Sources of bias in measures of allele-specific expression derived from RNA-seq data aligned to a single reference genome

Aligning sequence reads to both parental genomes can eliminate the systematic bias in relative allele-specific expression favoring the reference allele.

“two alternative strategies for aligning reads have been shown to eliminate the systematic bias in measures of relative ASE favoring the reference allele”

• Also stated in: § The systematic bias in measures of ASE correlates with the density of differentiating sites · boosted

  “aligning reads to the maternal and paternal genomes, which is a strategy that has previously been shown to produce unbiased measures of relative ASE [15,18-20,25]”

• cites [15] McManus CJ, Coolon JD, Duff MO, Eipper-Mains J, Graveley BR, Wittkopp PJ: Regulatory divergence in Drosophila revealed by mRNA-seq. Genome Res. 2010, 20: 816-825. 10.1101/gr.102491.109. sourcePartly supported

  The Results and Discussion section says that reads from mixed parental and hybrid samples were independently aligned to both parental genomes, supporting the alignment strategy. The paper reports checks on allele-assignment accuracy, but our copy does not show that this strategy eliminated systematic bias favoring a reference allele. (LLM, conf=0.84)

  Experimental — accuracy not yet measured
• cites [18] Rozowsky J, Abyzov A, Wang J, Alves P, Raha D, Harmanci A, Leng J, Bjornson R, Kong Y, Kitabayashi N, Bhardwaj N, Rubin M, Snyder M, Gerstein M: AlleleSeq: analysis of allele-specific expression and binding in a network framework. Mol Syst Biol. 2011, 7: 522- sourceSupported
• cites [19] Rivas-Astroza M, Xie D, Cao X, Zhong S: Mapping personal functional data to personal genomes. Bioinformatics. 2011, 27: 3427-3429. 10.1093/bioinformatics/btr578. sourceNot checked — no full text of the source
• cites [20] Satya RV, Zavaljevski N, Reifman J: A new strategy to reduce allelic bias in RNA-Seq readmapping. Nucleic Acids Res. 2012, 40: e127-10.1093/nar/gks425. sourcePartly supported

  The Results on simulated reads report that mapping to the paper’s enhanced reference almost completely eliminated the average bias favoring the reference allele. But the Methods describe an enhanced reference containing alternate alleles at known SNPs, not alignment to both parental genomes; the paper does not establish that specific strategy’s effect. (LLM, conf=0.88)

  Experimental — accuracy not yet measured
• cites [25] Graze RM, McIntyre LM, Main BJ, Wayne ML, Nuzhdin SV: Regulatory divergence in Drosophila melanogaster and D. simulans, a genomewide analysis of allele-specific expression. Genetics. 2009, 183: 547-561. 10.1534/genetics.109.105957. 61– 1SI–21SI sourceCould not verify

  The Materials and Methods and File S1 passages describe tiling-array probes, hybridization bias, and DNA controls—not alignment of sequence reads to both parental genomes. Searches for read alignment and reference-allele bias found no passage testing whether that strategy eliminates systematic bias. No page numbers were available in our copy. (LLM, conf=0.75)

  Experimental — accuracy not yet measured

Relative allele-specific expression is systematically biased when sequence reads are aligned to a single reference genome.

“relative ASE is systematically biased when sequence reads are aligned to a single reference genome”

• Also stated in: § Background · plain

  “Despite simulating an equal number of reads from each allele, a bias was observed causing reads to be assigned more often to the reference allele than the alternative allele.”

• cites Degner et al. [14] Degner JF, Marioni JC, Pai AA, Pickrell JK, Nkadori E, Gilad Y, Pritchard JK: Effect of read-mapping biases on detecting allele-specific expression from RNA-sequencing data. Bioinformatics. 2009, 25: 3207-3212. 10.1093/bioinformatics/btp579. sourceSupported

The alignment parameters used by Degner et al. allowed two or fewer bases within each read to differ from the reference genome.

