Skip to content
Tech Interview Prep
Technical interview guide

Gene Expression Analysis & RNA-Seq

Quantifying which genes are active and by how much: read counting, normalization, and differential expression testing.

Read
49 min
Practice MCQs
25
Interview Q&A
25
Edition
v5
Editorial status
Reviewed

Scope: DESeq2, edgeR, voom, STAR, Salmon, kallisto, featureCounts, tximport and MultiQC source publications; GENCODE, GA4GH and ENCODE guidance reviewed 2026-09-04.

Curated: · Written: · Reviewed:

25 questions

MCQ-7Worked example, hypothetical coordinates. A 100 bp single-end RNA-seq read is aligned to the genome with a splice-aware aligner (STAR or HISAT2, default settings) and the matching annotation. Its first 38 bases match the plus strand at chr17:43,044,200-43,044,237 exactly, and its remaining 62 bases match the plus strand at chr17:43,048,320-43,048,381 exactly. The annotation puts both ranges in the same gene, in that order, with a canonical GT-AG junction between them, and neither block has a second equally good match. Which alignment appears in the read's SAM record?
MCQ-5A liver mRNA-seq library is aligned once and counted with featureCounts three times on the same BAM, changing only the strand flag (-s 0 unstranded, -s 1 forward, -s 2 reverse): -s 0 assigns 86.9% of the aligned reads, -s 1 assigns 2.8% and -s 2 assigns 86.4%. A second library is then prepared from a comparable liver sample with a non-stranded kit, sequenced to similar depth and aligned with the same pipeline at a similar mapping rate. Which assignment rates should featureCounts report for the second library?
MCQ-10A team quantifies 12 RNA-seq samples against a GTF whose gene records carry `gene_id` and `gene_name` attributes, using a counting tool configured with `-s reverse -p --fracOverlap 0.5`. Two samples come back with roughly 40% of reads reported as unassigned, far above the ~10% the team sees on other projects. Inspection shows those two libraries were prepared with a different strandedness protocol than the rest. Which configuration change most directly addresses the unassigned-read spike?
MCQ-9A differential-expression pipeline joins its count matrix to a GTF by gene symbol. On the current release the join returns 19,340 rows for 18,000 input genes; 412 symbols map to more than one gene ID and together account for 1,340 extra rows. After the annotation is upgraded to the next release, 214 genes that had non-zero counts in the previous run come back with zero counts, and their symbols are absent from the new GTF. Which change addresses both symptoms?
MCQ-8A team quantifies two human RNA-seq replicates (48M and 52M paired-end reads) with Salmon against a versioned transcriptome. Gene G has 11 annotated isoforms. Isoform G-204 gets TPM 12.4 in replicate A and 41.8 in replicate B, while gene-level TPM is 58.1 and 59.6. Only 3% of the gene's reads fall in exons unique to G-204. What is the right reading of this result?
MCQ-11A team has RNA-seq samples quantified as TPM and wants DESeq2 to call differentially expressed genes between treated and control. Why is TPM usually not passed to DESeq2 as the count matrix?
MCQ-12An RNA-seq experiment compares control and treated samples, each sequenced to 20 million reads. In the treated sample, 50 genes that made up 1% of the control library now make up 40% of it. Gene X has 2,000 reads in the control library, and its true expression is unchanged. What does total-count (CPM) normalization report for gene X?
MCQ-13A bulk RNA-seq experiment has 6 replicates of one condition. For gene G the counts are 80, 120, 100, 160, 90, 170 (mean 120, sample variance 1400). A Poisson GLM fit to these counts gives a Pearson chi-square of 58.3 on 5 df; the 0.001 critical value for 5 df is 20.5. What should the analyst do next?
MCQ-14In a hypothetical RNA-seq experiment with 3 replicates per condition and 12,000 genes, Gene A has control counts 10, 12, 11 and treated counts 10, 12, 100. Its per-gene dispersion estimate is 1.57. The fitted mean-dispersion trend at Gene A's mean count is 0.05. What does DESeq2's dispersion shrinkage do to Gene A's dispersion estimate and its Wald test p-value for treated vs control?
MCQ-17A 3-vs-3 DESeq2 run reports a gene with 12 total counts, log2 fold change 6.2 and adjusted p-value 0.4. The analyst calls lfcShrink(type='apeglm') and re-sorts genes by shrunken log2 fold change for a heatmap. What did the shrunken value change?
MCQ-25A DESeq2 comparison of treated versus control returns raw p-values whose histogram has a large spike near 1 and almost no values below 0.05. What does this shape indicate about the analysis?
MCQ-18A differential-expression pipeline tests 20,000 genes with Benjamini–Hochberg and reports 300 hits at adjusted p < 0.05. BH controls the false discovery rate under independence or positive dependence. What does the 5% target guarantee about those 300 genes?
MCQ-22A team runs over-representation analysis on 500 differentially expressed genes from a liver RNA-seq experiment. They set the universe to all 20,000 annotated protein-coding genes in the genome, although only 10,000 of those are expressed in liver. The GO term 'fatty acid beta-oxidation' has 200 annotated genes genome-wide, 60 of which are expressed in liver; 30 of the 500 DE genes fall in the term. The team reports the term as enriched with a 6-fold observed/expected ratio. What is wrong with that report?
MCQ-16An RNA-seq experiment crosses genotype (WT as reference, KO) with treatment (vehicle as reference, drug), four samples per cell. The team fits `~ genotype * treatment` to log2 CPM and estimates: intercept 10.0, genotypeKO 2.0, treatmentdrug 3.0, genotypeKO:treatmentdrug −1.0. They need the drug effect within KO samples: the KO drug-minus-vehicle difference. Which value is it?
