# eQTL Mapping RNA-seq Microarray Integration: Common Problems and Fixes Expression quantitative trait locus (eQTL) mapping connects genetic variants to [gene expression](/blog/guides/gene-expression) measurements. When a study combines RNA-seq data with microarray data, either within one cohort or across cohorts that must be meta-analyzed, most failures trace back to a single root cause: the two platforms do not measure the same quantity on the same identifiers, under the same assumptions, with the same covariates. The integration then produces three familiar symptoms. Signals disagree between platforms, signals disappear when platforms are merged, or signals appear that collapse under a colocalization or replication check. The first thing to check is boring and decisive. Confirm that every expression feature on both platforms maps to the same gene definition, against the same genome build and annotation release, and that every expression sample pairs with the correct genotype sample. A large fraction of inconsistent eQTL results come from identifier drift (probe IDs that no longer match current gene symbols), annotation version skew between an old microarray annotation file and a fresh RNA-seq quantification, and silent sample swaps between the expression matrix and the genotype file. Fix those before touching statistics. This article walks through the failure modes of eqtl mapping rna-seq microarray integration from the most common to the rarest, explains the mechanism behind each one, gives fixes that hold up in reproducible pipelines, and ends with prevention practices and answers to the questions that come up most often. ## A Fast Diagnostic Sequence ![Ten step diagnostic sequence starting with feature identifier overlap and ending with the full eQTL scan.](/article-images/illustrations/eqtl-mapping-rna-seq-microarray-integration-diagnostic-sequence.webp) Run this sequence before any modeling. It takes an afternoon and resolves the majority of integration failures. 1. Print the expression feature identifiers from both platforms. Count how many match directly. If the overlap is below what the platforms should share, stop and fix the mapping first. 2. Record the annotation release, genome build, and probe annotation file version for every expression dataset. Write them into the project README. 3. Confirm that genotype and expression sample identifiers join one-to-one. Check sex concordance, reported kinship, and genotype-expression correlation structure if related individuals are present. 4. Plot the expression distribution of every sample. On microarrays, look for arrays with shifted intensity distributions. On RNA-seq, look for samples with abnormally low library size, abnormal GC content, or high percentages of reads mapped to rRNA. 5. Check expression value ranges. Log2 microarray intensities, raw counts, counts per million, transcripts per million (TPM), and variance-stabilized RNA-seq values all live on different scales. If values from the two platforms sit on visibly different ranges in the same plot, quantification harmonization has not happened yet. 6. Compute the first two principal components (PCs) of the combined expression matrix colored by platform. If platform dominates PC1, you have a platform effect that must be modeled or removed before joint eQTL testing. 7. Recheck cis-eQTL direction. For a handful of well-known strong cis-eQTLs, the effect allele should push expression in the same direction on both platforms. Flipped directions point to allele coding or strand errors on the genotype side. 8. Verify the variant reference allele against the correct genome build. A variant lifted from build 37 to build 38 without strand checking is a classic source of allele flips. 9. Run one gene, one variant, one model by hand. Trace the input values from raw data to the test statistic. Most pipeline bugs surface in this single trace. 10. Only then scale up to the full scan. ## Master Troubleshooting Table | Symptom | Likely cause | How to confirm | Fix | |--|--|--|--| | RNA-seq and microarray eQTLs for the same gene point in opposite directions | Gene identifier mismatch, allele strand flip, or probe targeting a different transcript | Map probe sequences to current annotation, check variant alleles against the reference build | Re-annotate all features to one gene model and one genome build, verify allele coding | | Platforms agree on effect direction but meta-analysis finds nothing | Different scales and variances injected into the combined test | Compare per-platform effect size distributions and standard errors | Meta-analyze on platform-specific standardized effect sizes, never on raw expression scales | | Many eGenes found on one platform, few on the other | Platform sensitivity difference, low expression genes dropped by probe filtering | Count eGenes detected on each platform at matched sample sizes and compare against known expression levels | Treat RNA-seq as the higher-resolution platform for discovery, use microarray data for replication where its probes are reliable [1] | | Effect sizes shrink in the RNA-seq cis-eQTL test