OVERVIEW

atSNP takes as input SNP information/genomic subsequences and a PWM for a transcription factor. atSNP uses a first order Markov model and an importance sampling technique to test the statistical significance of the best match between the PWM and the subsequences overlapping the SNP position with both the reference and the SNP alleles. More imporantly, it further tests the impact of SNPs on the matches to the PWM. The testing results are summarized with a table and a graphical output named composite logo plot. To learn more about the outputs, visit "Interpreting search output” and “Reading composite logo plot”.

Table 1 provides a comparison of functionality between atSNP's statistical method and alternative methods. atSNP provides the most comprehensive information on identifying the impact of SNPs on transcription factor binding. Only atSNP and is-rSNP provide statistical quantification of the impact of SNPs on the binding. Furthermore, atSNP addresses the scalability issue with an efficient computation enabled by an importance sampling technique. atSNP visualizes the best matches with both the reference and the SNP alleles and the change in those matches created by the SNP. atSNP Search results are obtained from the R package atSNP, that implements this procedure.

PROCEDURE

• Step 1: affinity score

Using a PWM as a set of parameters for a product multinomial model, atSNP computes log likelihoods for all subsequences overlapping the SNP location that have the same length as the motif. The best matches are those with the largest log likelihood with both the SNP and reference alleles. The scores of the best matches are used for statistical testing.

Example

We evaluate impact of a SNP, rs117959046, on the binding of the transcription factor GATA2, with GATA_known12 motif:



A C G T

0.245 0.000 0.981 0.000 0.472 0.245 0.094 0.000 0.000 0.132 0.340 0.906 0.019 0.000 0.283 0.170 0.000 0.000 1.000 0.113

The motif length is 5 and sequences of length 9 centered at the SNP position are extracted from both the reference allele and the SNP variation of the reference allele:

Reference allele:	AAGACAAGC
SNP allele:	AAGATAAGC

Next, we compute log likelihoods of all subsequences of length 5 spanning the SNP position and find the best match, which is associated with the largest log likelihood. The second subsequence on the forward strand (AAGACACG) has the largest log likelihood with the reference allele. (Table 2). The second subsequence on the forward strand also has the largest log likelihood with the SNP allele (Table 3).

Reference Allele	Forward strand score	Reverse strand score
AAGACAAGC	-53.45	-49.31
AAGACAAGC	-25.30	-50.10
AAGACAAGC	-70.91	-30.79
AAGACAAGC	-28.08	-72.66
AAGACAAGC	-49.50	-27.73

SNP Allele	Forward strand score	Reverse strand score
AAGATAAGC	-53.60	-49.64
AAGATAAGC	-2.28	-73.03
AAGATAAGC	-70.91	-26.84
AAGATAAGC	-48.74	-72.66
AAGATAAGC	-49.87	-27.22

• Step 2: p-value for the affinity score

The p-value for the affinity score with the reference allele (p-value Reference) is the probability that a random sequence has an affinity score larger than or equal to the best match score with the reference allele. atSNP simulates random sequences using a first order Markov model. It estimates the p-value as the frequency of randomly getting as large or larger affinity scores than the best match score with the reference allele. Similarly, the p-value for the affinity score with the SNP allele (p-value SNP) is estimated as the frequency of getting random scores as large or larger than the best match score with the SNP allele.

Example (Continued)

(Continued From Step 1) The log likelihood score for the reference allele is -25.30, and the log likelihood score for the SNP allele is -2.28. The p-value for the reference allele is the probability that a random sequence has an affinity score larger than or equal to -25.30, which is our target region (Figure 7). For such a small p-value, naive Monte-Carlo generates a small number of scores in the target region (Figure 8), and thus requires a large number of simulations. Alternatively, in atSNP, the importance sampling generates many more scores in the target region (Figure 9). The p-value for the reference allele is then estimated by the weighted frequency as p-value Reference = 0.499. The importance sampling algorithm significantly reduces the required number of simulations. Similarly, the p-value for the SNP allele, p-value SNP, is estimated as 0.009.







• Step 3: p-value for the affinity change between alleles

atSNP tests the significance of log rank ratio: log(p-value Reference)-log(p-value SNP). atSNP computes the p-value (p-value SNP Impact) based on a Monte-Carlo simulation that uses both the reference and SNP alleles together. This is done using the importance sampling technique as described in Step 2.

Example (Continued)

The log rank ratio for the impact of SNP, rs117959046, is 4.02. With the importance sampling, the p-value for the log rank ratio is estimated as p-value rank = 0.005. In atSNP Search, this p-value is referred to as p-value SNP Impact.

References

• Zuo, Chandler, Shin, Sunyoung, and Keleş, Sündüz. atSNP: transcription factor binding affinity testing for regulatory SNP detection. Bioinformatics (2015): 31, 3353-3355
• Andersen, Malin C., Engstrom, Par G., Lithwick, Stuart, Arenillas, David, Eriksson, Per, Lenhard, Boris, Wasserman, Wyeth W., and Odeberga, Jacob. In silico detection of sequence variations modifying transcriptional regulation. PLoS Computational Biology 4.1 (2008): e5.
• Grant, Charles E., Timothy L. Bailey, and William Stafford Noble. FIMO: scanning for occurrences of a given motif. Bioinformatics 27.7 (2011): 1017-1018.
• Macintyre, Geoff, Bailey, James, Haviv, Izhak, and Kowalczyk, Adam. is-rSNP: a novel technique for in silico regulatory SNP detection. Bioinformatics 26.18 (2010): i524-i530.
• Riva, Alberto. Large-scale computational identification of regulatory SNPs with rSNP-MAPPER. BMC genomics 13.4 (2012): S7.
• Thomas-Chollier, Morgane, Hufton, Andrew, Heinig, Matthias, O'keeffe, Sean, Masri, Nassim E., Roider, Helge G., Manke, Thomas, and Vingron, Martin. Transcription factor binding predictions using TRAP for the analysis of ChIP-seq data and regulatory SNPs. Nature protocols 6.12 (2011): 1860-1869.