MCQ-4A team has 16 RNA-seq libraries from 8 patients: every patient gave one tumor and one matched normal biopsy, all sequenced in one run, and 12,400 genes pass the expression filter. Fitting the counts with a negative-binomial GLM (edgeR) on ~ condition alone calls 21 genes at FDR 0.05; refitting the same counts on ~ patient + condition calls 890 at the same FDR. Per-gene log2 fold changes are nearly the same in both fits (Pearson r = 0.998; median absolute logFC 0.24 against 0.25), and PC1 of the filtered counts separates patients while PC2 separates tumor from normal. What accounts for the 21-to-890 change?
MCQ-15Six RNA-seq samples are fit with the design ~ batch + condition. The three treated samples were all prepared in batch 2 and the three controls all in batch 1. What does the design matrix show about the treated coefficient?
MCQ-3A team has 8 treated and 8 control RNA-seq samples. Every treated sample was prepared and sequenced in batch B2; every control sample in batch B1. They fit a model with treatment and batch as fixed effects and report the treatment coefficient. Which statement about the treatment coefficient is correct?
MCQ-2A lab must prepare 12 RNA-seq libraries, 6 from control tissue and 6 from treated tissue, but the reagent kit is only good for four preps, so the work is split over three days: day 1, four control; day 2, two control and two treated; day 3, four treated. They fit a model with condition and day as fixed effects. Which statement describes the condition effect this design estimates?
MCQ-21A 12-sample RNA-seq experiment compares 6 treated with 6 control samples across two sequencing batches. Batch 1 holds 4 treated and 2 control; batch 2 holds 2 treated and 4 control, so batch and condition are partially confounded. The team fits a negative-binomial GLM to call differentially expressed genes and must decide whether to put batch in the model or to remove the batch signal from the counts first. Which handling is right?
MCQ-1A lab runs an RNA-seq differential expression study comparing a drug-treated cell line against vehicle. Each condition has 4 RNA extractions from the same flask, split into 2 sequencing libraries each, giving 8 libraries per condition. The DE tool reports 3,100 genes at FDR < 0.05, but a collaborator's independent repeat of the same experiment with 4 separate flasks per condition finds only 40 genes at the same threshold. Which change to the design most directly addresses the cause of this discrepancy?
MCQ-20Sixteen bulk RNA-seq libraries from one tissue, 8 control and 8 treated, were prepared in two batches of 8, each batch holding 4 control and 4 treated libraries. Control libraries were sequenced to 42M reads each, treated libraries to 14M each. PCA on raw gene counts puts 86% of the total variance on PC1, which separates control from treated with no sample overlap and correlates with total mapped reads at Spearman rho = 0.98; its loadings concentrate on the 30 highest-count genes (means 20,000-80,000 counts per sample, per-gene CV about 0.15). Genes with means of 20-300 counts have CVs of 0.5-0.9. Repeating the PCA on log2(count + 1) drops PC1 to 27% of variance with its depth correlation at 0.05, and the conditions still separate, now weighted toward the low-count genes. Which reading of the raw-count PC1 is correct?
MCQ-19In a 24-sample RNA-seq differential-expression run (12 control, 12 treated), one treated sample has 4,800 counts for gene X, while the other 23 samples range from 120 to 260. Its mapping rate is 91%, rRNA fraction 0.8%, duplication 42%, and genotype check matches the recorded sample. The analyst wants to delete the sample before fitting the model. What is the best next step?
MCQ-6An RNA-seq core lab is quantifying eight samples, S1-S8, from one experiment. Standing acceptance rules: assigned reads >= 20M; rRNA fraction < 15%; duplication rate < 60%; correct-strand reads >= 90%; 3'/5' coverage ratio <= 1.5. A sample that fails two or more of these five families is excluded and re-prepared; a sample that fails exactly one family is held for review and still quantified; each family counts separately even when two failures share one root cause. The aggregated QC report gives: | sample | reads (M) | rRNA % | dup % | correct strand % | 3'/5' ratio | | --- | --- | --- | --- | --- | --- | | S1 | 34.1 | 3.2 | 28 | 96.1 | 1.2 | | S2 | 41.0 | 5.8 | 47 | 94.3 | 1.4 | | S3 | 18.4 | 6.1 | 31 | 95.2 | 1.3 | | S4 | 40.2 | 22.6 | 33 | 93.8 | 1.1 | | S5 | 9.6 | 41.3 | 44 | 92.0 | 1.3 | | S6 | 29.7 | 4.4 | 25 | 97.0 | 1.2 | | S7 | 33.5 | 8.9 | 68 | 71.4 | 1.4 | | S8 | 26.3 | 2.9 | 38 | 95.5 | 1.3 | Applying these rules, which samples are excluded for re-prep and which are held?
MCQ-23Bulk RNA-seq on lesional and non-lesional skin biopsies from the same 10 patients puts KRT14 at 0.40x in lesional tissue, a 2.5-fold drop. Matched snRNA-seq gives a mean of 400 KRT14 UMIs per keratinocyte in both conditions. Keratinocytes are 80% of nuclei in non-lesional skin and 32% in lesional skin, where the remaining 68% are T cells and macrophages that express no KRT14; total RNA per cell is comparable across the cell types. What should the team report as the effect on KRT14?
MCQ-24A lab reruns a differential expression analysis on the same 12 RNA-seq samples and gets a different set of significant genes than the paper reported. The paper's methods section lists the aligner, the reference genome build, and the differential expression tool with its version. The lab used the same aligner, the same build, and the same tool version, and confirmed the raw FASTQ files match the paper's checksums. Which single omission most plausibly explains the discrepancy?