compared with published results | Two-stage phasing of haplotypes treated phased inference as observed data | Compare estimates from a one-stage likelihood model against a two-stage pipeline on the same data | Use a maximum-likelihood framework that integrates haplotype inference and association testing in a single stage [2] | | Replicate samples cluster by platform, not by individual | Unremoved platform and batch effects | [Principal component analysis](/blog/data-analysis/principal-component-analysis-explained) colored by platform and processing batch | Model platform as a covariate or apply batch correction, then verify that true biological signal survives | | Apparent trans-eQTL hotspots with thousands of associations to one variant | Cell-type composition differences or a hidden batch variable driving expression genome-wide | Correlate the suspect variant with cell-type proportions or surrogate variables | Estimate and include cell-type proportions as covariates, for example with a deconvolution method such as DeCompress [3] | | eQTL mapping on a targeted expression panel produces unstable results | Small feature space defeats reference-free deconvolution | Check the number of genes on the panel against deconvolution method requirements | Expand the feature space with a reference dataset using compressed sensing, then deconvolve [3] | | Colocalization between GWAS and eQTL fails on one platform only | The microarray probe missed the relevant exon or the RNA-seq quantification captured it | Inspect probe targets and transcript coverage at the locus | Rerun colocalization with exon-level RNA-seq quantifications, which recover more candidate causal eQTLs [1] | | Associations detected with raw counts vanish after normalization | Library size and composition effects confounding counts | Compare library sizes and detect outlier samples before and after normalization | Use normalization appropriate to the test, typically raw counts with offset-based models or transformed values with robust scaling, and keep the choice consistent across platforms in meta-analysis | ## Problem 1, Gene Identifier and Annotation Version Mismatches This is the most common cause of failed integration and the cheapest to fix. **Mechanism.** Microarray annotations were frozen when the array shipped. Probe annotation files from years ago reference gene symbols that have been renamed, retired, or split. RNA-seq quantification pipelines use current gene models such as Gencode or Ensembl releases. A gene symbol like a HGNC alias can refer to different genomic coordinates across annotation versions, and one microarray probe can hybridize to transcripts that current annotation assigns to several genes. When integration joins on gene symbol, rows silently merge across different underlying measurements. The eQTL signal for a gene then mixes expression from transcripts that the variant does not regulate, diluting real cis signals and creating fake ones. **How to confirm.** Take ten probes from the microarray annotation file. Map each probe sequence to the current genome with an alignment tool such as Bowtie or an alignment utility from the Bioconductor ecosystem. Check whether the aligned coordinates fall inside the exon coordinates of the gene that the annotation file claims. Do the same for a handful of RNA-seq gene IDs against the same gene model. Any disagreement here predicts widespread problems. **Fix.** - Pick one gene definition. For human work, a current Gencode release with a stated version number is the standard choice. Map every microarray probe to that gene model by sequence, not by the vendor symbol column. - For multi-mapping probes, either drop them or keep them under a transparent rule, for example keep only probes that map uniquely to exons of a single gene. Write the rule down before looking at results. - Lift every dataset to one genome build. CrossMap or the UCSC liftOver tool handles coordinate conversion. For variants, use bcftools with the norm command and the appropriate chain file, then recheck reference and alternate alleles after the lift, because reference alleles can change between builds. - Keep the original annotation and the mapped annotation side by side in the output directory so the mapping is auditable. ## Problem 2, Different Quantities on Different Scales **Mechanism.** A two-color or one-color microarray produces background-corrected probe intensities, summarized and typically log2 transformed, that reflect hybridization signal relative to probe affinity. RNA-seq produces read counts that reflect molecule numbers relative to library size and gene length. These are related but not identical quantities. Gene length, probe affinity, 3-prime bias, isoform composition, and dynamic range compression all differ between platforms. Naively z-scoring both matrices and merging them treats these differences as if they were noise, which they are not. The practical consequence for eQTL work is subtle. Within a single platform, eQTL tests are usually robust to monotonic transformations because rank-based and linear tests on log-scale data behave similarly. Across platforms, the variance structure differs, so a combined analysis with pooled variance produces inflated test statistics for one platform and deflated statistics for the other. **How to confirm.