“The alignment parameters they used allowed two or fewer bases within each read to differ from the reference genome.”

• cites Degner et al. [14] Degner JF, Marioni JC, Pai AA, Pickrell JK, Nkadori E, Gilad Y, Pritchard JK: Effect of read-mapping biases on detecting allele-specific expression from RNA-sequencing data. Bioinformatics. 2009, 25: 3207-3212. 10.1093/bioinformatics/btp579. sourceSupported

A systematic bias occurs when reads from a heterozygous individual are aligned to a single reference genome.

“a systematic bias occurs when reads from a heterozygous individual are aligned to a single reference genome [14].”

• cites [14] Degner JF, Marioni JC, Pai AA, Pickrell JK, Nkadori E, Gilad Y, Pritchard JK: Effect of read-mapping biases on detecting allele-specific expression from RNA-sequencing data. Bioinformatics. 2009, 25: 3207-3212. 10.1093/bioinformatics/btp579. sourceSupported

Simulated datasets contained either ~10-fold or ~100-fold more differentiating sites than the human genotypes used to validate other methods for measuring relative allele-specific expression.

“Simulated datasets contained either ~10-fold or ~100-fold more differentiating sites than the human genotypes used to validate other methods for measuring relative ASE [14,18,20].”

• cites [14] Degner JF, Marioni JC, Pai AA, Pickrell JK, Nkadori E, Gilad Y, Pritchard JK: Effect of read-mapping biases on detecting allele-specific expression from RNA-sequencing data. Bioinformatics. 2009, 25: 3207-3212. 10.1093/bioinformatics/btp579. sourcePartly supported

  Methods §2.2 describes using human HapMap genotypes to measure allele-specific expression, and Results §3 reports 1,981 heterozygous SNPs meeting the read threshold. The paper does not establish that the citing document’s simulated datasets had approximately 10 or 100 times as many differentiating sites, so the numerical comparison is not supported by this source. (LLM, conf=0.78)

  Experimental — accuracy not yet measured
• cites [18] Rozowsky J, Abyzov A, Wang J, Alves P, Raha D, Harmanci A, Leng J, Bjornson R, Kong Y, Kitabayashi N, Bhardwaj N, Rubin M, Snyder M, Gerstein M: AlleleSeq: analysis of allele-specific expression and binding in a network framework. Mol Syst Biol. 2011, 7: 522- sourcePartly supported

  The Introduction (page not provided) says HapMap III SNP calls used in earlier studies were approximately 10-fold fewer than those available from the 1000 Genomes pilot II genotypes. It does not compare simulated datasets with those genotypes. Its approximately 100-fold figure concerns the number of possible comparisons between allele-specific binding and expression SNPs, not the number of differentiating sites. (LLM, conf=0.88)

  Experimental — accuracy not yet measured
• cites [20] Satya RV, Zavaljevski N, Reifman J: A new strategy to reduce allelic bias in RNA-Seq readmapping. Nucleic Acids Res. 2012, 40: e127-10.1093/nar/gks425. sourcePartly supported

  The Materials and Methods (Simulated data; page unavailable) describe reads simulated at human SNP loci, and the Results (Results on RNA-Seq data; page unavailable) describe evaluating mapping methods with data from two Yoruba HapMap individuals. The paper does not state that simulated datasets had approximately 10- or 100-fold more differentiating sites than those human genotypes. (LLM, conf=0.82)

  Experimental — accuracy not yet measured

Differences between alleles in sequences that appear more than once in the genome can cause reads to be excluded for one allele but not the other.

“Differences between alleles in sequences that appear more than once in the genome can also cause reads to be excluded for one allele but not the other [14].”

• cites [14] Degner JF, Marioni JC, Pai AA, Pickrell JK, Nkadori E, Gilad Y, Pritchard JK: Effect of read-mapping biases on detecting allele-specific expression from RNA-sequencing data. Bioinformatics. 2009, 25: 3207-3212. 10.1093/bioinformatics/btp579. sourceSupported

In simulated heterozygous human data, aligning reads to a single reference genome caused reads to be assigned more often to the reference allele than the alternative allele despite simulating an equal number of reads from each allele.