1. http://www.ncbi.nlm.nih.gov/books/NBK44417/#Content.what_is_a_reference_snp_or__rs_i
2. https://bioconductor.org/packages/release/data/annotation/manuals/SNPlocs.Hsapiens.dbSNP144.GRCh38/man/SNPlocs.Hsapiens.dbSNP144.GRCh38.pdf

Motif information
Motif	Motif name, e.g., MA0002.2 For JASPAR motifs, a link to the motif's entry in JASPAR database by clicking on the motif.
Transcription Factor	Transcription factor name, e.g., RUNX1 When available, a link to the FACTORBOOK page for the transcription factor is provided by clicking on the transcription factor.

atSNP Log Likelihoods (evaluated wrt PWM)
Rank Ratio (SNP Impact)	Log of ratio of p-value Reference to p-value SNP. This is the main statistic that the atSNP p-values are based on.
Reference	The log likelihood score of the best match (highest scoring subsequence) with the reference allele.
SNP	The log likelihood score of the best match (highest scoring subsequence) with the SNP allele.
Ratio	The difference between Log Likelihood Reference and Log Likelihood SNP.
Enhance Odds	Once atSNP identifies the best matching subsequence to the PWM with the reference allele, it evaluates the likelihood of that subsequence with the reference allele and the SNP allele. This quantity is the log of the ratio of these two likelihoods.
Reduce Odds	Once atSNP identifies the best matching subsequence to the PWM with the SNP allele, it evaluates the likelihood of that subsequence with the reference allele and the SNP allele. This quantity is the log of the ratio of these two likelihoods.

atSNP p-values
p-value SNP Impact	A significant p-value, e.g. 0.0001, statistically supports the potential gain or loss of function of the genomic region with the SNP in terms of transcription factor binding.
p-value Reference	p-value for score with the reference allele (Log Likelihood Reference). A significant p-value indicates that the match to the PWM with the reference allele is statistically supported.
p-value SNP	p-value for scores with the SNP allele (Log Likelihood SNP). A significant p-value indicates that the match to the PWM with the SNP allele is statistically supported.
*p-value Difference	p-value for the difference in scores with the reference and the SNP alleles (Log Likelihood Ratio). This is essentially the result from a likelihood ratio test. While we report this for completeness, the final atSNP results are based on p-value SNP impact which quantifies the significace of the Log Rank Ratio.
*p-value Condition Ref	Conditional p-value for scores on the reference allele based on Log Enhance Odds.
*p-value Condition SNP	Conditional p-value for scores on the SNP allele based on Log Reduce Odds.

^* While these p-values are reported for completeness, they are not used for querying.

Example

We evaluate the impact of the SNP rs117959046 on the binding of the transcription factor GATA2, using the GATA_known12 motif:

The SNP “rs117959046” is a genetic mutation from C to T in the forward strand at location Chromosome 1: 134,339,671. This creates a highly likely binding site for the transcription factor GATA2 as supported by the p-value SNP Impact of 0.00484. Furthermore, the p-value for the best match between the GATA2 motif and the DNA sequence with the SNP allele is 0.00898, whereas the p-value for the best match between the motif and the DNA sequence with the reference allele is 0.499. These p-values indicate that this genomic location, spanning the SNP position, does not harbor a GATA2 binding site with the reference allele, but the SNP creates a likely GATA2 binding site. Overall, atSNP statistically supports that the SNP leads to generation of an in silico predicted GATA2 binding site.

Examples

The SNP, rs117959046, is highly likely to create a binding site for the transcription factor GATA2 on the genome. The composite logo plot illustrates this with the comparison between the best matches of GATA2_2 PWM to the DNA sequence overlapping the SNP position with the reference and SNP alleles. Among subsequences of length 5 overlapping the SNP position with the reference allele, “AGACA” on the forward strand has the largest log likelihood score. This subsequence is referred to as the “best match with the reference allele”. This subsequence is depicted on the second row with the corresponding PWM on the first row. Similarly, among subsequences of length 5 overlapping the SNP position with the SNP allele, “AGATA” on the forward strand is referred to as the “best match with the SNP allele”. It is depicted on the third row with the corresponding PWM on the fourth row.

The SNP, rs199894601, is highly likely to create a binding site for the transcription factor RUNX1 on the genome. The best match position on the genome with the SNP allele is different from that with the reference allele. Specifically, the best matches occur on different strands and genomic coordinates (bp). The composite logo plot illustrates this with the comparison between the best matches of RUNX1 PWM from JASPAR to the DNA sequence overlapping the SNP location with the reference and SNP alleles. The beginning of the best match with the SNP allele is 3 bases before the SNP position in the 5' to 3' direction on the reverse strand. Among subsequences of length 11 overlapping the SNP position with the SNP allele, this subsequence has the largest log likelihood score. This subsequence is referred to as the “best match with the SNP allele” and is depicted on the third row with the corresponding PWM on the fourth row. Similarly, the “best match with the reference allele” is depicted on the second row with the corresponding PWM on the first row. The beginning of the best match with the reference allele is one base before the SNP position in the 5' to 3' direction on the forward strand.