Biol., 348: , Far, R.K., Sczakiel, G. The activity of sirna in mammalian cells is related to structural target accessibility: a comparison with antisense oligonucleotides. Nucleic Acids Research, 31: , 2003.