** For genes measured on both platforms, compute within-sample rank correlations between platforms. Most gene-gene correlations across platforms in well-run data should be strongly positive. Genes with near-zero or negative cross-platform correlation usually indicate a probe problem or an annotation mismatch rather than biology. Then plot the variance of each gene against its mean expression on each platform. If the mean-variance relationships have different shapes, combined linear models will misbehave. **Fix.** - Run eQTL tests within each platform first. This is the safest default and the approach used in practice in published comparisons. The TwinsUK lymphoblastoid cell line study of systemic lupus erythematosus variants ran microarray and RNA-seq eQTL analyses in parallel and then compared their ability to detect disease-relevant eGenes, rather than pooling raw values [1]. - For meta-analysis, combine standardized effect sizes, meaning the beta coefficient from each platform-specific regression divided by its standard error, using inverse-variance weighting. Do not pool expression values across platforms into one matrix and fit one model. - If a joint model is genuinely required, harmonize on a rank or quantile basis and include platform as a covariate, then verify the result on held-out samples. ## Problem 3, Batch Effects Masquerading as Platform Effects **Mechanism.** Batch effects come from processing date, library preparation lot, sequencer or scanner, operator, and reagent batch. They are present within a platform, not only between platforms. In eQTL work, a batch effect correlated with genotype structure can create false associations, and a batch effect correlated with disease status can mask true ones. Because batches are often correlated with platform in study designs where RNA-seq was done later than the microarray run, platform and batch are confounded unless the design breaks that correlation or the model separates them. **How to confirm.** Fit a linear model per gene with platform and batch date as predictors of expression. If batch explains substantial variance after platform, the two are separable. If a design has complete confounding, for example all RNA-seq samples processed in one batch and all microarray samples in another, no statistical method can fully separate the effects, and the honest fix is analytic caution or additional data. **Fix.** - Include batch variables directly in the eQTL model. Direct inclusion is preferable to pre-correction, because pre-correction with tools such as ComBat can remove genuine genotype-driven expression signal along with the batch signal when genotype and batch are correlated. - When pre-correction is used, apply it only to expression and never to genotypes, and always refit the eQTL model on corrected data with genotype included so that residual genotype signal is testable. - For RNA-seq counts, use model-based approaches such as a negative binomial or a linear model on transformed counts with surrogate variables estimated from the data, for example with the PEER framework commonly used in eQTL studies or the sva package. Reserve a portion of the strongest cis-eQTLs as positive controls to check that correction did not absorb the signal of interest. **Safety note.** Expression samples are frequently human-derived material, such as lymphoblastoid cell lines, whole blood, or tissue. Handle them under institutional biosafety procedures for human specimens, typically [biosafety level](/knowledge/diagnostics/emerging-tech/biosafety-levels-1-4-a-comparative-framework-for-diagnostic-laboratory-design) 2 practices, with appropriate training and personal protective equipment. Freezers containing RNA samples also hold Cryoblock or similar cryopreservation agents if mechanical freezing was used, so follow institutional hazard guidance when accessing ultra-low temperature freezers. ## Problem 4, Population Structure and Hidden Confounders in the eQTL Model **Mechanism.** Allele frequencies and expression levels both correlate with ancestry and with subtle technical structure. An eQTL model that omits ancestry covariates will detect variants associated with population labels rather than with expression. This is a textbook concern for any association study, and in eQTL work it is amplified because expression is high-dimensional, so confounding can create genome-wide false positives rather than a single spurious hit. **How to confirm.** Compute a genomic inflation metric on the eQTL scan, for example by examining the distribution of test statistics at variants far from the tested gene, which should behave like nulls under a cis-only scan. Examine quantile-quantile plots separated by platform. If the microarray arm shows inflation that the RNA-seq arm does not, the confounder is platform-specific, often a processing variable. **Fix.