“Despite simulating an equal number of reads from each allele, a bias was observed causing reads to be assigned more often to the reference allele than the alternative allele.”

• cites Degner et al. [14] Degner JF, Marioni JC, Pai AA, Pickrell JK, Nkadori E, Gilad Y, Pritchard JK: Effect of read-mapping biases on detecting allele-specific expression from RNA-sequencing data. Bioinformatics. 2009, 25: 3207-3212. 10.1093/bioinformatics/btp579. sourceSupported

Differences between alleles in mappability contribute to inaccuracy of allele-specific expression.

“Differences between alleles in mappability (i.e. the ability to align a read uniquely within the genome) also contribute to inaccuracy of ASE”

Excluding reads that overlapped one or more insertions or deletions (indels) between alleles increased the accuracy of allelic abundance.

“and reads that overlapped one or more insertions or deletions (indels) between alleles.”

• Also stated in: § Allele-specific differences in mappability and insertions/deletions affect measurements of ASE · plain

  “sites with perfect mappability that had an indel nearby showed more reads assigned to D. melanogaster than D. simulans allele when reads were aligned to only the D. melanogaster genome”

After filtering out problematic sites, measures of relative allelic abundance derived from aligning reads to a single reference genome were similar to those produced by aligning sequence reads separately to the maternal and paternal genomes.

“After filtering out these problematic sites, measures of relative allelic abundance derived from aligning reads to a single reference genome were similar to those produced by aligning sequence reads separately to the maternal and paternal genomes (Figure 5C-E).”

• Also stated in: § Conclusions · plain

  “After excluding these biased sites, as well as those affected by imperfect mappability and/or an indel(s) nearby, we found that RNA-seq data aligned to a single reference genome produced measures of relative ASE that were comparable to those resulting from separately aligning the same reads to allele-specific maternal and paternal genomes.”

After filtering based on the number of neighboring differentiating sites, 46.6%-86.9% of differentiating sites were retained in the 36-base intraspecific simulations when reads were aligned to a single reference genome and one, two, or three mismatches were allowed.

“After filtering based on the number of neighboring differentiating sites, we found that 46.6%-86.9% and 8.3%-50.5% of differentiating sites were retained in the 36-base intra- and interspecific simulations, respectively, when the reads were aligned to a single reference genome and one, two, or three mismatches were allowed (Figure 6).”

66.3%-95.2% of exons contained at least one reliable differentiating site in the interspecific simulations when the same alignment conditions were used.

“By comparison, 81.8%-91.8% and 66.3%-95.2% of exons contained at least one of these reliable differentiating sites when the same alignment conditions were used in the intra- and interspecific simulations, respectively.”

A bias toward the reference allele was still observed when two or three mismatches were allowed, but only for sites where the number of neighboring differentiating sites was greater than or equal to the number of mismatches allowed.

“A bias toward the reference allele was still observed, but only for sites where the number of neighboring differentiating sites was greater than or equal to the number of mismatches allowed during the alignment step (Figure 2B,C).”

Increasing the number of mismatches allowed to two or three improved measures of allelic abundance: 80.2% and 91.9% of differentiating sites were inferred to be equally abundant when two and three mismatches, respectively, were allowed.

“We found that increasing the number of mismatches improved measures of allelic abundance: 80.2% and 91.9% of differentiating sites were inferred to be equally abundant when two and three mismatches, respectively, were allowed.”

Clusters of differentiating sites prevented sequence reads from an alternate allele from aligning to the reference genome, causing a bias in relative allele-specific expression favoring the reference allele.

“We found that clusters of differentiating sites prevented sequence reads from an alternate allele from aligning to the reference genome, causing a bias in relative ASE favoring the reference allele.”