** - Include principal components of genotype as covariates in every eQTL test. The number of components is a design decision, commonly determined by the trace of the genetic relationship matrix or by testing the significance of each component. - Include expression-derived hidden factors, such as PEER factors, whose number scales with sample size in common practice for eQTL studies. - Keep the covariate set identical across platforms in any comparison, because differing covariate sets are a frequent explanation for platform disagreement. ## Problem 5, Cell-Type Heterogeneity Confounding Bulk Expression **Mechanism.** Bulk tissue expression is a mixture over cell types. Genetic variants that change cell-type proportions, or that correlate with composition differences across samples, produce expression differences that look like trans regulation. Conversely, a true cis-eQTL restricted to one cell type can be diluted in bulk tissue. This problem intensifies when the expression panel is small, because reference-free deconvolution methods need a broad feature space to estimate cell-type proportions. The DeCompress method addresses exactly this setting. Targeted mRNA panels measuring up to roughly 800 genes are common in academic and clinical settings because of low cost and high sensitivity for archived samples, but their limited feature space makes reference-free deconvolution difficult. DeCompress uses a reference RNA-seq or microarray dataset from similar tissue to expand the feature space of the targeted panel through compressed sensing, then performs ensemble reference-free deconvolution on the artificially expanded dataset to estimate cell-type proportions and gene signatures [3]. In simulated mixtures, four public cell line mixtures, and a breast cancer targeted panel of 1,199 samples measuring 406 genes, DeCompress reproduced cell-type proportions with less error than reference-free methods [3]. **How to confirm.** Correlate the leading expression PCs with estimated cell-type proportions. If the first PC is essentially a cell-type score, a large share of expression variation is composition rather than regulation. Also test whether suspect trans-hotspot variants correlate with the estimated proportions. **Fix.** - Estimate cell-type proportions and include them as covariates in the eQTL model. In the breast cancer application mentioned above, integrating compartment estimates into cis-eQTL mapping identified a tumor-specific cis-eQTL for CCR3 at a risk locus, an association that composition-confounded bulk analysis could obscure [3]. - Match the reference tissue to the study tissue when expanding a targeted panel. A blood reference will not deconvolve a tumor panel sensibly. - Where cell-type-specific references exist, prefer reference-based deconvolution, and use semi-reference-free methods such as DeCompress when only a similar-tissue reference is available [3]. ## Problem 6, Strand, Allele Coding, and Sample Pairing Errors on the Genotype Side **Mechanism.** eQTL association tests compare allele dosage against expression. Errors in allele coding invert or destroy the signal. Common causes include variants genotyped on the opposite strand between the genotype array and the reference build, ambiguous strand variants (A/T and C/G single nucleotide polymorphisms) whose strand cannot be determined from the allele alone, indels represented differently across platforms, and genome build differences that shift allele annotations after lifting. A separate failure class is sample identity. Genotypes belong to person A while expression belongs to person B, often from plate-level swaps that no one caught. **How to confirm.** - Check allele frequency agreement between genotypes and a relevant reference population. A variant whose allele frequency is roughly the complement of the [expected value](/blog/data-analysis/expected-value-definition-formula-examples) is strand flipped. - Test genotype-expression concordance for known strong cis-eQTLs, or for variants genotyped twice on the array, which must correlate perfectly with themselves. - For sample identity, compare genotype-derived sex against reported sex, and check pairwise identity-by-descent estimates against reported relationships. If RNA-seq data include genotyping-capable variants, expression-informed identity checks can also flag swaps. **Fix.** - Harmonize variants to a single reference, flip alleles as needed, and drop variants with ambiguous strand that cannot be resolved from frequency or neighboring variants. - Use allele dosages from a consistent calling pipeline, and document the allele coding convention in the pipeline configuration, not in someone's memory. - Before any scan, run a small identity check that joins genotype and expression sample keys and tests a handful of positive-control cis-eQTLs. If positive controls fail, nothing downstream matters. ## Problem 7, Two-Stage Phasing Weakens cis-eQTL Power **Mechanism.