• Also stated in: § The systematic bias in measures of ASE correlates with the density of differentiating sites · plain

  “This bias was caused by the failure of reads simulated from the alternative allele to align to the reference genome more often than those simulated from the reference allele.”

Prior to excluding any sites, 70.4%, 88.9%, and 93.3% of all differentiating sites showed equal allelic abundance when reads were aligned to a single genome with one, two, or three mismatches allowed, respectively.

“Prior to excluding any sites, 70.4%, 88.9%, and 93.3%, respectively, of all differentiating sites showed equal allelic abundance when reads were aligned to a single genome with one, two, or three mismatches allowed.”

The bias toward the reference allele was higher for the interspecific comparison (D. melanogaster vs D. simulans) than for the intraspecific comparison (two D. melanogaster strains) when reads were aligned to a single reference genome.

“As predicted, we found that the bias toward the reference allele was higher for the interspecific comparison than for the intraspecific comparison when reads were aligned to a single reference genome (Figure 2, compare I-K with A-C).”

Of sites with at least one neighboring differentiating site, 99.8% showed more reads assigned to the reference allele than to the alternative allele.

“Of these sites, 99.8% showed more reads assigned to the reference allele than to the alternative allele.”

When reads were aligned to both parental genomes, sites with perfect mappability showed equal representation of the two alleles regardless of the presence or absence of nearby indels.

“When reads were aligned to both parental genomes, sites with perfect mappability showed equal representation of the two alleles regardless of the presence or absence of nearby indels (Figure 4D).”

The extent of bias toward the reference allele increased with the number of neighboring differentiating sites.

“Furthermore, the extent of bias toward the reference allele increased with the number of neighboring differentiating sites (Figure 2A).”

After filtering based on the number of neighboring differentiating sites, 8.3%-50.5% of differentiating sites were retained in the 36-base interspecific simulations when reads were aligned to a single reference genome and one, two, or three mismatches were allowed.

“After filtering based on the number of neighboring differentiating sites, we found that 46.6%-86.9% and 8.3%-50.5% of differentiating sites were retained in the 36-base intra- and interspecific simulations, respectively, when the reads were aligned to a single reference genome and one, two, or three mismatches were allowed (Figure 6).”

After aligning reads to both parental genomes, 96.9% of differentiating sites showed evidence of equal allelic abundance.

“After aligning reads to both parental genomes, 96.9% showed evidence of equal allelic abundance.”

81.8%-91.8% of exons contained at least one reliable differentiating site in the intraspecific simulations when the same alignment conditions were used.

“By comparison, 81.8%-91.8% and 66.3%-95.2% of exons contained at least one of these reliable differentiating sites when the same alignment conditions were used in the intra- and interspecific simulations, respectively.”

Excluding differentiating sites with at least as many neighboring differentiating sites as the number of mismatches allowed increased the percentage of sites with equal allelic abundance to 96.3%-96.6% when aligning to a single reference genome.

“Excluding differentiating sites with at least as many neighboring differentiating sites as the number of mismatches allowed increased this percentage to 96.3%-96.6% when aligning to a single reference genome (Figure 5B).”

For 36-base reads simulated from D. melanogaster and D. simulans, only ~94% of sites with perfect mappability showed equal abundance when reads were aligned to a single (D. melanogaster) reference genome.

“but only ~94% of sites with perfect mappability showed such equal abundance when reads were aligned to a single (D. melanogaster) reference genome (Figure 3I-K).”

Excluding additional differentiating sites with imperfect mappability in both datasets, as well as sites with one or more nearby indels in the intraspecific dataset, had little effect on the proportion of differentiating sites and exons retained.

“Excluding additional differentiating sites with imperfect mappability in both datasets, as well as sites with one or more nearby indels in the intraspecific dataset, had little effect on the proportion of differentiating sites and exons retained (Figure 6).”

The bias in relative allele-specific expression favoring the reference allele increased with greater sequence divergence between alleles.