** RNA-seq provides information that microarrays do not. Beyond total gene expression, sequencing reads carry allele-specific expression (ASE), the relative abundance of transcripts derived from each allele. ASE distinguishes cis regulation, which acts on the allele linked to the variant, from trans regulation, which acts on both alleles equally. This gives RNA-seq additional power for cis-eQTL detection that microarrays cannot access [2]. The catch is phase. Assessing the cis effect of a candidate variant requires knowing which haplotype, meaning which combination of alleles on a chromosome, connects the candidate variant to the heterozygous sites in the gene that report ASE. Haplotype phase cannot be inferred with certainty from genotype alone. Many pipelines use a two-stage approach, first phasing the candidate variant against the gene and then treating the inferred phase as observed in the association analysis. Hu and colleagues showed that this approach attenuates the estimated cis effect and reduces the power to detect a cis-eQTL, because the uncertainty in phase is discarded at the boundary between stages [2]. **How to confirm.** On data with both genotypes and RNA-seq, compare cis-eQTL effect estimates from a two-stage pipeline against a joint likelihood model. If two-stage estimates are consistently smaller toward the null and marginally significant signals are common, phase uncertainty is eating power. **Fix.** Use a maximum-likelihood framework that integrates haplotype inference and the association analysis into a single stage, which is unbiased and statistically more powerful than the two-stage approach [2]. The same work provides a pipeline design for scanning all local eQTLs across all genes with RNA-seq data, which makes the joint approach practical at scale [2]. When integrating microarray and RNA-seq cohorts, apply the ASE-informed method to the RNA-seq arm, where the allele-specific information exists, and treat the microarray arm as a total-expression replication set. ## Problem 8, Platform Blind Spots Hide Real Signals **Mechanism.** Microarrays measure only the transcripts their probes target. Probe coverage reflects the transcript annotations available when the array was designed. RNA-seq measures whatever was sequenced, including novel transcripts, non-coding RNAs, and splicing patterns, subject to the reference annotation used for quantification. When a disease-relevant eQTL acts through an exon or transcript that the array does not measure, the microarray arm will miss it and any cross-platform comparison will conclude, wrongly, that the signal failed to replicate. The systemic lupus erythematosus study cited above illustrates the size of this blind spot. Working with the TwinsUK cohort in lymphoblastoid cell lines, the authors integrated genome-wide association data with both microarray and RNA-seq eQTL results and compared the platforms directly. They found the greatest frequency of candidate-causal eQTLs using exon-level RNA-seq and identified novel susceptibility genes, including NADSYN1 and TCF7, that were concealed using microarrays, along with four non-coding RNAs [1]. Many of these eQTLs influenced the expression of several genes, consistent with risk haplotypes carrying multiple functional effects [1]. The same study mapped splicing quantitative trait loci (asQTLs) using the Geuvadis cohort and identified novel splicing events in IKZF2, the T-reg restricted [transcription factor](/knowledge/molecular-biology/transcription-factor), and in other candidate genes including WDFY4 [1]. None of those splicing signals exist in microarray data by construction. **How to confirm.** For each eQTL found on RNA-seq but not on microarray, inspect the microarray probes overlapping the gene. Ask whether any probe covers the exon or transcript where the RNA-seq signal localizes. If no probe covers it, the non-replication is a coverage artifact, not a biological inconsistency. **Fix.** - Treat exon-level and transcript-level RNA-seq quantifications as first-class inputs to eQTL mapping, not as an afterthought, because they recover signals invisible at the gene level on either platform [1]. - When microarray data serve as replication, restrict the replication claim to genes with adequate probe coverage of the relevant transcript, and say so explicitly in the analysis plan. - Use conditional analysis and Bayesian colocalization to distinguish shared causal variants from nearby linked ones before declaring cross-platform confirmation or failure. This is the analytic design the SLE study applied when evaluating whether GWAS and eQTL signals share a causal variant [1]. - For splicing questions, microarray integration is simply out of scope. Use RNA-seq-based asQTL mapping [1]. ## Problem 9, Reference and Allele-Specific Mapping Bias This is the rarest of the routine problems, and the most technically demanding to fix. **Mechanism.** RNA-seq reads are aligned to a reference genome built from a limited panel of individuals. Reads carrying the non-reference allele align slightly worse, especially around variant sites. This differential alignment inflates apparent allele-specific expression toward the reference allele and, in an eQTL context, can distort ASE-informed cis-eQTL tests by correlating mapping success with genotype. Genotype arrays are unaffected by this bias, which means it shows up as platform-specific behavior. **How to confirm.