“This bias increased with greater sequence divergence between alleles.”

• Also stated in: § Read length and the amount of sequence divergence can also affect allelic bias · plain

  “we found that the bias toward the reference allele was higher for the interspecific comparison than for the intraspecific comparison when reads were aligned to a single reference genome”

When three mismatches were allowed for 36-base reads and four mismatches for 50-base reads, 91.9% and 92.1% of differentiating sites showed equal allelic abundance, respectively.

“91.9% and 92.1% of differentiating sites showed equal allelic abundance for 36- and 50-base reads when three and four mismatches, respectively, were allowed (Additional file 3).”

A similar bias toward the reference allele is caused by the presence of an indel near the focal differentiating site.

“A similar bias toward the reference allele is caused by the presence of an indel near the focal differentiating site.”

When reads were aligned to maternal and paternal genomes with zero mismatches allowed, 98.8% of differentiating sites showed equal representation.

“By contrast, 98.8% of differentiating sites showed equal representation when reads were aligned to the maternal and paternal genomes with zero mismatches allowed (Figure 2H).”

Of all sites considered, 49.8% had at least one neighboring differentiating site.

“Of all sites considered, 49.8% had at least one neighboring differentiating site (i.e., at least one other differentiating site within an overlapping read).”

Increasing the number of mismatches allowed increased the percentage of reads that failed to map uniquely: allowing one, two, and three mismatches, 2.2%, 2.5%, and 2.9% of all reads failed to map uniquely, respectively.

“Increasing the number of mismatches allowed reduced the bias toward the reference allele, but increased the percentage of reads that failed to map uniquely: allowing one, two, and three mismatches, 2.2%, 2.5%, and 2.9% of all reads failed to map uniquely, respectively.”

For both 36- and 50-base reads simulated from the two D. melanogaster genotypes, >97.9% of sites with perfect mappability showed the expected equal abundance of the reference and alternative alleles under all mapping conditions.

“For both the 36- and 50-base reads simulated from the two D. melanogaster genotypes, > 97.9% of sites with perfect mappability showed the expected equal abundance of the reference and alternative alleles under all mapping conditions (Figure 3A-H).”

Increasing the number of mismatches allowed when aligning sequence reads to the reference genome and restricting analysis to genomic regions with fewer differentiating sites than the number of mismatches allowed almost completely eliminated the systematic bias.

“Increasing the number of mismatches allowed when aligning sequence reads to the reference genome and restricting analysis to genomic regions with fewer differentiating sites than the number of mismatches allowed almost completely eliminated this systematic bias.”

• Also stated in: § Background · plain

  “We found that limiting analysis of relative ASE to regions of the genome with no more differentiating sites than the number of mismatches allowed eliminated the systematic bias toward the reference allele and produced measures of ASE similar to those inferred from aligning reads separately to the maternal and paternal genomes.”

When aligning reads to both parental genomes, sequence divergence had a negligible impact: the intra- and interspecific datasets produced nearly identical results.

“When aligning reads to both parental genomes, however, sequence divergence had a negligible impact: the intra- and interspecific datasets produced nearly identical results (Figure 2, compare L with D).”

The systematic bias is predominantly caused by additional differentiating sites located near the focal differentiating site that interfere with read alignment.

“We found that this systematic bias is predominantly caused by additional differentiating sites located near the focal differentiating site that interfere with read alignment.”

For 36-base reads simulated from D. melanogaster and D. simulans, 99.9% of sites with perfect mappability showed equal abundance when reads were aligned to both parental genomes.

“For the 36-base reads simulated from the D. melanogaster and D. simulans genomes, 99.9% of sites with perfect mappability showed equal abundance when reads were aligned to both parental genomes (Figure 3L)”

Increasing the number of mismatches allowed reduced the bias toward the reference allele.