** At heterozygous sites in highly expressed genes, compare the observed allelic ratio against the expected ratio of roughly one to one after accounting for the known cis-eQTL genotype. A systematic skew toward the reference allele across many heterozygous sites indicates reference bias. Also compare allelic ratios for the same sites when reads are aligned to a personal or diploid reference versus the standard single-genome reference. **Fix.** - Use a diploid genome approach, in which the reference is augmented with the individual's non-reference alleles at known variant sites, or use an alignment and genotyping scheme that reduces mapping bias. - Regress out or model the reference allele fraction as a bias term in ASE-based analyses. - Accept that some bias remains and interpret ASE-enhanced eQTL effects with that limitation stated. **Safety note.** Although this problem is computational, its diagnosis sometimes motivates targeted resequencing of candidate loci. Laboratory work on nucleic acid extraction commonly involves phenol-chloroform reagents or magnetic bead chemistries. Phenol and chloroform are corrosive and volatile, so use a fume hood, appropriate gloves, and eye protection, and follow the institutional chemical hygiene plan. ## Common Mistakes and Limitations ![Decision path checking probe coverage, allele coding, batch effects and cell-type composition before calling a signal false.](/article-images/illustrations/eqtl-mapping-rna-seq-microarray-integration-platform-disagreement-triage.webp) Several mistakes recur in real projects and deserve explicit treatment. **Pooling expression matrices across platforms.** The single most damaging mistake is concatenating a microarray matrix and an RNA-seq matrix after z-scoring each gene and fitting one joint eQTL model. This mixes different measured quantities and variance structures. The results are difficult to interpret and the failure is usually silent. **Correcting expression without genotype in the model.** Aggressive batch correction applied before association testing can absorb true genetic signal, particularly when batch and ancestry correlate. Correct only with the association model in hand, and confirm that strong cis-eQTLs survive. **Using gene symbols as join keys.** Symbols change. Ensembl and Entrez identifiers also version. Join on stable, current identifiers with documented provenance, and validate a random subset by sequence. **Treating non-replication as refutation.** A microarray failure to replicate an RNA-seq eQTL may reflect probe coverage, dynamic range, or transcript specificity. Interpret replication failures in light of platform design, as the SLE study did when comparing the two quantification types [1]. **Ignoring haplotype uncertainty in ASE-based mapping.** Treating inferred phase as known biases cis-effect estimates toward the null, which is precisely the problem the maximum-likelihood single-stage framework was designed to remove [2]. **Overlooking composition in bulk tissue.** A cis-eQTL detected after cell-type adjustment may be real in a specific compartment and invisible in unadjusted bulk analysis. Conversely, apparent eQTLs can be composition artifacts. Deconvolution-based covariates address both directions of the problem. **Limitations of the integrated approach.** Cross-platform integration cannot create information that neither platform captured. It cannot detect signals below either platform's detection limit, it cannot resolve transcript-level regulation absent transcript-level quantification, and its [statistical power](/blog/guides/statistical-power-what-it-is-and-why-it-matters-in-research) remains bounded by sample size and allele frequency. Individual datasets, particularly those with unusual tissue types, small panels, or non-European ancestry distributions, need analyst judgment and, where human-subject implications arise, appropriate institutional oversight, rather than a generic pipeline applied by default. ## Prevention, Building the Pipeline So These Failures Do Not Happen Prevention in eQTL integration is mostly about committing decisions to writing before data touch the pipeline. **Freeze the reference frame.** Choose one genome build, one gene annotation release, and one variant reference. Record versions in a configuration file that ships with the analysis. Every input, from microarray probe annotations to RNA-seq quantification tables, maps into that frame with a logged, testable step. **Join samples on audited keys.** Sample identifiers must join one-to-one between genotypes and expression, with sex checks, kinship checks, and positive-control eQTL tests run as a gate before the full scan. A gate that fails stops the pipeline, rather than producing a results table that someone will interpret at midnight. **Analyze within platform, combine by meta-analysis.** The default architecture is platform-specific eQTL scans followed by inverse-variance-weighted meta-analysis of standardized effect sizes, plus colocalization to test for shared causal variants [1]. Deviations from this architecture need explicit justification. **Match the method to the data modality.