“Increasing the number of mismatches allowed reduced the bias toward the reference allele, but increased the percentage of reads that failed to map uniquely”

Aligning reads to only the alternative allele produced complementary results.

“Aligning reads to only the alternative allele produced complementary results (Additional file 2).”

Differentiating sites with an excess of neighboring differentiating sites were the most common of the three types of problematic sites in both intra- and interspecific simulations.

“Differentiating sites with an excess of neighboring differentiating sites were the most common of these three types of problematic sites in both intra- and interspecific simulations (Figure 5A).”

Excluding problematic sites did not preclude measuring relative allele-specific expression for most exons.

“we showed that excluding these problematic sites did not preclude measuring relative ASE for most exons”

Differences between alleles in mappability do not systematically favor one allele or the other across the genome.

“but do not systematically favor one allele or the other across the genome.”

The precise density at which neighboring differentiating sites became problematic depended on the number of mismatches allowed during the alignment of sequencing reads.

“The precise density at which neighboring differentiating sites became problematic depended on the number of mismatches allowed during the alignment of sequencing reads.”

Sites affected by the systematic bias toward the reference allele could be identified and excluded prior to estimating allele-specific expression based on the density of differentiating sites.

“Using both simulated and real sequencing data, we found that sites affected by the systematic bias toward the reference allele could be identified and excluded prior to estimating ASE based on the density of differentiating sites.”

The observed findings are consistent with the hypothesis that the density of differentiating sites complicates the mapping of reads and leads to biased measures of relative allele-specific expression.

“These findings are consistent with our hypothesis that the density of differentiating sites complicates the mapping of reads and leads to biased measures of relative ASE.”

After aligning sequence reads to a single reference genome, excluding differentiating sites with at least as many neighboring differentiating sites as the number of mismatches allowed, imperfect mappability, and/or an indel(s) nearby resulted in measures of allelic abundance comparable to those derived from aligning sequence reads to both parental genomes.

“After aligning sequence reads to a single reference genome, excluding differentiating sites with at least as many neighboring differentiating sites as the number of mismatches allowed, imperfect mappability, and/or an indel(s) nearby resulted in measures of allelic abundance comparable to those derived from aligning sequence reads to both parental genomes.”

The simulation approach guaranteed that reads from each allele were present in equal amounts.

“Importantly, this approach guaranteed that reads from each allele were present in equal amounts.”

The major consequence of excluding sites based on the density of differentiating sites, mappability, and/or indels is that fewer allele-specific reads will be successfully mapped for each exon and for each gene.

“The major consequence of excluding sites based on the density of differentiating sites, mappability, and/or indels is that fewer allele-specific reads will be successfully mapped for each exon and for each gene.”

The bias toward the reference allele was caused by the failure of reads simulated from the alternative allele to align to the reference genome more often than those simulated from the reference allele.

“This bias was caused by the failure of reads simulated from the alternative allele to align to the reference genome more often than those simulated from the reference allele.”

Analyses using real and simulated reads to compare the same sets of alleles retain the same sites and exons when aligned to the same reference genome because differentiating sites are excluded based only on the genome sequence(s).

“Analyses using real and simulated reads to compare the same sets of alleles retain the same sites and exons when aligned to the same reference genome because differentiating sites are excluded based only on the genome sequence(s).”

The retention of more differentiating sites and exons in the intraspecific simulation than in the interspecific simulation is consistent with the lower sequence divergence within than between species.

“The retention of more differentiating sites and exons in the intraspecific simulation than in the interspecific simulation (Figure 6) is consistent with the lower sequence divergence within than between species.”

• AssumptionRests on an assumption: Sequence divergence is lower within D. melanogaster than between D. melanogaster and D. simulans.

The study determined the relative impact of these factors on measures of relative allele-specific expression.

“and determining the relative impact of these factors on these measures”

The study identified specific factors causing erroneous measures of relative allele-specific expression reported in prior work.

“By identifying the specific factors causing erroneous measures of relative allele-specific expression reported in prior work”