** ASE-informed single-stage likelihood testing for the RNA-seq arm [2], conventional total-expression testing for the microarray arm, exon-level and splicing-level quantifications where RNA-seq depth allows [1], and composition covariates from deconvolution for bulk tissue [3]. Each modality contributes what it can measure, and the integration step respects those boundaries. **Keep positive controls.** Maintain a short list of well-established cis-eQTLs and verify, on every pipeline run, that they are detected with the expected direction and rough magnitude on every platform in the analysis. This single check catches identifier errors, allele flips, sample swaps, and overcorrection faster than any other diagnostic. **Version and containerize.** Run the pipeline in a container with pinned software versions, and archive the expression matrices, genotype files, covariate tables, and results. Reproducibility is not decoration in eQTL work, because effect sizes are small and the difference between a real signal and a pipeline artifact often hides in a version number. ## Frequently Asked Questions ### Can I combine RNA-seq and microarray expression data in one eQTL analysis? Direct pooling of raw expression values is not advisable, because the platforms measure related but distinct quantities with different variance structures. The standard solution is to run eQTL scans separately on each platform and combine results through meta-analysis of standardized effect sizes, with colocalization used to evaluate shared causal variants. ### Why do my RNA-seq and microarray eQTL results disagree? The most frequent causes are gene identifier and annotation mismatches, allele or strand coding errors, unmodeled batch effects, and microarray probes that fail to cover the transcript where the signal lives. Check identifier mapping and genotype harmonization first, then probe coverage, before concluding that the biology differs. ### What normalization should I use for cross-platform eQTL integration? Normalize within platform using methods appropriate to that technology, then combine standardized regression coefficients across platforms. Avoid converting RNA-seq values to a pseudo-microarray scale or vice versa for pooled modeling, because the conversion hides rather than resolves the structural differences. ### Do I need PEER factors or principal components in my eQTL model? Yes, as a default. Genotype principal components control population structure, and expression-derived latent factors capture hidden technical and biological variation. Keep the covariate set identical across platforms so that cross-platform comparisons remain interpretable. ### Why does allele-specific expression improve cis-eQTL mapping? Allele-specific expression provides within-individual evidence of cis regulation, because a cis-acting variant changes the balance of the two alleles within a heterozygote, something total expression cannot show. A single-stage maximum-likelihood framework that integrates haplotype inference with association testing captures this information without the power loss caused by treating inferred phase as observed data. ### Why did my microarray study miss eGenes that RNA-seq studies report? Microarrays measure only probed transcripts, so genes, exons, non-coding RNAs, and splicing events absent from the probe design are invisible regardless of sample size. In a direct comparison within the same cohort, exon-level RNA-seq recovered more candidate-causal eQTLs and revealed susceptibility genes that microarray measurements concealed. ### How do I handle cell-type heterogeneity in bulk tissue eQTL mapping? Estimate cell-type proportions and include them as covariates in the association model. For targeted panels with a small feature space, semi-reference-free methods that expand the panel using a reference dataset from similar tissue outperform purely reference-free deconvolution and have been shown to support cis-eQTL discovery in bulk tumor samples. ### How many samples do I need for eQTL mapping? There is no single threshold, because power depends on effect size, allele frequency, expression quantification quality, and residual variance, and ASE-informed methods can extract additional power from RNA-seq data at a given sample size. Plan the study with a power calculation for the smallest effect you care about, and prioritize genotyping accuracy, expression quality control, and covariate modeling, which determine whether the samples you have behave like the samples you planned. ## Sources 1. [Mapping eQTLs with RNA-seq reveals novel susceptibility genes, non-coding RNAs and alternative-splicing events in systemic lupus erythematosus.](https://pubmed.ncbi.nlm.nih.gov/28062664/) 2. [Proper Use of Allele-Specific Expression Improves Statistical Power for cis-eQTL Mapping with RNA-Seq Data.](https://pubmed.ncbi.nlm.nih.gov/26568645/) 3. [DeCompress: tissue compartment deconvolution of targeted mRNA expression panels using compressed sensing.](https://pubmed.ncbi.nlm.nih.gov/33524140/)